EtoPlan a Concept for Concurrent Manufacturing Planning and Control

Transcription

1 EtoPlan a Concept for Concurrent Manufacturing Planning and Control - Building holarchies for manufacture-to-order environments - Mark Giebels

2 ISBN ISSN M.M.T. Giebels, 2000 Printed by PrintPartners Ipskamp B.V., Enschede, The Netherlands

3 ETOPLAN: A CONCEPT FOR CONCURRENT MANUFACTURING PLANNING AND CONTROL Building holarchies for manufacture-to-order environments PROEFSCHRIFT ter verkrijging van de graad van doctor aan de Universiteit Twente, op gezag van de rector magnificus, prof.dr. F.A. van Vught, volgens besluit van het College voor Promoties in het openbaar te verdedigen op donderdag 15 juni 2000 te 13:15 uur. door Mark Mathieu Theodorus Giebels geboren op 14 juni 1971 te Boxtel

4 Dit proefschrift is goedgekeurd door de promotoren prof.dr.ir. H.J.J. Kals en prof. dr. W.H.M. Zijm

5 It is not the strongest of the species that survives, nor the most intelligent; it is the one that is most adaptable to change Charles Darwin

6

7 vii Preface This thesis presents the results of a research project that started in October The project was initiated within the Centre for Production, Logistics and Operations Management (CPLOM) as part of the OSF project Integrale Productievernieuwing, which primarily aimed to bring together the research performed by various groups of the Applied Mathematics, Industrial Engineering & Management, and Mechanical Engineering departments at the University of Twente. The project has been supervised by Prof. H.J.J. Kals and Prof. W.H.M. Zijm, the chairmen of the former laboratories of Production & Design Engineering and Production & Operations Management, respectively. Since last year, these groups have merged into the laboratory of Design, Production and Management of the Mechanical Engineering department. The research project was originally entitled the integration of process planning and production control. This integration aspect has always been the point of departure when discussing research problems. In the title of this thesis, the adjective concurrent refers to the original title. In the course of the research project, we recognised that real integration requires fundamental changes in the planning and control architectures as well as in the way the information in the manufacturing companies is managed. The results presented in this thesis, therefore, have clear links to both the research project on Information Management performed in the same laboratory and the Holonic Manufacturing Systems research project of the Intelligent Manufacturing Systems (IMS) program. Furthermore, the FACT concept developed by Arentsen and Tiemersma has been an important source of knowledge. Throughout the project, the focus has shifted from shop floor control towards the higher manufacturing planning and control levels. The shop floor control research problems as well as the interaction with the micro process planning task have been surveyed thoroughly by many researchers during the last decade. The higher planning levels, viz. macro process planning and resource loading, however, are not studied in the same detail. This particularly holds for methods developed for the manufacture-to-order (i.e. make- or engineer-to-order) environments. In such manufacturing environments, the interaction between the resource loading and macro process planning tasks is of high importance. Also the interaction with the order acceptance task has been considered in the research project. After all, the higher level planning tasks are performed as soon as the information about the ordered products is known. Abstraction-planning methods that can handle stochastic information have appeared to be required to perform the higher-level manufacturing-planning tasks, due to both the incompleteness of the technological planning data in the macro process planning phase and the long-term unpredictability of the state of the shop floor. As can be concluded from the positioning of the research project described above, the integration of multiple research fields has been a leading thread throughout the project. During the years, we recognised that significant improvements can be achieved by

8 viii explicitly considering the aspects of integration of, particularly, the research field of production planning - which is dominated by mathematicians - and the research field of process planning, which is, of course, the domain of mechanical engineers. The proverbial walls between the various experts are threatened to be enlarged, while actually there are promising possibilities to pull these walls down. In this thesis, we try to clarify possible strategies and we define some solution methods to enable extensive integration. Enschede, April 2000 Mark Giebels

9 ix Summary An essential demand for future manufacturing control systems is the ability to deal with the increased complexity due to a higher product variety, smaller batches and shorter throughput times. In particular for manufacture-to-order (i.e. make- or engineer-to-order) environments it is generally recognised that future manufacturing planning and control systems have to become more flexible and integrative compared to the currently existing systems. The versatility in order characteristics calls for control systems that are able to evolve in time by continually restructuring their control hierarchies. This thesis presents a control concept, named EtoPlan, which meets these requirements by building temporary and multiple hierarchies (holarchies) of resources. The integration of the manufacturing planning tasks design, process planning, and production planning - is of special importance in manufacture-to-order environments. The questions of what product has to be produced, as well as how, where and when the various manufacturing activities should be executed, have to be answered practically simultaneously in such manufacturing environments. In order to deal with the required parallel processing of the manufacturing planning tasks, there is a need for both generic information structures and abstraction planning methods. A specific Information Management concept, which serves as the backbone for the EtoPlan concept, discerns three generic information structures for Products, Resources, and Orders, respectively. All manufacturing planning decisions are based on these three information structures, which makes real integration (not just interfacing) between design, process planning and production planning possible. Furthermore, an aggregate order planning method is presented in this thesis for higher-level integration of macro process planning and resource loading. The method explicitly models the existing uncertainty which results from the incompleteness and unreliability of the information due to a lack of detailed process plans in the macro process planning phase and the (long-term) unpredictability of the shop floor, respectively. The aggregate order planning method is based on interactively setting and changing stochastic values of processing times, start times, and lead times. Both the system architecture which defines the holarchies and the aggregate order planning method have been implemented in the EtoPlan prototype software system in order to show the applicability of the EtoPlan concept.

10

11 xi Table of contents PREFACE...VII SUMMARY... IX CHAPTER 1 INTRODUCTION Manufacture-to-order Manufacturing planning automation Manufacturing control Problem definition Overview of this thesis...5 PART I MANUFACTURING ENVIRONMENT...7 CHAPTER 2 MANUFACTURING PLANNING Process planning Feature interaction and sequencing Artificial Intelligence in Process Planning Production planning Enterprise Resource Planning (ERP) systems Capacity planning Capacity planning in the order acceptance phase Resource loading Production Scheduling Integration of Process Planning and Production Planning Integration concepts Non-Linear Process Plan (NLPP) Hierarchical Process Planning Blackboard-based systems Integration on shop floor level Integration of macro process planning and capacity planning Summary CHAPTER 3 ABSTRACTION PLANNING Literature on abstraction planning Uncertainty in manufacturing planning Conclusion...40

12 xii CHAPTER 4 MANUFACTURING CONTROL Hierarchical control Hierarchical control concepts Misconceptions about hierarchical control Some conclusions regarding hierarchical control Heterarchical control Multi-agent systems Evolution-based production control concepts Holonic Manufacturing Systems Holonic architectures Other evolution-based concepts Non-linear dynamics of manufacturing control systems Conclusions...60 PART II DESIGN OF MANUFACTURING PLANNING AND CONTROL CONCEPT...63 CHAPTER 5 REFERENCE MODELS Literature on modelling frameworks Literature on reference architectures A reference architecture for Integrated Manufacturing Planning and Control Conclusion...71 CHAPTER 6 INFORMATION MANAGEMENT A diversity of information structures A concept for information management Generic framework for the information structures Product Information Structure (PRIS) Resource Information structure (RIS) Order Information Structure (OIS) The external order life cycle domain The activity domain ORDER classification for aggregate planning Manufacturing planning views Conclusions...85 CHAPTER 7 SYSTEM ARCHITECTURE Temporary hierarchies of Applicability Groups (AGs) Method Applicability Groups (MAGs) AG Controller Autonomy and co-operation Scheduling Groups...94

13 xiii 7.5 Conclusions...95 CHAPTER 8 AGGREGATE ORDER PLANNING Uncertainty modelling for aggregate order planning Lead times Processing times Start times The order planning methodology Generic ORDER profile Resource loading views Due date determination Cost aspects of production planning Allocation of the resources Navigating through the planning views ORDER types Conclusions CHAPTER 9 A PROTOTYPE SOFTWARE IMPLEMENTATION Information Structures The order Information Order planning views Resource information The information structure of the Resource object The information structure of the Applicability Group object The information structure of the Method Applicability Group object Resource loading views AG availability view Resource availability view MAG availability view Cost views Navigating through the system Example Conclusions CHAPTER 10 CONCLUSIONS AND RECOMMENDATIONS Conclusions Recommendations REFERENCES TERMINOLOGY INDEX SAMENVATTING DANKWOORD...155

14 xiv

15 1 Chapter 1 Introduction In the second half of the twentieth century, the automation of planning and control tasks has evolved rapidly in manufacturing companies. Automation on the shop floor started with the advent of Computer Numerically Controlled (CNC) machine tools for which a machining plan (NC-plan) has to be drawn up off-line and can be used repeatedly by locally loading the program into the control system belonging to the machine tool. During the late seventies it was aimed to combine the advantages of flexible manufacturing in small batches with the efficiency of flow production. The grouping of several CNC machine tools in so-called Flexible Manufacturing Cells was seen as a way of achieving this aim. Later, in the eighties, so-called Flexible Manufacturing Systems (FMS) became a great focus of attention in manufacturing research and industry (O Grady, 1986). An FMS consists of several machine tools interconnected by a transportation system. Parts are transported on pallets to the machines where both the transportation system and machine processing are controlled by a central computer. It has been recognised later that FMSs are in fact not flexible at all. FMSs were mainly used in fixed lines for producing a small variety of products. Moriwaki states that the reasons for the somewhat paradoxical inflexibility of FMSs are the following (Moriwaki. 1999): The productivity has been reduced in order to machine without human attendance. FMSs are dedicated for a specific process. FMSs are centrally controlled. Since the last decade, the general emphasis in research shifted from gaining flexibility by automating machine tools and physically restructuring (e.g. by group technology) the shop floor to the integration of the manufacturing planning functions and building more dynamic control systems. Traditionally, the product design, process planning and production departments are strictly separated, which prevents achieving efficient plans that can cope with the increased product variety and shorter delivery dates demanded by customers today. The principles of Concurrent Engineering have received more and more attention in the industrial engineering research field (Sohlenius, 1992). During the design of products the processes should already be planned roughly in order to avoid inefficient design decisions. The need for integration also applies for the production planning and process planning processes, as e.g. stated by Alting and Zhang (Alting, 1989). It is noted, however, that it appeared to be difficult to translate the principles of concurrent manufacturing planning into practically useful planning systems. The differences in planning goals and information structures used in the various planning processes are the main causes of failure.

16 2 CHAPTER 1 The need for integration becomes apparent when the trends in manufacturing in the 1990 s are analysed. First of all, the revolution in Information and Communication Technology (ICT) with the eye-catching development of the Internet has resulted in a boost in global competition. Information and money is moved all around the world in less than no time. Relatively simple and bulk products can no longer be produced economically in the Western world countries. For instance, finished products imported from China are often cheaper than the blank materials used for in-house production. Hence, the industries in the Western countries only produce high-quality products, and more and more customised products, in, consequently, relatively small batches. The value added per product has to be increased continually in order to remain competitive in a global economy. A similar trend has been noticed regarding the length of the product life cycle. This trend puts increased pressure on the lead times in industry because products have to be delivered more quickly and more reliably. Finally, new technologies regarding products, processes and machine tools are proliferating. Consequently, organisational structures have to become more dynamic. The continuing versatility in products and processes together with the distribution of computing in machine tools imply increased local, rather than central, planning and control functionality. The term agility is often used to indicate the need for change in any direction. Instead of building something that anticipates a defined range of requirements based on ten or twelve contingencies, build it in such a way that it can be deconstructed and reconstructed as needed (Noaker, 1994) (Langer, 1999). The above trends in manufacturing have also lead to some important consequences for the internal management processes in the company. The forward visibility of, for example, sales is limited which restricts the role of forecasting. Forecasting is mainly of importance for budgeting purposes and not for drawing up master production plans anymore. The interactions between sales and production management are nowadays required at all levels in the planning, from capacity planning to shop floor control. New orders, including many so-called rush orders, are continually inserted in production plans and already planned orders are changed frequently due to changing customer demands in product features, delivery dates or quantities. 1.1 Manufacture-to-order The influence of customers on the manufacturing process is an important aspect for the classification of manufacturing environments. Make-To-Stock (MTS) production is based on forecasting of market demand. In a MTS manufacturing environment, the production orders are normally determined on the basis of specific stock levels of products. In Assemble-To-Order (ATO) production, the assembly of semi-finished parts into end products is initiated by a customer order, while the parts are made to stock. When the production activities are customer order driven instead of driven by forecasts, one speaks of make- or engineer-to-order manufacturing. In Make-To-Order (MTO) manufacturing, the production activities and, sometimes, the process planning activities are performed after a customer order has been received. In Engineer-To-Order (ETO) manufacturing, even the product design activities are performed to order. In this thesis, the focus is on manufactureto-order environments, which implies both make- and engineer-to-order. Also combinations of manufacture-to-order and make-to-stock production of semi-finished

17 INTRODUCTION 3 products are taken into account. After all, in today s manufacturing companies many (changing) combinations of production situations, from MTS to ETO, arise. In order to cope with such dynamic combinations of production situations, the manufacturing planning and control concepts must be made flexible for dealing with the large versatility of production activities. 1.2 Manufacturing planning automation Manufacturing planning embraces all planning tasks performed to enable the production of ordered products. These tasks can be subdivided in technological planning and production planning tasks. Technological Planning tasks determine what artifacts are produced and how. The what-question is solved by Product Design, while Process Planning concerns about how the product is produced. Production Planning relates to the timing of production activities and allocates the resources required for execution of these activities. The three mentioned manufacturing planning tasks (product design, process planning, and production planning) became even more separated under the influence of increased automation of subsystems. Computer Aided Design (CAD) systems have been developed for automating the product design process, although this was notably limited towards the drawing process. Attempts have been made to completely automate the process planning task by developing generative Computer Aided Process Planning (CAPP) systems. Due to the complexity of the process planning task, the variant based CAPP systems have, however, been applied in industry more frequently. The production planning task is generally subdivided in subsystems for long term planning (e.g. master production scheduling) and short term planning (e.g. shop floor planning). Just as the automatic process planning systems, also the advanced production planning systems that use automatic planning heuristics experience difficulties regarding the complexity of production situations. This problem even complicates when the integration of the manufacturing planning tasks is aimed for. Consequently, most developments in manufacturing planning automation can be characterised as island automation, which has resulted in enormous problems regarding the integration and even regarding the interfacing between the subsystems. 1.3 Manufacturing control The planning functions that are required to perform the manufacturing planning tasks must be incorporated in a control architecture. A control architecture defines the information and control flows between the subsystems so as to enable integration of the various planning and control functions. Planning and control decisions must be based on the requirements and constraints imposed by other planning and control functions. For example, a shop floor scheduler will have to reschedule activities when significant disturbances are reported by a workstation controller. Control architectures are often based on hierarchical control structures so as to decompose the planning and control problem. Furthermore, hierarchical control structures enable aggregate planning long before the actual production. Mainly due to the structural rigidity of hierarchical control architectures, control concepts based on

18 4 CHAPTER 1 heterarchical and, later, evolution based control, have become popular in manufacturing control research. In heterarchical control, only horizontal control flows between the control entities are allowed, which on the one hand results in greater flexibility, but on the other hand leads to a lack of global control. The principles of the evolution based control paradigms mainly differ from hierarchical control with respect to the flexibility of the hierarchies. The aim is to develop dynamic hierarchical structures so as to combine the flexibility of heterarchical control with the advantages of hierarchical control. 1.4 Problem definition The trends in manufacturing have resulted in a shift from large batch make-to-stock production towards manufacture-to-order with smaller lot sizes. Also, most companies nowadays combine make-to-stock production for standard parts with manufacture-to-order production to allow for delivery of client-specific products. The dedicated production planning and control systems for either make-to-stock or manufacture-to-order environments are not suitable anymore for these hybrid order management strategies. There is a need for as much flexibility in these systems as required for handling each order according to its specific demands. This thesis aims to provide a conceptual solution for the problem of managing a high versatility of order characteristics in a dynamic manufactureto-order market. Much attention is paid to the integration of production planning and control with the other planning processes in the manufacturing company, particularly process planning. The technological planning and production planning and control tasks are too often considered separately in the industrial engineering research community. In process planning, the situation on the shop floor is generally assumed to be static and, thus, known. On the other hand, all technological information is assumed to be deterministic and completely known in production planning, even on the higher levels of the production planning process. Obviously, these assumptions are detrimental to the final result of the planning process. In rather complex and dynamic production situations - as is common in manufacture-to-order environments such assumptions are simply not justified. An important problem for the integration of the various manufacturing planning tasks is the diversity of information structures used in the different concepts and systems. Production Planning and Control considers the whole order mix and is, therefore, generally structured according to the departmental and resource structures in the company. Process planning concerns the production of a single product where the process structure (routing) is taken as a basis for the information structure. Process structures, however, do not simply match the organisational resource structures as used by Production Planning and Control. Process structures are first and foremost build up along the lines of technological requirements and constraints. Particularly on the higher levels of aggregation, these process structures generally do not directly relate to the (structure of) resources in the company.

19 INTRODUCTION Overview of this thesis This thesis has been divided in two parts. The manufacturing environment is analysed in the first part. In the second part, a new concept for planning and control in a manufacture-toorder environment, named EtoPlan, is presented. Part I An analysis of the process planning and production planning functions is given in Chapter 2. This chapter particularly focuses on the integration of the manufacturing planning tasks. Literature on abstraction planning and uncertainty in manufacturing planning is discussed in Chapter 3. Chapter 4 discusses the various research streams on manufacturing control. Particularly, the need for hierarchies is elaborated. Part II Part II contains the design part of this thesis. Chapter 5 discusses reference models for modelling manufacturing systems. A distinction is made between modelling frameworks and reference architectures. A specific reference architecture for manufacturing planning and control is used as a reference in the remainder of this thesis. Chapter 6 discusses the role of information management in manufacturing planning and control and describes a specific concept for Information Management. This concept serves as the backbone for the EtoPlan concept. The system architecture of the EtoPlan concept is presented in Chapter 7. Chapter 8 presents the aggregate order planning methodology of EtoPlan. Chapter 9 describes a prototype software implementation based on the EtoPlan concept. In the final chapter, the conclusions are given as well as some recommendations for future research.

20

21 7 Part I MANUFACTURING ENVIRONMENT

22

23 9 Chapter 2 Manufacturing Planning This chapter focuses on the planning processes that have to be performed by manufacturing companies to enable and to co-ordinate the production activities on the shop floor. The term Manufacturing Planning is used in this thesis to include both technological planning (design, process planning) and production planning (planning of activities, planning of resources). The product model that comprises the geometrical and technical specifications of a product is generated by Product Design. The functional requirements of a product, the manufacturing capabilities, and the availability of resources determine together the requirements and the constraints for the design process. The interaction with the other manufacturing planning processes (Process and Production Planning) becomes more and more important, particularly in manufacture-to-order environments. The integration of Product Design and Process Planning has received significant research attention, under the name of Concurrent Engineering (Sohlenius, 1992). Ensuring the manufacturability of the designed products is an important objective of Concurrent Engineering. Typical examples of decisions that are of interest for both Product Design and Process Planning are: Material selection. Tolerance handling. Determination of dimensions and topology. Process Planning deals with the determination of the technical information that specifies how a product is produced. Process Planning decisions include macro planning (e.g. routing determination) and micro planning - just before the actual production takes place - like tool path calculations. The Process Planning task is elaborated in Section 2.1. Production Planning determines when the production activities will be executed and with which resources. Generally, there are some planning phases in Production Planning. The rough planning of aggregate activities on aggregate resources (e.g. production cells) is carried out with a relatively long time horizon. The determination of job sequences on individual machine tools is performed on a smaller time scale. The Production Planning task is discussed in Section 2.2. In manufacture-to-order environments, the interaction between Production Planning and Process Planning is essential. Many orders compete for the same resources, which often leads to the occurrence of bottleneck resources. Depending on the given situation on the shop floor, a different routing may be chosen, and, for example, cutting speeds may need to be adapted. An overview of research performed on the

24 10 CHAPTER 2 integration of Process Planning and Production Planning together with a general discussion is given in Section 2.3. More planning processes can be recognised in manufacturing planning. For instance, the Lay-out Planning of the resources is important for efficient routings. Small batch manufacturing is mostly characterised by a job shop lay-out structure containing groups of resources each consisting of a couple of general purpose machine tools. Other planning processes deal with the agreements with the external stakeholders of the company (customers, suppliers, etc.). The Sales department in manufacture-to-order companies is responsible for making quotations, putting up proposals for a delivery date, a price and product quality. The purchasing of the required materials is performed on the bases of both inventory positions and production plans. 2.1 Process planning According to Usher, process planning is the function that translates a set of design requirements and specifications into a set of technologically feasible instructions describing how to manufacture a piece part or assembly (Usher, 1996a). This function encompasses many different tasks including the determination of the routing of a part, the processes involved in its production, the process parameters, the machines, and the tooling. As can be deduced from the definition of process planning by Usher, process planning is positioned between the product design process and the production planning process. Most automated process planning systems need a detailed design as input information. The output of a process planning system is a technological plan that in turn serves as input for a production planning system. During process planning neither the competition of different orders for the available resources is taken into account nor the possibility of making design alterations. Process planning activities determines the production activities and the way in which they are executed. Process planning must take the characteristics of the products, orders and resources into account in order to generate a technological plan. Some typical process planning activities are (Houten van, 1991): interpretation of the product model, i.e. feature recognition, (machine) tools selection set-ups determination, fixture design, machining sequence determination tool paths calculation, cutting conditions calculation, NC program generation. Depending on the type of manufacturing environment considered, Computer Aided Process Planning (CAPP) systems can perform the process planning function in a generative way, a variant-based way or a combination of both. When the products, or parts of the products, to be manufactured can be grouped into families, it may be possible to draw up template process plans for each family. The eventual process plans are drawn up by modifying the

25 MANUFACTURING PLANNING 11 template process plans according to the specific needs. This is called variant process planning. For manufacture-to-order environments, this method is often not suitable, due to the high variations in products and orders. Also, the situation on the shop floor is changing frequently and significantly. For this reason, in manufacture-to-order environments, where production takes place in small batches, the generative process planning method - where process plans always are drawn up from scratch - is more suitable. While the variant method relies upon existing plans as a basis for process planning, the generative approach relies upon manufacturing planning knowledge. Manufacturing knowledge is stored in the planning systems by way of rules, decision trees, decision tables, and algorithms. The knowledge capturing processes required to build generative process planning systems is time consuming and costly, especially because it is often aimed to keep the human interventions in decisions making to a minimum. Some attempts have been made to completely automate the generative process planning task. An example is the PART system for small batch part manufacturing, developed at the laboratory of Design, Production and Management of the University of Twente (Houten, 1991a)(Houten, 1991b). PART has been implemented as a commercially available software system, marketed by Tecnomatix Technologies Ltd. The software architecture of PART, as displayed in Figure 2-1, contains several modules of which the Functional Modules are subdivided in Phases that are executed as part of a selected scenario. A loading function enabling the selection of machine tools considering their limited availability has been developed by Lenderink as an extension of the PART program (Lenderink, 1994). User User Interface Supervisor CAD Interface CI Volume Editor VE Feature Recognition FR Machine Tool Selection MTS Jigs & Fixtures J&F Machining Methods MM Tool Selection TS Cutting Conditions CC NC Output Compiler NC Planning PL Modeller GPM n o p q r s t User Interface Tuning Common Database Interface Database Figure 2-1 The PART architecture (Houten, 1991b)

26 12 CHAPTER 2 Due to the difficulty in capturing the enormous amount of manufacturing knowledge, most generative process planning systems are limited to small manufacturing domains, such as milling, assembly, etc. While products become ever more complex, it seems to be too demanding to develop generative process planning systems that can deal with broad and multiple manufacturing domains. Also, the incorporation of a rather restricted number of manufacturing features limits the applicability of current CAPP systems. The developments of a number of laboratory prototypes in the late eighties showed that automatic process planning requires a feature-based workpiece description as a starting point (Tönshoff, 1993). This information may be retrieved by means of the recognition of features, which requires a thorough analysis of the detailed product model. Subrahmanyam and Wozny review a variety of feature recognition techniques (Subrahmanyam, 1995). CAPP systems that combine properties from both generative and variant process planning are called semi-generative CAPP systems. By incorporating standard procedures, like decision tables or template operation sequences, these systems need a high degree of human interaction. A large number of different planning tasks have to be performed in order to obtain a complete process plan. The number of process planning activities is particularly huge in manufacture-to-order companies that produce many different types of products. Due to the high variety of parts and resources it is impossible to specify in advance a fixed sequence of planning activities (scenario) that is generically applicable for all products and/or processes. In other words, the different products need a planning method tailored to the productdependent requirements and constraints. For instance, for a specific product manufacturing process, it may be necessary to group a number of features in one set-up in an earlier planning stage, e.g. due to a time-consuming fixture-design activity. Larsen and Alting suggest that in general process planning includes two major phases, the first to be classified as time-independent and the second as time-dependent (Figure 2-2) (Larsen, 1992a): Analysis phase. Technical analysis of candidate solutions Selection phase. In literature the terms macro process planning and micro process planning are often used for the time-independent rough planning and the time-critical detailed planning part. Sheng Process Planning Analysis phase Selection phase Traditional constraints Production Planning Time independent Time dependent Figure 2-2 Analysis and selection phases in process planning (Larsen, 1992a)

27 MANUFACTURING PLANNING 13 Design Layer Manufacturability Layer Macro Level Process Planning Layer Micro-Level Manufacturing Layer Conceptual Design Selection of Manufacturing Processes Selection of Operations Scheduling Embodiment Design Manufacturability Analysis Operations Sequencing Shop Floor Control Detail Design Optimal Process Plan Generation and Selection Figure 2-3 Macro and micro process planning layers (Cay, 1997) and Srinivasan suggest that macro process planning refers to the determination of a feature sequence through the analysis of geometric and process interaction. Micro process planning refers to intra-feature planning, i.e. the optimisation at the feature level (Sheng, 1996). Cay and Chassapis define macro process planning as the manufacturability layer while the detailed process plans are generated through micro process planning (see Figure 2-3) (Cay, 1997). Other authors discern machine-independent (macro) and machine-dependent (micro) process planning. In this thesis, the prefixes macro and micro indicate the level of detail in the process plans. Macro process planning concerns the rough technological planning of production activities on an aggregate level. Micro process planning concerns the detailed technological planning of production activities resulting in the manufacturing instructions. Due to the diversity and product dependency of process planning sequences, the prefixes do not specify a fixed set of clearly defined process planning (sub)tasks Feature interaction and sequencing Many different approaches to operation (manufacturing feature) sequencing for CAPPsystems have been proposed by various researchers; see e.g. the reviews by Usher (1996b), Kiritsis (1995), and Lueng (1996). However, the allowable number of features and constraints is often too low to be practical for dealing with parts with many features in complex manufacturing environments. An example is the scenario presented by Sheng and Srinivasan for macro process planning based on feature commonalities (setups, tooling and cutting fluid) that considers environmental aspects (Sheng, 1996). This algorithm is hard to apply if more than 20 features, say, in for instance five set-ups on more than one machine tool are considered. Yut and Chang present a planning heuristic to automatically generate a process plan in the Quick Turnaround Cell (QTC) (Yut, 1996)(Chang, 1990). The process plan is represented as a containment hierarchy consisting of machine setups, fixture setups, tool setups, and operations (Figure 2-4). The heuristic aims at minimising the number of fixture and tool setups. Like most other heuristics in feature sequencing, the planning heuristic is still far from the ideal solution in automated process planning. The precedence graph represents a simplifying assumption that ignores many of the relevant feature interactions.

