Estimation Of Cutter Wear Of A Milling Machine Using A .

2y ago
16 Views
2 Downloads
708.26 KB
8 Pages
Last View : 17d ago
Last Download : 2m ago
Upload by : Tia Newell
Transcription

Estimation of Cutter Wear of a Milling Machine Using aSupport Vector Regression MethodYoung Do KOO1, Man Gyun NA*1, and Jung-Taek KIM21. Department of Nuclear Engineering, Chosun University, 309 Pilmun-daero Dong-gu, Gwangju 61452, Republic of Korea(magyna@chosun.ac.kr)2. Nuclear ICT Research Division, Korea Atomic Energy Research Institute, 989-111 Daedeok-daero Yuseong-gu, Daejeon 34039,Republic of Korea (jtkim@kaeri.re.kr)Abstract: The integrity of various machineries and equipment, which are used in several industrial fields, canbe considered as a factor determining safety and efficiency such as productivity and economy. In the sense,prognostics and health management (PHM) techniques are used to estimate the machine condition and life foreffective maintenance and risk minimization. It is known that PHM has already been applied in a wide rangeof industrial fields such as automotive industry, aeronautics industry, several energy industries, and military.PHM denotes diagnostics that monitors conditions of the mechanical system or device and detects its failuresymptoms, prognostics of remaining useful life (RUL), and effective health management using sensors.However, in a case of nuclear power plants (NPPs) that consist of large architecture and complex internalstructures, it is known that the actual failure data of its system and device are difficult to be obtained, comparedto other industrial fields. Thus, since the cutter in a milling machine can be considered as a rotor such as a pumpand a turbine in a NPP, the cutter wear data from PHM 2010 society conference data challenge were used in aneffort to study PHM technologies for NPPs.In this study, support vector regression (SVR) as a data-driven approach was used to estimate a total of 6 cutterwear of a milling machine. The basic concept of SVR is to map the input data into a high-dimensional spaceby nonlinear mapping to solve a linear regression problem in this space. Among the cutters, 3 actual cutter weardata were compared with estimated values obtained from the SVR method using the signals from dynamometer,accelerometers, and acoustic emission sensors built in the experimental device. Consequently, the proposedmethod can accurately estimate the degree of cutter wear in the experimental device and it is expected that SVRhas a capability to estimate wear of the rotating machines in NPPs in the future.Keyword: Prognostics and Health Management (PHM), Remaining Useful Life (RUL), Sensor,Support Vector Regression (SVR)1 IntroductionNuclear power plants (NPPs) are comprised of alot of various equipment and components. Amongthe total number of 24 NPPs operating in Republicof Korea, pressurized water reactors (PWRs) suchas optimized power reactor (OPR) 1000 andadvanced power reactor (APR) 1400 are mainreactor types in Korea.Generally, these PWRs are classified into aprimary system exposed to radioactivity and asecondary system without radioactivity. Theprimary system in NPPs has a role to transfer heatenergy of the reactor and this heat energy from theprimary system is changed into potential energy assteam through steam generator (SG) in thesecondary system.There are facilities such as reactor vessel (RV),reactor coolant pump (RCP), pressurizer (PZR),SG primary side, and so on in reactor coolantsystem (RCS), which is another name of theprimary system. They are under high pressure andtemperature conditions to prevent H2O coolantwith radionuclides from vaporization. Thesecondary system consists of turbine, condenser,pump, the rest side of SG, and so on, whichoperate under less harsher conditions.In an effort to keep the integrity of these variousNPP components, a study on application ofprognostics and health management (PHM)technology to NPPs was carried out[1] since theequipment have an important role on the safetyand efficiency of NPPs.PHM can be defined as accurately monitoringmechanical system, device, and facility, detectingthe fault symptom, and predicting remaininguseful life (RUL) using the features extracted fromthe data or sensor signals (refer to Fig. 1[1]). ThatISOFIC 2017, Gyeongju, Korea, November 26-30, 20171

