Multipurpose modelling and optimisation of production processes and process chains by combining machine learning and search techniques

Transcription

1 Multipurpose modelling nd optimistion o production processes nd process chins by combining mchine lerning nd serch techniues László Monostori, Zsolt J. Vihros Computer nd Automtion Reserch Institute, Hungrin Acdemy o Sciences Budpest, Hungry Abstrct The pper presents noel pproch or generting multipurpose models o mchining opertions combining mchine lerning nd serch techniues. These models re intended to be pplicble t dierent engineering nd mngement ssignments. Simulted nneling serch is used or inding the unknown prmeters o the models in gien situtions. It is expected tht the deeloped block-oriented rmework will be luble tool or modelling, monitoring nd optimistion o mnucturing processes nd process chins. The pplicbility o the proposed solution is illustrted by the results o experimentl runs.. Introduction Relible process models re extremely importnt in dierent ields o computer integrted mnucturing. They re reuired e.g. or selecting optiml prmeters during process plnning, or designing nd implementing dptie control systems or model bsed monitoring lgorithms. A wy is to implement undmentl models deeloped rom the principles o mchining science on computer. Howeer, in spite o progress being mde in undmentl process modelling, ccurte models re not yet ilble or mny mnucturing processes. Heuristic models re usully bsed on the rules o thumb gined rom experience, nd used or ulittie elution o decisions. Empiricl models deried rom experimentl dt still ply mjor role in mnucturing process modelling [7]. In the CIRP surey on deelopments nd trends in control nd monitoring o mchining processes, the necessity o sensor integrtion, sophisticted models, multimodel systems, nd lerning bility ws outlined []. Attching urther importnce to the issue, in 995 the CIRP Working Group on Modelling o Mchining Opertions ws estblished to promote the deelopment o models o chip remol opertions by deined cutting edges with the im to untittiely predict the perormnce o such opertions, nd to promote the use o such models in industry []. Diiculties in modelling mnucturing processes re mniold: the gret number o dierent mchining opertions, multidimensionl, non-liner, stochstic nture o mchining, prtilly understood reltions between prmeters, lck o relible dt, etc. A number o resons bck the reuired models: design o processes, optimistion o processes, control o processes, simultion o processes, nd design o euipment []. Artiicil neurl networks (ANNs), neuro-uzzy (NF) systems re generl, multirible, non-liner estimtors, thereore, oer ery eectie process modelling pproch. Such sot computing techniues seem to be ible solution or the lower leel o intelligent, hierrchicl control nd monitoring systems where bilities or rel-time unctioning, uncertinty hndling, sensor integrtion, nd lerning re essentil etures [8]. Successul ttempts were reported on in the literture [,, 7, 8, 9,, 5]. The ssignments to be perormed determined the input-output conigurtions o the models, i.e. the prmeters to be considered s inputs nd the ones s outputs. Dierent ssignments, howeer, reuire dierent model settings, i.e. dierent input-output model conigurtions. Considering the input-output ribles o gien tsk together s set o prmeters, the ANN model estimtes prt o this prmeter set bsed on the remining prt. The selection o input-output prmeters strongly inluences the ccurcy o the deeloped model, especilly i dependencies between prmeters re

