Can Advanced Analytics Improve Manufacturing Quality? Erica Pettigrew BA Practice Director (513) 662-6888 Ext. 210 Erica.Pettigrew@vertexcs.com Jeffrey Anderson Sr. Solution Strategist (513) 662-6888 Ext. 260 Jeff.Anderson@vertexcs.com April 13, 2016 1
OVERVIEW 15 Years experience with Fortune 100 CPG/FMCG Specialize in Business-centric, Enterpriselevel solutions, leveraging a variety of tools and technologies Focus on Partnership, Value, and Excellence, resulting in long-standing relationships with our Customers OUR KEY SERVICE AREAS & SOLUTIONS Business Analytics & Advanced Analytics Customer Relationship Management (CRM) Business Process Automation & Reinvention Collaboration and Content Management Systems (CMS) Custom Development Cloud & Mobile Computing Managed Services 2
Today s Agenda Complex Event Analytics Background: Complex Event Processing (CEP) and Complex Event Analytics (CEA) Project Management Approach for Discovery Projects Vertex Case Study: Steel Manufacturing One-Stop Data Shop Advanced Analytics Topics Quality Indicators: Signal vs. Noise Research Coil Quality Score Key Points & Takeaways 3
Background: Complex Event Processing (CEP) What is it & What can it do for me? 4
COMPLEX EVENT PROCESSING (CEP) The Concept Evaluation of a series of sensor, event, or state information to determine a relationship between events and provide situationally relevant assessments and action recommendations. The Technology Software & systems solutions which systematically relate and associate multiple (potentially disparate) events and data elements, deriving and applying context for business evaluation and action. Key Goals & Objectives: CONTEXT AWARENESS Enable Right-Time, Contextually Aware data understanding across multiple streams of information and inputs. CONNECTEDNESS Connect Systems and Audiences On a Single View of the Data & Understanding of Impacts. IMPROVED DECISION MAKING Discover, Identify, and Target challenges and opportunities, to support better business and operations decisions, moving toward data-driven frameworks to improve speed, reduce costs, manage risks, and more. OPTIMIZE IMPACT Create opportunities for impactful interactions with operational systems for maximum operational effectiveness. People are natural CEP processors. We constantly associate, derive, and apply context to the activities and occurrences around us. CEP technology works to automate this activity in a business context and make it actionable. 5
Traditional Complex Event Processing Event Flow and Complex Joins The traditional flow for Complex Event Processing is Event Capture, a set of Operations, and a resulting Action. Commercial off-the-shelf CEP products focus on Event Capture, Adapters, and providing interfaces to manually write and maintain business rules (models) for business processes. 6
Introduction to Complex Event Analytics (CEA) A Natural Evolution Complex Event Processing Implementing existing business rules with automation Transactional and operational reporting Real-time stream processing Context-awareness Business Intelligence (Traditional) Focus on reporting, information gathering, consolidation, & summarization Data warehouse copy data from transaction, legacy, and other systems onto single platform for reporting and analytics OLAP reporting slice & dice, drilldown to details Complex Event Analytics Discovery and refinement of rules using Data Science before, during, and after implementation Transformative approach with machine learning possibilities Real-time stream processing with threading and pre and post event analysis Business Analytics Range from single-point solutions to solve business challenges and pain points, to New holistic, business driven insights that address transformative challenges Solutions drive value by simplifying the messaging, reducing overhead and latency, building intelligence directly into solutions, and providing opportunity for automated and/or rapid action 7
Complex Event Analytics A Stronger Business Approach! An analytics-driven approach to Complex Event processing provides a shift to Datadriven solutions, by incorporating Data Science for Selective and Iterative Event Processing Operations and subsequent Analytic Opportunities. 8
BUSSINESS TRANSFORMATION Establishing Complex Event Analytics Moving from Islands of Data to an Intelligent Solution * Discovery Applications each take a lifecycle of their own and continue to bring unique benefits as the applications are applied, science is refined, and findings are iteratively embedded within the processes. Evaluate, Automate, Machine Learning Embed Data Science Thread the Data Discovery Applications * Exploratory Applications, Actions INCREASING ANALYTICS CAPABILITY Thread the Data Connections Connect and thread data; consider establishing contextual data lake for initial and ongoing analysis. Discovery Applications Indicators Discovery apps created / evaluate models offline, post event, and refine (where to look.) Discover initial root cause relationships. Exploratory Applications Actions Detailed root cause analysis, deploy operational system updates, increase velocity and variety of data included in models. Embed Data Science Deploy Deploy intelligent aggregation and threading into CEP system and enable business teams with visualization and predictive analytics tools. Evaluate, Automate, Machine Learning Continuous Improvement Measure models impact against KPIs for business processes. Refine models and consider Machine Learning 9