is, PHM system contains functions of conditionmonitoring, state assessment, fault detection,prognostics, operational decision support, and soon[2].regression analyses and showed goodperformance.However, a lot of data were needed to guaranteethe optimal performance using machine learningtechnique, and accordingly the problems oncalculation time and overfitting were raised.Therefore, subtractive clustering (SC) technique[3]and genetic algorithm[4][5] , which are for dataselection and optimization, were applied to SVRto confirm the performance through propergeneralization. For these reasons, the accuracy ofthe proposed SVR model increased and the preciseestimation performance of the cutter wear wasshown.2 Machine learning algorithmFig.1 PHM systems.PHM technologies have already been applied invarious industrial fields such as defense, aircraft,automobile, and wind turbines. Accordingly, inthe nuclear power generation field, it is consideredthat a PHM technology become a necessary needto enhance the safety and economy of NPPsmoving forward with both long term operation(LTO) and new builds[1].Unfortunately, however, it is known that there arerarely the real fault data, and thus it is difficult tobe obtained. Therefore, three types of sensorsignals such as force, vibration, and acousticemission (AE) data were used to estimate thecutter wear of a milling machine in this study.Although the cutter of a milling machine is not oneof the components in NPPs, it was selected forestimation target in this study since it can beconsidered a rotating machine such as pump andturbine in NPPs.Moreover, it can be regarded as a case study sincethe performance of the artificial intelligencetechnology used in this study was able to bechecked due to the fact that cutter wear wasestimated using the real sensor signals andcompared with the actual cutter wear data.Support vector regression (SVR), which is one ofthe machine learning method of artificialintelligence, was used in this study. It is knownthat this technique was applied to various2SVR used in this study is another name of supportvector machine (SVM) in regression analysis[6].SVM is a fundamentally machine learning methodand data-driven approach. This method seeks tobest fit the training data in establishing themapping function, while maintaining the ability togeneralize to unseen data.However, since the efficacy of regressionalgorithm is determined by data type, quality, andquantity, a lot of data have to be used to gain goodperformance[7]. Additionally, the performancediffers from the assumptions of errorminimization or training method inherent in thealgorithm[7].2.1 Support vector machines in regressionSVM as a model generally used in eventclassification or regression problem is analgorithm with a neural network structure basedon statistical learning theory. Current embodimentof these SVM was proposed by C. Cortes and V.Vapnik[6].It is noted that SVM and artificial neural network(ANN) techniques have similar structure.However, they have differences on trainingmethod or risk minimization[8]. SVM uses astructural risk minimization (SRM) principle(refer to Fig. 2) to minimize the upper bound onthe expected risk[8].In other words, it can be possible to establish theoptimized SVR algorithm by a SRM principleISOFIC 2017, Gyeongju, Korea, November 26-30, 2017

3finding the minimum of the bound on risk R(f*)defined as the sum of empirical risk and theconfidence interval. This SRM is depicted asfollows[9]:underfittingError orRiskThese weighting coefficient and bias arecomputed using the following regularized riskfunction with the ε-insensitive loss function[11]. Toacquire them, the following regularized risk functionhas to be minimized.overfittingN1R(W ) W T W C yk f ( x) 2k 1R(f*)Bound on risk orGeneralization errorwhereConfidence orEstimation errorh*hn Structureindex, hS*S1if yk f ( x) otherwise0 yk f ( x ) yk f ( x) Empirical risk orTraining errorh1(2)SnFig.2 SRM principle used in SVM.With the introduction of Vapnik’s ε-insensitive lossfunction[8][9][10], SVM has been finally extended tobe used in regression analysis. This can be appliedto a time series forecast and a nonlinear regression.To be specific, the basic idea of SVR is to map theinput data into a kernel-induced higher featurespace, and then perform linear regressionanalysis[12]. That is, nonlinear regression in theinput space can become linear regression in afeature space.In this study, the SVR model is established using(3)The term of the ε-insensitive loss function in Eq. (2)is determined according to the condition such asEq. (3). As shown in Fig. 3, the insensitiveness εhas a role to decide the ε-tube size and to stabilizethe estimation by controlling the number ofsupport vectors[11]. In addition, the parameter C asregularization parameter in Eq. (2) decides thetrade-off between the weight vector norm and theapproximation error[11]. In the SVR model, theparameters ε and C are design parameters and theyare related to generalization and overfitting.yk f ( x) N training data indicated as T ( xk , yk ) kN 1 inwhich xk is a sample data vector and yk denotes theactual output value, from which it learns arelationship between input and output values. TheSVR function is expressed as follows:Ryˆ f ( x) wk k ( x) b W T Φ( x) b(1)k 1where k ( x) is termed the feature which is thefunction nonlinearly from the input space x,W w1 w2wN T,Φ 1 2 N T.Parameters W and b are weighting coefficient andbias, respectively. y f ( x)Fig.3 A linear ε-insensitive loss function applied to SVR.The aforementioned generalized risk function isaltered into a constrained risk function with theslack variables as follows:N1R(W , , * ) W T W C k k* 2k 1(4)subject to the constraintsISOFIC 2017, Gyeongju, Korea, November 26-30, 20173