2 non-inertble. At dierent stges o production (e.g. in plnning, optimistion or control) tsks re dierent, conseuently, the estimtion cpbilities o the relted pplied models ry, een i the sme set o prmeters is used. The pper summrises the irst results o the reserch ctiity iming t inding multipurpose model or set o ssignments which cn stisy the rious ccurcy reuirements. A method or utomtic genertion o ANN-bsed process models by bck propgtion nd heuristic serch is described. The ppliction phse o the process models is lso detiled. A noel techniue bsed on simulted nneling serch is introduced to ind the unknown prmeters o the model in gien situtions. The pplicbility o the proposed solution is illustrted by the results o experimentl runs. The extension o the pproch to modelling nd optimistion o process chins is lso ddressed.. Automtic input-output conigurtion nd genertion o multipurpose ANN-bsed process models The utomtic genertion o pproprite ANN-bsed process models, i.e. models, which re expected to work with the reuired ccurcy in dierent ssignments, consists o the ollowing steps: - Determining the (mximum) number o output prmeters (No) rom the ilble N prmeters which cn be estimted by using the remining Ni = N - No input prmeters within the prescribed ccurcy. - Ordering the ilble prmeters into input nd output prmeter sets hing Ni nd No elements, respectiely. - Trining the network whose input-output conigurtion hs been determined in the preceding steps. The boe steps re perormed prllel, using the speed o the lerning process s n indictor or the ppropriteness o the gien ANN rchitecture to relise the reuired mpping. In order to ccelerte the serch or the ANN conigurtion, which complies with the ccurcy reuirements with the minimum number o input prmeters, seuentil orwrd serch (SFS) lgorithm is used. The irst two steps cn be ormulted s ollows. A serch lgorithm is needed to select ll the possible outputs rom the gien set o prmeters with regrd to the ccurcy demnds. The serch spce consists o ll the conceible possibilities. Usully, there is lrge number o input-output conigurtions to select No prmeters rom N, moreoer, No is unknown, indicting tht the serch spce is uite lrge. To elute whether gien conigurtion stisies the ccurcy demnds, the pproprite lerning process hs to be lso perormed. Using serch method without heuristics would tke too long time becuse o the size o the serch spce nd o the slowness o elution. This is the reson why the deeloped serch lgorithm uses the properties o the lerning stge o the ANN model s indictors or the elution. The importnce o the right input-output conigurtion is dominnt in the cse o noninertble dependencies where the input-output ordering o the prmeters is o undmentl importnce. Experiments show tht some complicted dependencies usully need lrger number o lerning steps then simple settings. The bsic ssumption o the proposed serch lgorithm is i enough runs re initited tht the speed o the lerning process cn be used s indictor or the ppropriteness o the chosen neurl pproch to relise the reuired mpping. The ppliction o the seuentil orwrd selection (SFS) [] lgorithm ws the compromise tking the lrge serch spce nd the time intensie ANN lerning into ccount. The serch process is ccomplished s ollows. The lerning dt set is gien by the user in the orm o N dimensionl ectors. To select the irst output prmeter, N ANNs re generted, ech hing one output nd N- input prmeters. Ater generting the ANNs, lerning begins by ll ANNs, concurrently. First, ech ANN perorms M lerning steps. The ANN with the smllest estimtion error is checked, whether it hs reched the reuired estimtion ccurcy. I not, nother lerning phse is strted with M epochs. I yes, then this mens tht n output ws ound which cn be estimted with the gien ccurcy bsed on the remining input prmeters. The next step o the lgorithm is to order this rible to the output set o prmeters nd to select urther output prmeter. This selection is relised by the sme method s or the irst output. For serching the second output, N- ANNs re generted becuse one output is lredy ixed, conseuently, there re N- possibilities to dd nother output to the set o output prmeters. The remining N- prmeters re used s inputs. Ater inding the second output, two outputs re ixed nd serch strts to ind third output, etc. Obiously, or dding new output to the set o

