- Email: [email protected]
NEW NEURAL NETWORKS APPROACH FOR OPTIMAL MACHINES TOOLS SELECTION
NEW NEURAL NETWORKS APPROACH FOR OPTIMAL MACHINES TOOLS SELECTION Romdhane Ben Khalifa, Noureddine Ben Yahia, Ali Zghal [email protected], [email protected], [email protected] ESSTT, University of Tunis, 5 avenue Taha Hussein, Montfleury,1008 Tunis, Tunisie
Abstract: This paper proposes an original method of optimization of machines tools selection during the generation of process planning. This method used the multi-layer neural networks which are able to give solutions for machine tools selection according to proposed machining features. In this work, we have elaborated three complementary parts which make it possible to choose the machine tools, as well as their characteristics; firstly, we have created a data base knowledge of know-how, and this here starting from the investigation with experts in mechanical companies of manufacture. Secondly, we have developed multi-layer neural networks whose desired outputs are classes of machine tools which are selected according to the characteristics of machining features. In end, we have created second multi-layer neural networks which produces the machine tools able and available to make machining according to classes of machine tools selected by the first neural networks, capacity of machining workshop, availability of the machine tools and cutting conditions. Copyright © 2006 IFAC Keywords: CAPP, Feed-Forward neural networks, CAD/CAM, machining features, machines-tools, cutting tools.
1. INTRODUCTION The automation of machining process planning activities by employing of Computer Aided Process Planning system collected much interest to these years course. It is always a challenge so that the engineers establish the bond created by conflict between requests for design and conditions of manufacture, like footbridge between CAD and CAM. Indeed, the automatic machining process planning (CAPP) «Computer Aided Process Planning » plays a significant role in integrity of the CAD/CAM systems. One of the principal objectives of automatic machining process planning system is to interpret the information of design and to prescribe the operations of machining appropriate and conformed to conditions determined by designer. However, the manufacturing companies usually try to reduce manufacturing costs and production times, as to increase the productivity. These objectives cannot be obtained without consideration of an optimal use of machine tools (Gologlu, 2004; Drstvensek et al., 2000). Indeed, the need for automatic of machine tool choice in mechanical manufacture is today a stage very interesting for effectiveness of machining operation as well as reliability of a manufacturing process planning. It is carried out according to workpieces characteristics as well as quantitative and qualitative informations (Yardakul, 2004).
Recently many authors have used different approaches for machines-tools selection, such as Chung (2004) has proposed a method of the selection of tools and machines using Web-based manufacturing environments (Chung et al., 2004) and Mustafa Yurdakul (2004) has used a strategic justification tool for machine tools. With the new strategic justification tool, the evaluation of investment in machines tools can model and quantify strategic considerations. AHP and ANP are applied in calculation of the contributions of machine tool alternatives to the manufacturing strategy of a manufacturing organization. Hierarchical decision structures are formed in the application of the AHP and ANP approaches (Mustafa, 2004). As well as Chiung Moon (2002) has proposed a model of integration of the machines tools and sequences of machining operations in CAPP systems by using genetic algorithm, this model with for goal to reduce the time and the loads of manufacture (Chiung et al., 2002). However, Neural Networks are powerful to replace the methods of classifications, like their high speed of resolution and their aptitude of training and significant adaptation. We benefited from these performances to apply the neural networks for the automatic choice of machines tools during generation of machining process planning. The first part of this paper relates to modelling of the machining features, like their characteristics. The second part is devoted to development of neuronal system structure for automatic machine-tools choice. In
the third part of this paper we will present the method of optimization to machine-tools choice by using a second network of neurons (NN2). We finish by interpretations of performances results of neuronal
system for automatic of machine-tools choice starting fro m a c as e s tud y, as w e ll as a con c lus ion . We present in figure 1 the module of the optimization process of machine tools choice by neural networks
Machining features Characteristics
Cutting conditions (Power ...) Number of work pieces …
M.O 1
NN1
M.O 2
Machines -tools classes
M.O 3
List of machine tools able to make machining
. .
NN2 Available machinetools
Machining workshop
Capability and characteristic of machine-tools
M.O 12
(Standard Nf E 60 010)
Fig.1. Optimization process of machines tools choice by neural networks 2. DEVELOPMENT OF NEURONAL SYSTEM NN1 FOR THE CHOICE OF MACHINES TOOLS CLASSES During the development of this module, we began by the creation of database of know-how starting from an investigation in manufacturing mechanical companies, with production specialists, experts, mechanical engineers and skilled workers production in machining, as well as teachers of mechanical production. Indeed, the method used in the expertise of the production specialists or equivalents is entitled ETED (Emploi Type Etudié dans sa Dynamique: Job kind in dynamic study), which makes it possible to carry out talks and to structure the results of the investigations (Mandon, 1991; Eversheim et al., 2001). Table 1 Classification of machines tools
We show on figure 2 the structure of neuronal system NN1 which is based on the choice of machines tools families in relation to machining features. The inputs of the network are the criteria of selection of machines tools which are extracted from the basic module of database of know-how (BEN Khalifa et al., 2003). There are several manners of classifying the machines tools, such as: by type of employment, by architecture (with horizontal spindle, vertical spindle and directional spindle), by type of order (conventional, automatic, numerical control... etc), and by dimensions (displacement following 3 axes X, Y, Z... etc). In our study we propose a classification of the machines tools according to the number of possible axes in a machine tool (one axis, two axes, three axes, four axes, five axes). This classification is related to morphologies of machining features, like their type of machining and their operations possible (table 1).