yk W T Φ ( x) b k , k 1, 2, T* W Φ ( x) b yk k , k 1, 2,* k , k 0,k 1, 2, ,N,N,N(5)Likewise, Eq. (4) is used to compute W and b, andthis can be solved using the Lagrange multipliermethod and a standard quadratic programmingtechnique.The slack variables shown in Fig. 4 are locatedoutside the ε-tube size and indicate the upper andlower constraints. These non-zero values have tobe minimized.yThe several Lagrange multipliers k k* havenon-negative and non-zero values and thecorresponding training data are regarded assupport vectors (SVs) lying on or outside ε-bound.2.2 Optimization of SVRTo optimize the SVR model performing linearregression in a kernel-induced space, the geneticalgorithm[4][5] was used in this study. The designparameters such as ε, C, and σ were optimizedusing the genetic algorithm as a techniqueimitating an evolutionary process of livingorganisms by the natural evolution mechanismssuch as selection, crossover, and mutation (refer toFig. 5).yi observed point iregression functionyˆ f ( x)StartGenerate initial chromosomes *jyjobserved pointEvaluate chromosomesxIs the maximum generationapproached?Fig.4 An illustration of the ε-insensitive loss function withBeyond linear regression, the SVR model can beapplied to nonlinear regression analysis using thekernel. In this study, the radial basis function isused as the kernel function and defined as follows:(6)The parameter σ in Eq. (6) indicates the sharpnessof the radial basis kernel function and it is one ofthe design parameters as the parameters ε and C,affecting on SVR’s performance.Finally, the SVR function using the kernelfunction becomes as follows:yˆ f ( x) k k* K xk , x bNGenetic operation such asselection, crossover, and mutationFig.5 Optimization procedure using the genetic algorithm.The genetic algorithm needs a fitness functionused to minimize the root mean square error(RMSE) and maximum error by assigning a scoreto each chromosome in the correspondingpopulation and evaluating how suitable achromosome is. The fitness function is used as Eq.(8).F exp( λ1 E1 λ 2 E2 )(8)where E1 and E2 are RMSE and maximum errorfor the development data set, respectively. λ1 andλ2 are weights for each error.(7)k 1E1 4StopNoslack variables. ( x x)T ( x xk ) K xk , x exp k 2 2 Yes1N devN dev yk 1ISOFIC 2017, Gyeongju, Korea, November 26-30, 2017k yˆ k 2(9)