28 14 CHAPTER 2 Process Plan machine-setup1 fixture-setup1 tool-setup1 tool-setup2 op1 op2 op3 op4 op5 fixture-setup2 tool-setup3 op6 op7 Figure 2-4 Process plan representation (Yut, 1996) For a CAPP-system to handle complex parts, representing a large number of interacting features, an efficient method is needed to explore and reduce the size of the search space containing valid operation sequences. Usher and Bowden have developed a genetic algorithm strategy for this purpose (Usher, 1996b). The results of their study show that improvements in the search for optimal sequences in process plans are possible through the incorporation of feasibility constraints. Five feasibility constraints (location reference, accessibility, non-destruction, geometric tolerance and strict precedence) provide the system with the capability to define a set of precedences between the features of a part, resulting in the construction of a feature precedence graph (FPG). Sequences that satisfy all the feasibility constraints can then be judged and ranked based upon optimisation criteria (number of set-ups, continuity of motion, loose precedence). Resource-independent optimisation criteria are used instead of calculating cost and time, since no resources have been selected yet at this point in the planning process. After the set-ups have been determined, the final sequencing of them is often determined based on either one or both of the following rules of thumb (Ong, 1996): The setup with the most critical tolerance specification is ranked last. The first setup is one with the least accurate tolerance specifications (Boerma, 1990) The setup with the most number of features is ranked first. Hwang and Miller describe a hybrid blackboard based model that interactively combines feature sequencing and process plan generation (Hwang, 1995). The blackboard approach enables the integration with other planning and control systems, e.g. the coupling with a shop floor control system for modifying the plan due to unexpected situations on the shop floor. The use of blackboard systems in manufacturing is further discussed in Section After a view on production planning has been presented in Section 2.2., the integration of the process planning and production planning processes is further discussed in Section Artificial Intelligence in Process Planning Artificial Intelligence (AI) is often used in attempts to develop so-called intelligent systems for CIM solutions, especially in the field of automated process planning. Early AI technologies which were, and still are, commonly used for knowledge representation in expert systems include production rules and uncertainty rules, semantic nets and frames (Gu, 1995).

29 MANUFACTURING PLANNING 15 Increased integration capabilities are pursued by applying more recent AI technologies like fuzzy logic, neural networks, genetic programming, case-based reasoning, and agent-based concepts. An extensive literature review on applications of AI methods in manufacturing engineering has been presented by Dini (Dini, 1997). Most AI research in process planning has been performed on the application of neural network systems for automatic process planning. Neural networks are massively parallel interconnected networks of simple (usually adaptive) elements and their hierarchical organizations which are intended to interact with objects of the real world in the same way as biological nervous systems do (Kohonen, 1988). An overview of neural network systems in manufacturing planning is presented by Zhang and Huang (Zhang, 1995). A more recent development regarding neural networks for automatic process planning is the methodology for the selection of machining operations and their within-operation sequence of features for rotational (axi-symmetrical) parts at the École Polytechnique, Montreal, Canada (Devireddy, 1999). Possible advantages of neural networks in comparison with expert systems are related to their potential ability to adapt to ever-changing and complex environments that cannot be planned on the basis of knowledge rules. The research developments on neural networks are, however, not that promising yet for handling complex manufacturing environments. Like expert systems, the problem domain for neural networks must be very narrow in order to achieve promising results. For instance, the neural networks approach of Devireddy, mentioned above, can only deal with rotational parts which is one of the easy domains in automatic process planning. It is expected that neural networks will at best fulfil only a supporting role in near future CAPP systems. An important drawback of most AI concepts, but particularly of neural networks, is its structural incomprehensibility. It cannot be understood how the output relates to the input. This makes it basically unsuitable for combined human-computer decision making. 2.2 Production planning Production planning concerns the time-based allocation of orders to resources. In the remainder of this thesis, a rough distinction is made between capacity planning and scheduling as two production planning functions on different levels of aggregation. Capacity planning in a manufacture-to-order environment comprehends the rough production planning activities with an average time fence of three months. Scheduling, performed on a lower level of aggregation, concerns the short term (one day to two weeks) allocation and the sequencing of the jobs to the resources on the shop floor. The distinction of different planning levels based on the time period considered and the amount of detail in the plans, is referred to as Hierarchical Production Planning (HPP) (Hax, 1975)(Hax, 1984). The main advantages of HPP are the reduction of complexity and the possibility to deal with incomplete information. Frameworks for HPP architectures differ greatly, depending on the type of planning. The best known framework is the MRP II

30 16 CHAPTER 2 Technological planning Company management Production planning Strategic Long-term capacity planning Tactical Macro process planning Order acceptance Resource loading Operational Micro process planning Scheduling Figure 2-5 Planning tasks in manufacture-to-order environments (Giebels, 2000) (see Section 2.2.1) framework which is based on material co-ordination. A hierarchical planning framework for pull systems based on the limitation of the workload (see Section ) has been presented by Hopp and Spearman (Hopp, 1996). In this thesis, the framework depicted in Figure 2-5 is taken as a reference for hierarchically classifying the various manufacturing planning tasks to be performed in a manufacture-toorder environment. In the Sections and , the depicted production planning tasks are discussed. Unfortunately, aggregate planning techniques are highly neglected in production planning and control (PPC) systems. Some mathematical programming models have been developed for the implementation of a HPP system based on three planning steps. First, the capacity requirements for aggregate product types are roughly planned. The product types are subsequently disaggregated into product families, and finally, the items are assigned to detailed time slots (Zijm, 1991)(Mehra, 1996)(Kira, 1997). In production planning, the critical factors that highly influence the planning of new production orders must early be recognised by analysis, simulation or prediction. Common critical factors are bottleneck machine tools, operator availability, stock levels, buffer capacities, etc. The changing relations between the critical factors, the differentiating probabilities of occurrence, and the various consequences, must be dealt with in solving the planning problem. The handling of dynamic combinations of critical factors needs special attention in future production planning and control systems. The critical factors do not only refer to the avoidance of bottlenecks in production. A method that also takes quality and cost related aspects into account is likely to perform better than one that handles these aspects separately. At the moment, it is extremely hard to design, let alone to mathematically solve realistic models of an all-encompassing production planning problem. Therefore, the goal must be to develop a manufacturing planning and control system that is capable to support the planning and the co-operation between various experts (either humans or computer systems).

31 MANUFACTURING PLANNING 17 Because planning is always performed far before the actual execution of the plans, both researchers and industrial planners often do not consider the operational aspects. Selforganisation by the workforce on the shop floor may, for instance, affect the usefulness of the plans (Ahrens, 1996). In particularly, the unpredictability of irregular effects sometimes decreases the fitness of pre-specified production plans significantly. Mainly due to the reasons mentioned above, early concepts for production planning were barely addressing models and methods from Operations Research (OR). Just recently, some more sophisticated OR models are being applied in real production planning systems. The lack of planning capability in ERP (and MRP) systems evokes the integration of tools for automatically planning into these systems, especially for the lower-level planning (viz. scheduling) functions (see Section 2.2.3). Recently, some new concepts for higher planning levels have been suggested (see Section 2.2.2) The increased versatility of the production planning problem makes it impossible to stick to the so-called pull control systems, like Just In Time (JIT), Kanban, and ConWip. The characteristics and usefulness of these pull systems have been described by Hopp and Spearman (Hopp, 1996). By not planning ahead, but just imposing generic control rules, these systems may function adequately in MTS and ATO environments where a small mix of products is produced. It is, however, impossible to deal with many different orders with their own (customer dependent) characteristics, as is the case in manufacture-to-order environments. In the following sections some aspects of production planning are discussed in more detail. First, a brief discussion on current Enterprise Resource Planning (ERP) systems is presented (Section 2.2.1). Next, the description of the production planning task is split up in the longer term capacity planning task (described in Section ) and the short-term scheduling task (Section ) Enterprise Resource Planning (ERP) systems Enterprise Resource Planning (ERP) systems have been developed to integrate and coordinate various business processes, such as sales, purchasing, production, engineering, financing, etc. An ERP system generally consists of various generic software modules and databases that are combined depending on the specific company characteristics and demands. For the implementation, also the company structures and procedures may have to be changed. A major aspect of the implementation is the determination of the data (Bills of Material, Lead Times, etc.). The accuracy and reliability of the data is often a problem. Lead times, for instance, depend on the utilisation of the shop floor, which highly limits a reliable determination on beforehand. ERP systems make use of so-called Work Flow Management for the structuring and coordination of task execution and information processing. The various information processing activities performed in the company have to be controlled and tuned in order to make interactive decision making possible. An important feature is the feedback of information to inform the planners about order progress and other performance indicators. A shop floor control system is necessary to acquire the data for such feedback information (see also Chapter 4). Unfortunately, adequate SFC systems, not being the same as shop floor monitoring systems, are often not included in an ERP implementation.

32 18 CHAPTER 2 Most ERP systems are based on a Manufacturing Resource Planning (MRP II) system. MRP II systems determine aggregate production plans on the basis of a forecast of the demand for end products, the Bills Of Materials (BOM), prespecified lead times, and the current inventory levels. The resource loading profiles are specified in the output of the MRP system by means of a Capacity Requirement Planning (CRP) module. In CRP, the cumulated required capacity is compared with the available capacity. Exceeding resource levels in ERP systems are generally shifted to the next loading period, regardless of precedence constraints. Hence, no intelligent finite capacity planning but just an afterward check on available capacity is performed in the CRP module. In practice, this leads to infeasible production plans, which is even made worse by assuming fixed lead times for establishing the production plans. Lead times should preferably be a result of the planning process due to the fact that waiting times, that mainly determine the lead times, are predominantly dependent on the amount of work, as well as the work mix, on the shop floor. Moreover, the planning and control of the amount of work is exactly one of the major concerns of Production Planning and Control (PPC). A more detailed description of the major drawbacks of MRP II has been outlined by Darlington and Moar (Darlington, 1996). Today s ERP solutions still lack the capability of handling complex and dynamic environments adequately. Although coping with versatile situations concerning external market demands and internal processes is an obvious requirement for future ERP systems, the traditional approach seems to be unsuitable for addressing these requirements. Versatility in external demands concerns product features, delivery time, quality, and price. Customers will more and more influence these input parameters for manufacturing. Furthermore, also customers diversify more and more, leading to a high variety of demands that future ERP systems must be able to handle. The dynamic market behaviour implies the need for more flexibility in planning and control of the internal process in order to cope with WIP fluctuations, varying product structures, plan reliability, aggregation of process information, shifting bottlenecks, a changing order mix, etc. There is a risk of rising stocks due to unjustly maximising local performance, particularly when there is a situation of excess capacity combined with one or more bottlenecks. Optimized Production Technology (OPT), later extended to what has become known as the Theory of Constraints (TOC), attempts to minimise this risk by balancing the bottlenecks in the production instead of just executing the plans from the Manufacturing Resource Planning (MRPII) system (Jacobs, 1984)(Goldratt, 1987)(Smith, 1994). A major disadvantage of the TOC is the assumption of a stable environment (stable order mix) with a single bottleneck. Therefore, the TOC method only achieves satisfactory results in line or semi-line production of a single product family. A more general disadvantage of current ERP systems is their extensiveness, which easily results in outdated data, expensiveness and lengthy implementations, and high maintenance costs. Likewise it is difficult to reconfigure a manufacturing system fast and easily in order to respond to changing production demands. Concluding, there is a need for finite capacity planning functionality that can deal with today s complex and dynamic manufacturing environments characterised by high product variety and short delivery times. Particularly in manufacture-to-order environments, there is a need for computer support in concurrent manufacturing planning, so as to integrate logistic and technological planning. In the remainder of this chapter, aspects of concurrent

33 MANUFACTURING PLANNING 19 manufacturing planning and the consequent requirements for developing planning methods are elaborated in detail. In the next chapters it is argued that a flexible control architecture and generic information structure are also needed for building manufacturing planning and control systems that can cope with complex and dynamic environments Capacity planning There are a number of different planning functions to be performed on the higher production planning levels. The term capacity planning is generally used for all these different planning functions collectively. When classifying the capacity planning functions for manufacture-to-order environments, the following distinction can be made: Long-term capacity planning: Investments in capacity (resources, stock-levels, operators) on the basis of market expectations (sales forecasts, strategic product types). Capacity planning functions in the order acceptance phase: Analysing the consequences to be expected for the workload by accepting a specific order. Analysing or determining the delivery dates for the individual orders, taking into account the expected costs related to a possible shift or interruption of other orders. For instance, it is analysed what the impact of (rush) orders is on the loading plans and what the expected additional costs are due to the resulting alterations of these plans. Resource loading functions in order to enable feasible lower-level schedules: The recognition of capacity problems and the planning of additional capacity for specific time periods at an early stage, when possible. E.g. to plan work overtime, hire additional operators, subcontract specific order(-step)s. Allocation of order(step)s to time periods. Interaction with process planning to determine routings. The strategic long-term capacity planning functions are beyond the scope of this thesis. Only the tactical (order acceptance) and the operational (resource loading) parts of the capacity planning task are considered here Capacity planning in the order acceptance phase The order acceptance phase differs significantly for various manufacturing environments. In a make-to-stock (MTS) manufacturing environment, the orders are generally accepted with a short delivery date if there is sufficient inventory of the demanded product. If not, a later delivery date is agreed with the customer, based upon the expected stock fill rate. In manufacture-to-order environments, on which this thesis focuses, orders should be accepted in such a way that the revenues exceed the extra costs for the company due to, for instance, material use, tool wear, or plan disturbances which may lead to e.g. overtime work. A capacity planning system that quickly can analyse the impact of potential orders on the capacity plans is therefore of major importance for the order acceptance phase. There are three important aspects to be considered for the order acceptance phase.

34 20 CHAPTER 2 The first is the capability to manufacture the ordered product. The resources in the company all have their (interdependent) technological capabilities that together determine the aggregate capability of the manufacturing company. These capabilities primarily determine the make or buy decision regarding the whole product or some parts of the ordered product. If some product parts or process steps must be subcontracted, this will generally increase the delivery time of the ordered product, which is the second important aspect of order acceptance. Next to the external processes like subcontracting, transport and materials acquisition, both the future availability of the resources and the processing times determine the possible delivery time of an ordered product. The lead time of the orders is influenced by the amount of Work In Process (WIP) on the shop floor, the routing of the product (possibility of avoiding bottlenecks), the production/transport batches, and the priority of the order. The relation between these aspects can not easily be analysed due to the dynamics and the complexity of manufacturing environments. This problem is further elaborated in the next section on resource loading. Not only the production lead times, but also the time required for performing the engineering activities determines the eventual delivery time. In many manufacture-to-order companies, the engineering lead time is longer than the production lead time. And, what is even more important, the due dates for the engineering departments are often not considered as hard constraints by the engineers. The relative unpredictability of engineering activities and the lack of concurrent manufacturing planning in most companies lead to an unnecessarily high unpredictability of the production processes. The third aspect of order acceptance is related to the costs and revenues that result from accepting the order. Setting the price of an order is limited due to competition. It is, however, expected that the relation between the price and the delivery time will receive more attention in the near future, certainly when also more sophisticated capacity planning and cost estimation systems become available that can uncover the strong relations between these aspects Resource loading Resource loading is necessary to enable feasible schedules on the shop floor. A badly balanced capacity plan leads to additional costs and/or to exceeding the due dates. In case of a lack of capacity for given production processes, many delays will occur, priorities will frequently change, additional overwork will have to be planned, customers will get dissatisfied, etc. On the other hand, if there is excess capacity, resources will be underutilised and there is a risk of rising stocks due to unjustly maximising the local performance. Until now, most manufacturing planning systems are based on the MRP II principles (see Section 2.2.1) without paying attention to finite resource loading techniques. A review of capacity planning techniques within standard software packages by Wortmann et al. (1996) does not even discuss existing finite resource loading techniques. The reason for this was that most standard software packages at that time were, and still are, based on MRP II using its Capacity Requirement Planning (CRP) and long-term Rough Cut Capacity Check (RCCC) modules. By only determining the capacity requirements without trying to shift

35 MANUFACTURING PLANNING 21 (sub-)orders in time, the costs and feasibility of the plans are negatively influenced. Planning overtime work is not popular and, therefore, expensive. The option of subcontracting is also costly, particularly on the short term, and it generally increases the total lead time which can easily result in exceeding the delivery date. It is of major importance for manufacturing companies to perform the resource loading function more effectively. Computer support is absolutely required for the handling of the huge amount of information concerning resource loading decision making. Thereby, it is the accuracy and completeness of the data that primarily determines the quality of the resource loading result. Due to a lack of accuracy of the data, current capacity planning systems perform badly. Perhaps the biggest problem is the assumption of fixed (load independent) lead times that are taken as input for the capacity plans. In this way, it is ignored that lead times actually are the result of the resource loading process. An increased workload in the shop will increase the average lead times due to increased queueing times. Because of the fact that detailed schedules are not yet drawn up by the resource loading function, it is impossible to get exact values for the lead times of the order (steps) in resource loading. The only way to determine the lead times is by prediction. If relatively much information is known about processing times and detailed routings of the ordered products, it is possible to apply queueing models for predicting the lead times. An extensive survey of queueing theories in manufacturing is recently published by Govil and Fu (Govil, 1999). The application of probabilistic queueing models for resource loading has been studied at the laboratory of Production and Operations Management, University of Twente (Zijm, 1996)(Buitenhek, 1998). An algorithmic framework for the integration of resource loading and material co-ordination based on queueing networks is presented by Zijm (Zijm, 2000). This model is applicable for all manufacturing environments except for manufacture-to-order environments due to the lack of detailed routings and processing times in these environments. Further research is required regarding the application of queueing models in these environments. The impact of workload on lead times has explicitly been recognised in the research on Workload Control systems, see e.g. Bertrand and Wortmann (1981). The success of JIT control implementations in Japan made researchers in the Western world even more enthusiastic about limiting the workload in the shops. Wiendahl proposes load-oriented order release and presents various methods for analysing the status of the shop regarding workload, utilisation, etc. (Wiendahl, 1993). A modified kanban control system for limiting the number of jobs in the system, named Constant Work In Process (CONWIP), is presented by Spearman et al. (Spearman, 1989) The fact that, in manufacture-to-order environments, many (detailed) engineering activities still have to be performed at the time resource loading is performed highly influences the requirements for building resource loading systems for such environments. First, the engineering activities themselves must be included in the capacity plans. Because product designers and process planners are relatively highly educated they are also rather costly for the company. It seems, therefore, rather logical that the engineering capacity often becomes a bottleneck in the life cycle of the orders. Second, the information that is available for drawing up the capacity plans is not planned in detail by the engineers yet. Therefore, one has not only to deal with the uncertainty that results from the stochastic characteristics of manufacturing (randomness, such as machine breakdowns), but also with the uncertainty

36 22 CHAPTER 2 that results from the rough status of the technological plans. The randomness in manufacturing can be anticipated by e.g. including slack times and alternative routings in the plans. Values for slack times can be determined on the basis of an analysis of parameters like Mean Time Between Failures (MTBF). The latter, uncertainty due to incompleteness of the technological plans, generally has a greater impact on the planning strategy. The completion of the information does not occur randomly, but is the result of the resource loading process itself and future planning processes that can hence be directed towards the best planning decisions. Consequently, the uncertainty is not only of negative influence to the resource loading process. It also leaves space for reacting to future events. In this thesis, a resource loading concept is proposed based on modelling the uncertainties as reliably as possible. One of the most difficult aspects of resource loading is the accurate prediction of the available capacity defined as hours per time period - of groups of resources due to the complex constraints and dependencies between the resources. In practice, multiple resources together determine the available capacity of a workstation which is actually rather dynamic. Especially the operators can hardly be modelled as static resources. In manufacture-to-order environments, the prediction of available capacity becomes even more difficult. When less information is available on routings and processing times, it becomes almost infeasible to predict future bottlenecks. And exactly the bottlenecks will determine the priorities for the later planning processes that subsequently determine the availability of workstations, e.g. by shifting operators. In other words, the variance of machine utilisation will be relatively high in manufacture-to-order environments, which again worsen the predictability during resource loading. Furthermore, the predictability is decreased by current trends on the labour market towards more flexible contracts. However, again these uncertainties caused by flexibility may also be used to improve the eventual performance of the company due to increased possibilities for problem solving. While the increased uncertainty and flexibility is not necessarily a problem in production planning in general, it is unarguably a problem for applying mathematical programming models like Linear Programming (LP). Much research on the application of LP models in manufacturing has been performed, but a realistic model of a manufacturing situation is difficult if not impossible to achieve (Hopp, 1996). As a result, only a few real life implementations of LP models in manufacturing have actually been reported successful. However, the research on solving LP models is still in progress, which may improve the applicability of such models in practice. Hans suggests an LP model for finite resource loading, in which a column generation scheme is developed to solve the problem (Hans, 2000). The number of columns in the formulation may be very large which makes it possible to formulate realistic models of a complex production planning problem. Adil suggests a large-scale LP model and a column generation scheme for production planning integrated with costs, product design features and process planning aspects (Adil, 1998). A final remark regarding resource loading concerns the transparency of resource loading techniques to the human planners. Because it is unthinkable today that all capacity planning tasks can be automated completely, it is important to develop resource loading systems that support the human planners in making their decisions. The decisions made by the human planner will be of a bad quality when the choices that are automatically generated cannot be understood by them. It is very likely, then, that these proposals are ignored completely.

37 MANUFACTURING PLANNING Production Scheduling Production Scheduling, performed at a more detailed level, concerns the short term (e.g. one week) allocation and sequencing of (auxiliary) jobs to and on resources on the shop floor so as to optimise overall performance objectives. Conditions other than resource availability (e.g. sequence-dependent machine set-ups, locations of transport devices, process planning effort, outsourcing, material purchasing) should also be taken into account. Some major difficulties with respect to scheduling in manufacture-to-order environments are: Incomplete and uncertain information regarding set-up and processing times due to missing detailed process planning information (on-line process planning) and a lack of experience (much variety in production activities). Unpredictability of the shop floor situation in the short term due to the impact of randomness in production activities (time and quality) and available resources (e.g. machine breakdowns). External sources of uncertainty, such as delays in the delivery of (raw) materials, changing customer demands, and arrivals of rush orders. The need to take into account multiple resources required for the execution of a production activity. The existence of conflicting objectives and the general complexity of a situation. In general, there are two types of scheduling methods: mathematical approaches for developing (deterministic or stochastic) scheduling algorithms that aim at optimising some performance index, and more heuristic approaches that are generally more applicable in a practical manufacturing environment. Most algorithms developed for deterministic and stochastic scheduling in the mathematical research field only deal with rather small problems in comparison with the problems observed in practical manufacturing environments, and are, therefore, hard to apply in complex job shop situations. The most important practical scheduling approaches are distributed scheduling, scheduling using priority rules, and scheduling based on local search algorithms. Scheduling methods based on priority rules (e.g. Earliest Due Date) are quite popular for application in practice. Such methods often use multiple simulation runs to choose the best (set of) priority rules for meeting the objectives. The performance of priority rules is, however, often quite weak. Due to the inherent unpredictability of shop floor situations, much research has been performed on distributed scheduling techniques. Hussain and Joshi present a rather generic distributed algorithm for scheduling jobs in a manufacturing system (Hussain, 1998). The deterministic algorithm provides a mathematical basis upon which independent entities of the shop compute their own local objectives and their contribution to the global one. A shortest path algorithm provides the minimum costs and the best machines to process an operation. Dewan and Joshi developed an auction model that can be used for distributed job shop scheduling (Dewan, 1999). By applying a market-alike approach, it is expected that the local decisions will lead to acceptable global performance.

38 24 CHAPTER 2 Although much research has been performed on the development of practical scheduling methods, Wiers (1997) states that the complexity and instability of production systems are still underestimated when applying most scheduling techniques. Hence, he concludes that the human scheduler is important in dealing with today s scheduling problems. Because of the complex mix of requirements and constraints, logistic optimisation of the manufacturing process in practice means an optimisation of some specifically chosen parameters. Examples of frequently used production parameters are machine utilisation and due date performance. Machine utilisation is often maximised in order to achieve an efficient manufacturing process. This is, however, not a task for the shop floor scheduling function, since the number of resources available and the number of orders to be produced are normally assumed to be fixed. This holds in particular if shop floor scheduling is executed as a part of a more comprehensive PPC architecture. Capacity requirements are, then, already anticipated by other modules. In particular, a sound resource-loading module is essential. It is, with these assumptions, more intelligent to balance machine utilisation in order to avoid the occurrence of bottleneck machine tools on the shop floor. The due dates can, then, be met more easily and it may be possible to load new orders to the shop floor so as to enhance the throughput of the factory. With respect to meeting the due dates, the minimisation of maximum tardiness can be considered as the most practical and successful one for job-shop environments. For this criterion, a multi-resource scheduling methodology has been developed, based on a resource decomposition procedure (Meester, 1996)(Schutten, 1996). Multi-resource scheduling aims at minimising the maximum tardiness of the suborders by applying a Critical Path Analysis method. The efficiency of the factory is negatively influenced when additional costs have to be made in order to meet the primary objective of the scheduling process (i.e. the due date). For meeting the due dates, it may be necessary to either plan some overtime work on bottleneck machine tools or to outsource some production activities. One may also decide to postpone delivery although this decision mostly implies the payment of penalty costs to the customer. When alternative solutions exist for the division of work over resources, the variable costs must be minimised (Liebers, 1998). For our purposes, the main objectives in scheduling are to minimise tardiness and variable costs. 2.3 Integration of Process Planning and Production Planning Before the introduction of the automation of manufacturing planning decisions, the technological and logistic planning tasks were normally planned simultaneously at various levels of aggregation. Supervisors on the shop floor determined the routing of the products through the factory on the basis of technological and logistic requirements and constraints. But the efficiency level was low, which expressed itself in a low machine utilisation, high storage levels, and an excess of auxiliary resources like tools and fixture parts. The continual developments in computer technology made it possible to perform planning tasks faster and more reliably. However, until now it has not been possible to develop planning

39 MANUFACTURING PLANNING 25 methods that take the broad scale of manufacturing planning constraints and requirements into account. There are a couple of reasons for this shortcoming. Firstly, computers can only deal with formal information, which demands that all relevant information for planning should explicitly be determined and specified before it can be used as input in the planning systems. Consequently, it is impossible to use meta-knowledge and context information in the planning (Kals, 1998). Secondly, due to the difficulty in determining the right information, most information in the system is either outdated or incomplete. Therefore, the information in most planning systems is rather unreliable. Thirdly, as a result of the increased developments in information processing, the complexity of manufacturing environments and the accompanying amount of information has increased drastically. Fourthly, mathematical solution methods are incapable in handling complex problems, particularly when it is also aimed to achieve multiple goals as is the case in practical manufacturing planning. Finally, the variance in the input information that comes available from other aggregate planning systems is nearly always neglected, which leads to unreliable output information. Consequently, since more planning tasks have been automated on islands of automation, concurrency in manufacturing planning has become more difficult to achieve. This evolution has lead to far-reaching local optimisation while at the same time the proverbial walls between the diverse planning processes, departments and systems grew higher. The attempts to pull down these walls by providing an increased number of interfaces between multiple planning systems have not been successful because interfacing itself does not solve the problems mentioned above. There is a clear need for integrally managing all the information in the company. Hence, the problem of information management is further elaborated in Chapter 6. The separation of the manufacturing planning task and the distribution of the remaining tasks during the years of automation in manufacturing has been confirmed by the analysis on the terms for planning systems as used in Germany: The traditional German term Arbeitsvorbereitung covers both Arbeitsplanung (process planning) and Arbeitssteuerung (production control). Since production planning and control software entered the market, the term Produktionsplanung & Steuerungssystem was used in stead of Arbeitssteuerung. While this was a small difference in wording, it nevertheless indicated the evolving separation between process planning and production planning (Tönshoff, 1993). Not only were the distributed manufacturing tasks executed by several planners, departments and systems, but also the execution happened in a strict sequential way. This caused and still causes a problem regarding the possibilities of taking the constraints of other (subsequent) planning processes into consideration. According to Tönshoff, sequential manufacturing planning causes the following problems (Tönshoff, 1993): The workshop loading is not considered at the time of process planning. The dynamic nature of today s workshops demands the possibility of quick replanning. Up to 30% of all orders are not processed according to the initially produced process plans. The reasons for such deviations are events such as bottleneck resources, rush orders, machine disruptions and tool unavailability.

