science.pdf

Journal of Materials Processing Technology xxx (2005) xxxxxxAbstractoperations.metal-cuttingforfeed-rateconductedK1.theofaremeragearederEvanthemachiningnecessarysatisfytioperatingAtainconditions.de0924-0136/$doi:10.1016/j.jmatprotec.2005.02.143Fuzzy control strategy for an adaptive force control in end-millingU. Zuperl, F. Cus, M. MilfelnerFaculty of Mechanical Engineering, University of Maribor, Smetanova 17, 2000 Maribor, SloveniaThis paper discusses the application of fuzzy adaptive control strategy to the problem of cutting force control in high speed end-millingThe research is concerned with integrating adaptive control with a standard computer numerical controller (CNC) for optimising aprocess. It is designed to adaptively maximise the feed-rate subject to allowable cutting force on the tool, which is very beneficiala time consuming complex shape machining. The purpose is to present a reliable, robust neural controller aimed at adaptively adjustingto prevent excessive tool wear, tool breakage and maintain a high chip removal rate. Numerous simulations and experiments areto confirm the efficiency of this architecture.2005 Elsevier B.V. All rights reserved.eywords: End-milling; Adaptive force control; FuzzyIntroductionA remaining drawback of modern CNC systems is thatmachining parameters, such as feed-rate, speed and depthcut, are programmed off-line. The machining parameterssimulations with the fuzzy control strategy are carried out.The results demonstrate the ability of the proposed system toeffectively regulate peak forces for cutting conditions com-monly encountered in end-milling operations.Force control algorithms have been developed and eval-usually selected before machining according to program-s experience and machining handbooks. To prevent dam-and to avoid machining failure the operating conditionsusually set extremely conservative.As a result, many CNC systems are inefficient and run un-the operating conditions that are far from optimal criteria.en if the machining parameters are optimised off-line byoptimisation algorithm 5 they cannot be adjusted duringmachining process.To ensure the quality of machining products, to reduce thecosts and increase the machining efficiency, it isto adjust the machining parameters in real-time, tothe optimal machining criteria. For this reason, adap-ve control (AC), which provides on-line adjustment of theconditions, is being studied with interest 3.InourC system, the feed-rate is adjusted on-line in order to main-a constant cutting force in spite of variations in cuttingIn this paper, a simple fuzzy control strategy isveloped in the intelligent system and some experimentalCorresponding author. Tel.: +386 2 220 7623; fax: +386 2 220 7990.E-mail address: uros.zuperluni-mb.si (U. Zuperl).uatedisnallyantrollerditions.trollerandatedallthetems,bysentedtesystems3controlhasmotion.for see front matter 2005 Elsevier B.V. All rights reserved.by numerous researchers. Among the most commonthe fixed gain proportional integral (PI) controller origi-proposed for milling by 4. Kim et al. 4 proposedadjustable gain PI controller where the gain of the con-is adjusted in response to variations in cutting con-The purely adaptive model reference adaptive con-(MRAC) approach was originally investigated by CusBalic 2. These controllers were simulated and evalu-and physically implemented by 1. Both studies foundthree-parameter adaptive controller to perform better thanfixed gain PI controller. As regards fuzzy control sys-an introductory survey of pioneering activities is givenHuang and Lin 3, and a more systematic view is pre-by in 4. Comparisons of fuzzy with proportional in-gral derivative (PID) control and stability analysis of fuzzyand supervisory fuzzy control are addressed in Ref.Much work has been done on the adaptive cutting forcefor milling 2. However, most of the previous worksimplified the problem of milling into one-dimensionalIn this contribution, we will consider force controlthree-dimensional milling.2 Processingscribesthesimulation/eposedimentalresearch.2.fuzzyseteThewhichplementcontrolmoreTherateasthecomparedcontrolFuzzyratecuttingcreasesrates,productionarebreakage.callytheforbelo.signingonactualcentagemisationalcorrectcontrolleraplepro2.1.aaboutinputoperatorthroughU. Zuperl et al. / Journal of MaterialsThe paper is organised as follows. Section 2 briefly de-the overall force control strategy. Section 3 coversCNC machining process model. Section 5 