Project Management Approaches Difference Between Discovery and Systems Delivery Why the difference? Initiatives designed to extract information from existing systems or new sources of data must acknowledge how messy and complex that process is. https://hbr.org/2013/01/why-it-fumbles-analytics Embed Data Science Evaluate, Automate, Machine Learning Thread the Data Discovery Applications * Exploratory Application, Actions System Driven Project Management: Initiating Planning Executing Monitoring & Controlling Closing Discovery Driven Project Stewardship: Develop theories Build hypotheses Identify relevant data Conduct experiments Refine hypotheses based on findings Repeat 10
VERTEX CASE STUDY Complex Event Analytics System Major U.S. Steel Manufacturer (& Recycler) 11
Case Study Summary Customer Highlights Highlights Profile Industry leading U.S. Steel Manufacturer Business Challenge Higher than acceptable unidentified late stage quality failures Solution Data Science-driven root cause analysis Threaded, accumulating CEA system Results Unified data view across mills (10+ TB) Advanced Analytic Insights and Applications Overview Identify events and factors affecting quality in the steel roll milling process, with a focus on Root Cause Analysis and Data-Driven Associations. Iteratively extend from current offline, manually intensive research and corrections, to data-driven statistical analysis, followed by online (real-time) monitoring of quality-affecting events that prevent defect propagation and increase quality, predictability, and product delivery confidence. > Establish unified, threaded, plant-wide product-level view of manufactured materials (start to finished good) for ad hoc, application, and advanced analytic consumption > > Explore new analytical methods, tools, and techniques to optimize quality improvements and target corrective actions Provide analytic applications, statistical methods, and visualizations for intuitive, effective plant operations 12
The Starting Focus: Two Primary Use Cases Building the Baseline and Eye on the Prize This project initiated with two primary focus points: Ability to view a single item across all mills, evaluate any point in the process (or across a series of processes), for any / all ad hoc uses. Incorporate current and emerging Statistical Methods on contextual information to support a range of use cases, from simplified Root Cause analysis to Predictive alerts. Use Case 1 1-Stop Data Shop Use Case 2 Advanced Analytics Merge and Connect Timestamp-based Item-level data from all mills for analytics consumption. Incorporate Statistical Methods with contextual Start-to-Finish dataset Provide Root Cause & Associative Analysis, Ability to Replay Across Mills Analytics to enable contextual, event-driven discovery, identification, targeting, and optimization. Supports improved decision making and creates opportunities for operational excellence. 13
Our Story: Complex Event Analytics for Mfg. The Journey to One-Stop-Data-Shop and Insights Through Advanced Analytics THE FUTURE Extend Analytics to support additional high-value use cases and audiences, ultimately incorporating Real Time Predictive Analytics to further optimize QA processes and improve yield while reducing costs and manufacturing delays. THE ANALYTICS Visualizing data, enabling end users to access the CEA data lake, adding and deriving context, tracing root causes, identifying patterns & outliers. Plant & Mill Operators, Engineers, Metallurgists. PROVING IT WORKS Beginning with two highly complex mills, and progressing to a fully loaded, performant, responsive, effective CEA system linking data from more than a dozen mills and manufacturing processes. DOWN WITH DATA & DESIGN From design to build, a custom solution created to house all key prime datasets in threaded fashion across mills, providing a CEA data lake for analytics. 14
The Data Layer: Unifying Entities Stream Processing and Batch Joins A Unified Data Warehouse was established to reconcile the various coil identification schemes and present the entirety of the coil s production history as a single joined view of the entities. A Unified Coil View was created using Accumulating Snapshot technique, incrementally expanding as the data from Melt, Slab, Hot Mill, and Cold Mill is incorporated. This traceable and threaded view, in addition to the atomic data, allows several Advanced Analytics techniques to truly transform the process and obtain valuable insights. 15