5E2 max yk yˆ k , k 1, 2,2k, N dev(10)where Ndev is the number of development datapoints, and yk and yˆ k are a target value and anestimated value using the SVR model.In this study, the development data were used todevelop the SVR model for estimation of cutterwear. All of the data for the SVR model as well asthe development data are described in Section 3.3 Data application to SVR modelThe machineries and facilities of NPPs areconsidered as important factors determining thesafety and economy. Since the SVR model isknown for a machine learning technique showinga good performance in various regression analysis,it can be used to estimate the state of NPPcomponents.However, this method is not easy to be applied toa nuclear power generation field due to a lack ofactual fault data and a difficulty on acquiring thedata. Therefore, the data from “PHM 2010 societyconference data challenge[13]” are applied to theSVR model to estimate the wear of three cuttersamong the six cutters with 3-flute of a computernumerical control (CNC) milling machine. ACNC milling machine is comprised of a cutter,workpiece, accelerometer sensor, dynamometersensor, and AE sensor as shown in Fig. 6.Fig.6 High speed CNC milling machine (Röders TechRFM760).Although the cutter of a milling machine as theestimation target is not directly related to thecomponents of NPPs, the effort for diagnosis andestimation of the state for equipment using thesensor signal data can be considered as improvingthe safety of NPPs. In addition, the SVR model isverified since the estimated cutter wear data canbe compared with the actual wear data.For these reasons, the cutter wear data from PHM2010 data challenge were applied to the SVRmodel, and the detailed description on the data andapplication result are stated as follows.3.1 Composition of cutter wear dataThe signals applied in the SVR model of this studywere obtained from the built-in sensors ofaccelerometer, dynamometer, and AE in a millingmachine testbed.A total of 315 data files comprised of these types ofreal-time sensor signals were used to estimate thewear of three cutters. It is noted that the number ofdata files mean the amount of wear after each cut forsix cutters. That is, each cutter made 315 cuts.Table 1 Sensor signals of a CNC milling machineNo. of channelsSignals1Force in X2Force in Y3Force in Z4Vibration in X5Vibration in Y6Vibration in Z7Acoustic emission (AE)The built-in sensor signals of a milling machine areconcretely divided into seven channels such as forceand vibration in X, Y, and Z directions, and AEsignals as indicated in Table 1. There are more than200,000 measured signals in every data file for eachcutter.Additionally, the actual wear data are given for threecutters (first, fourth, and sixth cutters) among theevery cutter, and accordingly these cutters was ableto be compared with the estimated cutter wear.However, only the sensor signals on the rest ofcutters (second, third, and fifth cutters) wereprovided for competition participants to comparewith their estimation results.3.2 Data selection scheme for the SVR modelAmong more than 200,000 points in the data, RMS,standard deviation (STD), and peak values of theISOFIC 2017, Gyeongju, Korea, November 26-30, 20175

points were used. This is to make the model track thetrend of the cutter wear well.That is, since one of the main factors to show theoptimal performance on fitting is the data, thesubtractive clustering (SC)[3] technique was used toeffectively train the SVR model by selecting the data,which are considered informative, as a cluster center.data points positioned near the pre-selected clustercenter considerably decreased. This is to unlikely tomake the data points near the pre-selected clustercenter a next cluster center.To find the data point with the highest revisedpotential as a next cluster center, the potentials ofevery data point are modified by the followingfunction:15Pi 1 (k ) Pi (k ) Pi*e 4 xk xi*2rβ2, k 1, 2,,N(12)12where xi* indicates the point for the i-th clustercenter and Pi* is the corresponding potential value.Creating the cluster centers was repeated until thenumber of them is equal to the number of each dataset, used to develop the estimation model of SVR.x296ClustersCluster centers300369123.3 Estimation result15x1Fig.7 Graphical description of SC technique.As expressed in Fig. 7, by calculating the potentialof data points as a determining factor of instructivedata, SC technique selects a data with the highestpotential as a cluster center.The SC technique regards every data point as acluster center and the amount potential of input datais defined as a function of Euclidean distances to allother data points as follows:NP1 (k ) e 4 xk xi2rα2, k 1, 2,,NTo verify the proposed model to estimate the cutterwear of a milling machine using the chosen data, thedata in this study were separated into the trainingdata, the validation data, and the test data. This is aneffort to prevent the SVR model from overfitting.The training data set and the validation data set,which is to optimize the SVR model, were directlyrelated to the development of the model. The testdata set had no effect on training and was used toindependently provide a measure of performanceafter training. Additionally, the development data setconsisted of the training and verification data asstated above.(11)i 1Table 2 Estimation performance of cutter wear using SVRwhere rα indicates a radius that defines the proximitybetween the points. This radius has a considerableinfluence on the input data potentials. Thus, after thepotentials of all points were calculated, the data pointwith the highest potential was selected as a firstcluster center. Generally, the potential of a data pointis high in case that there are many adjacent datapoints.Next cluster center was determined by defining thepotential of a data point as well. However,considerable potentials were subtracted for each datapoint as a function of its distance from the preselected cluster center. Eventually, the potential for6Data typeRMSE (%)No. of t3.59E-0415Development3.16E-05833 94Table 2 shows the RMSE of the cutter wearestimation and the number of the used data assignedto the SVR model. The estimation performance ofthe cutter wear of a milling machine using the SVRmodel can be considered outstanding.ISOFIC 2017, Gyeongju, Korea, November 26-30, 2017