describes thexperiments and implementation method of pro-control scheme. Finally, Sections 6 and 7 present exper-results, conclusions, and recommendations for futureAdaptivefuzzycontrollerstructureA new on-line control scheme which is called adaptivecontrol (AFC) (Fig. 1) is developed by using the fuzzytheory. The basic idea of this approach is to incorporate thexperience of a human operator in design of the controller.control strategies are formulated as a number of rulesare simple to carry out manually but difficult to im-by using conventional algorithm. Based on this newstrategy, very complicated process can be controlledeasily and accurately compared to standard approaches.objective of fuzzy control is keeping the metal removal(MRR) as high as possible and maintaining cutting forceclose as possible to a given reference value. Furthermore,amount of computation task and time can be reduced asto classical or modern control theory. Schematicrules are constructed by using real experimental data.adaptive control ensures continuous optimising feedcontrol that is automatically adjusted to each particularsituation. When spindle loads are low, the system in-cutting feeds above and beyond pre-programmed feedresulting in considerable reductions in cycle times andcosts. When spindle loads are high the feed rateslowered, safeguarding machine tools from damage fromWhen system detects extreme forces, it automati-stops the machine to protect the cutting tool. It reducesneed for constant operator supervision. Sequence of stepson-line optimisation of the milling process are presentedw.namicstheasvofcuttinglated,Delta1forceFig. 1. Comparison of actualTechnology xxx (2005) xxxxxxThe pre-programmed feed rates are sent to CNC controllerof the milling machine.The measured cutting forces are sent to the fuzzy con-troller.Fuzzy controller uses the entered rules to find (adjust) theoptimal feed-rates and sends it back to the machine.Steps 1 and 3 are repeated until termination of machining.The adaptive force controller adjusts the feed-rate by as-a feed-rate override percentage to the CNC controllera four-axis Heller, based on a measured peak force. Thefeed-rate is the product of the feed-rate override per-and the programmed feed-rate. If the feed-rate opti-models were perfect, the optimised feed-rate wouldways be equal to the reference peak force. In this case theoverride percentage would be 100%. In order for theto regulate peak force, force information must bevailable to the control algorithm at every controller sam-time. A data acquisition software (Labview) is used tovide this information.Structure of a fuzzy controllerIn fuzzy process control, expertise is encapsulated intosystem in terms of linguistic descriptions of knowledgehuman operating criteria, and knowledge about theoutput relationships. The algorithm is based on thes knowledge, but it also includes control theory,the error derivative, taking into consideration the dy-of the process. Thus, the controller has as its inputs,cutting force error Delta1F and its first difference Delta12F, andoutputs, the variation in feed rate Delta1f. The fuzzy controlariables fuzzification (see Fig. 2) as well as the creationthe rules base were taken from the expert operator. Theforce error and first difference of the error are calcu-at each sampling instant k, as: Delta1F(k)=FrefF(k) and2F(k)=Delta1F(k)Delta1F(k1), where F is measured cuttingand Frefis force set point.and model feed-rate.3.etalandforcesscribedmachinefeedingfitquencefromformcommandedtingmodel.mentalfeed-rateU. Zuperl et al. / Journal of Materials ProcessingFig. 2. Structure of a fuzzyCNCmachiningprocessmodelA CNC machining process model simulator is used tovaluate the controller design before conducting experimen-tests. The process model consists of a neural force modelfeed drive model. The neural model estimates cuttingbased on cutting conditions and cut geometry as de-by Zuperl 1. The feed drive model simulates theresponse to changes in commanded feed-rate. Thedrive model was determined experimentally by examin-step changes in the commanded velocity. The best modelwas found to be a second-order system with a natural fre-y of 3 Hz and a settling time of 0.4 s. Comparison ofxperimental and simulation results of a velocity step change7 to 22 mm/s is shown on Fig. 3.The feed drive and neural force model are combined tothe CNC machining process model. Model input is thefeed-rate and the output is the X, Y resultant cut-force. The cut geometry is defined in the neural forceThe simulator is verified by comparison of experi-and model simulation results. A variety of cuts withchanges were made for validation.changeFig.resultsTechnology xxx (2005) xxxxxx 3controller.The experimental and simulation resultant force for a stepin feed-rate from 0.05 to 2 mm/tooth is presented in4. The experimental results correlate well with modelin terms of average and peak force. The experimentalFig. 