3 output prmeters successul lerning step is reuired. Lerning is regrded successul i n ANN conigurtion cn lern the dependencies between input nd output ribles with the gien ccurcy. The lgorithm termintes i ter lrge number o lerning steps, none o the ANNs cn chiee the gien ccurcy, i.e. it does not tke the nturl ordering o the ilble prmeters into input nd output sets into ccount. During this serch lgorithm the lrgest number o outputs cn be ound, the ccurcy demnds re stisied nd the multipurpose ANN model is built up. It cn be seen tht this lgorithm hs regrd only or the gien ccurcy reuirement nd not or the gien ssignment. The pplicbility o the pproch ws tested by rtiicil dt (e.g. or hndling non-inertble dependencies), using dt deried rom nlyticl descriptions or set o engineering ssignments (dierent leels o plnning, optimistion nd control), nd by experimentl mchining.. Experimentl results To test the behiour o the deeloped lgorithm non-inertble dependencies were inestigted irst (x =x, x = x + x, x = x + x + x, sin(x )). Fourble results o these inestigtions promised rel world pplicbility, too. In the ollowing spce, results re presented with our engineering ssignments where the reuired models work on the sme prmeter set but the esible input-output conigurtions o these models re dierent.. The irst tsk is plnning. A surce hs to be mchined by turning to chiee roughness (prmeter: R [mm]) demnds o the customer. The engineer hs to determine the tool (prmeters: cutting edge ngle: χ[rd], corner rdius: r ε [mm]), the cutting prmeters (prmeters: eed: [mm/re], depth o cut: [mm], speed: [m/min]) nd predict phenomenon during cutting (prmeters: orce: F c [N], power: P[kW] nd tool lie: T[min]) conseuently model is needed where R seres s input nd other prmeters s outputs. Usully, the customer gies only n upper limit or the roughness.. The second tsk is to stisy the roughness demnds o the customer but with gien tool. In this cse the R, χ, r ε re inputs nd,,, F c, P, T re outputs.. The third tsk is to control the running cutting process with mesured monitoring prmeters such s orce nd power. Mesured lues o these prmeters cn be used s inormtion bout the current stte o the cutting process. In this cse R, χ, r ε, F c, P sere s input nd,,, T s outputs. The CNC controller hs to select the pproprite cutting prmeters to produce the reuested surce.. The ourth tsk is the sme s the third one but the CNC controller cn chnge only the nd prmeters becuse is prescribed. This cse needs model with inputs R, χ, r ε, F c, P, nd with outputs,, T. These ssignments show seerl input-output conigurtions or modelling dependencies between the dierent elements o prmeter set. The uestion rises: which model describes the cutting process in the best wy, i.e. with the highest ccurcy? The heuristic serch lgorithm cn nswer this uestion. In prcticl implementtion sensors, mchine controllers nd computers would proide prt o prmeters o n ANN opertion model. For simulting the mchining process in the inestigtions to be reported in this prt o the pper, ll inormtion were generted i theoreticl reltions, which re unctions o seerl input ribles. It should be stressed tht in prcticl implementtion these priori reltions re not necessry, the models re to be set up by using mesured lues. The lidity o the eutions is determined by the minimum nd mximum boundries o the prmeters. Four eutions (or orce, power, tool lie nd roughness) re used in this pper or the boe engineering tsks () [], F = 5 sin κ, () P =.9 T =.85 R c = ,.7.8 ( ( )).7.9 r.85.5 ε where the boundries o the eutions re s ollows (): :.L.[ mm / re], :L[ mm () κ :.L.[ rd], : 75L[ m / m r :.L.[ mm], T : 5L[min], ε conseuently, Fc : 8L[ N], P :.8L.5[ kw ], R :.5L.[ mm] With help o these strongly non-liner eutions, lues or tool lie, orce, power nd roughness cn be clculted bsed on the tool nd mchining prmeters.,,