Classes of MachinesTools
Class 1 (1 axis)
Class 2 (2 axes)
Class 3 (3 axes)
Class 4 (3 axes)
Class 5 (4 axes)
Motions Classes of MachinesTools
Z Class7 (4 axes) turning X, Y, Z, C
X, Z Class 8 (4 axes)
X, Y, Z Class 9 (5 axes)
X, Z, C Class 10 (5 axes)
X, Y, Z, B Class 11 (5 axes)
Class 6 (4 axes) milling X, Y, Z, C Class 12 (5 axes)
2x(X, Z)
2x(X, Z), C
X, Y, Z, A, C
X, Y, Z, B, C
X, Y, Z, A, B
Motions
Fig.3. Structure of neural networks NN1 The necessity of the automatic choice of a machinetool in mechanical manufacturing is today a very interesting stage for the efficiency of a machining operation as well as the reliability of a manufacture process planning. Indeed, to manufacture an entity of machining, it exists several possibilities of machinetool selection, however to optimize this choice us must respect certain number of criteria's of the machine-tool choice (the morphology of the piece to manufacture, measurements, the asked precisions, the dimensional, geometric and technological constraints, The automatic system of the choice of machine-tools which we used in this study is based on multi-layer artificial neural networks. They have the advantage to permit with a certain number of tests to select the appropriate machine-tool with the characteristics of proposed machining features (Chiung et al., 2002; Chrysoolours et al., 2001). The model of multi-layer neural networks is based on a simple representation of the biological neurons in form of a function of several variables. For this sort of networks, the activity of a neuron is modelled by a real number and the synapses by coefficients. As their name indicates it, the multi-layer neural networks are divided into layers; the first layer is a layer of inputs because it receives the vectors of inputs, reciprocally the last layer is a layer of outputs, it produces the results. The intermediate layers are called hidden layers, because states of neurons that they contain are not observable. The proposed neural networks is a self-adapting structure, it internally modified until attaining the desired result following the phase of training and generalization. Indeed, the training is a development phase of neural networks during which the behavior of the networks is modified until obtaining the desired behavior. It is done in the context of a task or a behavior to be learned. Information to be treated is coded in the shape of an inputs vector, which is communicated to the inputs neural networks. The answer of the network is
interpreted starting from the value of activation of its outputs neurons, of which the outputs vector. It is a procedure which consists in estimating the parameters of neural networks, so that this one as well as possible fills the task which is affected for him (Dunfied et al., 2004). At the end of this process, the network was to be able to generate the good solutions for examples which were not seen before it is the objective of the generalization phase. This process consists in generalizing the outputs results of the inputs to network do not belong to the training base. Indeed, for the multi-layer neural networks, the training algorithm used is the retro propagation of gradient (Zouidi et al., 2004). The application that we proposed here for the resolution to machining problems of prismatic interacting features of the type groove / step (Thomas, 2000; BEN Yahia et al., 2000; STEP, 1999). Indeed, to respect constraints, the geometric tolerances and the state of surface of workpiece registered to the definition drawing of machining feature (Yardakul, 2004; Ahmed et al., 2002), we affected these features in three different codes: (1) without constraint, (2) dimensional requirements, (3) geometrical requirements and/or state of surface. Besides, we have specified the material of manufacturing workpiece while regrouping materials in three families according to their hardness by three different codes : (1) alloys of copper and alloys of aluminium, (2) soft steels , (3) hard steels and cast irons. The outputs of the network consists of a matrix of dimension (6x18), the set of these outputs describes the proposed solutions by possible machine-tools classes according to the desired outputs (S1, S2, S3,…, S11, S12) (BEN Khalifa et al., 2005a). For this configuration of the neural networks, we propose the different types of training according to number of epochs, estimate of the gradient of errors and the training function (Dunfied et al., 2004; BEN Yahia, 2002).
Fig.3. Intersecting features groove / step
Table 2 Parameters of the neural networks NN1 a1 (mm) >10 >10 >10 >10 >10 >10 >10 >10 >10 >50 >200 ---->250 >250 >250 >250 >250
a2 (mm) >10 >10 >10 >10 >10 >10 >10 >10 >10 >50 >200 ---->250 >250 >250 >250 >250
Inputs a3 a1/a2 code Code C1 (mm) Cont. Mate. <1.5 <1 1 1 0 <1.5 <1.5 1 1 0 <1.5 <5 1 1 0 <1.5 <1 2 1 0 <1.5 <1.5 2 1 0 <1.5 <5 2 1 0 <1.5 <1 3 1 0 <1.5 <1.5 3 1 0 <1.5 <5 3 1 0 <1.5 <1 1 1 0 <1.5 <1 1 1 0 --------------------<1.5 <1.5 2 2 0 <1.5 <5 2 2 0 <1.5 <1 3 2 0 <1.5 <1.5 3 2 0 <1.5 <5 3 2 0
The graphs represented in the figure 4 shows the evolution of the training mean squared error (TMSE) according to the number of epochs and measurements of machining features. Indeed, it is noticed that the TMSE is weak for the different measurements of the machining features of our application and after 59 epochs according to the proposed measurements, it will have the stability of the network with a training mean squared error (TMSE) minimal lower to 0.1.