3. Comparison of actual and model federate.4resultsandthe3.1.dardlarimentsforceusedfederatedialforcesaryU. Zuperl et al. / Journal of Materials ProcessingFig. 4. Structure of a fuzzycorrelate well with model results in terms of averagepeak force.The obvious discrepancy may be due to inaccuracies inneural model, and unmodeled system dynamics.Cutting force modelingTo realise the on-line modelling of cutting forces, a stan-BP neural network (NN) is proposed based on the popu-back propagation leering rule. During preliminary exper-it proved to be sufficiently capable of extracting themodel directly from experimental machining data. It isto simulate the cutting process.The NN for modelling needs four input neurons for milling(f), cutting speed (vc) axial depth of cut (AD) and ra-depth of cut (RD). The output from the NN are cuttingcomponents, therefore two output neurons are neces-. The detailed topology of the used NN with optimal train-ingalso73.2.modelingferentberefnetwysed.difandtheinputconclusionsTechnology xxx (2005) xxxxxxcontroller.parameters and mathematical principle of the neuron isshown in Fig. 5. Best NN configuration contains 5, 3 andhidden neurons in hidden layers.Topology of neural network and its adaptation toproblemThe effect of topology is also studied by considering dif-cases. The topologies are varied by varying the num-of neurons in hidden layers. To evaluate the individualfects of training parameters on the performance of neuralork 40 different networks were trained, tested and anal-The network performances were evaluated using fourferent criteria 5: ETstMax, ETst, ETrn, and ETrnMaxthe number of training cycles. The number of neurons ininput and output layers are determined by the number ofand output parameters. From the results the followingcan be drawn.Processing4.equipmentsystemandwere9255)table.beelectricelectrictransmittedthemulti-channelamplifierquiredofpendingterfU. Zuperl et al. / Journal of MaterialsFig. 5. Structure of a fuzzyLearning rates below 0.3 give acceptable prediction errorswhile learning rates must be between 0.01 and 0.2 to min-imise the number of training cycles.To minimise the estimation errors, momentum rates be-tween 0.001 and 0.005 are good. However, the momentumrate should not exceed 0.004 if the number of training cy-cles is also to be minimised.The optimum number of hidden layer nodes is 3 or 6. Net-works with between 2 and 12 hidden layer nodes, otherthan 3 or 6, also performed fairly well but resulted in highertraining cycles.Networks that employ the sine function require the lowestnumber of training cycles followed by the ArcTangent,while those that employ the hyperbolic tangent require thehighest number of training cycles.DataacquisitionsystemandexperimentalThe data acquisition equipment used in this acquisitionconsists of dynamometer, fixture module, hardwaresoftware module as shown in Fig. 1. The cutting forcesmeasured with a piezoelectric dynamometer (Kistlermounted between the workpiece and the machiningWhen the tool is cutting the workpiece, the force willapplied to the dynamometer through the tool. The piezo-quartz in the dynamometer will be strained and ancharge will be generated. The electric charge is thento the multi-channel charge amplifier throughconnecting cable. The charge is then amplified using thecharge amplifier. In the multi-channel charge, different parameters can be adjusted so that the re-resolution can be achieved. Essentially, at the outputthe amplifier, the voltage will correspond to the force de-on the parameters set in the charge amplifier. The in-ace hardware module consists of a connecting plan block,analogueinterfloguetheThedirectionstheneouslyforceabletinglectedwithterialincoolanttrolandcationchinefeed-rateableof5.fuzzyulationthemilling45(cuttingselectedTechnology xxx (2005) xxxxxx 5controller.signal conditioning modules and a 16 channel A/Dace board (PC-MIO-16E-4). In the A/D board, the ana-signal will be transformed into a digital signal so thatLabVIEW software