7number 1, number 4, and number 6 cutters,respectively. A red line with ‘X’ symbol, which is anestimation line of a cutter wear, accurately catch upwith a trend for an actual cutter wear of a black line.In addition, the actual cutter wear data as the targetvalues were given for 3 flutes. Among them, the targetwear was the wear data of the most damaged flute foreach cut.0.18Estimated wearActual wear0.16Cutter wear (mm)0.140.120.100.080.064 Conclusions0.0404080120160200240280320Cutting numberFig.8 Estimation performance for cutter 1 using SVR.0.22Estimated wearActual wear0.20Cutter wear 0200240280320Cutting numberFig.9 Estimation performance for cutter 4 using SVR.0.24Estimated wearActual wear0.22Cutter wear (mm)0.200.180.160.14In an effort to apply a PHM technology to NPPindustries, the study on estimation of the state ofthe equipment using the SVR technique, which isgenerally applied to various regression analyses,by sensor signals was carried out.However, since it is known that the actual faultdata for NPPs and are hard to be obtained, the dataprovided from “2010 PHM data challenge” wereused to the SVR model with the genetic algorithm.Although a cutter wear of a milling machine as theestimation target is not directly related to the NPPs,it can be considered as a rotating machine such asa turbine and a pump in a NPP.In addition, the SVR model can estimate a cutterwear and be verified since the data from sensorsignals such as force, vibration, and AE, and theactual cutter wear data are given. RMSEs on everydata set of the SVR model, have very low values.Consequently, the proposed SVR model of amachine learning technique can be a suitablemodel as an on-line monitoring (OLM) and aPHM technology to diagnose and prognose thestate of the equipment using the sensor signals.Moreover, if the real data from NPPs are appliedto the proposed model, it is expected that thepresent study will be helpful to increase the safetyof 240280320Cutting numberFig.10 Estimation performance for cutter 6 using SVR.Furthermore, the estimation accuracy of the SVRmodel can be checked through Figs. 8-10 for theThis work was supported by the National ResearchFoundation of Korea (NRF) grant, funded by theKorean Government (MSIT) (Grant No.2016M2A8A2953046).References[1]J. Coble, P. Ramulhalli, L. Bond, J. Hines, and B.Upadhyaya, Prognostics and Health Management inISOFIC 2017, Gyeongju, Korea, November 26-30, 20177

