Modeling of LDO-fired Rotary Furnace Parameters using Adaptive Network-based Fuzzy Inference System

Transcription

1 The Internatonal Journal Of Engneerng And Scence (IJES) Volume 4 Issue 8 Pages PP ISSN (e): ISSN (p): Modelng of LDO-fred Rotary Furnace Parameters usng Adaptve Network-based Fuzzy Inference System 1 Gurumukh Das, 2 Ranjt Sngh 1 Assstant Professor, Department of Mechancal Engneerng, Faculty of Engneerng, DEI, Agra, UP, INDIA. 2 Professor Emertus, Department of Mechancal Engneerng, Faculty of Engneerng, DEI, Agra, UP, INDIA ABSTRACT In ths paper a novel approach.e. neuro-fuzzy technque s used for the frst tme n modelng rotary furnace parameters to predct the meltng rate of the molten metal requred to produce homogenous castngs. The relatonshp between the process varables (nput) vz. flame temperature, preheat ar temperature, rotatonal speed of the furnace, excess ar, meltng tme, and fuel consumpton and meltng rate (output) s very complex and s agreeable to neuro-fuzzy approach. The neuro-fuzzy model has been created out of tranng data obtaned from the seres of expermentaton carred out on rotary furnace. The results provded by neuro-fuzzy model compares well wth the expermental data. Ths work has consderable mplcatons n selecton and control of process varables n real tme and ablty to acheve energy and materal savngs, qualty mprovement and development of homogeneous propertes throughout the castng and s a step towards agle manufacturng. Keywords: Adaptve Network - based Fuzzy Inference System (ANFIS), Lght Desel Ol (LDO), Rotary Furnace Date of Submsson: 13 June 2015 Date of Accepted: 20 August I. INTRODUCTION To respond to changng demand scenaros, the system must be equpped wth a comprehensve manufacturng plannng and control system that ncorporates vast amounts of manufacturng knowledge n a form that s accessble rapdly. The desgn and mplementaton of these systems s one of the major challenges faced by today s manufacturng engneers [1-2]. The basc dea of Rotary furnace technque s of usng a dome rotatng contnuously to create homogenety n the castng. The rotary furnace conssts of a cylndrcal structure, whch rotates contnuously about ts axs. The furnace can be run by a varety of fuels but at present we are consderng a Lght Desel Ol (LDO) fred furnace. Ths technque suts the condtons and requrements of the local foundres n terms of the cost of castngs produced as well as ther qualty. Moreover the pollutants emtted by the furnace are well wthn the range specfed by the Central Polluton Control Board (CPCB) of Inda [3]. The Rotary furnace s the most versatle and economcal mode of meltng ron n ferrous foundres. But t s very strange that a very lttle nformaton s avalable n the form of lterature on ths furnace [4]. There are a number of varables controllable to varyng degrees whch affect the qualty and composton of the out-comng molten metal. These varables, such as flame temperature, preheat ar temperature, rotatonal speed, excess ar, meltng tme, fuel consumpton and meltng rate play sgnfcant role n determnng the molten metal s propertes and should be controlled throughout the meltng process. However, even an experenced operator may fnd t dffcult to select the optmum nput parameters whch would yeld deal molten metal and often he may choose them by guessng whch may not be effectve and economcal. In order to meet ths demand, an ANFIS model s developed that correlates well wth the expermental data [5]. II. NEURO-FUZZY SYSTEMS Neuro-fuzzy systems belong to a newly developed class of hybrd ntellgent systems that combne the man features of artfcal neural networks wth those of fuzzy logc, usng heurstc learnng strateges derved from the doman of neural network theory to support the development of a fuzzy system. Modern neuro-fuzzy systems usually are represented as a multlayer feed-forward neural network. In neuro-fuzzy models, connecton weghts, propagaton and actvaton functons dffer from common neural networks [6]. The IJES Page 38