4 To crete lerning nd testing prmeter sets rndom lues were determined in the llowed rnge o,, χ,, r ε considering lso the boundries o R, Fc, P, T while clculting their lues using the boe eutions. The dependencies between prmeters,, χ,, r ε, Fc, P, T, R were experienced s inertble in the gien prmeter rnge except the rible χ. Conseuently, to get n ccurte ANN model the rible χ hs to be lwys input. A hundred dt ectors were creted s stted boe. To test this type o problems the described input-output conigurtion nd model building pproch were repeted hundred times. The llowed erge estimtion error ws gien s ±.5%. Fiteen dierent ANN conigurtions were generted s results (Figure ). The rible χ is lwys on the input size o the ANN model s expected. (Figure : On the horizontl xis the resulted input-output conigurtions re listed represented by their output prmeters. The erticl xis shows the percentge conigurtion hs been selected in the hundred runs.) For testing estimtion cpbilities o the resulted ANN bsed models ll o the conigurtions were trined hundred times but by ech trining the relted physicl prmeters (,, χ,, r ε,) nd the strting weights were generted rndomly. The trget erge estimtion error ws ±. (±.5%). To test, nother set o hundred rndomly generted dt ectors were used nd the erge estimtion errors were clculted (Figure ). No signiicnt dierence could be ound between inputoutput conigurtions showing tht most o the dependencies mong prmeters re inertble. (Figure : The resulted input-output conigurtions represented by their output prmeters re listed on the horizontl xis.) The results indicte tht the deeloped techniue is ble to generte process models with the reuired ccurcy, moreoer, under gien circumstnces result is set o pplicble models ech gurnteeing the reuired ccurcy perormnce. As expected, the resulted input-output conigurtions cn not be used directly to the gien ssignments. The solution or this problem is presented in the next prgrph. Freuency[%] Possible outputs -- -re-p -- --P Fc-- T Output prmeters P - Figure : Resulted input-output conigurtions o the ANN models estimtion errors o dierent input-output conigurtions re- P outputs -- --P T P -- - Figure : Aerge estimtion errors o the models. Appliction o the multipurpose model or rious ssignments Usully, some prmeters re known, nd using the multipurpose model generted ccording to the preious prgrph, the tsk is to determine the other prmeters while stisying some constrints. Becuse o the generl nture o the multipurpose model, lmost in eery cse, prt o the input nd prt o the output ribles o the model re known by the user nd the unknown prt o the inputs is to be determined by tking the boe mentioned constrints into ccount. In the pper simulted nneling serch techniue is proposed or the ppliction phse o the multipurpose model. The serch process is guided by the ccurcy reuirements o the estimtion or the known output prmeters while holding the unknown input nd output prmeter(s) within its (their) rnge o ppliction boundries. The serch spce consists o unknown input prmeters. One point o the serch spce cn be

5 represented by one possible lue set o the unknown input prmeters. Ater plcing these prmeters together with the known input prmeters to the input side o the gien ANN n output ector cn be clculted (orwrd propgtion). During the serch process the unknown input prmeters re to be determined nd t the sme time three conditions re to be stisied:. Condition regrding the known output prmeters. This condition ssures tht only tht points o the serch spce cn be ccepted s result, which cn deutely estimte the known output prmeters by using orwrd clcultion. To mesure the deition between estimted nd known output prmeters n error cn be clculted (Error, on Figure ).. Condition regrding the unknown input prmeters. This condition is determined by the lidity o the ANN model. This lidity is usully speciied by the dt set used or the trining []. Boundries o the model cn be hndled by minimum nd mximum lues o the relted prmeters like in the engineering tsks presented boe. (The serch lgorithm cn tke lues or the unknown input prmeters only rom the relted llowed interls.). Condition regrding the unknown output prmeters. The third condition reltes lso to the lidity o the ANN. Vlues o the unknown input prmeters re only cceptble i the estimted lues o the unknown output prmeters re within their llowed rnge (Error, on Figure ). The serch lgorithm is terminted i ll o the three conditions boe re met. An error lue is ordered to ll isited points o the serch spce. In the deeloped lgorithm