40 26 CHAPTER 2 The technological characteristics of today s workshops make quick replanning more difficult. A major aim is to transfer responsibilities from upstream departments to the workshop (production islands). Long throughput times. Both process planning and production planning aim at minimising costs within the limits of the constraints, while meeting the requirements. For process planning these requirements refer to meeting the product quality, while production planning should make sure that the due dates are met. If process planning imposes major constraints on the production planning task, these planning processes may be seen to be conflicting, where ideally they should be co-operating. The conflict is mainly caused by the fact that these separate planning processes do only take their own requirements, costs function, and constraints into account. In order to achieve global optimisation a more co-operative way of planning should be developed, i.e. an approach that takes into account both the constraints and requirements from process planning as well as the constraints and requirements from production planning. Also the objective function (cost function) should reflect these common interests. The complexity going with the development of systems integrating process planning and production planning is enormous. It is not easy, if possible at all, to formulate an overall objective function. A quantification of cost aspects is complicated and debatable. Next to the objectives concerning cost aspects also the objectives with respect to time aspects (e.g., meeting due dates, minimising lead times) and quality aspects (e.g., tolerances) are often difficult to define and are conflicting due to the diversity of expert views on the same problem. The same holds for the handling of the information in the system, to be discussed in Chapter 6. Furthermore, the autonomy in decision making by experts should be limited because most decisions affect the planning decisions of other experts (e.g., machine allocation, subcontracting). Finally, the uncertainty due to incomplete plans and unpredictable events (see Section 3.2) is a main but unfortunately often disregarded factor that contributes to the complexity of the development of integrated planning systems Integration concepts Many different approaches have been published, considering the integration of the traditionally divided process and production planning processes. These research activities have often resulted in an implementation. Review of these research projects are given by (Alting, 1989)(Elmaraghy, 1993)(Huang, 1995)(Cho, 1994) (Usher, 1996a). In this section, the most important integration concepts are discussed. Some concepts suggest the development of complete process plans containing alternative routings (nonlinear process plans), others propose a real-time generation of process plans (e.g. the PART system mentioned in Section 2.1) that can be integrated in a hierarchical planning approach. These principles refer to the sequence and structure of planning the logistic and technological aspects.

41 MANUFACTURING PLANNING 27 CAPP generates process plans - single domain - single part at a time - static resources FUNCTIONAL GAP a need for CAPP & PPC to have common definitions of plans and resources, and synchronize and communicate their needs with each other in real time need for replanning when a bottleneck is encountered on the shop-floor PPC production scheduling & control - multiple domains - multiple products - capacity planning - dynamic environment Product specific product geometry parts specifications product structure manufacturing specific resources resources capability process planning knowledge constraints technical data cost data DATA GAP a need for transferring, mapping and synchronizing the database information e.g. Resource status & process plans Product specific product & parts (BOM) linear process plans & times manufacturing specific production targets resources & capabilities resource capacity & loading Figure 2-6 Functional and data gap between CAPP and PPC systems (ElMaraghy, 1993) There are, however, more obstacles to an effective integration. ElMaraghy and ElMaraghy rightly noticed that the reasons for the lack of compatibility between CAPP and production planning and control systems relate as well to the functionality of these systems as to their use of data (ElMaraghy, 1993). The so-called Functional and Data Gaps between CAPP and PPC systems are shown in Figure 2-6. They suggest a dynamic and reactive CAPP environment, called RPE (Reactive Planing Environment) and an integrator module to be interfaced with existing CAPP and PPC systems. Resource and product aggregations are similar for the CAPP and PPC systems, making it possible to integrate on various planning levels. The data gap is further discussed in Chapter Non-Linear Process Plan (NLPP) Many products can be produced via several different routings through a job shop. The possible routings can be linked together in a graph, called a non-linear process plan (NLPP) (Figure 2-7). The versatility of possible routings which can be selected during production scheduling is enhanced by drawing up these NLPPs. The NLPP approach has been introduced in ESPRIT Project 2457 FLEXPLAN (Tönshoff, 1989). The FLEXPLAN research project had been succeeded by the ESPRIT project Concurrent Manufacturing Planning and Shop Control in Small batch production (COMPLAN) (Detand, 1993)(Detand, 1995). The COMPLAN project resulted

42 28 CHAPTER 2 MT 4 MT 9 os MT 7 oj MT 1 as os MT 3 oj aj and split or split os MT 5 MT 2 oj or join MT 8 and join Figure 2-7 Non-Linear Process Plan graph in an architecture and a software system for generating and using NLPPs. The COMPLAN system contains two main functional groups of modules for process planning and workshop scheduling respectively. One of the major process planning modules is the NLPP Editor Control Module which offers the function to manually edit non-linear process plan data in a graphical editor or to automatically generate NLPPs for a limited set of workpieces. A central relational database is used for ensuring the consistency of the data. Park et al. (Park, 1998) have presented a software implementation for the integration of process planning and scheduling, which is based on feature-based process nets and a genetic algorithm. The system contains three modules: a process planning module, a scheduling module and an interface module. The process planning module employs the concept of a process net model for the representation of alternative plans. The ICEM PART generative process planning system (see Section 2.1) is used for constructing these process nets. The scheduling module applies a genetic algorithm to search the process net while considering the current shop floor status. Erkoc et al. have developed a NLPP were the final selection is based on the criterion of minimising the total weighted tardiness without altering the relative sequences of operations in the schedule. (Erkoc, 1998) Milatovic and Raman present a heuristic for generating alternate machining sequences which takes two kinds of feature relations into account (Milatovic, 1998). First, the mandatory precedence constraints, which indicate that one or more features must be executed before other features. Second, the soft precedence relations which include the costs that vary with the different sequences. A method originally developed for vehicle routing problems is implemented for the generation of the sequences. An algorithm for solving the problem of scheduling with alternative machines has been developed by Nasr and Elsayed (Nasr, 1990). Nasr and Elsayed simplified the NLPPapproach by assuming a single process plan, except that each operation has alternate machine tools. To tackle this problem, Kim and Egbelu have developed a scheduling algorithm, referred to as the preprocessing algorithm, which can handle NLPPs by combining branch and bound and more general integer programming techniques to search the entire solution space of the problem (Kim, 1999). Palmer proposes to apply neighbourhood search techniques, in particular simulated annealing, for the integration of

43 MANUFACTURING PLANNING 29 process plan optimisation with job shop scheduling (Palmer, 1994). Here, the possible flexibility in the process plans relates to both alternative machines for operations and alternative routings in the plan. Because of the effect of alternative routing, also the transportation and set-up times are taken into account. Although by some authors the NLPP concept is considered to be a flexible process planning concept, it in fact increases only the versatility of possible solutions rather than being flexible. Disturbances and delays of production jobs cannot be handled in a flexible manner (e.g. by adjusting relevant process parameters). The diversity of disturbances and process alterations is too large to be able to anticipate on all possible situations by creating many alternative routings. Especially within a manufacture-to-order environment with few repeat orders it is not a practical solution. The amount of time and costs required for creating the non-linear process plans will in general not counterbalance the workshop performance improvements, because the number of products produced is low which leads to much redundant process planning effort. It appears that NLPPs can only be applied efficiently in manufacturing environments with a significant number of repeat orders. Only then, it is likely that most routings in the net will ever be used. The main disadvantages of NLPP are due to its non-integration characteristic resulting from the persistence in the sequential planning of the technological aspects before the logistic aspects. This drawback has also been recognised by Kempenaers et al., leading to the introduction of feedback loops (Kempenaers, 1996). The introduction of feedback loops in NLPP is called closed-loop process planning (CLPP) (Larsen, 1990) (Khoshnevis, 1990). In CLPP process plans are created based on the monitored feedback from the production planning mechanism. By taking the dynamic behaviour of the manufacturing system into account, CLPP intends to create only feasible plans with respect to the production facilities. Real-time shop status monitoring is necessary for the implementation of a CLPP system Hierarchical Process Planning Due to the amount of work that is required for building the NLPP networks, attempts have been made to subdivide the process planning in several hierarchical levels of aggregation. First, rough plans are established that are subsequently planned in more detail during the latter phases in the planning. Larsen and Alting propose a Distributed Process Planning (DPP) concept that divides the planning process in three phases, namely, preplanning, pairing planning, and final planning, to handle the process and production planning tasks simultaneously (Larsen, 1992a)(Huang, 1995). The phases are depicted in Figure 2-8. The preplanning phase is a technological analysis of the product requirements and resource capabilities. This is an off-line static activity. The short-term dynamic constraints are considered during the pairing planning phase which matches the required job operations with the available resources. Consequently, this phase requires a close link to shop floor control. The final planning phase prepares the manufacturing instructions for the selected resources, like the NC programs for the machine tools and the information about the preparation of tools and fixtures.

44 30 CHAPTER 2 Preplanning Equipment Machine operation capabilities Job Pre-process planning (operations) Pairing planning Shop status Maintenance work calendar Available operations Operation matching Job operation sorting Order/jobs manager intervention Final planning Prepair production Figure 2-8 Distributed Process Planning (Larsen, 1992a) Another hierarchical approach, addressing the integration of CAPP and Shop Floor Control is developed at the Texas A&M University (Cho, 1994)(Wysk, 1995). A process planning task graph is decomposed into a hierarchical set of process plans, that reflect the resource requirements at each level (shop, workstation, equipment) in the production control hierarchy. This results in a hierarchical graph with several levels of aggregation, proposing alternative routings at each level of production control. The process planning graph is not generated in line with the hierarchical production planning activities. The graph is presented hierarchically in order to achieve a better interface with the production planning activities. A Process Planning Architecture for Integration with Scheduling (PARIS), also a concept based on DPP, has been developed at the Mississippi State University (Usher, 1996a). A two-phase approach is proposed, instead of the earlier mentioned three phases in dynamic process planning. The planning process is subdivided into a static and a dynamic phase. A list of alternative plans is generated during the static phase by planning functions that do not rely on information concerning the given status of the shop-floor resources. While the static phase is performed, no orders are available for these products. Therefore, the system is not applicable to manufacture-to-order environments. The detailed planning tasks, which consider the operational status of the resources and the actual orders, have to be performed during the dynamic planning phase Blackboard-based systems The application of blackboard architectures can be very useful for the integration of the planning of multiple knowledge domains in complex environments. The blackboard basically functions as a communication platform for collectively generating solutions by independent knowledge sources (experts). Using a blackboard is particularly useful in solving some specific problems in concurrent manufacturing planning and control: Due to a lack of communication and experience, the designer and process planner are often not aware of production constraints as e.g. a temporary unavailability of a resource.

45 MANUFACTURING PLANNING 31 Many manufacturing decisions concern more than one planning aspect. For example, the decision to machine with maximum cutting speed makes only sense, when the requirements and constraints from PPC and process planning (using the product model) are considered simultaneously. Some developments for manufacturing planning and control based on blackboard systems have been demonstrated in recent years of which the two most important ones, regarding the integration of process and production planning, are mentioned here. A hybrid blackboard model for process planning based on forward chaining for feature sequencing and backward chaining for process plan constructing has been developed by the department of Industrial and Management Systems Engineering of the University of South Florida (Hwang, 1995). The backward chaining strategy offers flexibility to modify the plan due to unexpected shop floor situations. The Integrated Process Planning / Production Scheduling (IP3S) Shell, designed around a blackboard architecture for Agile Manufacturing, has been developed by Raytheon Electronic Systems and the Intelligent Co-ordination and Logistics Lab of the Carnegie Mellon University (Laliberty, 1996). The IP3S blackboard architecture emphasises a more generic integration framework, rather than committing to a prespecified decision flow as in most other integration concepts. The resulting shell, shown in Figure 2-9, provides a customisable framework capable of supporting a wide range of integrated process planning IP3S Control KB Agenda Controller GUI Blackboard Current Working Context Process-Planning KS Production-Scheduling KS Contexts Knowledge Sources Analysis/Diagnosis KS Event Queue Communication KS Outgoing Messages Enterprise- Level Planning System External Systems Incoming Messages Supplier Tool Shop Shop Floor Figure 2-9 The IP3S shell (Hildum, 1996)

46 32 CHAPTER 2 and production scheduling decision flows. The IP3S system has been validated in the context of Raytheon Andover manufacturing facility which has led to significant performance improvements (Sadeh, 1998). Two important advantages of the IP3S blackboard architecture are the possibility of flexibly changing control mechanisms (manual/automated) and the encapsulation of existing problem-solving systems as knowledge sources. The IP3S project has demonstrated that a blackboard alike architecture for co-ordinating planning and event-handling is a promising research direction for concurrent manufacturing planning and control systems Integration on shop floor level On the shop floor level, the planning decisions concerning detailed scheduling and micro process planning, respectively described in Section and Section 2.1, are generally made. When more than one machine tool is capable of executing the same activity, the resource allocation function will select one of these machine tools. The same holds for the selection of operators and the auxiliary resources, such as tools and fixtures. Because the main goal of the scheduling function is to meet the due dates or at least to minimise the maximum lateness of the orders, it is important to know which activities are critical for reaching these goals. The selection of candidates for alternative processing ways can be performed by a Critical Path Analysis (Zijm, 1995). In this way, iteration between scheduling and process planning is easily applied. An excellent performance can be achieved, with a minimum of process planning effort, while maintaining the advantage that no alternative process plans need to be generated in advance. The process planning function benefits from this iterative approach as well. Traditionally, parameter setting in process planning was based on minimising costs. The costs of tool wear are balanced against the assumed costs of machining (per hour), based on average machine utilisation. It is assumed that fast production is required. But this will not always be the case in a manufacture-to-order job shop. The overall costs of performing an activity are much higher on a machine tool that functions as a bottleneck, than on the same machine tool during a time period in which it is under-utilised. By applying an iterative approach it is possible to obtain more actual information on the priority of the activities to be performed, i.e. whether the activity is part of the critical path of an order or the activity is to be executed on a bottleneck machine tool. In such critical situations it will very likely be more economical to set a higher cutting speed or to select a better but more expensive cutting tool for executing the activity. On the other hand, in an under-utilised period, a lower cutting speed in combination with using a cheaper cutting tool may be more economical. Next to the mentioned non-conformity due to the increased dynamics, also the current advances in the physical understanding of the process, integration of machine tools and concurrent engineering introduce new dimensions to the optimisation problem. Examples are the component quality, flexibility and environmental aspects as mentioned by Sheng

47 MANUFACTURING PLANNING 33 Operation Machining time (min) ATO Machine 1 Machine 2 Machine 3 PTO Priority Figure 2-10 Magic matrix (Halevi, 1993) and Srinivasan (Sheng, 1995). Sheng and Srinivasan point out that these contemporary dimensions cannot be adapted easily to machining economics, as many of the effects carry uncertainty, discontinuity or present difficulty with respect to internal costs representation. Halevi suggests the application of his magic matrix for final resource selection and operation sequencing on the basis of either minimising costs or maximising production (Halevi, 1993). A matrix, as shown in Figure 2-10, is drawn up for every product to be produced. An absolute theoretical optimum (ATO) process is determined based on technological constraints independent of the machine it will be processed on. A planner (computerised or manual) generates a generic process plan using imaginary machines, resulting in the values for ATO. Next, the time or cost values are determined for all machine tools in the company. A practical theoretical optimum (PTO) represents the minimum time or costs to machine the operation on one of the available machine tools. This matrix is used as input for the scheduling process. Halevi suggest to use the same magic matrix for machine performance measurement, determining the company level of competitiveness, and machine procurement decisions (Halevi, 1997). Song and Choi use a similar kind of matrix for integrating route selection and determining production conditions like machining speed and overtime work for flexible cellular manufacturing with multiple machines per cell (Song, 1993). They formulate a Nonlinear Mixed Integer Programming (NMIP) problem in order to minimise the total incremental production costs. A problem with these matrix approaches is to collect and to update all the required data. Also the automation of the process planning part (planning with imaginary machine tools) needs more research. The automation is necessary, due to the increased process planning work resulting from the approach. When producing many standard features, the approach may be suitable. Otherwise, the approach seems to be too complex Integration of macro process planning and capacity planning By performing rough production planning early in the process of concurrent manufacturing planning, i.e. on the higher levels of aggregation, some difficulties arise. The problems have been mentioned in this chapter. A brief overview of the various problems encountered is given in this section. First, the allocation of specific resources to the orders will often not be possible on the higher levels of aggregation, as the process plans are not sufficiently worked out in detail

48 34 CHAPTER 2 yet. It has always been questionable whether the allocation of the resources required to perform production steps should be performed by the process planner or by the production planner. Both of them have a substantial interest in the resource allocation process. Consequently, the only suitable approach is to consider both technological and logistic aspects simultaneously. This does not only hold for resource allocation but also for routing determination, outsourcing decisions, etc. Second, not all production orders, routings and processing times are already known for the relatively long time period which belongs to a higher aggregation level. Consequently, it is impossible to perform resource specific load planning or scheduling on the higher levels, at least in the traditional way. A new method for resource loading is proposed in this thesis based on the modelling of the uncertainties, in order to make concurrent planning possible. Third, next to the production activities also the engineering activities must be included in the capacity plans. The engineering activities often determine a large part of the lead times of the orders in manufacture-to-order production Summary Machine allocation and routing determination are of major concern for both process planning and production planning. The selection of materials and production equipment is generally considered as a technological aspect, while logistic planning concerns the availability of resources. Especially in manufacture-to-order environments these aspects are directly related. For example, the choice to use casting materials in order to cut down processing times primarily depends on the workload of the required resources. On the other hand, logistic decisions regarding, for instance, batch and transport sizes highly influence the product quality aspects. The Japanese industry has clearly demonstrated this relation by implementing just in time (Kanban) production control strategies. Summarising, it is clear that order acceptance must take both technological and logistic aspects into account, not as independent but as interacting processes. From an overall perspective, meeting due dates and the minimisation of additional costs should be the basis for the determination of routings and the selection of machine tools. Since the information about additional costs and logistic constraints caused by individual resources can only be obtained during the scheduling phase of production planning, machine allocation must be performed just before actual production. Consequently, the detailed (machine dependent) process planning information must either be generated after the scheduling process or on beforehand by working out a sufficient number of alternative routings and alternative machining jobs in NLPPs. For manufacturing environments dealing with a high product variety and few repeat orders, we feel that the NLPP approach is too costly and time consuming (see Section ). Hence, concurrent planning systems on multiple levels of aggregation are required for adequate manufacturing planning. Planning techniques for both process planning and production planning must therefore have the possibility of concurrent interaction. That is much more important than the optimisation of the local performances. In the next chapter, the planning on multiple levels of aggregation is discussed. The aim of that is to discuss the possibility of the concurrent planning of manufacturing activities.

49 35 Chapter 3 Abstraction Planning In manufacture-to-order environments, a significant uncertainty still exists after a price and a delivery date has been negotiated with the customer. The engineering department roughly estimates the routing and the expected costs after which the probability of meeting the delivery date is analysed on the basis of the resource loads. While the engineering activities proceed, more and more information about the required production processes becomes available. On the basis of that information, the production planning decisions are taken, that in turn influence the subsequent engineering decisions. For example, the process planners should preferably avoid making use of machine tools that tend to become bottlenecks in the production. This scenario broadly corresponds with the course of actions performed within manufacture-to-order companies. A more generic scenario for this type of manufacturing environment can be specified as follows. A client order that enters the company initiates planning activities, resulting in various subactivities that are subsequently planned in more detail. These planning activities include all actions that are to be performed for executing the order, like engineering, purchasing, sales, or production planning activities. Based on the time scale, a hierarchical order structure is build up. The aggregate production activities that are planned for execution on long term are roughly planned in a macro process planning system and a rough-cut capacity planning system. The activities subsequently are split in subactivities that can be planned in more detail later on. A micro process planner and a detailed scheduling system put up the detailed plans for the production activities that must be executed on short term. Level of abstraction in the order plans Order 1 Suborder Suborder at present 1 day 1 week 2 weeks 3 months Detailed scheduling Rough cut capacity planning NC programming Macro process planning Figure 3-1 Generic hierarchical order structure

50 36 CHAPTER 3 Such a hierarchical order structure is shown in a generic and simplified form in Figure 3-1. The x-axis represents the logarithmic time horizon of the planning activities. The level of abstraction, i.e. inversely proportional to the amount of detail, is represented on the y-axis. Order 1 is split in a couple of suborders, which are subsequently subdivided again. The amount of detail in the plans for the (sub-)orders increases and the time horizon decreases when descending within the order hierarchy. This obviously implies that aggregate planning tasks such as rough cut capacity planning and macro process planning deal with activities that generally are performed in the long run, whereas detailed scheduling and micro process planning concern a short planning horizon. However, this does not mean that all the aggregate planning activities are performed before the detailed planning activities. Although the order life cycle normally will follow the way from aggregate planning to detailed planning, it may also occur within one order that some parts of the order are already planned in detail at a certain time while other (e.g. less critical) parts are only roughly planned at that time. The application of several levels of abstraction in the planning process is necessary for being able to reach the tactical goals, while at the same time one can deal with the short term planning issues. Likewise, it may prevent the undesired situation of too much replanning work to be carried out. This is achieved by avoiding too much detail in the earlier planning phases. The production environment is too dynamic to plan newly entered orders immediately in a detailed manner. Nevertheless, the aim is to recognise the probable occurrence of problems as early as possible. These problems can relate to either the manufacturability or the violation of delivery dates due to capacity limitations. In manufacturing environments where the situation on the shop floor changes frequently as is the case in manufacture-to-order environments - more abstraction in the planning of the processes leads to more stable plans (Washington, 1994). Washington shows that less abstraction leads to increased lost work. This is typically true for the earlier phases in the planning process, i.e. on the higher levels of aggregation. As the required detail in the plan increases, i.e. on the lower levels of aggregation, the cost of changing the plan may outweigh the extra added value. 3.1 Literature on abstraction planning A wide range of planning and reasoning concepts is related to abstraction planning. ABSTRIPS was the first system based on abstraction planning (Sacerdoti, 1974). In ABSTRIPS the plans at a specific planning level are planned completely before the next level is planned more specifically. Knoblock formalised the ABSTRIPS approach in ALPINE (Knoblock, 1991)(Washington, 1994). ALPINE produces abstraction hierarchies to help control the problem solver's search. ALPINE is used as a learning method in Prodigy. PRODIGY is a fully-automated planner that combines generative state-space planning (Prodigy4.0) and case-based (variant-based) planning (Prodigy/Analogy). A limitation of the Prodigy system is the requirement of a reasonably predictable application domain, which is seldomly the case in the real world. A mixed-initiative planning framework for combined human and automatic planning has been developed as an interface extension to the Prodigy system (Cox, 1997). With the integration of the human planner it

51 ABSTRACTION PLANNING 37 is aimed to achieve better plans than either the human or an automatic planner can create alone. It may be interesting to develop an implementation of the mixed-initiative Prodigy system in a manufacturing environment. An early system based on abstraction planning is a progressive horizon planning system presented by Rymon for improving imperfect plans (Rymon, 1993). The system works by constructing intermediate plans via a combination of plan sketching and selection/optimisation sub-processes, and then adapting these plans to reflect new information and goals. Washington describes a generic planning method that constructs plans under time constraints while adapting to changes in the world (Washington, 1994). The planning method represents the dynamic and continuous nature of information and events in the real world by incrementally building up plans at multiple levels of abstraction. The approach merges planning and replanning into a seamless and inseparable process. A plan is a changing data structure, changed by the planner who is adding actions, by the world which changes in unanticipated ways, and by the planner again who revises the plans to match the world model. (Figure 3-2). Fujita and Lesser describe a method for deadline-based scheduling of subtasks to achieve the goal criteria specified for the higher-level task (Fujita, 1996) The characteristics of the (sub)tasks with respect to expected output quality and task duration are specified in terms of probability distributions. A five-step method, called Multi-Agent task Decomposition and Scheduling (MADS), has been described which decides how many subtasks are needed and how to partition and schedule the activities among these tasks. Experimental results show how different schedules and number of tasks are generated as a result of varying goal propagating world information world info world information sensors world abstract world info goals current state abstraction propagating initial goals abstract goals knowledge about the world effectors plan actions unmet expectations propagating intermediate goals abstract expectations intermediate goals unmet goals added steps replanning planning revised or removed steps Figure 3-2 Planning may be viewed as a number of processes modifying a global data structure (the plan) and reacting to changes in it (Washington, 1994).

52 38 CHAPTER 3 criteria in order to obtain a successful answer while meeting the task deadline within which the answer is needed. As part of the experimental results, they explore different criteria for rescheduling based on the degree of deviation that has occurred in actual task execution related to the expected quality and duration of the task. The MADS algorithm is closest to that proposed by Decker et al. (Decker, 1993) and Long and Lesser (Long, 1995), who clarified the categories of task relationships and created heuristics for parallel scheduling. The MADS algorithm extends their methods by exploiting information about the uncertainty of task duration and quality in the parallel scheduling of tasks. 3.2 Uncertainty in manufacturing planning I work on planning under uncertainty. That s the big field as far as I m concerned; that s the future. Maybe I m the only one who says that. George Dantzig, the Father of Linear Programming in OR/MS Today, October Manufacturing activities are to a significant extent unpredictable. Therefore, most estimates made for manufacturing planning are often less reliable than one assumes. The lack of detail in the plans that are developed in the first phases of the planning process implies incompleteness of information. This leads to the necessity for the other manufacturing planning activities to estimate the consequences for their own production plans. If a process planner does, for instance, not yet know on what machine tool an activity will be performed, the capacity plans will become less reliable. The uncertainty in planning activities becomes more influential on the higher level of aggregation, for the reason that less detailed information is available there. Galbraith defines uncertainty as the difference between the amount of information required to perform the task and the amount of information already possessed by the organisation (Galbraith, 1973). If a task cannot be understood completely in the preplanning phase, changes in the plans will occur regularly requiring information processing during task performance. Therefore the greater the task uncertainty, the greater the amount of information that must be processed among decision-makers during task execution in order to achieve a given level of performance. Galbraith clearly shows the impact of uncertainties on the planning strategy. An interactive planning process between the various experts is required for handling complex problems with great uncertainty as was also concluded in the previous chapter. Achieving integration in manufacturing planning is, however, not possible if the uncertainties in the information are not communicated. Generally, all information that is communicated between the various experts is only transferred in a deterministic way (e.g. only mean or most likely values), and thus ignoring the variances. In this way, the information becomes completely unreliable if the variances have a significant impact, which is naturally the case on the higher planning levels of aggregation in manufacture-to-order environments. Uncertainties about e.g. the processing times, lead times, starting times, or required resources cannot be ignored when drawing up the manufacturing plans. Hopp and Spearman show in their book Factory Physics that even in a less complex and less dynamic and rather simplified semi-line

53 ABSTRACTION PLANNING 39 manufacturing environment, the impact of variances on the performance parameters is often much greater than the mean values (Hopp, 1996). It can easily be understood that the variances, and thereby their impact, only increase in more complex and more dynamic environments like the ones considered in this thesis. A well-known approach for modelling uncertainty of job duration times is building PERT networks for project planning (Wiest, 1977). PERT is an acronym for Program Evaluation and Review Technique. On the basis of a stochastic Critical Path Methodology (CPM), an estimation (Normal distribution) for the total time in the critical path (i.e. lead time of the complete order) can be determined. Due to the stochastic values for job duration, more than one critical path may appear in PERT networks, making the lead time determination harder but more reliable. A Beta-distribution for the job durations in the network is fitted on a three-point estimation of the minimum, most likely, and maximum values. The three point estimation applied in PERT is further discussed in Section An important concept for decision making under uncertainty is the use of Bayesian networks as often advocated in the Artificial Intelligence research community. A Bayesian network is a graph that represents a probability distribution for a choice situation characterised by possible acts A, a probability distribution P over the set of possible domain states S, the outcome of each act in each possible state a(s), and a utility function u over the outcome space (Haddawy, 1999). The optimal act is the one that maximises expected utility: EU ( a) = p( s) u( a( s)) s S A drawback of using Bayesian networks relates to the required predictability of the environment concerning the acts, the domain states and the outcome. A probability function P and a utility function U must cover the totality of possible acts and states in the domain. This appears to be non-practical for a manufacturing planning domain. A technique for scheduling under uncertainty is design-to-time selection and scheduling of tasks in a task hierarchy (Garvey, 1995). The tasks are characterised by a discrete distribution of task quality and task duration which are selected and scheduled on the basis of optimising the overall quality of a problem solution. Design-to-time scheduling is related to earlier research on anytime algorithms and imprecise computation (Dean, 1988)(Liu, 1991). Finally, an approach for handling uncertainty in Constraint Logic Programming (CLP) has been developed for application in project planning for manufacture-to-order environments at the St.Petersburg State Electrotechnical University, Russia (Valkovsky, 1997). The preparation of the input data for the probabilistic models is, however, a difficult task. The same problem has been recognised above for methods using Bayesian networks. In fact, this is an inherent problem for all methods dealing with uncertainty.