2 Modelng of LDO-fred Rotary Furnace The neuro-fuzzy system s capable of extractng fuzzy knowledge from numercal data and lngustc data nto the system. The goal here s to avod dffcultes encountered n applyng fuzzy logc for systems represented by numercal knowledge (data sets), or n applyng neural networks for systems presented by lngustc nformaton (fuzzy sets). Nether fuzzy reasonng systems nor neural networks are by themselves capable of solvng problems nvolvng at the same tme both lngustc and numercal knowledge. A number of researchers have used the term hybrd systems to depct systems that nvolve n some ways both fuzzy logc and neural network features [7]. Neuro-fuzzy systems overcome the lmtatons of artfcal neural networks (ANN) and fuzzy system. A neuro-fuzzy system s traned by a learnng algorthm derved from neural network theory. The (heurstc) learnng procedure operates on local nformaton, and causes only local modfcatons n the underlyng fuzzy system. The learnng process s not knowledge-based, but data-drven [8]. A neuro-fuzzy system can be vewed as a specal mult-layer, feed-forward neural network. The frst layer represents nput varables, the mddle (hdden) layer(s) represent(s) fuzzy rules and the last layer represents output varables. Fuzzy sets are encoded as (fuzzy) connecton weghts. A neuro-fuzzy system can always be nterpreted (.e., before, durng and after learnng) as a system of fuzzy rules. It s possble both to create the system out of tranng data from scratch and to ntalze t by pror knowledge n the form of fuzzy rules [9]. A neuro-fuzzy system approxmates an n-dmensonal (unknown) functon that s gven partally by the tranng data. It s possble to vew a fuzzy system as a specal neural network and to apply a learnng algorthm drectly (hybrd models). Recently, several approaches were suggested for generatng the fuzzy rules from numercal data automatcally. Most notable s Jang s Adaptve Network - based Fuzzy Inference System (ANFIS). ANFIS Developed by Jang, s an extenson of the Takag, Sugeno and Kang (TSK) fuzzy model. ANFIS represents a neural network approach to the desgn of fuzzy nference systems. An ANFIS network makes use of a supervsed learnng algorthm to determne a non-lnear model of the nput-output functon, whch s represented by a tranng set of numercal data. Because, under proper condtons t can be used as a unversal approxmator, an ANFIS network s suted partcularly for solvng functon approxmaton problems n several engneerng felds. The present model allows the fuzzy system to learn the parameters usng hybrd learnng algorthm [10-12]. III. MODELING OF ROTARY FURNACE PARAMETERS In ths secton, the ANFIS modelng of rotary furnace parameters s descrbed. The data s obtaned from the experments conducted on a self-desgned and developed furnace as shown n the Fgure 1, at Foundry Shop, Faculty of Engneerng, DEI, Dayalbagh, Agra, INDIA and s used to tran the ANFIS model. Fg. 1: Self desgned and developed Rotary furnace. The IJES Page 39

3 Modelng of LDO-fred Rotary Furnace In the expermentaton, 200 kg. of the charge s melted n the rotary furnace. A Crcular burner s used for burnng LDO as a fuel. Total 201 numbers of experments were conducted at dfferent percentages of excess ar; varyng from 10% to 50% and varyng n the amount of ar preheat from 200 C to 400 C [13-14]. A sx nput ANFIS archtecture s shown n the fgure 2 wth fve layers. The descrpton of each layer s gven below: Fg. 2: ANFIS Archtecture [15] 1 Layer 1: Every node n ths layer s a square node, wth a node functon, O A ( ), where O s the nput to node, and A s the lngustc label assocated wth ths node functon. In ths archtecture, A s bell shaped wth maxmum equal to 1 and mnmum equal to 0. A ( ) c a 2 b ; where {a, b, c } are the premse parameters. The IJES Page 40

4 Modelng of LDO-fred Rotary Furnace Layer 2: The functon of node n ths layer s to multply the ncomng sgnals and produce the product of all nputs. For nstance, w A ( ) B ( ) C ( ) D ( ) E ( ) F ( ) j k l m n o 1, 2, 3,..., 201 and j, k, l, m, n, o 1, 2, 3, 4, 5, 6. Each node output represents the frng strength of a rule. Layer 3: The nput frng strength s normalzed n ths layer and output s called normalzed frng strengths. w w w 1 w 2... w ; =1, 2, 3,, Layer 4: Every node n ths layer s a parameterzed functon. The node functon s: 4 O w f w p q r s t u ; where =1, 2,, 201 and w s the output of prevous layer, and {p, q, r, s, t, u } s a parameter set. Parameters n ths layer are referred as consequent parameters. Layer 5: The sngle node n ths layer computes the overall output as the summaton of all ncomng 5 sgnals,.e. O overall output w f 1 The system s ntalzed wth a number of membershp functons and a rule base. Learnng conssts of two separate passes. In the forward pass, the consequent parameters are determned by least square method and antecedent parameters are updated by a gradent descent algorthm n the backward pass [16]. From the equatons obtaned n layer 4 and 5 t s observed that gven the values of premse parameters, the overall output s expressed as lnear combnatons of the consequence parameters. The output f can be rewrtten as: f f ( w ) p ( w ) q ( w ) r ( w ) s ( w ) t ( w ) u 1 wf Ths s lnear n the consequent parameters [17]. The forward pass of the learnng algorthm contnues up to nodes at layer 4 and consequent parameters are determned by the method of least squares. In the backward pass, the error sgnal propagates backward to update the premse parameters by gradent descent. The tranng nformaton s as follows: Number of nodes = 193 Number of lnear parameters = 405 Number of non-lnear parameters = 36 Total number of parameters = 441 Number of tranng data pars = 201 Number of checkng data pars = 100 Number of fuzzy rules = 81 Number of epochs = 500. IV. RESULT The predcted values by ANFIS model used n ths work are much closer to the expermental values as can be observed from the results. A partal set of the expermental results and the estmated values reported by ANFIS model are lsted n Table 1. The IJES Page 41