this lue is the mximum o Error nd Error presented boe. The lgorithm serches or the minimum error point. The lgorithm stops i no neighbour cn be selected nd the current error lue is below the prescribed error limit. This simulted nneling lgorithm works on the discrete points o the serch spce, thereore, the prmeters o unknown prt o the input ector consist o the discrete points o the relted interls. The distnce between two points o n interl is chosen to stisy the ccurcy reuirements o the estimtion prescribed by the user. As result, this lgorithm gies one solution or gien ssignment o the user. To look or lrger number o solutions the serch hs to be repeted. T unknown κ ( ) estimted Error ( ) unknown κ r ε known r ε T F c P Fc P R Error estimted Fc P R known R Figure : The generted ANN model nd its ppliction or the third tsk presented boe (control o the cutting process with mesured monitoring prmeters). Solution o the ssignments There re lrge number o solutions or ech o the enumerted ssignments. To represent the whole interl o solutions or ech prmeter the serch

6 lgorithm ws repeted hundred times t ech ssignment. To get simple iew bout the possible solution ield, the mximum nd minimum lues o the results were selected or ll prmeters, or ech tsk. These prmeter ields re listed in Figure. (The horizontl xis represents the number o the gien tsks.) Results in this tble show the descending interls o cceptble prmeters rom the plnning phse to the CNC control. The reuested lue o prmeter R is specil becuse the user gies only upper limit or this prmeter. In the ssignments the llowed highest lue or the roughness o the produced surce is. mm. The tool used or cutting is determined in the second tsk, lues o relted prmeters re χ=.59 rd, r ε =.79 mm. In monitoring, mesured lues o orce nd power were Fc=7N nd P=8.9kW, respectiely. In the ourth engineering tsk the prescribed speed lue ws = m/min. In eery cse the tsk o the modelling ws to stisy the roughness demnd o the user through choosing pproprite lues o relted prmeters. The diersity o solutions indictes the opportunity to incorporte optimistion into the decision mking processes bsed on the generted multipurpose models Fc P T kpp..8.. re Figure : Descending interls o llowed prmeter ields in the our engineering tsks presented beore. Optimistion o mchining processes by using the multipurpose model Optimistions cn be relised to stisy some constrins or gols where there re seerl solutions o gien ssignment. There re dierent pproches to optimise gien process or process chin []. At the Computer nd Automtion Reserch Institute block-oriented sotwre ws deeloped nmed ProcessMnger to optimise opertions nd/or production chins orm rious points o iew t the sme time. Multiple o objecties cn be hndled by the usul weighting techniue. The pplicbility o the progrm system is illustrted here through the optimistion o the plte turning ssignment. Optimistions were perormed rom the twoold iewpoints o the customer (surce roughness minimistion), nd the producer (minimistion o production time). To relise optimistions rom both o these iewpoints weighting ctors were ried to result in dierent compromises. Figure 5 shows possible compromises through lues o the relted prmeters belonging together. These results cn be lso used directly to support business decisions nd compromises.

7 Optimiztion with two stndpoints ( possible compromises ), (mx=.5 mm/re.) (mx =.89 m/sec) (mx =.99 micron) t (mx = 5. sec),8,,,t,,, <-(min) both (compromises) t(min)-> Figure 5: Prmeters resulted by the optimistion o the plte turning opertion. On the let side the iewpoint o the customer (R - min.) on the right side the iewpoint o the producer (t - min.) is stisied. Cures show possible compromises between the two iewpoints. Figure nd Figure 7 illustrte the ppliction o ProcessMnger or the threeold optimistion o the iewpoints o the customer (minimistion o the surce roughness), owner o the compny (proit/productiity mximistion) nd the employed engineer (mximistion o process stbility through the / rtio). Figure shows the building up phse o