54 40 CHAPTER Conclusion An important step for successfully achieving integration between the manufacturing planning process is the modelling of the uncertainty in the information that is communicated between the various planners. A method for handling several kinds of uncertainties in the aggregate planning levels of manufacture-to-order environments has not been developed yet. Therefore, a concept for order planning with stochastic information is presented in Chapter 8.

55 41 Chapter 4 Manufacturing Control In the previous two chapters, manufacturing planning methods and strategies have been examined. Technological planning methods, production planning methods as well as abstraction planning have been addressed. In order to supply the planning tasks with up-todate and reliable information, a broad manufacturing control strategy is needed. Solberg and Heim have already addressed the need for a strategic approach for dealing with the tangled web of information in manufacturing companies (Solberg, 1989). They described four strategic approaches for the management of manufacturing information and the accompanying planning activities: Subsystem optimisation, total integration, hierarchical decomposition, heterarchical decomposition. Subsystem optimisation is the oldest approach, often followed rather unconsciously or without recognition that there is an alternative. The main disadvantage refers to the lack of co-ordination resulting into islands of automation. The fact that most CAD, CAPP, CAM and MRP systems do not interactively process the information, but only interface at the front and back-end, indicates the popularity of this approach. Recognising the deficiencies of subsystem optimisation, the Computer Integrated Manufacturing (CIM) philosophy advocated the need for total integration resulting in a centralised planning approach; all requirements and constraints that had to be met in manufacturing planning could be considered simultaneously. However, this approach appeared to be unworkable, because it could not deal with the increasing complexity of manufacturing companies causing huge amounts of information, many dependencies, and continual change. In order to be able to cope with the increasing complexity some decomposition methods were elaborated, that have led to an evolution in control architectures (Dilts, 1991). Figure 4-1 shows the evolutionary steps in manufacturing control. The hierarchical form has often been proposed for reduction of the complexity of the centralised form by decomposing the problem solving tasks. Tasks are decomposed into subtasks that are subsequently decomposed into smaller sub-subtasks etc. These (sub)tasks are controlled on various hierarchical levels, together constructing a hierarchical structure. This structure fits well on the existing hierarchical structures of most manufacturing companies. Therefore, the control levels normally correspond with the departmental structure of the company concerned. This has resulted in the following rather generic hierarchical structure: a controller on factory level for company-wide planning, controllers on cell level for the departments, workstation controllers and equipment controllers. In order to minimise the vertical information flows, some researchers suggest to allow

56 42 CHAPTER 4 Centralised Control Hierarchical Control Modified Hierarchical Control Temporary hierarchies Co-operation Heterarchical Control Evolution-based Control controller entity communication Figure 4-1 Taxonomy of manufacturing control (Dilts, 1991). horizontal communication between the controllers on the same level, i.e. without having to consult the higher level controller. This modification of hierarchical control is often referred to as modified hierarchical control (Arentsen, 1995). It has the advantage that the chance of overloading the information processing entities on the higher levels is reduced. In heterarchical control, the hierarchical levels that were introduced in the hierarchical control principle are removed again. The difference with centralised control is that the control function is distributed over many local controllers, mostly belonging to individual machine tools. The heterarchical control principle is based on co-operation of distributed controllers without the support of higher level control modules. Co-ordination between the local controllers is often achieved by applying a market like approach, mostly based on a price bidding (auction) method. The main advantage of the heterarchical approach compared to the hierarchical approach refers to the flexibility with regard to problem solving and product routings. On the other hand, meeting global goals and integrating abstraction planning techniques seem to be very difficult, if not impossible, to achieve in heterarchical control. The need for both flexibility and aggregate planning in modern manufacturing environments, resulting from today s market demands, asks for the development of control methods that are able to combine the advantages of both heterarchical (flexibility) and hierarchical (aggregate planning) features. Keywords are self-configuration, co-operation, modification and high flexibility. Different methods have recently been developed that aim to achieve these goals, which are often indicated as evolution based control concepts. Important developments are the fractal company, holonic control, bionic and biological control. This chapter describes the hierarchical approach (Section 4.1), the heterarchical approach (Section 4.2), and the mentioned evolution based approaches (Section 4.3). In Section 4.4 the influence of non-linear dynamics in manufacturing on the control of the activities is briefly discussed.

57 MANUFACTURING CONTROL Hierarchical control The term hierarchical control needs a clear definition due to the fact that completely different meanings are being used in the literature. In the ISO International Standard (Concepts and rules for enterprise models) the concept of hierarchy is defined as follows (ISO, 1999): Hierarchy is a principle by which real world items and abstractions can be ranked and ordered. There are two kinds of hierarchies: part-of hierarchies and kind-of hierarchies. Part-of hierarchies represent the composition of elements or the decomposition of systems. Kind-of hierarchies represent levels of abstraction that are ordered by generalization and specialization. In manufacturing control research, the three principles of hierarchical control, described by Albus for a hierarchical robot control system, are often taken as a reference. The principles are (Albus, 1981)(Jones, 1990): Different levels are introduced to enable the decomposition of problems in order to reduce the complexity and to limit the responsibility and authority at each level. Each level has a distinct planning horizon, the length of which decreases when going down the hierarchy. Control resides at the lowest possible level. Another important characteristic of hierarchical manufacturing control systems refers to an increased level of detail in the plans when descending within the hierarchy. Hierarchical control makes it possible to perform aggregate planning with incomplete information on the higher control levels. According to Arentsen, hierarchical control is characterised by several levels of control with different time scales, containing control modules arranged in a pyramidal architecture (Arentsen, 1995). A distinction is also made between proper and modified hierarchical control. In the so-called proper hierarchical form, there are no control flows between the control entities on the same hierarchical level. Modified hierarchical control allows some co-ordination between these control entities, without having to consult a higher level control entity (see Figure 4-1). This modification decreases the chance of overloading the higher level control entities, a risk that was already noticed by Galbraith in the early seventies (Galbraith, 1973). On the other hand, the control entities, then, need information about the other entities in the system and a co-ordination protocol must be developed. It is therefore questioned whether an overload is more likely to occur in proper hierarchical control systems than in modified hierarchical control systems.

58 44 CHAPTER Hierarchical control concepts The first well-known hierarchical architecture for production control is the Automated Manufacturing Research Facility (AMRF) architecture developed at the National Institute of Standards and Technology (NIST) (Jones, 1986). The AMRF architecture consists of five hierarchical control levels with vertical command control flows and feed-back information between the control entities (Figure 4-2). A remarkable feature of the AMRF architecture is the existence of virtual cells that vary in configuration according to the type of product to be produced (McLean, 1982). The restriction that only one single virtual cell can be active at a given time, limits the possibility of producing different parts simultaneously. A virtual manufacturing cell is formed when the demand for a specific part appears, and the cell fades away when that demand disappears. The later Manufacturing Systems Integration (MSI) project, that aimed at the development of an integration architecture for the integration of production engineering and control within a manufacturing shop, was part of the AMRF effort at NIST (Senehi, 1992). Senehi et al. primarily aimed to remove the human from the control loop, by defining a control architecture, information models and formal interfaces between the various computer systems. Interaction between the process planning and production control tasks is facilitated by applying the ALPS standard language for process specification (Catron, 1991). The control architecture consists of a single top-level shop controller, one or more levels of workstation controllers and an equipment control level (Figure 4-3). The configuration remains fixed after the implementation, i.e. while the shop is in production. Thus, the virtual cell concept of McLean (discussed above) has unfortunately not been applied in the subsequent architecture of AMRF. Facility control input Shop status feed-back Shop Cell Cell Workstation Workstation Equipment Equipment Figure 4-2 Automated Manufacturing Research Facility (AMRF) architecture (Jones, 1986).

59 MANUFACTURING CONTROL 45 Shop Controller Workcell Controller Workcell Controller Equipment Controller Equipment Controller Workcell Controller Equipment Controller Equipment Controller Equipment Controller Figure 4-3 The Manufacturing Systems Integration (MSI) control architecture (Senehi, 1992). Another hierarchical control concept is the Production Activity Control (PAC) model and the accompanying Factory Co-ordination (FC) architecture of the ESPRIT project 477, COntrol Systems for Integrated MAnufacturing (COSIMA) (Bauer, 1991). The FC architecture contains two control levels, namely shop level for multi-cell control and a cell level for controlling the workstations in a manufacturing cell (see Figure 4-4). The functional architecture of the Cell Controllers contains the five generic functions Scheduling, Dispatching, Monitoring, Mover, and Producer, together constituting the PAC Production Environment Design Factory Control Shop level Manufacturing System Analysis Process Planning Product Based Layout Management performance Scheduler measures factory level request schedule status Dispatcher plan release Monitor information from cellcontrollers Cell level Cell Controller (PAC) Cell Controller (PAC) Cell Controller (PAC) Cell Controller (PAC) performance Scheduler measures schedule request Monitor Dispatcher status instructions data collection Mover Producer Cell level Equipment level Execution Layer (Devices etc.) Figure 4-4 The PAC model and FC architecture (Bauer, 1991).

60 46 CHAPTER 4 model. On shop level, two planning and control modules are distinguished. The Production Environment Design (PED) module includes the system engineering and process planning functions. The Factory Controller contains three of the five functions of the PAC model, while the Producer and Mover functions which establish the interfaces with the physical resources are substituted by the Cell Controllers. At the Research Group on Integrated Shop Floor Control of the Chalmers University of Technology, Sweden, (CRISC) further extensions of the PAC module for cell control have been suggested. The Mover and Producer functions are extended with monitoring tasks and the possibility of peer-to-peer communication (Maglica, 1996). From a discussion on the industrial applicability of the PAC architecture it can be concluded that the functional architecture is applicable for various types of automated manufacturing systems, like FMS cells (Gullander, 1997). Arentsen (1995) presented a generic architecture for Factory Activity Control (FACT). All production and auxiliary activities (e.g. tool assembly, quality control) that take place on the shop floor are integrally controlled. A representation of a shop floor is given by the Manufacturing Cell reference model (Figure 4-5). The shop floor activities are controlled on four hierarchical levels, viz. factory level, cell level, station level and equipment level. A control functions building block consisting of the four functions Planning, Dispatching, external storage of fixtures materials storage local storage of fixtures fixture preset station fixture room (un)load station pallet storage materials room transport of pallets, fixtures and materials milling machine work station turning machine inspection device work station measuring machine transport of tools and measuring devices local storage of tools tool preset station tool room storage of measuring devices quality room external storage of tools = (un)load device Figure 4-5 Manufacturing Cell reference model (Tiemersma, 1990)(Arentsen, 1995).

61 MANUFACTURING CONTROL 47 Feed-forward Higher Level Building Block Feed-back Planning Diagnostics Off-line Dispatching Monitoring On-line Lower Level Building Block Figure 4-6 Control functions building block (Arentsen, 1995). Monitoring and Diagnostics is generically applied for all control functions on all control levels in the hierarchy (Figure 4-6). The same building block is applied in the EtoPlan system architecture presented in Chapter 7 of this thesis. In Section 7.2, the building block is described in more detail. Next to the control building block, Arentsen suggests an information framework, which comprises the object classes Orders, Products, Processes, and Resources. An extensive implementation of the FACT architecture has been developed by Arentsen and Tiemersma, of which the realisation of the system in a manufacturing company has resulted in considerable improvements (Arentsen, 1995). The FACT architecture as well as the PAC architecture can be classified as modified hierarchical control. This approach may improve the effectiveness of disturbance handling and the reliability of plan execution, see the next section Misconceptions about hierarchical control With the appearance of heterarchical and evolution based control concepts, which are respectively described in Sections 4.2 and 4.3, the hierarchical principles are often misrepresented in comparing analysis. This has led to some often used misconceptions regarding hierarchical control. First, hierarchical control has often been misinterpreted as (nearly) equal to centralised control by some researchers. It is, then, assumed that only two hierarchical levels exist, where the entities on the lowest level function as slaves and all the planning activities are performed by one central planner, the master. In this way, no effect is obtained with respect to the decomposition of the planning problem itself, which notably was the main reason for the initial development of hierarchical control. The notion of hierarchical control as master-slaves relations can only be held if hierarchical control is assumed to be equal with centralised control. Centralised control is, however, in no way in conformity with the principles of hierarchical control as defined by Albus. A hierarchy with more than one hierarchical planning and control level will always be based on the distribution of the

62 48 CHAPTER 4 planning task, and can, therefore, never exist of one central planner (master) and slaves with no planning functionality. The following example, based on the modified hierarchical FACT concept (Section 4.1.1), clearly shows that the aforementioned interpretation is wrong. The highest level (factory control) performs the capacity planning task considering a time horizon of approximately 3 months in an average job-shop. Cell control performs the detailed scheduling task (+/- 2 weeks), workstation control performs the job sequencing task (+/- 1 day) and the task of equipment control is to control the execution of the production activities. Consequently, no central production plan is drawn up, but various planners on multiple levels of abstraction generate multiple plans in parallel. In some sense related to this master/slave misconception is the assumption that lower level control entities are subordinate to the higher-level control entities in hierarchical control. Although the decision-making responsibilities are often determined on the basis of the hierarchical control levels, this is not a necessity in hierarchical control systems. Van Aken states that organisational structures usually involve a combination of stratification and hierarchy (Aken, 1978)(Zwegers, 1998a). These features should, however, not be confused. Hierarchy is characterised by problem decomposition (in order to deal with complexity) and aggregation levels in the planning (in order to deal with incomplete information). Stratification refers to the order of control entities with regard to conflict solving between entities interests. Another misconception regards the possibility of replanning when disturbances occur. Within a hierarchical control system, the problems are dealt with at the lowest possible level, i.e. bottom-up. When, for example, a machine breakdown of short duration has occurred, it will probably not significantly disrupt the schedule that has been developed on cell level. Such a problem is recognised by the equipment controller and solved on workstation level by the workstation controller. Problems are only reported to the next higher control level if the problem is so drastic that the subordinate level cannot meet the demands received from the higher-level control module. In other words, the problemsolving autonomy of the control entities is limited if the problem also negatively influences the progress of the production plans on the higher control levels. It is often stated that a negative feature of hierarchical control systems is the inability of dealing with regularly occurring disturbances, for example by Bongaerts et al. (1996). This is unquestionably true if the aforementioned centralised interpretation of hierarchical control is assumed, as is done by Bongaerts et al. in their tests from which they draw this conclusion. Viz. a detailed plan is centrally drawn up and executed by 'slaves' without the possibility of replanning in real-time. Most hierarchical control systems they refer to, a.o. the FACT system (Arentsen, 1995), are, however, implementations that meet the hierarchical control principles discussed in the previous section. Real hierarchical control systems are modelled according to the decomposition and aggregation principles of hierarchical control, which makes them very suitable to deal with disturbances.

63 MANUFACTURING CONTROL Some conclusions regarding hierarchical control As mentioned in this section, much confusion about hierarchical control exists in the literature. The aspect of multiple levels of aggregation, based upon different time horizons and different levels of abstraction in the plans, are often not taken into account although these are the most important aspects of hierarchical control. The main problem of hierarchical control systems is their structural rigidity. Solberg and Heim noticed this resistance to evolution by observing that in most cases, there is no single correct decomposition of authority into units (Solberg, 1989). This rigidity is strengthened by the fact that hierarchical control structures are conceptually similar to the organisational structure of a company. This organisational structure does normally contain departments and physical grouping of resources (e.g. workstations), which makes it rather resistant to evolution. It is, however, not a conceptual necessity in hierarchical control that the organisational structure of a company matches the control structure. Nevertheless, the need for more flexibility in hierarchical control is clear. Special attention must be paid to the possibility of in-time (i.e. at the time the system is in operation) modification of the hierarchical structures of the system. Rearranging, adding and deleting resources must be part of the tactical and operational planning process. Another problem in hierarchical control as used in production control is that the information and control structures derived from the organisation structure does normally not match the information and planning structures as used in technological planning. The gap between the information structures is further elaborated in Chapter Heterarchical control Mainly due to the difficulties regarding information management in centralised and hierarchical control together with the rigid structure of these control concepts, in 1985 the heterarchical approach has been proposed (Hatvany, 1985). The basic philosophy is that information management and the planning decisions are distributed over the resources and the orders involved. Various attempts have been made to implement control systems based on this philosophy. Hatvany promoted the heterarchical approach characterised by fully distributed control, retention of a minimal amount of information, and the co-operation between autonomous communicating entities (Hatvany, 1985)(Duffie, 1988). During the early 1990 s, more concepts based on autonomous entities and agents have been developed, see for example (Duffie, 1990)(Dilts, 1991)(Upton,1992)(Duffie, 1996).

64 50 CHAPTER 4 An example of heterarchical control as formulated by Upton is quoted in the textbox below (Upton, 1992). Consider a machine shop with fifty CNC machine tools and automatic vehicles for transportation of materials. The heterarchical approach is best outlined by describing an example of the production procedure which is followed in order to produce an individual item. A production-control computer requests that a casting, which has arrived at the shop, will be machined. The casting is manually bolted in a flexible fixture on a pallet, and a small computer is fitted to the pallet. This computer contains a processor, some memory and a radio. This assembly will be referred to as "the part". The production control system loads the memory of the part s computer with the processing requirements of the product. In other words, the control system tells the part what it needs to look like. The product enters a short system entry buffer, and while it is waiting, its computer broadcasts its description throughout the system to the flexible CNC machine tools. Some machines examine the job s description and decide that they are unable to machine the product because it is too big for their bed, say. Others simply have the wrong type of geometry for the job. Still others decide that they can do the processing work on the product. The machines computers plan a process for the job, and decide how long it would take them (and how much it might cost in terms of toolwear) to do the work. Having determined how much processing time is needed, the machine checks its local buffer, determines how many jobs it has waiting for it, and forms a "bid" on the job. It transmits this bid across the network to the waiting part. The part waits until a system-set deadline to receive bids and having collected them, selects a winner from the bidding machines. It sends a message to that machine that it has selected it and expects to arrive for processing. The next arrangement the part needs to make is for transportation to the machine. Here there are a number of possibilities but let us say, for now, that the part essentially "calls a cab": It requests transportation from an automated vehicle dispatcher which sends a vehicle to it, and it arrives at the machine. After waiting in line it is machined. While it is waiting it arranges subsequent machining and ultimately leaves the system, once all tasks are complete. Having been processed the part moves out of the system to be assembled into its final product. The part has thus been processed without a central control computer, and with simple, modular, physically decentralised hardware and software. Compared to the hierarchical approach, the local controllers act entirely autonomously, the amount of information processing is reduced, the modifiability is increased, and the response time to disturbances will normally be shorter. Entities in a heterarchical system should not need to know the plans and the intentions of the other entities in order to minimise complexity, minimise global information, and improve fault-tolerance (Prabhu, 1999). An undesirable consequence of this characteristic is that meeting global goals becomes uncertain and planning on multiple levels of aggregation (abstraction planning) is nearly impossible in heterarchical control.

65 MANUFACTURING CONTROL 51 By not allowing higher level controllers in the system, it seems impossible to deliberate about competing goals as well as contradicting so-called soft constraints. Local storage of the information is an important feature of heterarchical control, which also means that entities in the system do not have access to the information of other entities. Much effort should therefore be paid to rules and facilities to ensure an overall satisfactory behaviour of the local decision makers in the system. They should meet each other in the system and exchange the right information. Auction-alike procedures based on market principles in society are often applied. Realising that market rules and control structures in society are enormously complex, it becomes clear that this approach can only be applied successfully in rather simple and uniform environments. This makes heterarchical control difficult to implement in companies that have to deal with frequently changing situations, e.g. in case of shifting bottlenecks, high product variety, significant influences of clients and other stakeholders. An oversimplification can also result in, again, rigidity in control, something that heterarchical control should in fact prevent in comparison with hierarchical control systems. Zwegers notices this risk when describing the situation where each individual controller only knows and negotiates with its successor and predecessor in a production line or has no negotiation capabilities to handle disturbances (Zwegers, 1998a). In such a situation, the system is heterarchical, but not robust; it is not structurally stable. The main problem is that off-line central schedulers are useless in heterarchical control systems because the local controllers make their decisions autonomously. The interdependencies between decisions made by the scheduler are significant in a schedule, which means that if the schedule is not obeyed by some of the local controllers the whole schedule is not suitable anymore. Today, the problem of meeting the overall company goals is increasingly recognised by researchers in the field of heterarchical control. At the DISCRETE laboratory of the Pennsylvania State University, real-time control loops are integrated in a heterarchical control architecture (Prabhu, 1999). The challenge is to avoid explicit co-operation among entities while letting the entities implicitly co-operate by modifying their local behaviour based on observations of the effect of their behaviour on lateness, work in process, queuing time, etc. Prabhu suggests a part-driven heterarchical manufacturing system that can be controlled by explicitly adjusting the time at which parts start seeking machines. A linear integral control-theoretic analytical model has been developed to represent the dynamical behaviour of closed-loop distributed controllers in this system (Figure 4-7). An adaptive arrival-time control algorithm improves the systems behaviour by varying the gain factor according to the due-date deviation (Cho, 1998). Various part-driven arrival-time controller algorithms have been compared with each other and with common dispatching rules using numerical simulation for the single machine case. The control algorithms perform slightly better than the dispatching rules on both mean and variance of tardiness (Prabhu, 1997)(Cho, 1998).

66 52 CHAPTER 4 Due date d(t) 1 d(t) 2 d(t) n Due date deviation z(t) 1 z(t) 2 z(t) n Controller Controller Controller Arrival time a(t) 1 a(t) 2 a(t) n Part Processing Predicted completion time c(t) 1 c(t) 2 c(t) n Figure 4-7 Arrival time control for n parts (Prabhu, 1995). A framework based on a heterarchical contract-net-style auction protocol for integrating process planning and shop floor control is developed at the Penn State University (McDonnell, 1999). The cascading auction protocol facilitates distributed real-time completion of process planning functions when a part progresses through the system. The process plans are expressed at different levels of abstraction by applying the method of Wysk et al. (1995) (see also Section ). It can be concluded that heterarchical control systems are specifically developed for low level control, where resources compete for jobs on the shop floor. Simple co-operation and competition rules are applied and no long-term planning is required. Especially in manufacturing environments characterised by a rather unpredictable behaviour, a heterarchical control system for low-level control may yield acceptable performance. The application of heterarchical control for high-level planning and control problems is, however, principally impossible. A heterarchical control system should therefore be embedded in a broader manufacturing control system in order to be able to control the input to the shop floor, as well as the production preparation functions like (macro) process planning, material procurement, etc. The low-level heterarchical control system will then be a part of a more extensive control structure with hierarchical features. Therefore, the question arises whether heterarchical control is primarily a way of structuring the lower level manufacturing control systems, instead of a manufacturing control concept on its own. If this is true, there is no clear difference between heterarchical control and distributed scheduling. In this perspective, heterarchical control is just a way of managing lower-level planning and control that can feasibly be part of a hierarchical overall control concept Multi-agent systems Many distributed (heterarchical) control systems have recently been implemented as Multi- Agent Systems (MAS) because a multi-agent architecture allows decisions to be taken in a decentralised way. Decisions are taken locally by the agents via interactions between the different agents. An agent is an autonomously operating software program, which contains a partial representation of its environment and which has its own objectives. In production control systems, intelligent agents often represent the physical resources, parts and jobs in the system.

67 MANUFACTURING CONTROL 53 A large industrial implementation of an agent-based shop floor control architecture is the Autonomous Agents for Rock Island Arsenal (AARIA) project being developed for an army manufacturing facility in the USA (Parunak, 1998b). Both internal company entities (operations, resources, operators, parts) and external stakeholders (customers, suppliers) are modelled as agents. Parunak et al. (1998b) define seven requirements for the AARIA implementation, of which three requirements particularly point toward the application of an agent-based system. These are: Empowerment, which requires that human s function as peer elements in the system rather than being run by the system. This requirement refers to the autonomous character of agents. Frequent change of the system in response to environmental change. Agents can be modified without the need to modify or even notify other components. Metamorphosis, meaning that the system maintains continuity between different entities that represent different stages in a common life cycle (e.g. order Å part Å production history). Agents can change through time. After a broad industrial survey, Parunak suggests that the use of an agent-based system meets best some important manufacturing systems requirements. Five such requirements are particularly salient: agents are best suited for applications that are modular, decentralised, changeable, ill-structured, and complex (Parunak, 1998a). Wiendahl and Ahrens also use the agent theory for implementing a distributed planning and control system (Wiendahl, 1995). They model generic structures for an order agent and a resource agent (figure 4-8). Three subsystems are distinguished: the implementation system combines the so-called special abilities of the agents, the information system carries out the communication with the other agents, and the objective system specifies the local objectives of the agents. Carver and Lesser address the problem of global inconsistencies in distributed situation assessment when sharing incomplete information or information at different levels of detail objectives system maximum profit, minimum risk objectives system maximum profit, minimum risk information system message sending message receive message management information system message sending message receive message management implementation system determination of product structure offer calculation network plan design account management order agent implementation system feasibility check offer calculation location search account management machine agent a) structure of an order agent b) structure of a machine agent Figure 4-8 Generic structures for order and resource agents (Wiendahl, 1995).

68 54 CHAPTER 4 between agents (Carver, 1995). The agents create so-called sources of uncertainty (SOU s) whenever it is determined that a local hypothesis can obtain evidence from another agent, i.e. whenever a subproblem interaction (constraint) is detected. These SOU s are viewed as sources of uncertainty about the correctness of an agent s local solution because they represent unresolved questions on the global consistency of the solution. Unfortunately, like in other distributed planning systems, a communication explosion will occur if too many subproblems are mutually dependent. Like the above mentioned concepts, Multi-Agent Systems are often applied in one-level (i.e. heterarchical) control systems, due to their negotiation capability and distributed decision making. Recently, however, the concept has also been suggested by some researchers for use in multi-level control. At the Department of Computer Science, University of Massachusetts, Amherst, a special type of agent has been developed which incorporates three different levels for multi-level conflict resolution (Wagner, 1999). Agents, as shown in Figure 4-9, have multiple goals and tasks. It may not be possible to resolve the conflicts satisfactorily to the local agent or some set of agents - depending on whether the model is self-interested or co-operative - through the inter- and intra-agent negotiation processes. Then, the agents may need to move the negotiation to a new level, a meta-level, and change the tasks or goals on which they are negotiating. This may also induce a change of the objective function(s). This method can, however, not be called hierarchical because no agent grouping is performed. Kouiss et al. (1997) present a distributed dynamic scheduling approach based on a multiagent paradigm. Each agent makes its own local decisions via the application of dispatching rules, under the control of a supervisory agent. The supervisory agent selects the dispatching rules that it assumes to be most suited to meet the global production objectives. Meta-Level Inter-Agent Level Domain Expert - Manages state and domain view - Responsible for selection of agent s high-level goals subset of problem solving options desired solution characterization non-local information commitments proposed commitments received feasibility results characteristics of possible solutions Co-ordination Module - Manages agents non-local view - Responsible for negotiation with other agents selected schedule satisfied commitments violated commitments Dialogue to establish or revise high level goals Dialogue to achieve existing goals Domain Expert Co-ordination Module Intra-Agent Level Local Agent Scheduling Module - Manages local agent control - Responsible for evaluating local and non-local processes and determining a local course of action agent A Local Agent Scheduling Module agent B Figure 4-9 Multiple interacting levels of conflict resolution (Wagner, 1999).