Outputs C2 C3 C4 C5 C6 C7 C8 C9 C10 C11 C12 0 0 0 0 0 0 0 0 0 0 1 ----1 1 1 1 1
1 1 1 1 1 1 1 1 1 1 1 ----1 1 1 1 1
1 1 1 1 1 1 1 1 1 1 1 ----1 1 1 1 1
0 0 0 0 0 0 0 0 0 0 1 ----1 1 1 1 1
1 1 1 1 1 1 1 1 1 1 1 ----1 1 1 1 1
0 0 0 0 0 0 0 0 0 0 0 ----0 0 0 0 0
0 0 0 0 0 0 0 0 0 1 1 ----0 0 0 0 0
0 0 0 1 1 1 1 1 1 0 0 ----1 1 1 1 1
0 0 0 1 1 1 1 1 1 0 0 ----1 1 1 1 1
0 0 0 1 1 1 1 1 1 0 0 ----1 1 1 1 1
0 0 0 1 1 1 1 1 1 0 0 ----1 1 1 1 1
and we have compared it with the one of the training (BEN Khalifa et al., 2005b; BEN Yahia et al., 1999).
Fig.5. Evolution of the training error and of generalization of the NN1
Fig.4. Result of training of the network NN1 In order to fix an optimal structure of neural network, we must choose the parameters of network well such as the number of the hidden layers, the number of neurons in the hidden layer and the function of training and adaptation. Indeed, the best structure of network is obtained starting from a very weak TMSE and an optimal number of epochs "Kopt" so that the generalization mean squared error (GMSE) is minimal. However, we have presented at the graphs represented in figure 5 show the neural networks system an input vector doesn’t belong to the basis of training then we have examine the generalization mean squared error
The evolution of training error (TMSE) and generalization error (GMSE) of our neuronal system for the automatic choice of machines tools. We notice that the generalization error (GMSE) decreases until a number of epochs (or iterations) to be determined well, by this value error (GMSE) increases. This translated it on training of the network. Indeed, it is obliged to stop the training for an optimal number of epochs (Kopt=59). In order to choose the number of optimal neuron, we made a series of training for various numbers of neurons in the hidden layers. The choice of the optimal number of epochs corresponds to a minimal GMSE and TMSE. The number of neurons making it possible to have a minimal TMSE in this case is 20 neurons. This phase enabled us to choose the optimal structure of the
neuronal system starting from information of the studied machining features (figure 3) and the characteristics of machines tools available in machining workshop according to the French standard (NF E 60-010). In this optimization case of machines tools choice, we are placed in the context of a machining workshop producing small and mean sets of pieces. The machining feature that we propose through this survey of optimization is the same that the one for the choice of machines tools classes, it is a intersect machining features (groove / step) (BEN Khalifa, et al., 2005c). The best structure of the network is obtained from a very weak training mean squared error (TMSE) and an optimal iteration number so that the error of generalization (GMSE) is minimal (Ding1 et al., 2004). The graphs represented in the figure 6 shows the evolution of the training mean squared error (TMSE) of NN2 according to the number of epochs and measurements of machining features. We observe that the generalization error decreases until a number of epochs very determined. After this value of epochs, GMSE increases. This translates the on-training of network. Indeed, it is obliged to stop the training for an optimal epochs number (29 epochs).
network. Lastly, the following structure was fixed: - inputs and outputs: 6 inputs and 12 outputs, - number of hidden layers: 1 only hidden layer, - number of neurons in the hidden layer: 20 neurons, - activation function: hyperbolic tangent for the set of the neurons in the network. 3. DEVELOPMENT OF NEURONAL SYSTEM NN2 FOR OPTIMIZATION OF MACHINES TOOLS SELECTION To make sure a correct machines tools choice during generation of machining process planning, we must ensure existing of production means and resources in machining workshop, such as machines tools types (existing and available), like their characteristic and their capacity of machining. However, to optimize this kind of preparation operations of the machining process planning, quantitative methods have only been developed with consideration of a simple objective, as minimization of cost or maximization of profit... etc. For process of simple objective optimization, several researchers have proposed different techniques as differential calculation, regression analysis, linear, geometric and stochastic programming. However, optimization of machines tools choice is an optimization no linear with constraints, so it is difficult with conventional algorithms of optimization to solve this problem by reason of problems of convergence speed or accuracy. Indeed, we present a new method of optimization of machines tools selection by using neuronal approach. Indeed, to optimize this choice we have used a second network of multi-layer neurons (NN2) whose desired outputs are machines tools able for suggested machining feature according to of the machine tools classes selected by first neural networks (NN1), as well as machining workshop capacity (figure 1). The training base of NN2 is consisted inputs vectors representing machines tools classes selected by NN1 and machine tools existing and available in considered machining workshop. As well as the wished outputs are machine tools able to machine of suggested machining features (M1, M2, M3..., M12). The set of these outputs describe the possible solutions for optimized choice of machines tools by observing the cutting conditions of studied feature, there capacity of machining workshop and the characteristics of available machines tools (Table 3). Moreover, we have took in dedication the cutting conditions which are calculated outside from the
Fig.6. Evolution of the training error and of generalization of the NN2 The optimal structure that we have fixed during the two Phases of training and generalization, the following structure was fixed: - inputs and outputs: 24 inputs and 12 outputs, - number of hidden layers: 1 only hidden layer, - number of neurons in the hidden layer: 15 neurons, - activation function : hyperbolic tangent for the set of the neurons in the network.