Solution Metrics Noteworthy Highlights 10+ TB TOTAL SIZE (Data Lake) Summarized sensor readings, organized, threaded and connected by batch and product ID across all mills, from scrap metal start to finished steel coil product. Thousands of attributes HIGH SPEED DATA PERFORMANCE Fact tables between 5-8 Billion Records with over a thousand columns per record. Stored procedures provide complex filtering using any number of combinations. 10+TB 2TB/hr SCALABILITY AND FLEXIBILITY Supports both columnar (relational) view of data, as well as Key Value Pairs optimizing the data layer for specific analytic use cases. VERTEX SOLUTION & VALUE: Flexible, Scalable, Efficient HIGH VALUE ANALYTICS From 2 months research for 2 persons (pre-vertex Complex Event Analytics solution) to 2 hours with a purpose-built application where speed, accuracy, value (and potential) is evident. TECHNOLOGY STACK: Microsoft Centric A scalable big-data responsiveness solution and capability, built within a standard stack ecosystem. 16
Quality Indicators: Signal vs. Noise Research OVERVIEW This application provides an Advanced Analytics example for systematically identifying differentiating factors in expansive datasets. Advanced Analytics Showcase analysis and comparison of coil attributes using proven statistical methods. Perform data-driven noise filtering and classification / ranking of attributes. Leverage Unified Coil View for threaded dataset across mills. Conduct Root Cause Analysis. Application Overview Data Selection Pre-filtered/selected data set with Coil ID s and attributes will provide input for Statistical Model. Dimension Reduction Statistical Model relates and evaluates available attributes, comparing Good vs. Bad, and reducing dataset to Top Informative Variables. Charts & Visualizations Evaluate model results with Scatterplots and Spotfire other charts in story-board Root Cause analysis fashion. 17
Results: Signal vs. Noise Research Application Separating Signal from Noise (Finding the Most Informative Variables in Large Datasets) BEFORE: Diagram shows k-means clustering using ALL variables Total of 8 clusters Blue marks are Good Coils Black marks are Bad Coils No evident patterns no obvious clustering according to good/bad status IMPACT 2 months 2 people AFTER: Ranked list of 20 most informative variables Variables known to be irrelevant can simply be removed, and analysis rerun. The emergent pattern: At least one group contains bad coils exclusively! 1 Application 2 Hours Repeatable 18
Coil Quality Score OVERVIEW Transform feature values into normalized functions of variance. Then combine the transformed values into a single omnibus measure of product quality. Ranking & filtering coils on the basis of this score is the first step to identifying and correcting quality problems related to process variance. Advanced Analytics Data table containing transformed attribute values + combined quality score Interactive table allows user to sort, filter & select coils based on score. Drill-down histogram & correlation plot based on user select/filter Linear regression plots for diagnostics; believability Application Overview Objective Attribute Selection Data Transform Charts & Visualizations Reduce the entire coil profile down to a single number, normalized to lie between zero and 1. Attributes are identified and selected as the basis for computation of the CQS. Many transformations are possible. This instance uses a standard normal (Gaussian) conversion to z-score probabilities. Addition of simple histogram and scatterplot can support drill-down view of CQS predictors. 19
Results: Coil Quality Score Summarizes Events Activity Across the Entire Line Then deep dive into questions such as What proportion of today s items were above the 96% mark? What is this month s trend on the total quality index? What effect did last week s installation of a have on the index? What effect did the most recent recipe change have on total quality score? How well does the index predict results from Test Lab and vice versa? What were common profile characteristics among the lowest percentile quality items (last week, last month, etc.)? 20
Complex Event Analytics Key Points Highlights & Takeaways Creating a unified view of the data allows for new insights across the entire process though Maintain the atomic data for analysis within the processes itself (and allow for replay and modeling) Embedding Data Science in the process to constantly refine the event processing rules (transformation, not just automation!) Root Cause Analysis Problem Identification Compare & Contrast Replay Trace Across Mills Evaluate Associations Project Management Takeaways Allow time for building domain expertise Be aware of the iterative nature of analytic modeling Account for times to fix data quality issues and other data related issues Worry more about improving business understanding than about deploying technology 21
Closing Thought - Don t fumble it! People don t think in a vacuum; they make sense of situations on the basis of their own knowledge, mental models, and experiences. They also use information in different ways, depending on the context. https://hbr.org/2013/01/why-it-fumbles-analytics THANK YOU TIME FOR DISCUSSION & QUESTIONS 22