69 MANUFACTURING CONTROL Evolution-based production control concepts The increase in internal and external changes asks for the development of control concepts and systems that can easily adapt themselves. Like organic structures have to adapt to changing influences in order to keep alive, also the structure of a company should be dynamic so as to be able to react effectively to the changing environment. The actual rigidity of hierarchical control systems and the limiting non-existence of higherlevel control functionality in heterarchical control systems engendered some research projects on control concepts that may combine the advantages and avoid the disadvantages of hierarchical and heterarchical control. An important research stream is the Intelligent Manufacturing Systems (IMS) program that originally was proposed by Professor H. Yoshikawa (Yoshikawa, 1993). This section discusses some important developments in evolutionary control concepts. Most attention is paid to studies on Holonic Manufacturing Systems, because it has provided the most promising results with regard to control architecture concepts Holonic Manufacturing Systems The Holonic Manufacturing Systems (HMS) group is one of the research groups within the Intelligent Manufacturing Systems (IMS) consortium that aim to develop concepts for future manufacturing systems. The Holonic Manufacturing Control paradigm as proposed by the HMS research group serves as the basis for developing control concepts that combine the best features from both hierarchical and heterarchical control. The idea of holonic manufacturing is based on the concept of holonic systems, developed by Arthur Koestler (Koestler, 1967). The term holonic is a derivation from the word holon which is a combination from the Greek holos (meaning whole ) with the suffix - on (suggesting a part ). Koestler suggests that an entity in a complex system is simultaneously a part of a larger whole and a whole consisting of more detailed parts. According to Koestler, holons are simultaneously self-contained wholes to their subordinated parts, and dependent parts when seen from the inverse direction (Wyns, 1996). A hierarchy of holons in e.g. social organisations and living organisms is named a holarchy. Koestler s definition of holarchy is used as a reference by the HMS group for defining manufacturing holarchies (Valkenaers, 1994): A holarchy is a hierarchy of self-regulating control building blocks (holons), which function (a) as autonomous wholes in supra-ordination to their parts, (b) as dependent parts in subordination to controls on higher levels, (c) in co-ordination with their local environment.

70 56 CHAPTER 4 Considering the description of hierarchical control in section 4.1, the differences between holonic and hierarchical control are not immediately apparent. According to Bongaerts, et al. (1996), holonic control differs from traditional hierarchical control on the following aspects: Holons can belong to multiple hierarchies. Holons can form temporary hierarchies. Holons don t rely on the proper operation of each holon in the hierarchy to get their work done. The latter of the three differences refers to the increased autonomy of the holons compared to the control modules in hierarchical control. The work plans that are provided to the lower-level holons are considered as an advice. Furthermore, production rules are applied in order to prevent the natural behaviour of aiming at local optimisation by the holons. Unfortunately, this modification potentially introduces unreliability in the system. Lowerlevel holons are not aware of the arguments that have resulted in the actual work plan drawn up by the higher level holon. It is, therefore, possible that ineffective decisions are taken by a lower-level holon that disrupt the execution of some work plans dispatched to the other lower-level holons. Therefore, applying abstraction planning, as described in Chapter 3, may prove to be a more suitable approach. Abstraction planning means that the higher-level holons do define rough goals, requirements and constraints only. The question how to meet these goals and requirements is left to the lower-level holons. These holons define the more detailed goals, requirements and constraints themselves, which, in turn, are handed over to their subsequent lower-level holons, etc. A lower-level holon reports to the higher-level holon(s) if it is not able to meet the specified requirements, e.g. due to occurring disturbances. Subsequently, the higher level holon initiates (roughly defined) corrective actions if necessary. The above-mentioned control rule, necessary for the integration of abstraction planning in manufacturing control, complies with the definition of an invariant (i.e. fixed rules) in Holonic Manufacturing Systems (Bongaerts, 1998). According to Bongaerts et al. an invariant is a logical proposition that expresses the requirements on the behaviour of a holonic control algorithm. They apply the strategy of increasing the autonomy for the lower-level holons by considering the (detailed) work plans obtained from a higher-level holon as an advice. By applying abstraction planning, the autonomy of the lower-level holons is increased by leaving the detailed planning decisions to the subordinates, while simultaneously demanding that the roughly described requirements should be met. The other two differences between holonic and hierarchical control, namely multiple and temporary hierarchies, as mentioned by Bongaerts et al. (1996), are believed to be essential characteristics for future manufacturing control systems. Multiple hierarchies are necessary for achieving concurrency in the various manufacturing planning processes. Applying an object-oriented approach based on the holonic principles of simultaneously being a whole and part of other wholes may lead to a final demolishment of the walls that still exist between e.g. process planning and production planning departments. According to Valkenaers et al. (1996), each hierarchy can reflect a different view on the system, e.g. a scheduling view, a process planning view, a maintenance view, etc. A specific Information Management concept (see Chapter 6), as

71 MANUFACTURING CONTROL 57 currently developed in the laboratory of Design, Production and Management at the University of Twente, serves as a backbone for creating multiple views in multiple domains of the order-, resource- and product information structures, (Kals, 1998). Temporary hierarchies can solve the problem of increased dynamics in manufacturing companies. Traditional hierarchical systems are based on a function-oriented static control structure in which all orders and products are handled similarly. Due to the changing characteristics of the manufacturing environments (viz. high product variety, small batches, short throughput times, tight due dates) these static control structures are not suitable anymore. There is a need for continually re-arranging the control structure of the company, according to the dynamically changing situation in the factory. This especially holds for manufacture-to-order environments Holonic architectures A structural model of a holonic manufacturing control system, as shown in Figure 4-10, has been developed at the PMA department of the Catholic University of Leuven (Bongaerts, 1995)(Bongaerts, 1998). The model consists of an On-line Manufacturing Control (OMC) holon for co-ordinating the work of the resource holons and the order holons. It is possible to implement the holonic architecture in a pure heterarchical form, because the lower-level resource and order holons are self-organising, co-operative, and autonomous. In that case, the OMC holon is not really needed. The co-ordination of the work by the OMC holon is achieved by co-operation with one or more centralised Scheduler Holons. A scheduler holon is included for global performance optimisation. When disturbances occur, the OMC holon can react immediately without having to wait for a revised schedule from the Scheduler Holon On-line Manufacturing Control Holon Neighbour Holon data Shop floor control rules Schedule entry Monitoring Shop floor organiser Register new resources Register new orders Register new schedulers Request Work Request Manufacturing Resource Holon Order Holon Figure 4-10 Structural model of a holonic manufacturing control system (Bongaerts, 1998).

72 58 CHAPTER 4 scheduler holon. The scheduler holons are continually updating their plans, which are regularly communicated to the OMC holon. The structural model actually contains two control levels, a centralised controller (the OMC holon) that receives its input from one or more scheduling holons, and several basic holons for distributed control of the resources and orders. Unfortunately, the holonic principles of building multiple and temporary hierarchies cannot be recognised in the structural model. It seems worthwhile to extend the holonic concept with a structural model for building such dynamic hierarchies with multiple control levels for problem decomposition purposes. A holonic alike approach based on multi-agents is the MetaMorph I architecture developed at the University of Calgary (Maturana, 1996)(Maturana, 1997). To meet the dynamic requirements in manufacturing, the control system s architecture is capable to change frequently by using so-called mediator agents which facilitate the co-ordination of the resource agents. The mediators create virtual clusters of agents that are dynamically formed or dissolved with the emerging requirements of the manufacturing system. At the Technical University of Denmark, Langer introduces the Holonic Multi-cell Control Systems (HoMuCS) architecture using the PROSA reference architecture (Langer, 1999). The structural elements and their relationships are described based on a set of objectoriented models. Langer describes the functional models and the information structures in detail. It is intended to implement the HoMuCS system with a multi-agent system. The integration of process and production planning has also received some research attention in HMS. Sugimura et al. (1998) present a process planning and scheduling integration method based on a holonic architecture. Further, a Holonic Machining Unit (HMU) is described for dynamic, NC-internal self-planning of cutting conditions (Bengoa, 1996). Some research on developing a holonic task based NC controller for in-process reaction to disturbances (reactive process planning) has been performed (Tanaya, 1995)(Kruth, 1996)(Kruth, 1998) Other evolution-based concepts The fractal company The fractal company concept developed at the Fraunhofer-Institute for Manufacturing Engineering and Automation (IPA) uses fractal theories for dynamically structuring autonomous organisational units (Warnecke, 1992). The term fractal company comes from fractal geometry for describing and analysing objects in multi-dimensional spaces. The main characteristic of fractals is self-similarity, implying recursion, pattern-inside-of-pattern (Tharumarajah, 1996). Fractals function as companies in the company. By providing greater flexibility for response at the operating level, individual staff members and teams are in a position to independently organise and execute complete work tasks. The concept is particularly applicable for manufacture-to-order manufacturing companies, which require continuous adaptation of flexible and agile organisational units to changing conditions.

73 MANUFACTURING CONTROL 59 The demand for production planning and control systems that are able to respond to changes in the manufacturing environment, has initiated the development of the 3-L-PPC concept at Fraunhofer-IPA (Westkaemper, 1999). The concept is characterised by the fact that it abandons complex planning methods and instead uses simple, clear planning procedures and tools. A capacity planning technique that uses simple baskets with traffic light finite capacity indications should prevent the occurrence of unfeasible workloads. This rough planning tool links the sales department to the production area to ensure that the delivery dates promised to customers can actually be met. An implementation of the 3-L- PPC concept has been introduced at two SME companies in Germany (Sihn, 1999). It is stated that the results are promising, e.g. a 50% leadtime reduction has been achieved. Biological manufacturing control While holonic control is inspired by sociological structures and the fractal company is based on mathematical structures (fractals), another research stream on manufacturing control takes the structures of natural life as a reference. From this biological viewpoint, control entities are modelled like organs that are at the same time autonomous but coordinate their actions in order to maintain harmony. Like Koestler s holon which is simultaneously a part of a larger whole and a whole consisting of more detailed parts (Section 4.3.1), organs are part of the life-form and consist of more detailed components (e.g. cells). Research on such biologically based systems has been performed by Okino and Ueda who developed concepts for a bionic manufacturing systems and biological manufacturing systems, respectively (Okino, 1993)(Ueda, 1997a)(Ueda, 1997b). Somewhat related research on the creation of Artificial Life in computers is being performed at the Santa Fe Institute on the basis of the theories of Chris Langton (Langton, 1995)(Lewin, 1993). Although the various biological concepts are promising, until now, just laboratory implementations of biological manufacturing systems are known. The biological approach has been applied to address dynamic scheduling problems (Ueda, 1997a). Further, a simulation method for shop floor layout configuration has been described by Vaario and Ueda (Vaario, 1997). 4.4 Non-linear dynamics of manufacturing control systems As manufacturing environments have become increasingly complex and the interaction with the external stakeholders of the company becomes more and more dynamic, the problem of unpredictability in manufacturing grows proportionally. An important cause of unpredictability is the non-linearity of manufacturing processes. Traditionally, manufacturing problems where modelled with clear end-points, for example the generation of a NC-plan. The assumption of a well-defined endpoint is however not appropriate for manufacturing planning and control tasks such as shop floor control and capacity planning. This is due to the multiple goals and the global objective of business continuation (Parunak, 1999). Parunak suggests the model of going concerns which is

74 60 CHAPTER 4 characterised by maintenance goals and interaction mechanisms instead of the traditional well-defined end-point problem solving. Understanding the dynamics of manufacturing environments goes beyond optimising mean value based performance indicators and requires variance analysis as described by Hopp and Spearman (Hopp, 1996). Parunak even goes one step further by applying concepts from dynamical system theory (informally, chaos theory ) to understand the heartbeat of the factory (Parunak, 1998c). A non-linear control-theoretic analytical model has been developed by Prabhu and Duffie for analysing and modelling non-linear dynamics in autonomous heterarchical manufacturing systems control (Prabhu, 1995). They apply closed-loop distributed controllers to a heterarchical control system in order to improve the robustness to uncertainties such as new part arrivals and machine failures. Ahrens addresses the differences in views of the planners and the self-organised production groups that can be recognised in most manufacturing companies (Ahrens, 1996). According to Ahrens, it happens so often that the operations level is not impressed by the orders of the planning and control authorities and behaves quite obstinately from their point of view. He mentions the problem that in non-linear dynamic systems, like manufacturing systems, abnormalities determine what happens during phases of instability, and such phases are the order of the day in many branches of industry. 4.5 Conclusions The evolution of manufacturing control concepts has been discussed in this chapter. Some conclusions can be drawn. Applying hierarchies in manufacturing planning and control both enhances the predictability and is necessary for meeting the overall goals of the company. It can also be stated that hierarchies are necessary for handling disturbances adequately. Compared to heterarchical control, the influence of disturbances on other activities is handled more effectively in hierarchical control systems. The main problem of hierarchical control systems is their structural rigidity. A well-known saying is that the first constant of modern manufacturing systems is that nothing is constant (Parunak, 1998b). Future manufacturing control concepts can therefore only be successful if they can cope with the enormous complexity and frequent changes resulting from the dynamic nature of the manufacturing environment. The static rigid composition of hierarchical control structures as well as the lack of co-ordination in heterarchical control structures both obstruct a solution that meets the demands. It is needed to create a certain degree of organisational fluidity to effectively handle changing needs (Glance, 1995). The evolution-based control concepts attempt to achieve such a dynamic control structure. Holonic control, in particular the multiple and temporary hierarchies (holarchies), seems to yield a promising approach towards future manufacturing control systems. A system architecture based on the principles of holonic control is, however, still missing. In the remainder of this thesis, such a system architecture based on holarchies is presented (Chapter 7).

75 MANUFACTURING CONTROL 61

76

77 PART II 63 Part II DESIGN OF A MANUFACTURING PLANNING AND CONTROL CONCEPT In manufacture-to-order environments, the required process steps and routings through the factory are normally not known at the time of order acceptance. Due to a lack of process planning information, detailed planning of ORDERS and the allocation of these ORDERS to RESOURCES is impossible then, which makes the estimation of realistic delivery dates and costs difficult and unreliable. Applying aggregate technological and logistic planning techniques is a necessity for such manufacturing environments. The Engineer-to-order Planning (EtoPlan) concept has been developed in order to deal with the problem of incompleteness of technological information and of uncertainty of logistic information about ORDERS that are planned on long term. As such, the concept aims at bridging the gaps between order intake, process planning and production planning. The EtoPlan concept has been built on the reference architecture for Integrated Manufacturing Planning and Control described in Chapter 5, in which the central Information Management pillar (further described in Chapter 6) enables concurrent manufacturing planning by effectuating integrated information processing. The system architecture of EtoPlan, described in Chapter 7, contains a description of the functional control architecture and clarifies the way in which temporary control hierarchies are built. The EtoPlan concept for aggregate order planning is presented in Chapter 8. A prototype software implementation of the EtoPlan concept has been developed for demonstrating the practical relevance of the concept. A brief description of the implementation is given in Chapter 9.

78

79 65 Chapter 5 Reference Models The design of a manufacturing system proceeds through several design phases. For each design phase specific models are developed for representing the design choices. In the literature, different terminology is used for these models. Wyns clearly discerns between a modelling framework, a reference architecture and a system architecture (Wyns, 1999). A modelling framework describes the possible methodologies, methods, tools, etc. to develop a CIM-system for a specific problem, while a reference architecture focuses on providing a generic rough model of the manufacturing environment. A system architecture is designed on the basis of a reference architecture. It specifies the structure of a component of the manufacturing system, as well as its functions and interactions with other components. After the functional description of the system, a generic implementation of the system includes the software modules and data structures. The generic implementation is actually realised by filling in the company specific structures, resources and data. This last step is referred to as the realisation. In this chapter, a brief literature review of modelling frameworks (Section 5.1) and reference architectures (Section 5.2) is given. In Section 5.3, the reference architecture, that is taken as a reference for the design of the EtoPlan concept for integrated manufacturing planning and control, is described and motivated. 5.1 Literature on modelling frameworks Modelling frameworks define the design steps, the modelling methodology that serves as a guidance on how to go through a manufacturing system design process. An understandable (also for non-computer scientists), complete and serviceable (in terms of standard languages, models and methodologies) modelling framework is essential for successfully completing an enterprise integration project. The most well-known modelling framework is the Computer Integrated Manufacturing Open Systems Architecture (CIM-OSA) enterprise modelling framework of the ESPRIT project AMICE (AMICE, 1993). An evaluation of the CIM-OSA framework is presented by Zwegers et al. (Zwegers, 1997a). The CIM-OSA framework is described in a highly formal way aiming at full computer automation of all functions of the developed manufacturing systems. Two other important and less formal modelling frameworks are the GRAI Integrated Methodology (GIM) developed at the University of Bordeaux, and the Purdue Enterprise

80 66 CHAPTER 5 Reference Architecture (PERA) developed by the Purdue Laboratory for Applied Industrial Control. An important feature of GRAI-GIM is the definition of four co-operating systems (decision system, information system, operating system, physical system) in the GRAI-GIM model (Chen, 1997). The role of human aspects in an enterprise integration program is specifically emphasised in the PERA modelling framework (Williams, 1994). On the basis of the function performed by humans, the manufacturing architecture and the information architecture are split up into an information system architecture, a human and organisational architecture, and a manufacturing equipment architecture (figure 5-1). Contrary to the CIM-OSA framework, GRAI-GIM and PERA cover the whole life cycle of a manufacturing system development program in a structured way. In 1994, the IFAC/IFIP Task Force on Architectures for Enterprise Integration has recommended the selection and combination of the best features of the CIM-OSA, GRAI- GIM and PERA modelling frameworks so as to achieve a best modelling framework (Williams, 1994). Consequently, the Generic Enterprise Reference Architecture and Methodology (GERAM) modelling framework is jointly developed in order to organise existing enterprise integration knowledge instead of redefining it (Bernus, 1999)(Chen, 1997)(Kosanke, 1995). The above mentioned modelling frameworks are not accompanied by a reference architecture. This lack of a generic template architecture limits the applicability of such modelling frameworks for developing functional architectures for manufacturing planning and control systems. Information functional network Manufacturing functional network Information Architecture Manufacturing Architecture Information Systems Architecture Extent of automation Human and Organisational Architecture Human component of the information architecture Human component of the manufacturing architecture Extent of automation Manufacturing Equipment Architecture Figure 5-1 The PERA modelling framework (Williams, 1994)

81 REFERENCE MODELS Literature on reference architectures A rough description of the structure of the manufacturing system is required for effectively developing subsystems of the whole manufacturing system. Enterprise integration is of major importance for the success of future manufacturing systems. A reference architecture lays the foundation of an integrated manufacturing system if it is structured in such a way that each subsystem can be embedded in the whole system without explicitly or implicitly putting up walls between the subsystems. Biemans uses the term reference model with respect to the representation of a complex system as a configuration of components (Biemans, 1989). Each component performs its own, globally defined, distinct tasks but at the same time interacts with other components to jointly complete the overall system tasks. The Manufacturing Planning and Control System (MPCS) reference model described by Biemans, shown in Figure 5-2, corresponds to the definition of a reference architecture in this thesis. The MPCS reference model contains the two interacting parts MPCS Management and MPCS Executor. MPCS Executor consists of six static hierarchical control entities for controlling the production operations. MPCS Management contains four components (Master Planner, Product & Process Developer, Supervisor, and Monitor). The Supervisor module relates to the MPCS Executor by configuring the control entities and determining their procedures. The Monitor component provides the other MPCS Management components with Executor feedback information. An important feature of the MPCS reference model is the recognition of the need for integral control of the whole manufacturing environment. Company Controller Master Planner Process Planner Configuration Manager Product Designer Machine Designer Logistic Procedure Developer Logistic Controller Developer Maintenance MPCS Management Monitor MPCS MPCS Executor Figure 5-2 The Manufacturing Planning and Control System (Biemans, 1989)

82 68 CHAPTER 5 customer orders Company Management System Engineering Technological Planning Product Design Resources, products and processes Process Planning Production orders Logistic Control Factory level Cell level Station level Equipment level Factory Activity Control (FACT) Figure 5-3 Manufacturing Planning & Control reference model (Arentsen, 1995) On the basis of both the MPCS reference model and the CIM reference model introduced by Tiemersma (1992), an adapted Manufacturing Planning & Control (MP&C) reference model, depicted in Figure 5-3, had been developed at the laboratory of Design, Production and Management of the University of Twente (Arentsen, 1995). The MP&C reference model consists of three main components, namely Company Management, Technological Planning, and Logistic Control. Strategic decisions concerning resources and products, together with decisions concerning order intake are made by Company Management. Technological Planning contains the planning tasks for specifying the features of the products (Product Design), deciding how a product should be made (Process Planning), and determining which manufacturing resources should be incorporated in the system (System Engineering). The hierarchical control concept has been adopted for the Logistic Control component (see Section 4.1). The Logistic Control component determines when the manufacturing activities should be performed and with which resources. A system architecture and an implementation for Factory Activity Control (FACT) which concentrate on the Logistic Control tasks on the Cell and Station levels and the interaction with Process Planning has also been presented by Arentsen (1995). The possibility of interaction between the logistic and technological planning components on multiple planning levels is an important improvement in the MP&C reference model. The strict distinction between the logistic and technological planning components implies, Holonic Manufacturing System Order holon Process execution knowledge Production knowledge Resource holon Product holon Process knowledge Figure 5-4 PROSA reference architecture (Wyns, 1999)

83 REFERENCE MODELS 69 however, that these planning tasks cannot be integrated properly. Interfacing between these tasks is the only possible way of tuning the separate plans. The same holds for the explicit distinction between the three subtasks of Technological Planning (System Engineering, Product Design, and Process Planning). A reference architecture for holonic manufacturing control systems has been developed at the PMA department of the Catholic University of Leuven (Wyns, 1999). This reference architecture, called PROSA, consists of three types of basic holons: Order holons, Product holons, and Resource holons (Figure 5-4). These basic holons are structured using objectoriented concepts like aggregation and specialisation (see also Section 4.3.1). The logistic planning and control tasks are performed by the Order holons, technological planning tasks are performed by the Product holons, and the Resource holons are responsible for resource allocation and resource management. PROSA is an abbreviation of Product-Resource- Order-Staff Architecture, which indicates the existence of staff holons as an addition to the basic holons. These staff holons (e.g. a scheduler) can assist the basic holons with global information or expert knowledge. Although the PROSA reference architecture does not show an explicit distinction between the technological planning and logistic control components, this distinction is implicitly modelled by assigning these tasks to the Product and Order holons respectively. 5.3 A reference architecture for Integrated Manufacturing Planning and Control A manufacturing reference architecture must first and foremost function as a basis for developing system architectures for the various manufacturing processes in such a way that the integration of the eventually developed systems is guaranteed. Hence, as an example, it should not take the design process or the production planning process as the starting point, as is the case with most reference architectures of which some have been mentioned in the previous section. A reference architecture for integrated manufacturing planning and control must emphasise the equal importance of all planning processes. Another important requirement of a reference architecture is its generic applicability regarding the range of manufacturing environments. Current manufacturing environments can not easily be classified strictly anymore in terms like make-to-stock or make-to-order, due to the increasing variety which can be observed on the shop floors. Both standardisation and direct customer influence are increasing which has lead to a combination of manufacture-to-order production of (end-)products and make-to-stock production of mainly semi-finished products. Because it is assumed that this development will continue, there is a need for a generic reference architecture that can accommodate different manufacturing situations.

84 70 CHAPTER 5 Resource information structure Order information structure Information Management Product information structure Figure 5-5 Manufacturing Engineering Reference Model and the position of Information Management (Lutters, 1999a) Mainly for these reasons, the Manufacturing Engineering Reference Model shown in Figure 5-5 has been developed in the laboratory of Design, Production and Management of the University of Twente (Lutters, 1997)(Lutters, 1999a). The reference model contains the following functions. Company Management Company Management is placed at the top of the reference model, dealing with the strategic decisions concerning the range of products, long-term resource investments, marketing and sales strategies. The strategic decisions of Company Management are beyond the scope of this thesis, in which the focus is on tactical and operational planning and control. The tactical and operational decisions are dealt with by the next three functions: Product Engineering, Resource Engineering and Order Engineering. In the figure, these functions are represented by the vertical pillars. Product Engineering Product Engineering refers to all the engineering activities related to the product life cycle of a specific type of product (Lutters, 1997). It is concerned with the design and development of a product and its variants, starting from functional requirements up to final recycling/disposal. Order Engineering All the production activities as well as all the supporting activities (e.g. the NCprogramming) necessary to make the products are planned and controlled by Order Engineering. Customer- or stock-driven production orders are decomposed into a hierarchy of planning jobs, main production jobs and auxiliary jobs. Typical tasks are internal duedate setting, determination of batch sizes and resource allocation. Next to controlling the activities that are performed internally in the company, Order Engineering also deals with the external life cycle of the orders. This task involves the communication with the external stakeholders (e.g. clients, suppliers) of the company.

85 REFERENCE MODELS 71 Resource Engineering Resource Engineering deals with all the life cycle aspects of the resources in relation to the execution of the production tasks. It therefore includes the (capability) specification, design, development, acquisition, preparation, use and maintenance of the resources of a company. Information Management Information Management is the central pillar in the Manufacturing Engineering Reference Model. Information Management effectuates all information processing requirements of the engineering tasks (Lutters, 1999a). In fact, information structures are used as the integration medium for the different design and manufacturing functions. The manufacturing information is represented in three information structures, dealing with product-, order- and resource information, respectively. Each of the structures supports the tasks of the corresponding engineering pillar. Information Management is discussed in more detail in Chapter 6. Production Production is concerned with the actual execution of the plans generated by the engineering tasks. From production, information is fed back to the engineering tasks. 5.4 Conclusion In this Chapter, two types of reference models have been presented briefly: modelling frameworks that specify the design process of a CIM system and reference architectures that provide a generic rough model of the manufacturing environment. The Manufacturing Engineering Reference Model - a reference architecture for integrated manufacturing planning and control - encompasses all manufacturing planning tasks that are to be performed in a manufacture-to-order environment. Unlike most other reference architectures, it does not focus on either the technological or logistic planning tasks, and it even does not make a distinction between these planning tasks. Information Management serves as the kernel in the reference model to streamline the concurrent generation of the product-, order-, and resource information (see the next chapter). The aforementioned features make the Manufacturing Engineering Reference Model greatly suitable to function as the basis for the development of a concurrent planning and control concept.

86

87 73 Chapter 6 Information Management In the first part of this thesis, it has been concluded that future manufacturing planning and control systems must be able to perform the planning of technological and logistic aspects concurrently (Chapter 2) without a prespecified scenario of planning steps. The variety of demands and tasks as they occur in the manufacturing environment asks for interactive planning on multiple levels of aggregation (Chapter 3). This implies that the sequential execution of the manufacturing planning processes (e.g. design, process planning, production planning) is not suitable anymore as a basis for the design of manufacturing systems. The need for more concurrency in the manufacturing planning tasks on various abstraction levels can only be achieved if different experts can make use of the same information. This requires the development of a generic information structure and an information management system for support in the co-ordination of the planning activities. In this chapter, the role of information management in manufacturing planning and control is discussed and a specific Information Management concept is described. 6.1 A diversity of information structures There is a fundamental difference between the information processing traditionally applied for either process planning or production planning & control. In Process Planning, a technological plan for realising the product ordered by the customer is generated, consisting of several levels of abstraction. First, the methods required for producing the product are selected. Second, the set-ups are determined and, third, the tool paths are calculated. Traditionally, these planning steps are performed without any concern of capacity problems that may occur due to other orders with which resource capacity must be shared. Often, even the resource allocation decision is not taken into account, in order to leave room for decision making in the production planning department. In conclusion, process planning is typically performed in a single-order oriented manner. The usual aim is to match the capabilities of the resources with the machining requirements of the product. On the other hand, Production Planning and Control (PPC) is traditionally structured in a resource-oriented manner. A factory controller considers all the available resources at the factory level. A cell controller controls the subordinate departments and workstation controllers control the workstations in the department. PPC focuses on matching resource capacities with the required processing times and the sequence of operations.