Table 3 Parameters of the neural networks NN2 Inputs a1 (mm) >10
a2 (mm) >10
a2 (mm) >10
>50
>50
>50
>100
>100
>100
>200
>200
>200
>300
>300
>300
Classes of machines-tools; machining workshop [0 0 1 1 1 1 0 0 0 0 0 0; 1 5 3 0 1 0 0 0 0 1 0 0] [0 0 1 1 0 1 0 0 1 0 0 0; 1 5 3 0 1 0 0 0 0 1 0 0] [0 0 1 1 0 1 0 0 1 0 0 0; 1 5 3 0 1 0 0 0 0 1 0 0] [0 0 1 1 0 1 0 0 1 0 1 0; 1 5 3 0 1 0 0 0 0 1 0 0] [0 0 0 1 0 1 0 0 1 1 0 0; 1 5 3 0 1 0 0 0 0 1 0 0]
Outputs M1
M2
M3
M4
M5
M6
M7
M8
M9
M10
M11
M12
0
0
2
0
1
0
0
0
0
0
0
0
0
0
2
0
0
0
0
0
0
0
0
0
0
0
2
0
0
0
0
0
0
0
0
0
0
0
2
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
4. CONCLUSION We have showed in this paper the interest of neural networks for optimization of machine tools choice during generation of machining process planning. The validation has been proposed here for the relative machines tools choice to interaction machining features of type groove / step. The neuronal approach remains promising compared to the approaches group technology and alternatives, especially in speed of establishment and update in manufacturing industry, as well as the precision in the automatic of machine tools choice of the data base (BEN Khalifa et al., 2005d). REFERENCES Ahmad, N. and A.F.M. A.Haque (2002), Artificial Neural Networks Based process selection for cylindrical surface machining, Proceedings of the Int. Conf. on Manufacturing, ICM 2002 09-11 August, , Dhaka. pp.321 – 326. Ben Yahia, N., F. Fnaiech et B.Hadj Sassi (2000), Application des réseaux de neurones multicouche pour l’élaboration de phase d’usinage, Conférence International Francphonie d’Automatique CIFA, pp 640 645, 5-7 Juillet. Ben Yahia, N. et B.Hadj Sassi (1999) Elaboration automatique de phases d’usinage basé sur les réseau de neurones artificielles » Conférence International de Productique (CPI 99) 25-26 Nov. au Maroc (FST de Tanger). Ben Yahia, N.(2002), Processus d’élaboration de gamme automatique d’usinage : application aux entités prismatiques Thèse de doctorat, ENIT,. Ben Khalifa, R., N. BEN Yahia et A. Zghal (2003), Elaboration d’un système neuronal pour le choix automatique des outils tournants, Journées Scientifiques et Pédagogiques de Mécanique et Energétique JSPME ISET Gafsa, 1, 2, 3 Décembre p101-107. Ben Khalifa, R,. N. BEN Yahia et A. Zghal (2005a) Choix automatique des machines-outils basé sur les réseaux de neurones multicouches, Premier Congrès International Conception et Modélisation des Systèmes Mécaniques, CMSM’2005 23-25 Mars, Hammamet Tunisie. Ben Khalifa, R., N. BEN Yahia and A. Zghal (2005b), Automated selection of machines-tools using artificial neural networks, 2nd AUS International Symposium on Mechatronics, AUS-ISM05, April 19–21, American University of Sharjah Sharjah, United Arab Emirates. Ben Khalifa, R., N. BEN Yahia et A. Zghal (2005c), Intégration des réseaux de neurones multicouches pour le choix automatique des machines-outils, 6e Congrès international de génie industriel – 7-10 juin, Besançon, France. Ben Khalifa, R., N. BEN Yahia et A. Zghal (2005d), Optimisation du choix des machines-outils par un système neuronal, Conception et Production Integrées CPI’2005, 09-11 Nouvembre Casablanca, Maroc.
Gologlu, C. (2004), Machine capability and fixturing constraints-imposed automatic machining set-ups generation, Journal of Materials Processing Technology 148, 83–92. Chiung, M., M., Lee, Y., Seo and Y., Hae Lee (2002), Integrated machine tool selection and operation sequencing with capacity and precedence constraints using genetic algorithm , International Journal of Computers & Industrial Engineering 43 605–621. Chryssolours, G., S.Zannis, C.Derdas and K.Tsirbas, (2001), Dimensional Accuracy of FDM Parts, The 34th CIRP International Seminar on Manufacturing Systems, 16-18 May 2001, Athenes, Greece. Chung, C. and Q.Peng (2004), The selection of tools and machines on web-based manufacturing environments, International Journal of Machine Tools & Manufacture 44 317–326 Drstvensek, I. and M. Brezocnik (2000), On intelligent CAD/CAPP Integration Base on Feature Recognition and Evolutionary Computation, 3rd International Conference on Integrated Design and Engineering, Canadian Society for Mechanical Engineering IDMME 2000. Dunfied, J., M. Tarbouchi and G. Labonte (2004), Neural Network Based control of a Four Rotor Helicopter, IEEE icit 2004 Yasmine-Hammamet, December 8-10. Ding1, L. and Y. Yue Novel (2004), ANN-based feature recognition incorporating design by features, International Journal of Computers in Industry xxx (2004) xxx–xxx. Eversheim, W., M. Koschig and N. Michalas (2001), Knowledge-based Technology Data Management, The 34th CIRP International Seminar on Manufacturing Systems, 16-18 May, Athenes, Greece. Mandon, N. (1991), L’emploi-type étudié dans sa dynamique, CEREQ, Collection Documents. Mustafa, Y., (2004), AHP as a strategic decisionmaking tool to justify machine tool selection, Journal of Materials Processing Technology 146, 365–376. STEP (1999), Standard for External representation of Product data, ISO 10303-224, Industrial automation systems and integration. Product data representation and exchange: mechanical product definition for process plans using machining features. Thomas, H. (2000), Génération de séquences d’alésage par approche neuronal, , 3rd International Conference on Integrated Design and Manufacturing for Mechanical Engineering, Canadian Society for Mechanical Engineering IDMME 2000. Zouidi, A., A.Chaari M. Stambouli and F. Fnaiech (2004), Nonlinear Continuous Time Modeling of a High Pressure Mercury Vapor Discharge Lamp Using Feed ForwardBack-propagation Neural Networks, IEEE icit 2004 Yasmine-Hammamet, December 8-10.