[2][3][4][5][6][7][8][9][10][11][12][13]8Nuclear Power Plants: A Review of Technologiesand Applications, Richland, Washington, PacificNorthwest National Laboratory (PNNL), 2012.J. Sheppard, M. Kaufman, and T. Wilmering, “IEEEStandards for Prognostics and Health Management”,Proceedings of AUTOTESTCON, Salt Lake City,Utah, United States, Sep. 8 11, 2008.S. Chiu, "Fuzzy Model Identification Based onCluster Estimation", Journal of Intelligent & FuzzySystems, Vol. 2, pp. 267-278, 1994.D. Goldberg, Genetic Algorithms in Search,Optimization, and Machine Learning, Reading,Massachusetts, 1989.M. Mitchell, An Introduction to Genetic Algorithms,Cambridge, Massachusetts, 1996.C. Cortes and V. Vapnik, “Support-VectorNetworks”, Machine Learning, Vol. 20, pp. 273-297,1995.J. Hines, J. Coble, and B. Upadhyaya, “Applicationof Monitoring and Prognostics to Small ModularReactors”, Proceedings of Future of InstrumentationInternational Workshop (FIIW), Oak Ridge,Tennessee, United States, Nov. 07 08, 2011.S. R. Gunn, Support Vector Machines forClassification and Regression, Southhampton,University of Southampton, 1998.V. Kecman, Learning and Soft Computing: SupportVector Machines, Neural Networks, and FuzzyLogic Models Cambridge, Massachusetts, 2001.V. Vapnik, The Nature of Statistical LearningTheory, New York, 1995.K. Y. Chen and C. H. Wang, "Support VectorRegression with Genetic Algorithms in ForecastingTourism Demand", Tourism Management, Vol. 28,pp. 215-226, 2007.M. G. Na, W. S. Park, and D. H. Lim, "Detectionand Diagnostics of Loss of Coolant Accidents UsingSupport Vector Machines", Nuclear Engineeringand Technology, Vol. 55, pp. 628-636, 2008.PHM Society, PHM data challenge, Oct., ,2010.ISOFIC 2017, Gyeongju, Korea, November 26-30, 2017

ISOFIC 2017 , Gyeongju, Korea November 26-30, 2017 1 Estimation of Cutter Wear of a Milling Machine Using a Support Vector Regression Method Young Do KOO1, Man Gyun NA*1, and Jung-Taek KIM2 1. Department of Nuclear Engineering, Chosun University, 3

Related Documents:

The laser cutter is a potentially dangerous machine. The most common safety risk while using the laser cutter comes from fire. The laser cutter can also produce dangerous fumes (e.g., chlorine), if used with improper materials. Operators must receive training prior to use of the laser cutter. The cutter must remain under constant supervision .

Traditionally, spiral bevel gear teeth are cut by a circular cutter. In general machining it may call a “fly cutter”. But the gear cutter has, not one, but may blades attached to a disc cutter head. The distance from the cutter to the axis of the tool rotation determines the curvature of the tooth line on the finished gear.

4. CUTTER WHEEL CONTROL: To start the cutter wheel, hold down the thumb button on the control handle. Then pull the trigger back. After the wheel is started, the thumb button can be released, but the trigger must be held to keep the cutter wheel turned on. To raise the cutter head, pull the cutter height control bar back.

Using the Techno Plasma Cutter About the Techno CNC Plasma Cutter The plasma cutter used in this book is a Techno CNC Plasma Cutter with a Hypertherm Powermax 45 torch and a moving gantry. This means that the part of the machine that holds the torch, the gantry, moves instead of the table. This saves space, and in many cases, increases speed and

Cutter Replacement Parts Item Number Description Part Number Kit Quantity 1 Kit Option Cutter ZM400 79841 1 1 Kit Option Cutter ZM600 79842 1 2 Kit Replacement Static Brush & Cutter Guard with the Angled End Sensor Flag (old design) ZM400. Order this replacement item if the printer currently has an Angled End cutter guard installed. 79864M 1 2

wear of insignia, wear of the shoulder sleeve insignia (current organization and former wartime service), and wear of the dress uniform insignia; adds wear of insignia on the old and new versions of the female blue and white coats, wear of the full-color U.S. flag cloth replica, and a new paragraph on the wear .

Lower Wear Pad L 204mm Doosan 9 - Lower wear pads. LOWER WEAR PADS CODE PART DESCRIPTION BRAND F11B01 Upper Wear Pad L 156mm Doosan Daewoo 8508376 Lower WearPad Class II L 160mm Hyster Yale 10 - Lower wear pads. MAST WEAR PADS AND HOOKS CODE PART DESCRIPTION BRAND 61356-23600-71 Mast Wear Pad

ASME A17.1, Safety Code for Elevators and Escalators, International Building Code, and other non-governmental safety standards identify minimum design requirements for elevators and for building systems that interface with the elevator controls. The performance language