88 74 CHAPTER 6 Thus, a static hierarchical resource structure constitutes the usual basis for planning and control of the logistic aspects of the orders, while a hierarchical single order structure forms the basis for planning the technological aspects. The mismatch between these information and planning structures makes it almost impossible to perform these tasks simultaneously. The information gap between process planning and production planning has been recognised by some researchers as a problem for integrating the two planning processes. An important development is a language for process specification (ALPS) performed at the Automated Manufacturing Research Facility (AMRF) of the National Institute of Standards and Technology (NIST) (Catron, 1991). The development of ALPS is part of the AMRF manufacturing control research project of which the hierarchical control framework has been discussed in Section ALPS is used as an interface between process planning and production control, and internally between production control processes. The process specification language is based on the directed graph notation, which allows specification of parallel activities, event synchronisation, alternative processes, resource management, and task decomposition. As shown in Figure 2-6 a data gap has been recognised also by ElMaraghy and ElMaraghy together with a functional gap (ElMaraghy, 1993). Macro planning is used to describe the multi-domain (e.g. milling and turning) manufacturing operations, each of them detailed later on at the micro planning level. They propose an Integrator for bridging the functional and the data gap between existing CAPP and PPC system. The integrator module addresses the time dependent issues related to event handling, communications, and database updating. In this way, it is aimed at the generation of an effective response to actual shop floor disturbances. A drawback of the methods mentioned above is that they merely deal with interfacing rather than with the integration of the various planning processes. As such, the communication is facilitated and implicitly maximised, while it is better to minimise the need for communication by re-arranging the planning activities (Kals, 1998). Furthermore, there is still the problem of matching the process planning hierarchy based on technological requirements with the PPC hierarchy based on logistic requirements. The experiences from the MSI architecture of NIST (see Section 4.1.1) confirm this fundamental problem. The MSI architecture allows both centralised and distributed storage of data. It assumes that information to be shared between multiple systems is globally accessible and is not embedded in the systems that generate it. This imposes a similar information structure to be used for all the systems. Due to the fact that a fixed hierarchical control structure is required by the MSI architecture, the same hierarchy is used for the process plans. This causes rigidity in both production and process planning. For that reason, Senehi et. al. recognise the need for dynamic process and production planning systems (Senehi, 1992). Zhang and Zhang propose an object-oriented solution to bridge the gap in information structures (Zhang, 1996). The Object-oriented Modeling Technique (OMT) of Rumbaugh has been used for modelling the object models, the dynamic models for identifying the behaviour of the system, and the functional models to illustrate how the operations are performed (Rumbauch, 1991). Three basic object classes have been defined by Zhang and Zhang: the resource module, the product module, and the planning module. The most

89 INFORMATION MANAGEMENT 75 important functional models are process selection and machine selection. The detailed planning steps and thus the complete scenario of manufacturing planning should be prespecified while setting up the system. This is an important drawback if the aim is to design an adaptable, self-organising manufacturing planning and control system. The mismatch between information structures not only occurs between Process Planning and Production Planning. One of the major difficulties in achieving concurrent engineering is that Product Design makes use of geometrical product information structures, where Process Planning uses a process-based information structure. For this reason, much research effort has been paid to the recognition of manufacturing features in CAPP systems, that receive a complete product design, containing design features, as input (see Section 2.1.). 6.2 A concept for information management The functional planning departments (or expert groups) in a company all make their own decisions in order to meet the requirements and constraints of the company and its environment. These departments deal with the same orders, products and available resources, but all in their own specific domain. However, most decisions taken in a specific department are also of concern to other departments, because these decisions limit the solution space of subsequent decisions makers. Aiming at concurrent manufacturing engineering will increase the need for feedback and inter-departmental communication, which may lead to extremely complex and uncontrollable flows of information between the separate departments (Lutters, 1997). An additional problem is the difficulty of backward transformation of information through the planning stages (Figure 6-1). For instance, it is impossible to transform manufacturing instructions drawn up by the process planning department into the original detailed design specification. Mainly for the aforementioned reasons, a new concept for Information Management is proposed by Lutters et al. (Lutters, 1999b). The information required for making decisions is pulled by the departments instead of pushed from one department to the next in the planning cycle, as is the case in traditional manufacturing planning. The central issue in this approach is the need for information which drives the control and the course of the manufacturing planning processes. Functional specification Conceptual design Embodiment design Process planning Production planning Figure 6-1 The manufacturing planning cycle consisting of one way transformation processes (Kals, 1998).

90 76 CHAPTER 6 Resource information structure Order information structure Information Management Product information structure Figure 6-2 The three information structures as part of Information Management (Lutters, 1997). All information generated by the separate departments is attached to a general model containing the following three information structures (Figure 6-2): Product Information Structure (PRIS). Order Information Structure (OIS). Resource Information Structure (RIS). The three information structures contain all the information generated during the whole life cycle of the products, orders, and resources, respectively. These classes of objects are regarded as the basic elements present in a manufacturing environment. All the information required can be stored in one of the three information structures or as attributes attached to relations between the structures. Which planning processes are performed and in which sequence depends on the need for information initiated by external or internal events that subsequently initiate the planning processes. This approach differs fundamentally from the traditional scenario-based planning strategies where a pre-defined sequence of planning activities has to be performed. A comparison between this approach and the going concerns mentioned by Parunak (see Section 4.4) can be drawn. Both recognise the limitations of end point problem solving. In elaborating this feature for the Order Information Structure, the order life cycle is oriented towards the generation or updating of the required order information instead of towards e.g. the scheduling process. In this way, Information Management consists of an integrated collection of tasks that can be used as a basis to initiate, accompany, control and evaluate all the manufacturing processes in a structured and transparent way (Lutters, 1999b). These tasks are based on the (evolving) information contained in the three information structures, as is shown in Figure 6-2. All manufacturing planning processes - from functional design to actual production make use of the same three information structures, which allows for an effective and efficient integration of these planning processes. The object classes (products, resources, orders) that are the basis for the information structures are structured by means of the fundamental information structure depicted in Figure 6-3. The attributes only differ from the elements through the fact that they have only one link to an element or relation in the structure.

91 INFORMATION MANAGEMENT 77 attributes relation n n element attributes Figure 6-3 The fundamental information structure (Lutters, 1997). Hierarchies of the objects are built up during their life cycle, based on the fundamental information structure. This is generically done, independent of the type of object (e.g. operator, design feature, or production activity) where the information refers to. No predefined hierarchies exist in the generic information structures. The information structures evolve in time regarding both the hierarchies of the objects and the values of the attributes. The generic framework for the information structures is described in Section after which the Product-, Resource-, and Order Information Structures are presented in the subsequent sections. In this thesis, most attention is paid to the Order Information Structure (OIS), because all logistic aspects of the activities to be performed inside the company and the communication with the external stakeholders are dealt with by Order Engineering for which the OIS is the primary information structure. The OIS forms the information architecture for the concept of concurrent manufacturing planning and control, which will be further elaborated in the remainder of this thesis Generic framework for the information structures The generic framework for the information structures, shown in Figure 6-4, was first presented by Lutters at the 29 th CIRP International Conference on Manufacturing Systems, Osaka, Japan (Lutters, 1997). It contains a core model that is built up according to the fundamental structure depicted in Figure 6-3. Domain I Filter Aspect View Core model Figure 6-4 Generic framework for the information structures (Lutters, 1997).

92 78 CHAPTER 6 The elements are part of a certain aspect system of the objects, referred to as domains in the information structures. For instance, the geometric product model is a specific aspect of the product information. Therefore, a physical product definition domain is specified in the Product Information Structure (PRIS). A planner who uses and generates the information has a specific view on a certain domain of an information structure. For example, a process planner has a view on the physical product definition domain of the PRIS, and simultaneously a view on the capability domain of the Resource Information Structure (RIS). The views on the domains provide the right interpretation of the information for the process planner, in order to provide the best support for decision making. The process planner gets, for instance, a view on the manufacturing features of a product for determining the set-ups for machining the product. The views contain no information themselves, but provide a view on a domain in the core model. This means that multiple views on the same domain are mutually dependent. A view furnishes a focused, partial representation and specification of the information in a certain domain of the Product-, Resource-, or Order Information Structure (Lutters, 1997). Filters are put over the views in order to focus on a specific part of the domain. For instance, when a process planner performs a detailed process planning activity (e.g. toolpath planning), he or she will only be interested in a small part of the geometric product information. Filters are applied to leave out all information, which is at that time not relevant to the decision maker. In the following three sections, the information structures for products, resources, and orders are modelled in more detail to provide information architectures for building concurrent manufacturing planning and control systems Product Information Structure (PRIS) The Product Information Structure (PRIS) was the first of the three information structures that was described in more detail (Lutters, 1997). All information related to a product model that is of concern for and/or results from a manufacturing planning decision is managed by the PRIS. This information can relate to any stage of the product life cycle; it can address e.g. function, quality, or geometry. The following three domains of the PRIS have been discerned by Lutters et al. (1999b): The objective domain. The physical product definition domain. The control domain. The objective domain comprises of the information that delineates the specifications and requirements of the product (e.g. functional structure, conceptual design). The physical product definition domain is used to establish the physical elements of the product, together with the attributes that further specify these elements (e.g. shape, material). All the information describing the adaptable behaviour of the product is contained in the control domain (e.g. software). The domains themselves are rather independent; e.g. one and the same function can be realised by several (combinations) of geometric elements.

93 INFORMATION MANAGEMENT 79 Based on the PRIS, a method for variant based cost estimation is proposed by Ten Brinke et al. (Brinke, 1999). The structure defines a product in terms of elements and their relations, which can be used for comparing products regarding costs. Because the other information structures contain information for cost estimation as well, the concept is further developed to include the Resource Information Structure and the Order Information Structure. An extension of the cost estimation methodology towards the inclusion of production planning aspects has been proposed (Huttinga, 2000) Resource Information Structure (RIS) The Resource Information Structure (RIS) contains all the information about the physical resources in the company. A physical resource can e.g. be a machine tool, an operator, a fixture, a tool, a blank material, or a (semi-)finished physical product. All physical entities in a company that are required for performing an activity are modelled as resources in the RIS. This includes the physical product(part)s; the product model is stored and managed in the PRIS. All the information that is generated during the whole life cycle of the resource, from the initial investment considerations until final disposal, is managed in the RIS. Next to the physical resources there are also aggregate resources in a company. A milling department can, for instance, be specified as an aggregate resource consisting of multiple similar or different milling machine tools. The same holds for an FMS workstation containing multiple resources with various capabilities. The hierarchical grouping of the resources can vary over time in order to be able to draw up temporary hierarchies as will be described in Chapter 7. Two domains can be discerned in the RIS (Kals, 1998): The physical resource domain. The method domain. The physical resource domain contains two important views for manufacturing planning and control. The capability view refers to the processing capabilities of a physical resource. With respect to a machine tool, the capability specification may, for instance, include information on the type of machining processes (e.g. milling), the dimensions of the possible workspace (sizes of the bed), accuracy range, or the number of axes. The capacity view provides information on the time-related aspects of the resources. It includes information about the mean and variance of the utilisation, mean time between failures, tool change times, etc. The method domain contains information on the processing types (methods) such as milling, turning, or assembly. The methods are in themselves independent of the physical resources in the company. However, their application is directly related to the resources in the company or to the processing capabilities of other companies with which the company has a relation regarding outsourcing. The capability and capacity information of a certain group of resources determine the actual method description.

94 80 CHAPTER Order Information Structure (OIS) The Order Information Structure deals with all the information that provides the basis for, and results from, decision making throughout the life cycle of the orders. That means, from the initiation by the client until the actual production or even until the moment that the ordered product is turned into scrap again. Because this thesis focuses on the production and planning activities that take place in the factory, especially the part of the order life cycle that covers the activities to be performed between the order acceptance and the final delivery is elaborated further. An order contains two different kinds of information, related to either external or internal activities to be performed. First, the information related to the external activities of the order comprises e.g. client data, quotations, prospects, prices, supplier data, delivery dates, number of products. Second, a client order initiates an aggregate activity that has to be executed by the company. The aggregate activity is subdivided in a hierarchy of activities for e.g. working out a manufacturing plan, purchasing the required materials and, at the end, really producing the ordered products. The framework of the Order Information Structure (OIS) is shown in Figure 6-5. In conformance with the PRIS and the RIS, also the OIS contains more than one domain. Information about production activities and manufacturing planning activities that are to be performed or have been performed in the past is stored and managed within the activity domain of the Order Information Structure. An external order life cycle domain has been defined for handling the information regarding the contacts with the stakeholders outside the company (e.g. clients, subcontractors). These domains are described in the next two sections. In Section a classification of orders is given in order to identify the order aspects that are significant for the planning of orders in manufacturing companies. Activity Domain Aspect View: Capacity Planning Aspect View: Resource Requirement Core model Figure 6-5 Order Information Structure (OIS)

95 INFORMATION MANAGEMENT The external order life cycle domain An important aspect of order management concerns the communication with the stakeholders outside the company. The stakeholders determine the existence of the company. They do, either directly or indirectly, substantially influence decisions of the company. An adequate communication with the stakeholders is, therefore, vital for the company. The most important stakeholders outside the company are the clients, the suppliers and the subcontractors. But also stakeholders like potential (temporary) employees, shareholders, trade unions, government authorities or pressure groups can be of great influence to the policy of the company. Having a clear understanding of the (potential) clients and their needs is essential to create the right conditions for profitable manufacturing. It becomes more and more unacceptable that the sales and marketing departments autonomously decide about quotations and acceptance of the orders. The specific knowledge of the engineers and the production managers is indispensable for making deliberate estimates about manufacturability, costs and lead times. On the basis of the information obtained from the specialists, together with their own information about the market, the sales department should specify a competitive and feasible quotation. Next to the process of obtaining profitable orders (the company as a supplier), there is also a process that deals with subcontracting and the purchasing of materials or semi-finished products (the company as a client). Again, these decisions cannot be taken autonomously by the purchaser. Information about former orders is kept in the Order Information Structure for reasons of, for example, analysing the performance of the company, forecasting, making investment decisions, and traceability. Although many of these tasks will use information from the activity domain, also information about the external life cycle of the order is used for performing these tasks. History information that concerns product and resource aspects is not stored in the OIS, but reference is made to this information by linking it to the orders (activities) for which or during which this information was created. In this way, the information is kept in the Information Structures and accompanying domains it belongs to and reference is made to the former executed activities in the OIS for traceability purposes. Examples of views on the external order life cycle domain are views for supporting sales, purchasing, subcontracting, marketing, and investments The activity domain Many different kinds of activities are performed in a manufacturing company. Next to the activities on the shop floor for producing (series of) a product before a specified due date, there are supporting activities in order to be able to satisfactorily execute the production activities. A supporting activity can, for instance, be a maintenance activity on a resource, a tool assembly activity, a cleaning activity, an engineering activity, or a scheduling activity. While on the one hand the optimisation of the production plans and on the other hand the minimisation of the planning costs generally appear to be conflicting goals, it is necessary to find the right balance between them. The costs for making the production plans do often

96 82 CHAPTER 6 exceed the costs for actually producing the products. This is especially true for engineer-toorder manufacturing in small job shops. Consequently, it is best not to discriminate between the activities to be performed in the company. Just like a machine tool can become a bottleneck in the company, also a process planner can become a (temporary) critical resource. It doesn t matter whether it is a production or a planning activity that causes a delay. All logistic information related to the activities are, therefore, stored in the same activity domain of the Order Information Structure. Logistic information is all the information about time, place, required resources and batch size. As mentioned in Chapter 3, the activities are planned and executed on various abstraction levels. The level of abstraction is high on the upper levels of the planning hierarchy. The lower level activities have more details included in their plans, hence the level of uncertainty in the planning data decreases when going down the hierarchy. This hierarchy in order abstraction is laid down in the activity domain of the OIS. The establishment of these order hierarchies and the relations with the resources required for executing the activities is elaborated further in Chapter 7. In the remainder of this thesis, the objects Order, Resource, and Products are henceforward written in SMALL CAPS. This is done to clearly state that the generic object is meant. Instantiations of these objects can either occur in the most detailed form or in a highly aggregated form. For example, a RESOURCE can be a single tool but also a workstation containing several machine tools. Likewise, an ORDER can refer to a tool assembly activity, an engineering activity, a job, or a production order containing several jobs ORDER classification for aggregate planning In manufacture-to-order environments, the ORDERS in the order mix greatly differ with respect to the planning goals, priorities, and constraints. This makes order planning and control more complex and asks for dynamic planning strategies that take the individual characteristics of the ORDERS into account. In this section, a generic ORDER classification model is presented that is used to recognise these ORDER characteristics. ORDERS can be of various kinds, e.g. client orders, internal orders, suborders, maintenance activities, or engineering activities. The classification model comprises seven generic dimensions that are of importance for determining the objectives, priorities, and constraints in the aggregate planning of individual ORDERS. The focus is on the order acceptance and resource loading planning processes (Giebels, 2000). Below, the seven dimensions are mentioned and further explained. Versatility The versatility indicates how many possible routings, alternative resources, and alternative NC-plans (or manual instructions) are available to actually execute the ORDER. Alternative

97 INFORMATION MANAGEMENT 83 technological plans may, for instance, be represented by way of Non-Linear Process Plans (NLPP), see e.g. (Tönshoff, 1989). Abstraction level The abstraction level indicates the level of detail of the information in the plans (see Chapter 3). State of acceptance The dimension state of acceptance indicates the ORDER s state of acceptance in the different manufacturing plans. An ORDER may, for example, still be in the state of order acceptance (i.e. drawing up an offer), or it is already accepted and included in the resource loading plans, or the ORDER has already been dispatched to the shop floor. Consequently, the acceptance shows the ease of cancelling the ORDER or changing its due date. External stakeholder priorities This dimension refers to the priorities the company agrees upon with its external stakeholders. Concerning resource loading and order acceptance, the important external stakeholders are clients, subcontractors, and suppliers. Typical external stakeholder priorities are due dates, quality, and prices. Impact on resource loading The impact on resource loading depends on the following three aspects. Resource-availability with respect to the ORDER This aspect refers to the availability of the applicable resources needed to execute the ORDER in its planning period. The planning period is defined as the time period between the internal start time and the internal due date of the ORDER. Relative size ORDERS vary in impact on resource loading due to their sizes with respect to processing times, set-up times, and number of products. Relatively large ORDERS have a greater impact on the overall resource load than smaller ORDERS. Risk of negatively affecting loading performance Various ORDER complexity issues may affect the loading performance. For example, ORDERS that must be partially subcontracted may disturb the resource loading plan(s) due to uncertainty in delivery dates. ORDERS with a high chance of failure may also disturb the actual execution of the loading plan. Furthermore, complex ORDER structures and troublesome constraints/demands on certain ORDERS may disturb planning process, the loading performances, and the actual execution of the loading plan(s). Relative slack on critical path The slack for an ORDER is the time between its internal release time and internal due date (i.e. planning period) minus its processing time (including set-ups). The relative slack on critical path equals the slack divided by the length of the planning period. The relative slack represents the flexibility to shift the ORDER in time within the constraints of its planning period. For instance, the relative slack shows that an ORDER with a processing time of four

98 84 CHAPTER 6 hours and a planning period of two weeks can be planned more flexibly than an ORDER with a processing time of four hours and a planning period of one week. Strategic importance This dimension refers to the strategic importance of the ORDER. Although hard to quantify, there will undoubtedly be ORDERS that have a higher strategic importance to the company than other ORDERS. As an example, one may consider an ORDER of an important client, or an ORDER that may result in a subsequent (profitable) ORDER from the same client. The ORDER classification makes clear that ORDERS cannot be treated uniformly in the various aggregate planning functions. The seven dimensions serve as a template to define the objectives and the constraints of the ORDERS concerned. E.g. an ORDER with little relative slack on the critical path should be treated as a rush ORDER. Or, a certain ORDER that has a good chance of successive profitable repeat ORDERS is of relatively high strategic importance Manufacturing planning views This thesis focuses on the process planning and production planning activities in a manufacture-to-order environment. Consequently, two important views, for capacity planning and macro process planning, are briefly discussed in this section. The reader, however, may note that the actual implementation of these views depends on the specific manufacturing situation. Furthermore, many other views on the information structures are necessary for overall manufacturing planning and control, a.o. a scheduling view and a cost view. Capacity planning view The capacity planning view presents combined information from the activity domain of the OIS and the physical resource domain. For example, current estimations about the availability of a certain RESOURCE on the longer term are presented to the user, together with an overview of the ORDERS that have already been planned roughly. With this information, the user can - manually or with the help of a decision support tool adjust the plans, plan additional RESOURCES, or add newly entered ORDERS. In Chapter 9, some views related to capacity planning are presented which have been implemented in the prototype implementation of the EtoPlan system. Macro process planning view The macro process planner determines the rough routing of a product through the factory. The process planner is potentially interested in several views depending on the tasks he or she is performing at the time. A view on the physical product domain of the PRIS, which shows the required manufacturing features, together with a capability view on the method and/or physical resource domain of the RIS is required, when the decisions have to be made on the set-ups and their sequence. In the case that the PRODUCT is not yet completely designed, it may be worthwhile to investigate some what-if scenarios regarding the manufacturability or the costs of the PRODUCT. Especially the use of a variant-based

99 INFORMATION MANAGEMENT 85 approach, where similar product types with regard to the processing methods are compared, can be useful for achieving this important step in Concurrent Engineering (Lutters, 1999a). 6.3 Conclusions The differences in the information structures used in the various manufacturing planning processes (design, process planning, production planning) is a major problem on the way to achieve integrated manufacturing planning and control. We conclude that the information structures must be generic and of an evolving nature, the latter for being able to support evolving manufacturing planning processes and to be able to adapt to changing situations in the company. This holds for all the information structures used in the company. Especially the hierarchies used in the planning of manufacturing activities must be dynamic, in order to prevent a rigid planning and control structure. Real integration can only be achieved if the hierarchies used for process planning correspond to or can be translated into the hierarchies used for production planning. In this way, interactive manufacturing planning becomes possible. The Information Management concept presented in Section 6.2 fulfils the requirements for concurrent manufacturing planning and control. By applying an information pull approach for initiating planning processes, the need for communication between the planning processes is minimised, facilitating concurrent planning activities (Lutters, 1999b). The generic information structures for ORDERS, RESOURCES, and PRODUCTS can comprise all the information generated in the manufacturing companies. The planning processes are provided with dedicated views on relevant information in support of the decisions to be made. The Order Information Structure is presented in this chapter. It deals with the information about the logistic aspects of the ORDERS (activities), which is of importance for the further development of EtoPlan in the next chapters. The order classification presented in this chapter makes clear that orders cannot be treated uniformly in the various aggregate planning functions. The seven dimensions of the ORDER classification scheme serve as a template to define the objectives and the constraints of the ORDER

100

101 87 Chapter 7 System Architecture In Chapter 6, it has been concluded that the gap between the fundamentally different information structures applied for manufacturing planning can only be overcome by developing generic and adaptable information structures. To this end, we propose the use of three generic information structures for PRODUCTS, RESOURCES and ORDERS in the EtoPlan concept. A second conclusion, drawn from the discussion in Chapter 6, stated that the planning hierarchies used for process planning and production planning and control must be similar. This can only be achieved if the hierarchies themselves are adaptable. As a result, the system architecture proposed in this chapter is based on building temporary hierarchies for manufacturing planning and control. In Section 7.1, the basic elements of this concept, i.e. Applicability Groups (AGs) as a way to model temporary hierarchies, are introduced. The functional architecture of the AG Controllers is presented in Section 7.2. Aspects of autonomy and co-operation are next discussed in Section 7.3. In Section 7.4 a special resource group for performing centralised scheduling is introduced. Finally, in section 7.5 it is investigated how the EtoPlan system architecture differs from the already existing control concepts that were described in Chapter Temporary hierarchies of Applicability Groups (AGs) As has been concluded in the previous chapters, there is a need for building flexible temporary hierarchies (holarchies) of manufacturing RESOURCES in order to be able to handle the diversity in characteristics of the ORDERS that are executed simultaneously. In this section, the structural model of EtoPlan is presented. The structural model describes the structure of the control entities and the relations between them. In the EtoPlan concept, holarchies are built by dynamically grouping the RESOURCES according to the requirements of individual ORDERS. This allows many different kinds of PRODUCTS to be manufactured simultaneously while still being able to consider the multilateral interactions between the ORDERS. This is an important feature for small batch manufacturing, where the routings of the diverse production ORDERS may vary highly. The multiple and temporary hierarchies (holarchies) of RESOURCES are built up as follows.

102 88 CHAPTER 7 For every individual ORDER, a unique group of applicable RESOURCES is drawn up. A RESOURCE is considered applicable if: The RESOURCE is considered capable of meeting the already known technological requirements in the roughly defined process plans for executing (a part of) the ORDER. The RESOURCE is considered to be available during a time period that is roughly planned for executing (a part of) the ORDER. An Applicability Group (AG) is a virtual group of all RESOURCES which are considered applicable for the (partial) execution of a given ORDER. Figure 7-1 shows the entity relation diagram of an AG. An AG consists of at least one RESOURCE. A RESOURCE can relate to n AGs. Every AG belongs to a single ORDER. However, an ORDER does not necessarily relate to an AG. Namely, it is possible that the AG configuration process has not been performed yet. In the above definition of an Applicability Group, the word virtual refers to the fact that the configuration of an AG does not imply a physical grouping of RESOURCES, as is e.g. the case in Group Technology. RESOURCES may be anything, e.g. a machine tool, a tool, a fixture, material, an operator etc. Only the RESOURCES that are relevant for the planning and control decisions, performed on the right aggregation level for a given ORDER, are grouped in an AG. On the higher levels of aggregation, mainly machine tools will be considered. For some specific production ORDERS it can, however, be recommended or even may be necessary to take other kinds of RESOURCES into account. E.g. in many manufacture-toorder companies, machine utilisation rates are relatively low and the capacity constraints mainly depend on the operators available on the shop floor. In that case it is necessary to include the planning of operators into the capacity planning task, which also makes it necessary to determine what and how many operators should be allocated to the AGs. When more detail is included in the manufacturing plans, i.e. on the lower levels of abstraction, also the need to include more RESOURCE types (e.g. tools) in the plans increases. Then, these RESOURCES are added to the AGs. This makes the planning of for instance tool paths in micro process planning or the logistic planning of auxiliary jobs (e.g. tool assembly) possible. The AGs are drawn up, split up and changed by different planners depending on the type of planning activities that are performed at that time. For instance, when an ORDER for a new PRODUCT enters the company, an engineer will specify some (aggregate) process steps that together describe a rough routing of the ORDER through the factory. For every determined process step, the engineer initiates and configures an AG that will probably contain many applicable RESOURCES, e.g. all milling machine tools. When the aggregate process steps AG 0..n 1..n RESOURCE ORDER Figure 7-1 The AG entity-relation diagram

103 SYSTEM ARCHITECTURE 89 are subsequently planned in more detail, the process planner splits up the AG into a couple of (sub)ags that individually belong to a definite manufacturing activity (e.g. a set-up). Also the logistic aspects in production are taken into account when planning the AGs. Thus, a capacity planner can for instance remove a specific RESOURCE from an AG when this RESOURCE threatens to be overloaded due to allocation of other ORDERS in the same time period. The AGs exist throughout the life cycle of the ORDER for which they are created. An AG is initiated by either a higher level AG or a newly entered ORDER. It exists until all related jobs are completed. Depending on the complexity of the PRODUCT to be produced, the number of hierarchical levels in the ORDER structure will differ. Although the AGs are directly related to the life cycle of the ORDERS, the AGs are dealt with by the Resource Information Structure (RIS, see section 6.2.3), because the AGs consist of RESOURCES. AGs are, therefore, specified in the physical resource domain of the RIS. A new view on the physical resource domain, named Applicability View, is established by combining the capability and capacity views for managing AGs. RESOURCES are part of multiple AGs on multiple levels of aggregation as there are often numerous ORDERS being planned and executed simultaneously. This very likely occurs in a manufacture-to-order environment. For example, a resource can simultaneously be part of 1) a small AG connected to a job which is presently executed on a workstation and 2) a large AG connected to a production ORDER for which only the macro process planning has been performed and a detailed routing has not yet been generated. In practice, a resource can be part of up to hundreds of AGs simultaneously (see Chapter 9). The EtoPlan concept aims to take into account both the optimisation of global goals in production and the solution of short-term production problems. The solution of global planning problems normally covers a relatively long time period, whereas detailed operational activities concern a significantly shorter time period. This is, however, not always the case. Some planning activities already require considerable detail in an early planning stage. For example, when the production of a non-modular fixture for a certain operation can only be performed outside the company, then the detailed fixture design and, thus, the detailed process planning must be performed in an early stage. The levels of detail and the time of planning largely depend on the characteristics of the ORDERS considered. Therefore, the abstraction levels in EtoPlan are determined in an order-oriented way.