Abstract: This paper proposes an original method of optimization of machines tools selection during the generation of process planning. This method used the multi-layer neural networks which are able to give solutions for machine tools selection according to proposed machining features. In this work, we have elaborated three complementary parts which make it possible to choose the machine tools, as well as their characteristics; firstly, we have created a data base knowledge of know-how, and this here starting from the investigation with experts in mechanical companies of manufacture. Secondly, we have developed multi-layer neural networks whose desired outputs are classes of machine tools which are selected according to the characteristics of machining features. In end, we have created second multi-layer neural networks which produces the machine tools able and available to make machining according to classes of machine tools selected by the first neural networks, capacity of machining workshop, availability of the machine tools and cutting conditions. Copyright © 2006 IFAC Keywords: CAPP, Feed-Forward neural networks, CAD/CAM, machining features, machines-tools, cutting tools.
1. INTRODUCTION The automation of machining process planning activities by employing of Computer Aided Process Planning system collected much interest to these years course. It is always a challenge so that the engineers establish the bond created by conflict between requests for design and conditions of manufacture, like footbridge between CAD and CAM. Indeed, the automatic machining process planning (CAPP) «Computer Aided Process Planning » plays a significant role in integrity of the CAD/CAM systems. One of the principal objectives of automatic machining process planning system is to interpret the information of design and to prescribe the operations of machining appropriate and conformed to conditions determined by designer. However, the manufacturing companies usually try to reduce manufacturing costs and production times, as to increase the productivity. These objectives cannot be obtained without consideration of an optimal use of machine tools (Gologlu, 2004; Drstvensek et al., 2000). Indeed, the need for automatic of machine tool choice in mechanical manufacture is today a stage very interesting for effectiveness of machining operation as well as reliability of a manufacturing process planning. It is carried out according to workpieces characteristics as well as quantitative and qualitative informations (Yardakul, 2004).
Recently many authors have used different approaches for machines-tools selection, such as Chung (2004) has proposed a method of the selection of tools and machines using Web-based manufacturing environments (Chung et al., 2004) and Mustafa Yurdakul (2004) has used a strategic justification tool for machine tools. With the new strategic justification tool, the evaluation of investment in machines tools can model and quantify strategic considerations. AHP and ANP are applied in calculation of the contributions of machine tool alternatives to the manufacturing strategy of a manufacturing organization. Hierarchical decision structures are formed in the application of the AHP and ANP approaches (Mustafa, 2004). As well as Chiung Moon (2002) has proposed a model of integration of the machines tools and sequences of machining operations in CAPP systems by using genetic algorithm, this model with for goal to reduce the time and the loads of manufacture (Chiung et al., 2002). However, Neural Networks are powerful to replace the methods of classifications, like their high speed of resolution and their aptitude of training and significant adaptation. We benefited from these performances to apply the neural networks for the automatic choice of machines tools during generation of machining process planning. The first part of this paper relates to modelling of the machining features, like their characteristics. The second part is devoted to development of neuronal system structure for automatic machine-tools choice. In
the third part of this paper we will present the method of optimization to machine-tools choice by using a second network of neurons (NN2). We finish by interpretations of performances results of neuronal
system for automatic of machine-tools choice starting fro m a c as e s tud y, as w e ll as a con c lus ion . We present in figure 1 the module of the optimization process of machine tools choice by neural networks
Machining features Characteristics
Cutting conditions (Power ...) Number of work pieces …
M.O 1
NN1
M.O 2
Machines -tools classes
M.O 3
List of machine tools able to make machining
. .
NN2 Available machinetools
Machining workshop
Capability and characteristic of machine-tools
M.O 12
(Standard Nf E 60 010)
Fig.1. Optimization process of machines tools choice by neural networks 2. DEVELOPMENT OF NEURONAL SYSTEM NN1 FOR THE CHOICE OF MACHINES TOOLS CLASSES During the development of this module, we began by the creation of database of know-how starting from an investigation in manufacturing mechanical companies, with production specialists, experts, mechanical engineers and skilled workers production in machining, as well as teachers of mechanical production. Indeed, the method used in the expertise of the production specialists or equivalents is entitled ETED (Emploi Type Etudié dans sa Dynamique: Job kind in dynamic study), which makes it possible to carry out talks and to structure the results of the investigations (Mandon, 1991; Eversheim et al., 2001). Table 1 Classification of machines tools
We show on figure 2 the structure of neuronal system NN1 which is based on the choice of machines tools families in relation to machining features. The inputs of the network are the criteria of selection of machines tools which are extracted from the basic module of database of know-how (BEN Khalifa et al., 2003). There are several manners of classifying the machines tools, such as: by type of employment, by architecture (with horizontal spindle, vertical spindle and directional spindle), by type of order (conventional, automatic, numerical control... etc), and by dimensions (displacement following 3 axes X, Y, Z... etc). In our study we propose a classification of the machines tools according to the number of possible axes in a machine tool (one axis, two axes, three axes, four axes, five axes). This classification is related to morphologies of machining features, like their type of machining and their operations possible (table 1).