5 Modelng of LDO-fred Rotary Furnace S. No. Table 1: Comparson of Meltng Rate obtaned by Expermentaton on Rotary furnace & by ANFIS Model Excess Flame Rotatonal Meltng Preheat Ar Fuel Expermental Values Ar Temperature Speed Tme Temperature Consumed of Meltng Rate (%) ( C) (RPM) (Mn) ( C) (Lters) (MT/hr.) Estmated Values of Meltng Rate (MT/hr.) The IJES Page 42

6 Modelng of LDO-fred Rotary Furnace V. CONCLUSION The developed neuro-fuzzy model n ths paper can effectvely estmate the meltng rate based on nput process varables vz. flame temperature, preheat ar temperature, rotatonal speed of the furnace, excess ar, meltng tme, and fuel consumpton that correlates well wth the expermental values. Ths technque easly captures the ntrcate relatonshp between varous process parameters and can be easly ntegrated nto exstng manufacturng envronment and also opens new avenues of parameter estmaton, functon approxmaton, optmzaton and onlne control of complex manufacturng systems. ACKNOWLEDGEMENT The authors gratefully acknowledge the nspraton provded by the Most Revered Prof. P.S. Satsang Sahab, Charman, Advsory Commttee on Educaton, Dayalbagh, Agra, INDIA. REFERENCES [1] S.L. Goldman, R.N. Nagel, and K. Press, Agle Compettors and Vrtual Organzatons: Strateges for Enrchng the Customer (New York: Van Nostrand Renhold, 1995) [2] Y.Y. Yusuf, M. Sarhad, and A. Gunasekaran, Agle Manufacturng: The Drvers, Concepts and Attrbutes, Internatonal Journal of Producton Economcs, 62, 1999, [3] R. Sngh, G. Das, and R.K. Jan, Modelng and optmzaton of rotary furnace parameters usng regresson and numercal technques, Proc. 68 th World Foundry Congress & 56 th Indan Foundry Congress, Chenna, Inda, 2008, Paper ID OP 62, [4] R. Sngh, G. Das, and M. Radha Krshna, Economc and mcrostructure analyss of ferrous castngs from coke-less cupola, coke fred cupola and rotary furnace. DEI Journal of Scence and Engneerng Research, 14 (1&2), 2007, [5] R. Sngh, G. Das, and R. Seta, Parametrc modelng of a rotary furnace for agle producton of castngs wth artfcal neural networks, Internatonal Journal of Agle Manufacturng, 10 (2), 2007, [6] H.R. Berenj, and P. Khedkar, Learnng and Tunng Fuzzy Logc Controllers through Renforcements, IEEE Trans. Neural Networks, 3(5), 1992, [7] J.J. Buckley, and Y. Hayash, Fuzzy Neural Networks: A Survey, Fuzzy Sets and Systems, 66, 1994, [8] J.J. Buckley, and Y. Hayash, Neural Networks for Fuzzy Systems, Fuzzy Sets and Systems, 71, 1992, [9] S.K. Halgamuge, and M. Glesner, Neural Networks n Desgnng Fuzzy Systems for Real World Applcatons, Fuzzy Sets and Systems, 65, 1994, [10] D. Nauck, and K. Kruse, Desgnng Neuro - fuzzy Systems through Back-Propagaton, Fuzzy Modellng: Paradgms and Practce (Boston: Kluwer, 1996). [11] D. Nauck, and K. Kruse, Neuro-fuzzy Systems Research and Applcatons Outsde of Japan: Fuzzy-Neural Networks (n Japanese) (Tokyo: Asakura Publcatons, 1996). [12] J.S.R. Jang, ANFIS: Adaptve-Network-Based Fuzzy Inference Systems, IEEE Trans. on Systems, Man & Cybernetcs, 23, 1993, [13] R. Sngh, R. Seta, and G. Das, Modelng and optmzaton of rotary furnace parameters usng artfcal neural networks and genetc evolutonary algorthms, Proc. 31 st Natonal Systems Conference, MIT, Manpal, Inda, 2007, Paper ID 75. [14] R. Sngh, G. Das, and R. Seta, An evolutonary approach to parametrc optmzaton of rotary furnace usng lght desel ol, Internatonal Journal of Agle Manufacturng, 10(3), 2008, [15] G. Das, and P. Das, Cuttng Forces n Drllng Operaton: Measurement and Modelng for Medum-scale Manufacturng Frms, Internatonal Journal of Computer Applcatons, 121(8), 2015, [16] T. Takag, and M. Sugeno, Fuzzy Identfcaton of Systems and ts Applcaton to Modelng and Control, IEEE Transactons on Systems, Man & Cybernetcs, 15, 1985, [17] M. Sugeno, and G.T. Kang, Structure Identfcaton of Fuzzy Model, Fuzzy Sets and Systems, 28, 1998, The IJES Page 43