ProcessMnger, where the model o the plte turning is relised by n ANN nd the other ribles to be optimised, e.g. cutting intensity nd / or stbility, re gien by eutions. Prmeters resulted by the optimistion o the plte turning opertion re illustrted by D-plots in Figure 7. tios o the weighting ctors o the three ribles to be optimised re represented long the xes. The surces re to be used together, i.e. the moing long the plne mrked by R nd / occurs on ech o the digrms t the sme time. The corner mrked by indictes the position, where the iewpoint o the compny owner is the most importnt nd by moing long the xes R nd / represents tht the iewpoints o the customer nd the engineer become more nd more importnt with respect to. Chin building Externl Models Plte turning R PROCESSMANAGER R Intensity = Chip orm (/) Trgets o optimistion ( / ) = Externl connections Figure : Chin model or optimistion o the plte turning opertion with optimistion criteri

8 / R R. R / R / R / / / / R R Figure 7: Prmeters resulted by the threeold optimistion o the plte turning opertion 5. Modelling nd optimistion o process chins As it ws pointed out in [], it is not enough to concentrte on the inl tolernces usully deined by design. The inl tolernces re determined not only by the inishing opertions, but re the results o the initil tolernces o the workpieces nd the intermedite tolernces reched by the elements o the process chin resulting in the inished prt. The output o one opertion is the input o nother one or it is eture o the end product. To build model or production chin, models he to be ordered to eery stge o production. The seuence o production opertions cn be modelled by chin o opertions connected by their input-output prmeters []. To he process models with the reuired ccurcy is especilly importnt in the cse o process chins where the errors cn cumulte (Figure 8). (The eect o indiidul models on their output prmeter is indicted with }.) i. Input prmeter j. Output prmeter } i. Input prmeter j. Output prmeter } } i N. Input prmeter j N. Output prmeter } } Other input prmeters o the. model MODEL. Other output prmeters o the. model Other input prmeters o the. model MODEL. Other output prmeters o the. model Other input prmeters o the N. model MODEL N. Other output prmeters o the N. model Estimtion o prmeters long the production line Figure 8: Errors o prmeter estimtions long the whole production chin

9 The tolernce chnnel through which the mnucturing process is to be led is inluenced by number o prmeters: mteril properties, nominl nd ctul mchine prmeters, cutting conditions, tool stte, etc. The non-deterministic nture o mnucturing processes is the undmentl brrier tht preents us rom determining this chnnel nd mpping it to NC progrms beore mchining. Systemtic nd ccidentl non-conormities cn be enumerted tht contribute to this stochstic nture [] Production chin Chin building Trgets o optimistion Elution, optimistion Opertion Prmeter Prmeter Opertion p Prmeter Prmeter Elution Prmeter Prmeter Elution N Prmeter Prmeter Prmeter n Prmeter n Prmeter n p Prmeter n n PROGRAM - PROCESSMANAGER Model Prmeter Prmeter Prmeter o Prmeter Prmeter Externl connections Prmeter Prmeter Model Prmeter Prmeter Prmeter o n Model m Prmeter Prmeter Prmeter o Prmeter n Externl Models Prmeter n n Model m n Prmeter Prmeter Prmeter o n Figure 9: Errors o prmeter estimtions long the whole production chin The inl prt o the pper dels with the problem o modelling nd optimistion o process chins through the extension o the modelling nd serch techniues introduced or single processes. The ProcessMnger block-oriented rmework or modelling, monitoring nd optimistion o mnucturing processes nd process chins reerred boe incorportes (Figure 9): - deinition o the elements o the chin, - determintion o the process models by integrting nlyticl eutions, expert knowledge nd exmple-bsed lerning, - connecting the single models into process chin by coupling input-output model prmeters not limited to models o successie processes in the chin, - deinition o eligible interls or limits or the process prmeters nd monitoring