Classes of MachinesTools
Class 1 (1 axis)
Class 2 (2 axes)
Class 3 (3 axes)
Class 4 (3 axes)
Class 5 (4 axes)
Motions Classes of MachinesTools
Z Class7 (4 axes) turning X, Y, Z, C
X, Z Class 8 (4 axes)
X, Y, Z Class 9 (5 axes)
X, Z, C Class 10 (5 axes)
X, Y, Z, B Class 11 (5 axes)
Class 6 (4 axes) milling X, Y, Z, C Class 12 (5 axes)
2x(X, Z)
2x(X, Z), C
X, Y, Z, A, C
X, Y, Z, B, C
X, Y, Z, A, B
Motions
Fig.3. Structure of neural networks NN1 The necessity of the automatic choice of a machinetool in mechanical manufacturing is today a very interesting stage for the efficiency of a machining operation as well as the reliability of a manufacture process planning. Indeed, to manufacture an entity of machining, it exists several possibilities of machinetool selection, however to optimize this choice us must respect certain number of criteria's of the machine-tool choice (the morphology of the piece to manufacture, measurements, the asked precisions, the dimensional, geometric and technological constraints, The automatic system of the choice of machine-tools which we used in this study is based on multi-layer artificial neural networks. They have the advantage to permit with a certain number of tests to select the appropriate machine-tool with the characteristics of proposed machining features (Chiung et al., 2002; Chrysoolours et al., 2001). The model of multi-layer neural networks is based on a simple representation of the biological neurons in form of a function of several variables. For this sort of networks, the activity of a neuron is modelled by a real number and the synapses by coefficients. As their name indicates it, the multi-layer neural networks are divided into layers; the first layer is a layer of inputs because it receives the vectors of inputs, reciprocally the last layer is a layer of outputs, it produces the results. The intermediate layers are called hidden layers, because states of neurons that they contain are not observable. The proposed neural networks is a self-adapting structure, it internally modified until attaining the desired result following the phase of training and generalization. Indeed, the training is a development phase of neural networks during which the behavior of the networks is modified until obtaining the desired behavior. It is done in the context of a task or a behavior to be learned. Information to be treated is coded in the shape of an inputs vector, which is communicated to the inputs neural networks. The answer of the network is
interpreted starting from the value of activation of its outputs neurons, of which the outputs vector. It is a procedure which consists in estimating the parameters of neural networks, so that this one as well as possible fills the task which is affected for him (Dunfied et al., 2004). At the end of this process, the network was to be able to generate the good solutions for examples which were not seen before it is the objective of the generalization phase. This process consists in generalizing the outputs results of the inputs to network do not belong to the training base. Indeed, for the multi-layer neural networks, the training algorithm used is the retro propagation of gradient (Zouidi et al., 2004). The application that we proposed here for the resolution to machining problems of prismatic interacting features of the type groove / step (Thomas, 2000; BEN Yahia et al., 2000; STEP, 1999). Indeed, to respect constraints, the geometric tolerances and the state of surface of workpiece registered to the definition drawing of machining feature (Yardakul, 2004; Ahmed et al., 2002), we affected these features in three different codes: (1) without constraint, (2) dimensional requirements, (3) geometrical requirements and/or state of surface. Besides, we have specified the material of manufacturing workpiece while regrouping materials in three families according to their hardness by three different codes : (1) alloys of copper and alloys of aluminium, (2) soft steels , (3) hard steels and cast irons. The outputs of the network consists of a matrix of dimension (6x18), the set of these outputs describes the proposed solutions by possible machine-tools classes according to the desired outputs (S1, S2, S3,…, S11, S12) (BEN Khalifa et al., 2005a). For this configuration of the neural networks, we propose the different types of training according to number of epochs, estimate of the gradient of errors and the training function (Dunfied et al., 2004; BEN Yahia, 2002).
Fig.3. Intersecting features groove / step
Table 2 Parameters of the neural networks NN1 a1 (mm) >10 >10 >10 >10 >10 >10 >10 >10 >10 >50 >200 ---->250 >250 >250 >250 >250
a2 (mm) >10 >10 >10 >10 >10 >10 >10 >10 >10 >50 >200 ---->250 >250 >250 >250 >250
Inputs a3 a1/a2 code Code C1 (mm) Cont. Mate. <1.5 <1 1 1 0 <1.5 <1.5 1 1 0 <1.5 <5 1 1 0 <1.5 <1 2 1 0 <1.5 <1.5 2 1 0 <1.5 <5 2 1 0 <1.5 <1 3 1 0 <1.5 <1.5 3 1 0 <1.5 <5 3 1 0 <1.5 <1 1 1 0 <1.5 <1 1 1 0 --------------------<1.5 <1.5 2 2 0 <1.5 <5 2 2 0 <1.5 <1 3 2 0 <1.5 <1.5 3 2 0 <1.5 <5 3 2 0
The graphs represented in the figure 4 shows the evolution of the training mean squared error (TMSE) according to the number of epochs and measurements of machining features. Indeed, it is noticed that the TMSE is weak for the different measurements of the machining features of our application and after 59 epochs according to the proposed measurements, it will have the stability of the network with a training mean squared error (TMSE) minimal lower to 0.1.
Outputs C2 C3 C4 C5 C6 C7 C8 C9 C10 C11 C12 0 0 0 0 0 0 0 0 0 0 1 ----1 1 1 1 1
1 1 1 1 1 1 1 1 1 1 1 ----1 1 1 1 1
1 1 1 1 1 1 1 1 1 1 1 ----1 1 1 1 1
0 0 0 0 0 0 0 0 0 0 1 ----1 1 1 1 1
1 1 1 1 1 1 1 1 1 1 1 ----1 1 1 1 1
0 0 0 0 0 0 0 0 0 0 0 ----0 0 0 0 0
0 0 0 0 0 0 0 0 0 1 1 ----0 0 0 0 0
0 0 0 1 1 1 1 1 1 0 0 ----1 1 1 1 1
0 0 0 1 1 1 1 1 1 0 0 ----1 1 1 1 1
0 0 0 1 1 1 1 1 1 0 0 ----1 1 1 1 1
0 0 0 1 1 1 1 1 1 0 0 ----1 1 1 1 1
and we have compared it with the one of the training (BEN Khalifa et al., 2005b; BEN Yahia et al., 1999).