104 90 CHAPTER 7 AGs can appear in various states of aggregation. Below, two different examples of AG creation situations are briefly discussed. Example 1. If it is currently only known that a large number of cutting operations in probably three set-ups are required to execute an ORDER, it is likely that three AGs consisting of specific groups of cutting machine tools that meet the required processing capabilities will be created. In Figure 7.2 such an example is shown. An aggregate ORDER 1 is subdivided in three child ORDERS. For each child ORDER, an accompanying AG is drawn up. The arrows between the child ORDERS indicate that the child ORDERS must be performed sequentially. AG 1 AG 1.1 All milling machine tools Order 1.1 2UGHUÃ Order 1.3 AG 1.3 All CNC-milling machine tools Order 1.2 AG 1.2 Figure 7-2 AG example 1 Example 2. If a special group of operations can only be executed on a dedicated machine tool with pre-assembled cutting tools and fixtures, an AG consisting of the aforementioned RESOURCES and two AGs for performing the auxiliary ORDERS (tool assembly and fixture mounting) will be created by the planners. Such a situation is shown in Figure 7.3. Two different operators are part of the AGs for performing the operations; one operator for controlling the CNC-milling machine tool (Operator 1, ORDER 2.3) and one operator for performing the tools and fixture mounting activities (Operator 2, ORDER 2.1 and ORDER 2.2). A process planner is also added to the AG belonging to ORDER 2.3, because the NC-program for the CNC machine tool has not yet been defined. AG 2.1 Order 2.1 AG 2.3 Order 2.3 Order 2.2 AG 2.2 Figure 7-3 AG example 2

105 SYSTEM ARCHITECTURE 91 Figure 7-4 Time overlap within an ORDER hierarchy In summary, no static hierarchies of RESOURCES are defined, but for every production order a temporary hierarchy (holarchy) of RESOURCES is created. This temporary hierarchy of RESOURCES (AGs) corresponds to the hierarchy of the order structure. The manufacturing planning and control functions are performed by the AG Controllers. Every AG in the hierarchy, and thus every ORDER, is controlled by an accompanying autonomous AG Controller. When an AG is instantiated from a higher-level AG, the responsibility for the planning and control of the activities to be performed by the new AG in a specific time period is shifted on to the AG controller belonging to this new AG. The planning of the activities technological as well as logistic is then performed by the planning function of that AG Controller. Figure 7-4 schematically shows the overlap in time periods considered by the AG Controllers on different aggregation levels. The dark Gantt bars represent the ORDERS that are already created from a higher-level ORDER by the AG Controller of the AG belonging to the higher-level ORDER. The light-grey Gantt bars have already been planned but not yet dispatched by the higher-level AG Controller, which means that no AG Controllers for these AGs have been initiated yet. Until an autonomous AG Controller is initiated for such a light-grey ORDER, all planning activities considering this specific ORDER are still performed by the higher-level AG (ORDER) Controller. The functional architecture of the AG Controllers is discussed in section Method Applicability Groups (MAGs) Applicability Groups are created around all ORDERS that have to be executed in the company. In a manufacture-to-order environment the process information is at best roughly defined when the orders enter the company. This results in a large number of RESOURCES that are considered applicable for the execution of the highly aggregate ORDERS, cf. Figure 7-2 for an example. In most cases, a technological abstraction of the required capabilities of the RESOURCES is used to determine the applicability of the RESOURCES for such an aggregate ORDER. For instance, all milling machine tools are allocated to the AG as was the case for Order 1.1 in Figure 7-2. In order to facilitate the creation of the AGs by the planners, the RESOURCES are statically grouped in Method Applicability Groups (MAGs). A method is a technological abstraction of the capability of RESOURCES as described in section

106 92 CHAPTER AG Controller The functional architecture of the AG Controller, shown in Figure 7-5, is based on the generic building block for production planning and control as described by Arentsen (1995). An AG Controller contains the four functions Planning, Dispatching, Monitoring and Diagnostics. Interaction with the other Controllers occurs in three different ways (see Figure 7-5): Vertical interaction with the AG Controllers of the parent ORDERS and child ORDERS. Horizontal interaction with the AG Controllers of the ORDERS with the same parent ORDER. Bottom-up interaction via the RESOURCES. The first two interactions roughly correspond to the way interaction occurs in traditional modified hierarchical control systems (see section 4.1) A difference is that the AG Controllers only interact with AG Controllers within the same ORDER hierarchy. The third type of interaction, bottom-up co-ordination from the RESOURCES, is needed because AGs are configured in dynamic and temporary hierarchies. Hence, it is not possible to configure control hierarchies on the basis of static groups of resources, as is traditionally applied in hierarchical control. There will be a lack of overall control when bottom-up interaction from the resources is not made possible. In the latter case, the control flows are restricted to the co-ordination of ORDERS within the same ORDER structure. It is, hence, necessary to define ways of interaction between AG Controllers belonging to different ORDERS. In order to prevent an overload in control communication which may happen if AG Controllers interact with too many other AG Controllers in the system - it is decided to delegate to the RESOURCE Controllers the task to intervene when conflicting plans with respect to resource utilisation are proposed by the AGs concerned. An AG Controller is divided into an off-line part and an on-line part. The two functions belonging to the off-line part are Planning and Diagnostics. The off-line Planning function receives its input from the higher-level AG Controller. Siblings AG Controllers Feed-forward AG Controller Planning Diagnostics Dispatching Parent Order AG Controller Feed-back Monitoring Off-line On-line Resource Controllers Child Order AG Controllers Figure 7-5 Functional architecture of the AG Controller

107 SYSTEM ARCHITECTURE 93 The following two main planning tasks are performed by the Planning function: Initiation and configuration of the lower level AGs. Determination of the requirements and constraints for the lower level AGs. A wide variety of manufacturing planning tasks is performed by the Planning function. The requirements and constraints considered by the AGs are not limited to logistic aspects. They may include all the manufacturing information, i.e. information from all three information structures (PRODUCT, ORDER, and RESOURCE information). So, next to the logistic planning activities, also the design and process planning activities are performed by the Planning function of the AG Controller. The same holds for the other three functions of the AG Controller, e.g. process monitoring, performed by the Monitoring function, regards logistic aspects (e.g. lateness), product quality aspects, and process execution aspects (e.g. tool wear). Interaction with the Diagnostics function serves to inform the planning function about the current progress of the activities of the already dispatched lower-level AGs. The output of the planning function (the work plan) is handed over to the on-line Dispatching function for execution of the work plan. Together with the Monitoring function, the Dispatching function belongs to the on-line part of the control building block. On-line means that the considered time horizon coincides with the time horizon considered by the planning and diagnostics (off-line) functions of the lower-level AG Controller. This means that the dispatching function only dispatches those tasks in the work plan that can already be planned by the lower-level AG Controller. The planning of all the tasks in the work plan that are to be executed later on still remains the task of the off-line planning function of the current AG Controller. On the other hand, no planning activities can be applied for the tasks that have already been dispatched to a lower level AG Controller; they will be executed by the lower-level AG Controller. The higher-level AG Controller monitors the progress of the planning and execution activities, of the lower-level AG Controllers, by its on-line Monitoring function. The diagnostics function of the lower level AG Controller informs this monitoring function about execution problems, such as job tardiness, quality problems or resource breakdowns. The monitoring function in turn may then order the dispatching function to inform the other lower-level AG Controllers about the problem in case of an emergency situation. Also the Diagnostics function is informed so as to analyse the impact of the problem related to the current work plan. The planning function may subsequently decide to perform a replanning activity in order to fix the problem or at least to minimise the propagation effects.

108 94 CHAPTER Autonomy and co-operation Decision making is performed in a bottom-up approach with respect to the RESOURCES together with a top-down decomposition of the ORDERS. This combined approach is most suitable due to the relative long life cycle of the RESOURCES and the considerable shorter life cycle of the ORDERS as is characteristic for manufacture-to-order environments. The ORDER hierarchy is build up by repeatedly decomposing parent ORDERS into child ORDERS that subsequently become parent ORDERS of their child ORDERS. After the child ORDERS are dispatched and their accompanying AG Controllers are created, they themselves are responsible for achieving the objectives within the imposed constraints. As mentioned earlier, only in case when the AG Controllers of the child ORDERS cannot solve occurring problems, the planning and dispatching functions of the AG Controller of the parent ORDER are activated again. In other words, the AG Controllers of the child ORDERS are completely autonomous in local decision making within the constraints that are imposed for achieving the overall goals. The objectives that relate to all instantiated objects that are relevant for production control, both ORDERS and RESOURCES, are achieve by co-operation in AGs. As described in Section 7.1, AGs are configured for all the ORDERS in the manufacturing system. The AGs consist of the main RESOURCES that are relevant for the planning and control decisions concerning the ORDER. These RESOURCES are also part of many other AGs, which makes it possible to notice availability conflicts between the plans of the ORDERS in the systems. The RESOURCE Controllers are responsible for initiating planning activities for solving such conflicts. A solution can either be found by one or more of the involved AG Controllers or by initiating an additional planning entity, such as a Scheduling Group, to be described in the next section. 7.4 Scheduling Groups At shop floor level a scheduling function must be performed. Scheduling means the timebased allocation of ORDERS (jobs) to RESOURCES (workstations). Often, there is a need for automatic scheduling at the shop level (see Section 3.2.3). If a centralised scheduling algorithm is required, it is necessary to define and limit the number of RESOURCES and ORDERS considered. The use of a time horizon is a common way of limiting the number of ORDERS considered for the scheduling. The RESOURCES are often, but not necessarily, grouped according to their location in the factory, e.g. a cutting department. Subsequently, the objectives and constraints must be defined. Scheduling Groups are integrated in the EtoPlan architecture in order to be able to perform such scheduling tasks. A Scheduling Group (SG) contains RESOURCES, PRODUCTS and ORDERS. The PRODUCTS are of concern for the scheduling function due to possible batch merging and splitting considerations. A well-known example of the importance of product information on batch merging is the nesting problem of sheet metal products (Vries, 1994).

109 SYSTEM ARCHITECTURE 95 The task of a SG on shop level is the time-based allocation of ORDERS to RESOURCES. This is done within the limits of the constraints determined on the basis of the resource loading plans and the technological process information. Furthermore, the SG monitors the progress of the work so as to reschedule or otherwise to give advise to the lower-level control entities. A SG does not plan the detailed operations. Instead, the AG Controllers on the lower control levels perform these planning activities. The AG Controllers, in turn, are completely autonomous regarding the planning of the operations (e.g. process parameters, tool paths) as long as the constraints set by the SG are not violated. 7.5 Conclusions As mentioned earlier, the EtoPlan concept comes closest to holonic manufacturing control. A traditional resource-dependent hierarchical structure containing e.g. factory, cell, workstation and equipment levels where the planning function are statically specified at the control levels, is not applied in the EtoPlan concept. The temporary planning hierarchies in EtoPlan are dependent on the ORDERS currently considered in the planning. These will dynamically be created and deleted, when an ORDER is created or deleted, respectively. In this way, a flexible control system is built consisting of autonomous, cooperating control entities of the RESOURCES and ORDERS, structured in temporary orderoriented hierarchies (AGs). The system configuration is thus continually adapted according to the dynamically changing manufacturing situation. An important aspect of the concept is the paradigm shift from a RESOURCE focused manufacturing control structure to an ORDER focused control structure. In other words, the control structure is continually reconfigured around the activities that are to be performed.

110

111 97 Chapter 8 Aggregate ORDER Planning The EtoPlan ORDER planning methodology supports the integration of macro process planning, resource loading, and order acceptance. Resource loading has been defined as a view on the aggregate planning of ORDERS, considering the availability of the RESOURCES. Macro process planning concerns the rough technological planning of aggregate ORDERS. Preferably, the resource loading process is performed concurrently with the macro process planning process. In this way, the decisions to be taken can be tuned to one another. The goals of the EtoPlan decision support concept for aggregate ORDER planning are: To provide a reliable view on the availability of RESOURCES, Applicability Groups (AGs), and Method Applicability Groups (MAGs). To support the aggregate planning of ORDERS with regard to start times, lead times, and variable costs. To provide decision support for the order-acceptance process concerning: Due date setting and feasibility analysis. The estimation of variable costs that result from critical situations in production planning. The aim of aggregate ORDER planning in EtoPlan is to provide resource loading views, on the basis of the order-oriented planning hierarchy of AGs that are primarily created by the macro process planning process. In doing so, it becomes possible to support the macro process planning task with the logistic consequences of technological decisions. Due to the significant incompleteness of technological plans in the macro process planning stage, it is inevitable to deal with the uncertainties in the information regarding processing times, necessary RESOURCES, and routings. In other words, the input information for the planning of ORDERS in time cannot be considered deterministic. Therefore, in Section 8.1 it is described how the uncertainties are modelled in the EtoPlan ORDER planning concept. The ORDER planning methodology itself is elaborated in Section 8.2. Next, the resource allocation process is discussed in Section 8.3. Finally, the differences in orders types and the consequent influences on the planning procedure are discussed in Section 8.4.

112 98 CHAPTER Uncertainty modelling for aggregate ORDER planning The level of abstraction in the plans decreases as the plans are further filled in (see Chapter 3). On the higher levels of abstraction, where the resource loading task is generally performed, mostly rough planning decisions, related to method selection, subcontracting, and route determination, are taken. Because of that, the information input for resource loading is often incomplete. It is not exactly known how much time a certain ORDER will take, as a result of: the lack of detailed process planning information, and regularly occurring disturbances. In applying resource loading early in the process of concurrent manufacturing planning some difficulties are encountered. The allocation of specific RESOURCES to the ORDERS is often not possible in such an early stage. In particular, for order acceptance one has to quote an initial price and delivery date on short notice, leaving no time for a detailed process plan to be made at that stage. As a consequence, the quotation has to be based on rough capacity plans. Similarly, not all production ORDERS are already known for the relatively long time period that is considered. Hence, it does not make sense to apply the traditional methods for resource specific loading and scheduling, viz. based on deterministic input data, at such an early planning stage. The output of a planning system will never be better than the quality of its input data. In spite of the uncertainty in, and the incompleteness of the plans, reliable information must be obtained about e.g. the contingency of capacity problems in the near future. Consequently, the planning methodologies for production planning must take the unreliability of the information into account. In EtoPlan, the uncertainties are accounted for explicitly in the planning method. Below, it is described how the uncertainties regarding lead times, processing times and start times are modelled Lead times The actual lead time of an ORDER results from the whole manufacturing planning and control process and will, therefore, only be known at the end of the internal life cycle of the ORDER. However, lead times are particularly useful in resource loading when little information about the ORDERS is known. In that planning phase, it is, for example, often not yet known how many processing steps an ORDER has to go through, what the precise processing times of these steps will be and, most importantly, how many other ORDERS will compete for the same scarce resources. The latter mainly determines the waiting times between the processing steps. Estimates of the lead times - with which the plans can be established - are needed, because of the fact that the information is not yet known in detail. The estimation of the lead times is based upon estimations of the processing times, the number of process steps, and the waiting times (mainly caused by congestion due to capacity constraints). Due to the uncertainty in the processing times and especially the

113 AGGREGATE ORDER PLANNING 99 Figure 8-1 Beta density function waiting times, rough estimations of the lead times are made by the planners. Both the process planners and the logistic planners make a specific contribution to the estimation process. The process planner can provide the information regarding processing times and routings, whereas the logistic planner generally has a better insight into the waiting times. It is important to note that the waiting times are based on the resource loading plans which themselves are based on the lead times. The lead times must, therefore, evolve with the utilisation levels in the resource loading plans that will subsequently alter as a result of the lead time alterations. This type of circular reasoning is of considerable importance for the implementation of a reliable resource loading system. A reliable way of modelling the uncertainty of lead times is by fitting Beta-distributions on estimates of the minimum, maximum, and most likely lead times (Figure 8-1) (Lau, 1995). The Beta density function can be determined in the following way: Let a denote an estimation of the minimum lead time, b an estimation of the maximum lead time, and finally m an estimation of the most likely lead time. The mean value µ of the lead time is defined by µ = 1 ( a + 4m + b) 6 (1) Similarly, the variance ν is defined as 1 ν = ( b a) 6 2 (2) It can be shown that the following Beta density function f(x) has exactly the above mean and variance, Γ( α + β ) α 1 f ( x) = ( x a) ( b x) α + β 1 Γ( α) Γ( β )( b a) β 1, a x b, α, β > 0 (3) with α and β chosen such that α µ = a + ( b a), (4) α + β

114 100 CHAPTER 8 2 αβ ( b a) ν = 2 ( α + β ) ( α + β + 1) (5) These formulas for the determination of Beta distributions of lead times are also used in the classical PERT procedure. Many authors have disputed the logic behind these formulas (Lau and Somarajan, 1995). There is considerable confusion in literature on what a and b must correspond to. For the original PERT developers (Malcolm, et al., 1959), a and b correspond to the absolute endpoints T 0 and T 1, respectively. However, the majority of current operations management textbooks state that that a and b must be T s 0.01 and 0.99 fractiles (Lau and Somarajan, 1995). Lau and Somarajan propose an improved procedure, which by preference, however, is based on more than three fractile estimates. For the planning problem described in this thesis, it seems most practical to apply the classical PERT procedure, as defined by the above mentioned formula s. It is, for the performance of the system, of minor importance whether the absolute endpoints or the 0.01 and 0.99 fractiles are used for the estimation of a and b Processing times Similar to the procedure for the lead times, also the processing times are modelled by a Beta-distribution, fitted on the minimum, maximum and most likely value estimation Start times The planned start time depends on both the lead times of the preceding process steps and the number of other ORDERS that compete for the same RESOURCES. When it is, for instance, not yet possible to determine whether an ORDER will be executed in week 12 or week 13, this decision may be postponed to a later planning phase. In this way, maximum flexibility is maintained in the plans. Normally, planning procedures with discrete time buckets are applied for this kind of aggregate resource loading (Figure 8-2a). However, these loading periods are set for mean-time margins and do not discriminate between critical and noncritical ORDERS. In order to be able to handle both critical ORDERS with relatively little slack and e.g. production-to-stock ORDERS with more slack, a different continuous planning strategy is applied (Figure 8-2b). A mean start time is set for the ORDERS and the necessary slack is integrated by assuming a normally distributed deviation over the start time. A small deviation is allowed for critical ORDERS and larger deviations may hold for ORDERS with more slack, depending on the individual characteristics of the ORDERS. Apart from the possibility to distinguish between critical and non-critical ORDERS, also the level of completeness of the technological plans and the level of uncertainty can be taken into account when determining the deviations over the planned start times.

115 AGGREGATE ORDER PLANNING 101 expected load (hours per week) expected load (e.g. hours 8 per day) maximum capacity expected availability 6 expected availability Figure 8-2 a) Capacity planning with b) continuous capacity planning discrete time buckets As argued before, it is not useful to attempt to generate exact (deterministic) values of lead times and start times of the ORDERS. However, for the planning of very critical and therefore high priority ORDERS, the coefficient of variation (CV), i.e. the standard deviation divided by the mean, can be set close to zero. Other ORDERS will be planned around such early fixed ORDERS in a later phase of production planning, e.g. in the scheduling phase. 8.2 The ORDER planning methodology The distributions of the input parameters for aggregate ORDER planning in EtoPlan have been introduced in the previous section. In this section, the way in which the resource loading plans are set up in EtoPlan is discussed (Section 8.2.2), together with the due date determination procedure (Section 8.2.3) and the management of cost aspects that are derived from the resource loading process (Section 8.2.4). First, the generic determination of the ORDER profile that forms the basis for aggregate ORDER planning in EtoPlan is described in Section Generic ORDER profile In traditional planning a rectangle is mostly used to represent an activity in a plan. These activities are represented by exact values for the processing or lead times. A well-known representation of activities in a plan is the Gantt chart which is commonly used in scheduling systems (Figure 8-3). Since EtoPlan uses distributions of lead times, processing Machine 1 Machine 2 Machine 3 time Figure 8-3 Gantt chart

116 102 CHAPTER 8 probability of being in process : / : P L : PFP : P : P = mean processing time : = mean lead time L : PF P mean start time min. lead time max. lead time time Figure 8-4 Generic ORDER profile times and start times, instead of deterministic values, a new representation form for the activities had to be developed. This has resulted in a generic ORDER profile as depicted in figure 8-4. The shape of the probabilistic ORDER profile results from the combination of the Normal-distribution of the start time (S) and the Beta-distributions of both the lead time (L) and the processing time (P). The profile represents the ORDER s probability (percentage) of actually being processed at a moment in time, in Figure 8-4 represented for processing WLPHVHTXDOWRWKHLUPHDQ WKHYDOXH DQGWKHYDOXH 7KHPHDQYDOXHDQGWKH YDULDQFH 2 ) of the processing time are calculated by means of formulas (1) and (2), see Section The area below the ORDER profile should correspond to the processing time, which makes it possible to evaluate the utilisation of the resources required for the execution of the ORDER. Consequently, the profile becomes smaller in vertical direction, proportional to the ratio P/L. For example, consider an exact lead-time value of 10 hours and an exact processingtime value of 2 hours. The height of the ORDER profile can, then, maximally reach the 20% probability of being in process at a moment in time. The function that describes the order profile (O) related to the mean processing time is defined by g Oµ P µ P ( x) = ( F ( x) F ( x)), FE ( x) FS ( x) for all x µ L S E with the distribution function of the end time (E) defined by E = S + Note that: L 0 g µ P µ P O ( x) dx = FE x dx FS x dx µ E µ µ = S = P µ (1 ( )) (1 ( )) ( ) L µ 0 0 L µ + E = µ S µ L µ P, since

117 AGGREGATE ORDER PLANNING 103 The dotted lines in Figure 8-4 are graphical representations of the following functions: g µ P ± 2σ P ( x) = Oµ P ± 2σ P µ L ( F ( x) F S E ( x)) E can be approximated with either a mixture of Erlang distributions or a hyperexponential distribution, dependent on the value of the coefficient of variation C E, see Van Houtum and Zijm (Houtum, 1991). Let the mean and variance of the distributions be represented by and 2. C E is defined by 2 C E = σ + σ 2 2 S L 2 ( µ S + µ L ) If C E 2 WKHGLVWULEXWLRQIXQFWLRQRIE can be approximated by a mixture of an Erlang-(k-1) and an Erlang-k distribution, with the following density function: f E ( x) = pλ k 1 k 2 x e ( k 2)! λx k 1 k x + (1 p) λ e ( k 1)! λx with k chosen such that 1/k C E 2 k-1), and p and defined by p = [ kc { k(1 + C ) k C } 1/ ]/(1 + C ), E λ = ( k p) / µ. E E The distribution function of E then becomes F ( x) = p(1 E k 2 n= 0 e λx ( λx) n! n E ) + (1 p)(1 E k 1 n= 0 e λx ( λx) n! If C 2 E WKHQ WKH GLVWULEXWLRQ IXQFWLRQ RI E is approximated by a hyperexponential distribution with the following density function: λ1x λ2x f E ( x) = pλ1 e + (1 p) λ2e with 1, 2 and p defined by λ {2/ µ }{1 + [( C 0.5) /( C 1)] 1/ }, 1 = E E E + λ = ( E λ, 2 4 / µ ) p λ λ µ 1) /( λ ). = 1( 2 E 2 λ1 1 The distribution function of E then becomes λ ( ) (1 1 x λ2 F x = p e ) + (1 p)(1 e x ) E With the distribution function of the end time known, it is now possible to calculate the ORDER-profile function depicted in Figure 8-1. This function includes the stochastic values for the start time, lead time and processing time. In the next section, it is described how the resource loading profiles are calculated on the basis of the ORDER-profile functions. n )

118 104 CHAPTER Resource loading views The resource loading task in EtoPlan is performed as a continual process of ORDER and RESOURCE planning decisions. Typical planning decisions concerning ORDERS are: adding new ORDERS, shifting start times, changing the standard deviation of the start times, outsourcing ORDERS, and updating processing and lead times. The planning of RESOURCES includes e.g. the planning of overtime work, material requirements planning, and the (re)configuration of Applicability Groups (AGs). The various planning decisions are initiated by either external or internal events. Examples of such events are newly entered ORDERS, changes in the process plans, or the occurrence of RESOURCE capacity shortage in a certain time period. Information Management (see Chapter 6) co-ordinates the initiation of the required planning activities. Loading profiles, i.e. the estimated utilisation of a RESOURCE over the time, are drawn up for all main RESOURCES in the company that may potentially become a bottleneck in the production (see Figure 8-2b). In practice, the generation of loading profiles will probably be restricted to the machine tools but they may also include operators (e.g. for assembly operations), special tools, transportation devices, or planners (e.g. process planners). These RESOURCES will probably be part of many AGs from which their loading profiles can be calculated on the basis of the ORDER profiles that belong to these AGs. Referring to Section 7.1, an AG has a one-to-one relation with an ORDER. The amount (or percentage) of the ORDER profile that is accounted for when building the loading profile of a certain RESOURCE primarily depends on the number of other RESOURCES that are part of the AG considered. In the following examples, some possible procedures are elaborated: The AG contains three RESOURCES, but only one RESOURCE is required for executing the ORDER. It is not yet known which one of the three RESOURCES will eventually be selected. If all RESOURCES do have the same probability of being selected, then 33% of the accompanying ORDER profile is assigned to each resource. It is, of course, also possible that one RESOURCE is more likely to be selected. In that case, a higher percentage may be assigned to this RESOURCE and lower percentages to the other two RESOURCES. The ORDER will later be split up into three SUBORDERS. Each RESOURCE is required for executing one of the SUBORDERS. As long as details about the SUBORDERS are not yet known, it is presumed that the execution of all three SUBORDERS will require the same amount (i.e. 33%) of the processing time. Of course, variations are also possible here. A combination of the previous two situations. For calculating the resource loading profiles, the cumulative mean processing time values and the cumulative variances are taken into account. Figure 8-5 illustrates the calculation of a resource loading profile by accumulating the ORDER profiles. Let k denote the index of an ORDER that is allocated to resource R i with probability P k (R i ), defined as explained above. The resource loading profile (R i LP) function related to the mean processing times ( P ) of all ORDERS 1,..,n, possibly in process at some time x, is defined by

119 AGGREGATE ORDER PLANNING 105 g Ri LPµ P ( x) = n k= 1 µ Pk ( Ri ) µ Pk Lk ( F Sk ( x) F Ek ( x)) This function is drawn in Figure 8-5 by a thick line. The thick dotted lines in Figure 8-5 are graphical representations of the following two functions: g n Pk R LP ( x) = g ( ) ± 2 [ ( ) ± 2 R LP x Pk R i µ σ i µ i µ P P P k= 1 Lk σ ( F Sk ( x) F Ek ( x))] Note that the mean values of the processing times of the individual ORDERS often exceed the initially specified most likely values, since typically b-m m-a, indicating that it is more likely that the process is delayed rather than completed early. Hence, some buffer against processing delays (e.g. quality faults) has already been included in the loading profile based on the mean processing time values. When the loading profile of a RESOURCE has been determined, a loading plan of that RESOURCE can be drawn up. A loading plan contains the loading profile, a maximum capacity line and a line that represents the estimated availability (Figure 8-2b). Because the resource loading plans will be drawn up for a much longer time period than a working day (e.g. for a month), the loading profiles must be adapted to the number of hours the resource is available for production during a given time period. For example, the loading profile of a resource will have an average height of 4 hours per day, if a resource has a 50% average chance of production during a whole working day of eight hours. The maximum capacity line indicates the usual number of hours per time period that the RESOURCE is available for production, including the estimated idle times. This may, for instance, be 8 hours a day for machine tools that require operator assistance and 20 hours a day for an FMS. The estimated availability line indicates the number of hours per time period that the RESOURCE 2 expected utilisation : PFP : P : PF P time Figure 8-5 Resource loading profile

120 106 CHAPTER 8 is expected to actually execute ORDERS, also known as the effective capacity. The estimated availability line is derived from the maximum capacity of the RESOURCE, the expected down times due to a temporary lack of ORDERS, and possibly the relations with other RESOURCES. The latter is clarified in the following example. A workstation consists of one CNC lathe and one CNC milling machine tool, both being operated by one person. The two machine tools can work in parallel (maximum of 2 * 8 hours a day). The operator is needed for set-up, transport, cleaning, workpiece changing and control operations. As a result, the machine tools will each work just 6 hours a day on average. However, if one of the machine tools appears to be a bottleneck during a certain time period, the operator will attempt to minimise delays on that machine tool. In this manner, the machine tool will probably be used for 8 hours a day effectively. Therefore, both the maximum capacity per day and the estimated availability to the user is represented in the EtoPlan concept. In this way, the user is informed about both the soft and the hard constraints in the plans. The resource loading plans of the AGS are derived from the loading plans of the main RESOURCES the AG consists of by summing up their loading profiles, maximum capacity lines and estimated availability lines. The same procedure applies when determining the loading plans of the Method Applicability Groups (Section 7.1.1). Examples of resource loading views for RESOURCES, AGs and MAGs, are described in Chapter 9 where the prototype software implementation of EtoPlan is presented Due date determination The delivery date which has to be settled with the customer is determined primarily based on the (rough) routing of the production ORDER through the factory and the existing loading plans. A process planner specifies the routing for producing the product, after which the ORDER profiles for all ORDERS in the ORDER hierarchy are calculated. The ORDER profiles depend on the loading plans of the accompanying AGs and on the ORDER profiles of the preceding and succeeding ORDERS in the production graph. The start time (mean and variance) is mainly derived from the ORDER profile of its predecessor(s) in the production graph, because an ORDER profile indicates the end time of the corresponding preceding ORDER. Because the ORDER profile(s) of the predecessor(s) indicate(s) no exact value of the end time, the start time (mean and variance) is derived from the distribution(s) of the end time of the preceding ORDER(S). Due to the need for uniformity in the determination of the ORDER profiles, the distribution of the end time(s) of the preceding ORDERS is transformed into a Normal distribution of the start time of the current ORDER. In this way, the generic ORDER profile can be drawn up again for the succeeding ORDER. Resuming from Section , the distribution function of the end time of an ORDER is composed from the Normal distribution of the start time and the Beta distribution of the lead time. The mean and the variance of the resulting distribution of the end time are used for determining the mean and variance of the Normal distribution of the start time of the succeeding ORDER. The ORDER profiles of all ORDERS in the ORDER hierarchy of a customer order are determined by the procedure described above. When this process is finished, the due date for the customer order can be determined. This is done by setting a target probability. A target probability of 95% means that the chance of eventually meeting the due date is 95%.