is able to read and receive the data.voltages will then be converted into forces in X, Y and Zusing the LabVIEW program. With this program,three axis force components can be obtained simulta-, and can be displayed on the screen for analysingchanges. The ball-end-milling cutter with interchange-cutting inserts of type R216-16B20-040 with two cut-edges, of 16 mm diameter and 10helix angle was se-for machining. The cutting inserts R216-1603 M-M12rake angle were selected. The cutting insert ma-is P30-50 coated with TiC/TiN, designated GC 4040P10-P20 coated with TiC/TiN, designated GC 1025. TheRENUS FFM was used for cooling. The fuzzy con-is operated by the intelligent controller module (Labview)the modified feed-rates are send to the CNC. Communi-between the force control software and the NC ma-controller is enabled through memory sharing. Theoverride percentage variable DNCFRO is avail-to the force control software for assignment at a rate1 kHz.SimulationsandfuzzycontrolmillingexperimentTo examine the stability and robustness of the adaptivecontrol strategy, the system is first examined by sim-using Simulink and Labview fuzzy Toolset. Thensystem is verified by various experiments on a CNCmachine (type HELLER BEA1) for Ck 45 and Ck(XM) steel workpiece with variation of cutting depthFig. 6).The ball-end-milling cutter (R216-16B20-040) with twoedges, of 16 mm diameter and 10helix angle wasfor experiments. Cutting conditions are: milling6 U. Zuperl et al. / Journal of Materials Processing Technology xxx (2005) xxxxxxwidthveTo use the fuzzy control structure on Fig. 1 and to opti-mise the feed-rate, the desired cutting force is Fref = 280 N,pre-programmed feed is 0.08 mm/teeth and its allowable ad-justing rate is 0150%.Fig. 7 is the response of the cutting force and the feed-ratewhen the cutting depth is changed. It shows the experimentalresultcuttingisFig.adaptiFig. 6. Workpiece profile.RD= 3 mm, milling depth AD= 2 mm and cutting speedc= 80 m/min.The parameters for fuzzy control are the same as for thexperiments for the traditional system performance.twaxial2enceoftionsrunditions.axial2000depthmainsperformance.7. Experimental results with variable cutting depth. Response of MRR, resultingve fuzzy control.where the feed-rate is adjusted on-line to maintain theforce at the maximum desired value.Simulated control response to a step change in axial depthpresented in Fig. 8. The simulation represents a 16 mm,o-flute cutter, at 2000 rpm, encountering a step change indepth from 3 to 4.2 mm. The step change occurs ats and the controller returns the peak forces to the refer-peak force within 0.5 s. In this research the stabilityfuzzy controller is evaluated by simulation. Test simula-with small and large step changes in process gain areto ensure system stability over a range of cutting con-Small process gain changes are simulated with andepth change from 3 to 4.2 mm at a spindle speed ofrpm. Large gain changes are simulated with an axialchange from 3 to 6 mm at 2000 rpm. The system re-stable in all simulation tests, with little degradation incutting force, feed-rate. (a) Conventional milling and (b) milling withProcessing6.tionalinpiecemuchcontrolfeed-ratetraditionalpointisefproingsystemadjustment;tryandThetheitytrollers.tantvU. Zuperl et al. / Journal of MaterialsFig. 8. Simulated fuzzy control response to a step change in axial depth.ResultsanddiscussionIn the first experiment using constant feed rates (conven-cutting, Fig. 7a) the MRR reaches its proper value onlythe last step.However, in second test (Fig. 7b), machining the samebut using fuzzy control, the average MRR achieved ismore close to the proper MRR.Comparing Fig. 7a and b, the cutting force for the neuralmilling system is maintained at about 240 N, and theof the adaptive milling system is close to that of theCNC milling system from point C to point D. FromA to point C the feed-rate of the adaptive milling systemhigher than for the classical CNC system, so the millingficiency of the adaptive milling system is improved.The experimental results show that the MRR can be im-ved by up to 27%. As compared to most of the exist-end-milling control systems, the proposed fuzzy controlhas the following advantages 3: 1. multi-parameter2. insensitive to changes in workpiece geome-, cutter geometry, and workpiece material; 3. cost-efficienteasy to implement; 4. mathematically modeling-free.simulation results show that the milling process withdesigned fuzzy controller