Fig.5. Evolution of the training error and of generalization of the NN1
Fig.4. Result of training of the network NN1 In order to fix an optimal structure of neural network, we must choose the parameters of network well such as the number of the hidden layers, the number of neurons in the hidden layer and the function of training and adaptation. Indeed, the best structure of network is obtained starting from a very weak TMSE and an optimal number of epochs "Kopt" so that the generalization mean squared error (GMSE) is minimal. However, we have presented at the graphs represented in figure 5 show the neural networks system an input vector doesn’t belong to the basis of training then we have examine the generalization mean squared error
The evolution of training error (TMSE) and generalization error (GMSE) of our neuronal system for the automatic choice of machines tools. We notice that the generalization error (GMSE) decreases until a number of epochs (or iterations) to be determined well, by this value error (GMSE) increases. This translated it on training of the network. Indeed, it is obliged to stop the training for an optimal number of epochs (Kopt=59). In order to choose the number of optimal neuron, we made a series of training for various numbers of neurons in the hidden layers. The choice of the optimal number of epochs corresponds to a minimal GMSE and TMSE. The number of neurons making it possible to have a minimal TMSE in this case is 20 neurons. This phase enabled us to choose the optimal structure of the
neuronal system starting from information of the studied machining features (figure 3) and the characteristics of machines tools available in machining workshop according to the French standard (NF E 60-010). In this optimization case of machines tools choice, we are placed in the context of a machining workshop producing small and mean sets of pieces. The machining feature that we propose through this survey of optimization is the same that the one for the choice of machines tools classes, it is a intersect machining features (groove / step) (BEN Khalifa, et al., 2005c). The best structure of the network is obtained from a very weak training mean squared error (TMSE) and an optimal iteration number so that the error of generalization (GMSE) is minimal (Ding1 et al., 2004). The graphs represented in the figure 6 shows the evolution of the training mean squared error (TMSE) of NN2 according to the number of epochs and measurements of machining features. We observe that the generalization error decreases until a number of epochs very determined. After this value of epochs, GMSE increases. This translates the on-training of network. Indeed, it is obliged to stop the training for an optimal epochs number (29 epochs).
network. Lastly, the following structure was fixed: - inputs and outputs: 6 inputs and 12 outputs, - number of hidden layers: 1 only hidden layer, - number of neurons in the hidden layer: 20 neurons, - activation function: hyperbolic tangent for the set of the neurons in the network. 3. DEVELOPMENT OF NEURONAL SYSTEM NN2 FOR OPTIMIZATION OF MACHINES TOOLS SELECTION To make sure a correct machines tools choice during generation of machining process planning, we must ensure existing of production means and resources in machining workshop, such as machines tools types (existing and available), like their characteristic and their capacity of machining. However, to optimize this kind of preparation operations of the machining process planning, quantitative methods have only been developed with consideration of a simple objective, as minimization of cost or maximization of profit... etc. For process of simple objective optimization, several researchers have proposed different techniques as differential calculation, regression analysis, linear, geometric and stochastic programming. However, optimization of machines tools choice is an optimization no linear with constraints, so it is difficult with conventional algorithms of optimization to solve this problem by reason of problems of convergence speed or accuracy. Indeed, we present a new method of optimization of machines tools selection by using neuronal approach. Indeed, to optimize this choice we have used a second network of multi-layer neurons (NN2) whose desired outputs are machines tools able for suggested machining feature according to of the machine tools classes selected by first neural networks (NN1), as well as machining workshop capacity (figure 1). The training base of NN2 is consisted inputs vectors representing machines tools classes selected by NN1 and machine tools existing and available in considered machining workshop. As well as the wished outputs are machine tools able to machine of suggested machining features (M1, M2, M3..., M12). The set of these outputs describe the possible solutions for optimized choice of machines tools by observing the cutting conditions of studied feature, there capacity of machining workshop and the characteristics of available machines tools (Table 3). Moreover, we have took in dedication the cutting conditions which are calculated outside from the
Fig.6. Evolution of the training error and of generalization of the NN2 The optimal structure that we have fixed during the two Phases of training and generalization, the following structure was fixed: - inputs and outputs: 24 inputs and 12 outputs, - number of hidden layers: 1 only hidden layer, - number of neurons in the hidden layer: 15 neurons, - activation function : hyperbolic tangent for the set of the neurons in the network.