121 AGGREGATE ORDER PLANNING 107 If there are various internal routings within an ORDER structure, it is not always obvious which internal routing will become the actual critical path of the ORDER. A critical path refers to the internal routing in an ORDER structure that determines the overall lead time of the ORDER. Within stochastic PERT-networks more internal routings may, with a specified probability, appear to be critical paths (Wiest, 1977)(Elmaghraby, 1995). Integration of the so-called near critical paths with relatively large standard deviations makes the planning of the start times more complex, but also more reliable. An influential characteristic of ORDER planning in manufacture-to-order environments is the fact that the engineering activities must be planned to precede the actual production activities. Generally, a wall is placed between the preparatory engineering activities and the production activities by setting a milestone, which indicates the deadline for the engineering department. Nowadays, detailed engineering activities are, however, performed more often shortly before the actual production takes place, i.e. during the execution of preceding production steps. Therefore, it is more logical to also plan the engineering activities as auxiliary ORDERS in the ORDER graph. In EtoPlan, the engineering activities are logistically planned in the same way as production activities, viz. by configuring AGs and determining ORDER profiles Cost aspects of production planning The goal of production planning regarding cost aspects is to minimise the variable costs that are a result of production planning decisions. The following variable costs can be recognised: Extra payments for overtime work. The price for subcontracting minus the variable costs for in-house production. Inventory costs due to excessive Work In Process (WIP). Lateness costs, in particular penalty costs. Earliness costs for short time delivery of blank materials. Furthermore, also the variable costs determined by process planning as a result of solving a problem that occurred in production planning must be taken into account. An example is the selection of a more expensive machine tool when a cheaper machine tool appears to be overloaded. A generic concept for cost based decision support in product design has been developed by Liebers at the laboratory of Design, Production and Management of the University of Twente (Liebers, 1998). The concept is currently extended in order to identify the required costing information for cost estimation and decision support in process planning and production planning (Brinke, 1999)(Huttinga, 2000). It is expected that a more sophisticated cost model can be of great use for the order acceptance process. In the EtoPlan concept, the user is advised about the relevant cost aspects with regard to production planning. Chapter 9 describes a cost view in the prototype implementation.

122 108 CHAPTER Allocation of the RESOURCES Like the process concerned with the determination of the production times, also the resource allocation task has to deal with incompleteness of information. It is, for instance, often not known whether a resource is available at the desired time of production, due to e.g. operator unavailability, maintenance ORDERS or rush ORDERS. It is also difficult or sometimes even impossible to judge the capability of the RESOURCES for processing the production ORDER, when the technological requirements (e.g. tolerances) are only roughly known. Therefore, we apply the approach of least commitment for building Applicability Groups in the following manner. The Method Applicability Groups (MAG, see Section ), are generally used for initially configuring AGs for highly aggregate ORDERS of which no more than rough capability descriptions are available yet. When more information becomes available and the highly aggregate ORDERS are split up into more detailed ORDERS, the AGs will contain less RESOURCES, until, at the end, only the RESOURCES that are actually used for production (e.g. machine tool, operator, and cutting tool) are part of the AG. Although an AG will normally decrease its number of RESOURCES when more information becomes available, there are two reasons why an AG can also be extended with new RESOURCES. First, because only those RESOURCES that are relevant for the aggregate planning decision concerning a certain ORDER are part of the accompanying AG (see Section 7.1), the group of RESOURCES (AG) will increase when the ORDER is eventually planned in more detail. For example, a cutting tool will probably be of no relevance when it is even not yet known which machine tool will perform a certain ORDER. However, after the RESOURCE allocation process, one may probably wish to include the timely availability of a cutting tool. Then, the cutting tools that can be used for executing the production ORDER will become part of the accompanying AG. A second reason for extending an AG is the possibility of bad former process assessment. For example, when - e.g. after more detailed process planning - a certain tolerance value appears to require a decrease in the number of set-ups, this may require another type of machine tool. 8.4 Navigating through the planning views In accordance with the principles of Information Management as described in Chapter 6, the information is pulled by the manufacturing planning experts only when it is needed for making a given planning decision. Navigation through the planning views on the basis of occurring events is, therefore, an important aspect of the EtoPlan concept. These events may be related to external influences, like e.g. the entry of a new customer order, an offer request, an alteration of the delivery date, or other contract changes. Most events will, however, relate to internal causes. In most cases, they arise as a result of previous planning decisions. An example is the initiation of new AGs. However, events may also arise due to the occurrence of problems, like capacity deficiencies, late ORDERS, or rejections due to quality problems. In order to initiate automatic problem solvers in the case of trivial events, or to provide the users with the right information for manual decision making, the development of a

123 AGGREGATE ORDER PLANNING 109 navigation tool for the EtoPlan concept is required. Such a navigation tool is currently being developed at the Laboratory of Design, Production, and Management of the University of Twente. It seems worthwhile to combine the EtoPlan framework with the use of agents and a multi-agent system. Such an extension of the EtoPlan concept may add some important functionality. First, if agents are coupled to the AG or Resource Controllers, some basic rules can be applied for automatically taking trivial planning decisions and for determining which planning views will be presented to the user and in what sequence. Second, the interaction between the AGs via the RESOURCES can be modelled by using agents due to their inherent features of co-operation and autonomous decision making ORDER types The ORDER planning decisions highly depend on the characteristics of ORDERS concerned. For instance, the solution method, for dealing with an ORDER that most probably will exceed its due date, will depend on answers to questions like: Does the due date violation also imply a violation of an already accepted delivery date? Can we postpone the delivery dates and to what costs? Can we increase the priority of the ORDER so as to decrease the lead time? Are there possibilities to shorten the processing times or to reduce the number of process steps by adjusting the process plans? Can we decrease the lead times by batch splitting and parallel processing? A number of solutions are potentially available. Which one to choose depends on the ORDERS concerned. The seven dimensions of the ORDER classification scheme, presented in Section , serve as a template to define the objectives and the constraints of the ORDERS. E.g. an ORDER with relatively little slack on the critical path should be treated as a rush ORDER. Or, a given ORDER that has a good chance of successive profitable repeat ORDERS is of relatively high strategic importance. In future research we will study how to quantify the order dimensions in order to use these in the aggregate planning functions. 8.5 Conclusions In this chapter, an ORDER planning method has been presented to concurrently plan the technological and logistic aspects of the ORDERS on an aggregate level. Because of the lack of detailed planning data due to the abstraction planning approach and the shop floor randomness, all planning variables are modelled stochastically. For every ORDER, process planning defines stochastic processing times and a rough routing. The start times and lead times are determined in the resource-loading process, again defined as stochastic variables. These variables are together represented by means of a generic ORDER profile that is used to determine the in-time capacity requirements of the ORDERS. Incorporating the uncertainty into the aggregate plans enhances the reliability and, thereby, the feasibility of the production plans. It supports the integration of the company management function order

124 110 CHAPTER 8 acceptance and the higher level planning tasks in both process planning (i.e. macro process planning) and production planning (i.e. resource loading).

125 111 Chapter 9 A Prototype Software Implementation A prototype of a decision support system for integrated order planning (EtoPlan) has been developed. Its purpose is to analyse and test the practical applicability of the system architecture (described in Chapter 7) and the accompanying concept for aggregate order planning (Chapter 8). The implementation covers the planning on the higher aggregation levels where both macro process planning and resource loading take place. Figure 9-1 displays the position of the EtoPlan system related to some important manufacturing planning and control functions. It supports the decision-making tasks of both resource loading and macro process planning in order to achieve concurrency in the execution of the planning tasks. The aim is not to completely automate the planning tasks in the EtoPlan implementation. Because of that, various types of planning strategies (both technological and logistic), as well as a situation of frequently changing strategies can be supported by the EtoPlan decision support system. The system provides the user with information about resource availability estimates with respect to individual resources and the (Method) Applicability Groups (see Section 7.1). It also reports about the resource loading problems that are expected in the near future and it provides insight in the consequences of the current resource loading decisions for the additional costs. Subsequently, a human planner can, possibly with the help of other planning tools, take the decisions. In particular, it supports the planner in determining start times, processing times, lead times, Applicability Groups, due dates, suborders, costs, and overtime work. As indicated in Figure 9-1, the EtoPlan system does not cover the lower levels of manufacturing planning and control, i.e. shop floor control and micro process planning & control. The EtoPlan system has to be interfaced with a shop floor control system. The different time horizons allow for a strict division between the two systems. All suborders order acceptance EtoPlan DSS macro process planning micro process planning & control resource loading Shop Floor Control Figure 9-1 The position of the EtoPlan prototype implementation

126 112 CHAPTER 9 that are to be executed, e.g. in the next two weeks, are dealt with by the SFC system. The EtoPlan system continuously reports the new suborders that reach this time line to the SFC system. Conversely, the SFC system reports about the progress in executing the plans. The EtoPlan system has been developed using Microsoft Visual C++ Developer Studio 6.0 (Microsoft, 1998b). The dynamic system structure, build up in an object-oriented way, allows for the continual creation, deletion and change of ORDERS and Applicability Groups (i.e. groups of resources allocated to an ORDER). The ORDER and Applicability Group (AG) objects are characterised by a relatively short lifetime compared to the individual resources and Method Applicability Groups (MAGs). A number of views on ORDERS, AGs and RESOURCES are created in order to supply the user with helpful information in a user-friendly manner. The views can be classified in order planning, resource loading, and cost views. The order planning views, described in Section , relate to the order hierarchy (routing) and the planning of the ORDERS in time. Section 9.2 describes the implementation of the views that provide insight in the consequences of capacity constraints on planning the ORDERS. The cost aspects that are relevant for taking planning decisions when alternatives are available are dealt with by the cost views described in Section 9.3. Before the different views are described, the information structures of the main object classes Resource and Order are specified in Section Information Structures In the implementation of EtoPlan, all the information about the manufacturing environment is encapsulated in the objects representing the ORDERS and RESOURCES. These objects are instantiations of the main object classes Order and Resource. In this section, the information structures of these main object classes are described. The object classes are based on the generic information structures, which have been presented in Chapter 6. A generic graph structure implemented using the CObList Class of the Visual C++ MFC development environment (Microsoft, 1998b) is used for facilitating the integration with an Information Management software implementation which serves as the kernel of an extensive software integration project that is currently under development. The project aims at integrating all the manufacturing planning tasks, from the first customer contact until final delivery, including design engineering and cost estimation. Progress is also made to establish an internet-based front office for concurrent manufacturing planning and communication with clients and suppliers, see e.g. (Popma, 1999).

127 A PROTOTYPE SOFTWARE IMPLEMENTATION 113 Figure 9-2 Order List Dialog Graphical User Interface (GUI) The order Information The ORDER objects are of primary significance for establishing the planning views in EtoPlan. Client orders enter and leave the company, and therefore the EtoPlan system. They are subsequently split into suborders structured in an order hierarchy. Every time a new ORDER - which can be any type of order (see Section ) - enters or is initiated in the system, the information about the order is filled up by way of the OrderListDialog Graphical User Interface (GUI), shown in Figure 9-2. The order data is grouped in five categories: Order identification; specifying a unique OrderID, an order description, the type of order (e.g. client order, engineering order, production job, etc.), and the accompanying internal or external client. Time related planning data; all the data resulting from the in-time planning of the order. This data is categorised in constraints (minimum start time and maximum end time) and intentions (stochastic start time, processing time and lead time parameters). Product related data; specifying the ProductID(s), the type of product(s), the product description, and the number of products the ORDER applies for. Resource related data; the capacity requirements of the ORDER by way of a pointer to an AG in the Resource Information Structure. The data related to the applicable resources is not included in the OrderListDialog GUI because some other GUI s are implemented for the management of this information. The views on the resource information, including the information structure of an AG, are further described in Section The structure of the order; specifying the order hierarchy and precedence relations. Complex order networks are not yet possible in the prototype implementation, due to the limited development time. This will, however, not be a difficult extension of the system. At present, an infinite number of hierarchical levels can be specified together with the precedence relations between orders at the same level. The indicator percentage planned in the OrderListDialog GUI provides the user with information about the part (percentage) of the mean processing time defined for a parent

128 114 CHAPTER 9 order that has already been specified in its suborders. For example, the processing time for order 673 has an estimated mean value of 10 hours. Furthermore, two suborders of order 673 are already specified in more detail with an estimated mean value for their processing times of, say, 2 and 5 hours respectively. Then, the percentage planned of order 673 will be 70%, namely ((2+5)/10)*100%. This can, for instance, imply that a third suborder has still to be specified with a mean processing time value of approximately 3 hours. Because the processing time is estimated, it may also happen that no further suborder will be initiated, which implies that the mean value of 10 hours can be changed into 7. The same procedure is followed for the variance values in the estimation of the processing time Order planning views Some additional views on the structure and the data of the ORDERS are provided by the system for facilitating the order planning. The Order Information View Microsoft Project98 (Microsoft, 1998a) has been used for the representation of the data of the ORDERS belonging to (a part of) the order hierarchy. Figure 9-3 shows such an order hierarchy consisting of a main ORDER and three SUBORDERS. The interface between EtoPlan and MS-Project98 has been established by means of OLE Automation linking. The following information about the ORDERS is represented in the order information view: Planning constraints The feasible planning period bounded by the minimum start time and the (internal) due dates (numbered 1 in figure). Precedence relations (arrows). Planning data An estimation of the mean processing time (2). An estimation of the mean lead time (3). The mean planned start time (4). The time period of which it is known that with a 95% chance the complete ORDER can be executed within this time period (5). This bar is derived from the order profile, which is derived from the Beta-distributions of the lead and processing times and the Normal-distribution of the planned start time (see Section 8.2.1) Figure 9-3 An example of a view on the order hierarchy

129 A PROTOTYPE SOFTWARE IMPLEMENTATION Resource information The implemented RESOURCE objects are limited to three types: the machine tools in the factory, the Applicability Groups and the Method Applicability Groups. The operator availability is implicitly handled via an estimation of the availability of the machine tools (see Section 8.2.2). Furthermore, subcontractors to which specific orders can be allocated are defined as an alternative for including them in the capacity plans of the factory. The other resources required for manufacturing - like tools, fixtures, materials, or planning systems (e.g. CAPP software) are not taken into consideration in the EtoPlan prototype implementation. An extension of the system to include the manufacturing planners (e.g. engineers) and their planning systems is required when the EtoPlan concept is further implemented for commercial usage. This, however, does not fundamentally change the structure of the system due to the generic Resource object class definition (see next section) The information structure of the Resource object In this section, the Resource object that represents the physical resources in the factory is outlined. As mentioned above, only the machine tools are considered in the prototype implementation. The Resource object itself is, however, not designed for representing machine tools only. The main differences between handling the Resource objects and the Order objects in the system refer to the life times of the objects. Whereas the order mix continually changes, the Resource objects that directly relate to the physical RESOURCES - can be considered as relatively static objects that only change due to strategic and tactical decisions. Nevertheless, notice that the grouping of the RESOURCES into AGs - but not the Resource objects themselves - is roughly changing as frequently as the Order objects, due to the direct relations that exist between the ORDERS and the AGs. The ResourceListDialog GUI is shown in Figure 9-4. The maximum capacity (hrs/day) and the estimated availability (hrs/day) of the physical resources can be entered or altered in the ResourceListDialog window. These parameters are important for setting up the resource loading views described in Section 9.2.

130 116 CHAPTER 9 Figure 9-4 Resource List Dialog GUI Information is provided about the AGs and the logical resources the resource is part of. A logical resource refers to a static grouping of resources, like a production unit. Logical resources are often physically grouped together on the shop floor. This is however not a necessity. A group of transport resources can, for instance, also be modelled as a logical resource. Logical resources are particularly useful for co-ordinating specific planning tasks, such as the Scheduling Group described in Section 7.4. At present, only an aggregate description of the resource capabilities has been implemented by using Methods (see section for a discussion of the Method domain). A detailed description of resource capabilities (e.g. maximum power) has not (yet) been enclosed in the implementation, because the goal of the prototype implementation was to focus on the logistic aspects of manufacturing planning. The detailed capability description will receive more attention as soon as the actual integration with a process planning system is established The information structure of the Applicability Group object Applicability Group objects establish the relation between the Resource objects and the Order objects. The Order objects in the system are each linked to a uniquely created AG object, which deals with the planning decisions concerning the allocation of Resource objects to the Order object. Figure 9-5 shows the Dialog window for configuring the Applicability Groups. Resources are either appointed applicable or inapplicable for executing a given ORDER. The definition of applicable has been discussed in Section 7.1. In the Dialog window it is also possible to specify (estimate) some parameter values that are necessary for estimating the lead time of the accompanying ORDER. These parameters include the number of process steps (i.e. suborders), the number of different departments

131 A PROTOTYPE SOFTWARE IMPLEMENTATION 117 Figure 9-5 Applicability Group Dialog GUI these process steps will be executed by, and the number of process steps to be subcontracted. Due to the major influence of these parameters on the actual lead times, estimations are to be made in an early phase, i.e. during macro manufacturing planning The information structure of the Method Applicability Group object The Method Applicability Group (MAG) object class contains a list of Resource objects, which are applicable for performing the types of activities that characterise a given Method. As has been described in Section 7.1.1, a Method is a technological abstraction of the resource capabilities required for performing a given type of activity (see also Section 6.2.3). In Figure 9-6, three MAGs have been inserted in the listbox. As can be noticed, it is also possible to define sub-mags. A sub-mag contains a particular selection of the Resource objects that are part of its parent MAG. For example, only the CNC Milling machine tools are appointed applicable for the Milling.CNC MAG object, where the parent MAG object Milling contains all resources that are capable of performing milling operations, thus including the conventional milling machine tools. In the EtoPlan prototype implementation, the MAGs are mainly used for configuring new Applicability Groups. When only aggregate information about the required capabilities is known - which is often the case in the macro process planning phase - the existence of MAG objects in the system facilitates the allocation process of the applicable resources for newly entered Order objects.

132 118 CHAPTER 9 Figure 9-6 Method Applicability Group Dialog GUI 9.2 Resource loading views The order planning views presented in the previous section merely contain information about the individual ORDERS they are created for. The most important objective for aggregate production planning, however, is to decide upon the allocation of multiple ORDERS in the system, given a limited capacity availability in a limited time period. In this section, the implementation of views for capacity loading and resource requirements are presented. These implemented views are based on the resource loading methodology presented in Section AG availability view As has been described in Section 8.2.2, the capacity plans of the AGS are derived from the capacity plans of the main RESOURCES the AG consists of by summing up their loading profiles, maximum capacity lines and estimated availability lines. Figure 9-7 shows an example of an AG availability view related to a specific order ( ). An AG with two milling machine tools (MT002 and MT007) has been created for this order. The maximum capacity of this group of resources is 15 hours/day. The estimated availability (13 hrs) represents the maximum capacity minus an estimation of the idle times of the resources due to the unavailability of other resources required for executing the ORDERS. The order profile drawn on the x-axis, represents the estimated capacity requirement for executing order The order profile is derived from the generic order profile, described in Section

133 A PROTOTYPE SOFTWARE IMPLEMENTATION 119 Figure 9-7 AG Availability View In this implementation, only the loading profiles based on the mean processing times are drawn. Instead of representing the variance in the processing times, as was mentioned in Section , we chose to indicate the consequences of significantly higher lead times due to relatively high Work In Process (WIP) levels. The thick line in Figure 9-7 represents the estimated loading profile of the group of resources (AG) that corresponds to order This line is determined by putting together the order profiles of all the orders that need one or more resources (which hence are part of their AGs) that belong to the AG considered - in this case the AG of order This procedure has been described in detail in Section The area under the line between two dates (area A) corresponds to the estimation of the required total processing time at the resources that belong to the AG during the specified time period. The thin line shows the loading profile if the estimations of the maximum lead times are considered instead of the Beta-distribution of the lead time. The latter will probably be relevant during the time periods with a high Work in Process (WIP). A peak in the thick line indicates that there is a high probability that a capacity problem will occur at that time. This will go hand in hand with a high WIP. As a result, the lead time of most suborders involved will increase. As expected, the thin line (maximum lead times) shows a smoothing effect compared to the thick line (lead time distribution). If peaks in the thick line occur and no sufficient corrective actions can be taken, it is likely that the thin line represents the actual situation that will arise. The list boxes, input buttons and data on the right hand side of the Graphical User Interface, as shown in Figure 9-7, are either used for informing the user or for the alteration of some ORDER parameters belonging to the Applicability Group considered. The control boxes indicated with a number in the figure are explained: 1) A combo box with the suborders of the ORDER considered that are already initiated. 2) All ORDERS in the system. The selected ORDER ( ) is the ORDER that is currently being planned. 3) A combo box with all ORDERS that influence the loading profiles. These are all ORDERS which share one or more of the RESOURCES of the AG and which are to be executed, in

134 120 CHAPTER 9 this example, between June 8 and July 5. If the user indicates a smaller time period (e.g. period A), then the orders involved in this time period are shown in this combo box. This can be useful if the planner wants to be informed about which orders are critical, and, therefore, cause peaks in the loading profiles. 4) Used for start time/date determination. 5) Used for the determination of the standard deviation over the start time. 6) Setting the range of the time period that is shown on the x-axis. 7) Some information about the ORDER for supporting the planner in making the planning decisions. 8) The percentage of the processing time of the ORDER which has been planned already in more detail through its suborders. 9) The RESOURCES contained in the Applicability Group that corresponds to the ORDER considered Resource availability view In order to avoid a critical overload on the individual resources, views on the workload expected in the near future are provided by the system. Figure 9-8 shows an example of such a Resource availability view. The view looks somewhat the same as the AG availability view, except that no orders can be planned in the Resource availability view. Some other planning decisions can be taken on the basis of the information provided in this view, like: Extending capacity by planning work overtime or hiring additional operators. Shifting operators between machine tools. Subcontracting some critical orders. In the list box at the right hand side of the screen, the orders involved in the loading profiles are presented. Again, it is possible to narrow the visible time period so as to detect the critical orders, which cause the peaks in the loading profiles. Figure 9-8 Resource Availability View

135 A PROTOTYPE SOFTWARE IMPLEMENTATION MAG availability view Most macro process planning decisions mainly concern the rough planning of the methods (e.g. CNC milling) for producing the ordered PRODUCTS. Generally, most attention is paid to the technological requirements and constraints, less attention is paid to the capacity requirements and no attention to the availability aspects. Because of this one-sided view, many process-planning decisions are not as cost effective as is often assumed. Too many replanning activities due to the occurrence of capacity problems are the result of this approach. Method Applicability Group (MAG) availability views are created in order to inform the macro process planners about estimations of availability of the method-related RESOURCE groups (see Section 7.1.1). An example of a MAG availability view is shown in Figure 9-9. As can be seen, the availability views of the MAGs look almost the same as the availability views of the RESOURCES, presented in the previous section. With the use of MAG availability views, it can, for instance, be decided early to subcontract particular suborders of the product manufacturing process. Also, the people responsible for order acceptance are informed about the availability of the required production methods for the time period requested. The MAG availability views are particularly useful in a manufacture-to-order environment, because of a lack of detailed information about the required routing through the factory. 9.3 Cost views The cost aspects that are considered in the EtoPlan order planning method have been described in Section In the prototype implementation, the variable costs that result from critical situations in resource loading are presented to the user in various ways to support decision making (Huttinga, 2000). Not only the mean values of the costs, but also the distributions are presented in graphics. Together with the graphics of actual cost estimations, also the estimations of extra costs that will be required if alternative solutions for solving capacity problems are chosen are presented. In this way, the user is advised about the best choice from a cost point of view. Figure 9-9 MAG Availability View

136 122 CHAPTER 9 Figure 9-10 Cost Diagnostics View Figure 9-10 shows the Cost Diagnostics GUI which represents an estimation of the additional costs that result from the planning of a single ORDER. The grey bars show the amount of costs arising from overtime work and the thin black bars indicate the lateness costs. Note that the cost values are derived from the capacity plans which are represented in a stochastic way. Also the cost views, therefore, present stochastic values of costs. 9.4 Navigating through the system A tool for navigating through the decision-making procedures is not yet available. At the time this thesis was written, all navigation functions are to be performed by the user, without any intelligent support by the system. When, for instance, a peak in the AG availability views occurs, the solution approach is to be determined completely by the user. In order to integrate intelligent navigation support in the system, additional research on developing a navigation tool is currently performed. This tool will support the user in achieving the planning goals mainly by showing the required route for decision making. The system must point out and represent a particular view on the information structure of the Resource-, Method- and/or Order objects, whenever it is useful or required for making the desired planning decisions. Keeping the example of a peak occurrence in the AG availability view in mind, such a tool can, for instance, prioritise the ORDERS so as to replan those orders that are primarily responsible for this peak in demand. The priority of the orders with respect to problem solving will be based on the classification of the orders discussed in Section In this way, the user is supported in making the decisions regarding resource loading. For the purpose of solving an arising problem, not only logistic alterations, but also process planning decisions can be initiated.