indices, - deinition o cost unction to be optimised, etc. Conclusions The pper presented noel pproch or generting multipurpose models o mchining opertions which combines mchine lerning nd serch techniues. Simulted nneling serch ws used or inding the unknown prmeters o the multipurpose model in gien situtions. It is expected tht the deeloped ProcessMnger will be luble tool or modelling, monitoring nd optimistion o mnucturing processes nd process chins. Tking the globlistion issues nd the incresing role o irtul enterprises into ccount, the distributed ersion o the system will show up urther beneits. Acknowledgements This work ws prtilly supported by the

10 Ntionl Reserch Foundtion, Hungry, Grnt Nos. F nd T8. A prt o the work ws coered by the Nt. Comm. or Techn. De., Hungry Grnts (EU-9-B-5 nd EU-97-A- 99) promoting Hungrin reserch ctiity relted to the ESPRIT LTR Working Groups (IiMB 8 nd IMS 995). Reerences. G. Chryssolouris, M. Guillot, M. Domroese, An pproch to intelligent mchining, Proc. o the 987 Americn Control Con., Minnepolis, MN, June -, pp. 5-. (987). P.A. Deijer, J. Kittler, Pttern recognition, sttisticl pproch, Prentice-Hll Interntionl Inc., London, (98). F. Krupp Gmbh, Widi-Richtwerte ür ds drehen on Eisenwerkstoen, Fried. Krupp Gmbh, Essen, (985). S. Mrkos, Zs.J. Vihros, L. Monostori, Qulityoriented, comprehensie modelling o mchining processes, in: Proceedings. o th ISMQC IMEKO Symposium on Metrology or Qulity Control in Production, September 8-, Vienn, Austri, pp (998) 5. M.E. Merchnt, An interpretie look t th century reserch on modelling o mchining, Inugurl Address, Proc. o the CIRP Interntionl Workshop on Modelling o Mchining Opertions, Atlnt, Georgi, USA, My 9, pp. 7-. (998). L. Monostori, A step towrds intelligent mnucturing: Modelling nd monitoring o mnucturing processes through rtiicil neurl networks, CIRP Annls,, No., pp (99) 7. L. Monostori, Hybrid AI pproches or superision nd control o mnucturing processes, Key-note pper, Proc. o the AC'95, IV Int. Con. on Monitoring nd Automtion Superision in Mnucturing, Miedzeszyn, Polnd, Aug. 8-9, pp (995) 8. L. Monostori, D. Brschdor, Artiicil neurl networks in intelligent mnucturing, Robotics nd Computer-Integrted Mnucturing, Vol. 9, No., Pergmon Press, pp. -7. (99) 9. L. Monostori, A. Márkus, H. Vn Brussel, E. Westkämper, Mchine lerning pproches to mnucturing, CIRP Annls, Vol. 5, No., pp (99). S.S. ngwl, D.A. Dorneld, Lerning nd optimistion o mchining opertions using computing bilities o neurl networks, IEEE Trns. on SMC, Vol. 9, No., Mrch/April, pp (989). T. Tóth, F. Erdélyi, The inner structure o computer ided process plnning hing regrd to concurrent engineering, nd Interntionl Workshop on Lerning in Intelligent Mnucturing Systems, April -, Budpest, Hungry, pp. -7. (995). H.K. Tönsho, J.P. Wulsberg, H.J.J. Kls, W. König, C.A. n Lutterelt, Deelopments nd trends in monitoring nd control o mchining processes, CIRP Annls, Vol. 7, No., pp. -. (988). C.A. Vn Lutterelt, T.H.C. Childs, I.S. Jwhir, F. Klocke, P.K. Venuinod, Present sitution nd uture trends in modelling o mchining opertions, CIRP Annls, Vol. 7, No., (998). Zs. J. Vihros, L. Monostori, S. Mrkos, Selection o input nd output ribles o ANN-bsed modelling o cutting processes, Proc o the X. Interntionl Workshop on Superising nd Dignostics o Mchining Systems, Innotie nd Integrted Mnucturing, Krpcz, Polnd, Mrch -, pp. -. (999) 5. G. Wrnecke, R. Kluge, Control o tolernces in turning by predictie control with neurl networks, Proceedings o The Second World Congress on Intelligent Mnucturing Processes & Systems, June -, 997, Budpest, pp. -7. nd Journl o Intelligent Mnucturing, Vol. 9, No., August 998, Specil Issue on Sot Computing Approches to Mnucturing, Chpmn & Hll, pp E. Westkämper, Superision o ulity in process chins by mens o lerning process models, Proc. o the Second Int. Workshop on Lerning in IMSs, Budpest, Hungry, April -, pp (995) 7. S. Yerrmreddy, S.C.-Y. Lu, K.F. Arnold, Deeloping empiricl models rom obsertionl dt using rtiicil neurl networks, Journl o Intelligent Mnucturing, Vol., pp. -. (99)