Table 3 Parameters of the neural networks NN2 Inputs a1 (mm) >10
a2 (mm) >10
a2 (mm) >10
>50
>50
>50
>100
>100
>100
>200
>200
>200
>300
>300
>300
Classes of machines-tools; machining workshop [0 0 1 1 1 1 0 0 0 0 0 0; 1 5 3 0 1 0 0 0 0 1 0 0] [0 0 1 1 0 1 0 0 1 0 0 0; 1 5 3 0 1 0 0 0 0 1 0 0] [0 0 1 1 0 1 0 0 1 0 0 0; 1 5 3 0 1 0 0 0 0 1 0 0] [0 0 1 1 0 1 0 0 1 0 1 0; 1 5 3 0 1 0 0 0 0 1 0 0] [0 0 0 1 0 1 0 0 1 1 0 0; 1 5 3 0 1 0 0 0 0 1 0 0]
Outputs M1
M2
M3
M4
M5
M6
M7
M8
M9
M10
M11
M12
0
0
2
0
1
0
0
0
0
0
0
0
0
0
2
0
0
0
0
0
0
0
0
0
0
0
2
0
0
0
0
0
0
0
0
0
0
0
2
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
4. CONCLUSION We have showed in this paper the interest of neural networks for optimization of machine tools choice during generation of machining process planning. The validation has been proposed here for the relative machines tools choice to interaction machining features of type groove / step. The neuronal approach remains promising compared to the approaches group technology and alternatives, especially in speed of establishment and update in manufacturing industry, as well as the precision in the automatic of machine tools choice of the data base (BEN Khalifa et al., 2005d). REFERENCES Ahmad, N. and A.F.M. A.Haque (2002), Artificial Neural Networks Based process selection for cylindrical surface machining, Proceedings of the Int. Conf. on Manufacturing, ICM 2002 09-11 August, , Dhaka. pp.321 – 326. Ben Yahia, N., F. Fnaiech et B.Hadj Sassi (2000), Application des réseaux de neurones multicouche pour l’élaboration de phase d’usinage, Conférence International Francphonie d’Automatique CIFA, pp 640 645, 5-7 Juillet. Ben Yahia, N. et B.Hadj Sassi (1999) Elaboration automatique de phases d’usinage basé sur les réseau de neurones artificielles » Conférence International de Productique (CPI 99) 25-26 Nov. au Maroc (FST de Tanger). Ben Yahia, N.(2002), Processus d’élaboration de gamme automatique d’usinage : application aux entités prismatiques Thèse de doctorat, ENIT,. Ben Khalifa, R., N. BEN Yahia et A. Zghal (2003), Elaboration d’un système neuronal pour le choix automatique des outils tournants, Journées Scientifiques et Pédagogiques de Mécanique et Energétique JSPME ISET Gafsa, 1, 2, 3 Décembre p101-107. Ben Khalifa, R,. N. BEN Yahia et A. Zghal (2005a) Choix automatique des machines-outils basé sur les réseaux de neurones multicouches, Premier Congrès International Conception et Modélisation des Systèmes Mécaniques, CMSM’2005 23-25 Mars, Hammamet Tunisie. Ben Khalifa, R., N. BEN Yahia and A. Zghal (2005b), Automated selection of machines-tools using artificial neural networks, 2nd AUS International Symposium on Mechatronics, AUS-ISM05, April 19–21, American University of Sharjah Sharjah, United Arab Emirates. Ben Khalifa, R., N. BEN Yahia et A. Zghal (2005c), Intégration des réseaux de neurones multicouches pour le choix automatique des machines-outils, 6e Congrès international de génie industriel – 7-10 juin, Besançon, France. Ben Khalifa, R., N. BEN Yahia et A. Zghal (2005d), Optimisation du choix des machines-outils par un système neuronal, Conception et Production Integrées CPI’2005, 09-11 Nouvembre Casablanca, Maroc.
Gologlu, C. (2004), Machine capability and fixturing constraints-imposed automatic machining set-ups generation, Journal of Materials Processing Technology 148, 83–92. Chiung, M., M., Lee, Y., Seo and Y., Hae Lee (2002), Integrated machine tool selection and operation sequencing with capacity and precedence constraints using genetic algorithm , International Journal of Computers & Industrial Engineering 43 605–621. Chryssolours, G., S.Zannis, C.Derdas and K.Tsirbas, (2001), Dimensional Accuracy of FDM Parts, The 34th CIRP International Seminar on Manufacturing Systems, 16-18 May 2001, Athenes, Greece. Chung, C. and Q.Peng (2004), The selection of tools and machines on web-based manufacturing environments, International Journal of Machine Tools & Manufacture 44 317–326 Drstvensek, I. and M. Brezocnik (2000), On intelligent CAD/CAPP Integration Base on Feature Recognition and Evolutionary Computation, 3rd International Conference on Integrated Design and Engineering, Canadian Society for Mechanical Engineering IDMME 2000. Dunfied, J., M. Tarbouchi and G. Labonte (2004), Neural Network Based control of a Four Rotor Helicopter, IEEE icit 2004 Yasmine-Hammamet, December 8-10. Ding1, L. and Y. Yue Novel (2004), ANN-based feature recognition incorporating design by features, International Journal of Computers in Industry xxx (2004) xxx–xxx. Eversheim, W., M. Koschig and N. Michalas (2001), Knowledge-based Technology Data Management, The 34th CIRP International Seminar on Manufacturing Systems, 16-18 May, Athenes, Greece. Mandon, N. (1991), L’emploi-type étudié dans sa dynamique, CEREQ, Collection Documents. Mustafa, Y., (2004), AHP as a strategic decisionmaking tool to justify machine tool selection, Journal of Materials Processing Technology 146, 365–376. STEP (1999), Standard for External representation of Product data, ISO 10303-224, Industrial automation systems and integration. Product data representation and exchange: mechanical product definition for process plans using machining features. Thomas, H. (2000), Génération de séquences d’alésage par approche neuronal, , 3rd International Conference on Integrated Design and Manufacturing for Mechanical Engineering, Canadian Society for Mechanical Engineering IDMME 2000. Zouidi, A., A.Chaari M. Stambouli and F. Fnaiech (2004), Nonlinear Continuous Time Modeling of a High Pressure Mercury Vapor Discharge Lamp Using Feed ForwardBack-propagation Neural Networks, IEEE icit 2004 Yasmine-Hammamet, December 8-10.