An On-Machine Measurement Method for Touch-Trigger Probe Based on RBFNN

Transcription

1 Research Journal of Appled Scences, Engneerng and Technology 5(3): , 2013 ISSN: ; E-ISSN: Maxwell Scentfc Organzaton, 2013 Submtted: June 19, 2012 Accepted: July 04, 2012 Publshed: January 21, 2013 An On-Machne Measurement Method for Touch-Trgger Probe Based on RBFNN Xaomng Qan, Peng Zhao and Pehuang Lou Department of Mechancal Engneerng, Nanjng Unversty of Aeronautcs and Astronautcs, Nanjng, Jangsu, , PR Chna Abstract: The touch-trgger probe s a knd of sensor nstalled n a machnng center to measure the dmenson of the work pece. A method for On-Machne Measurement (OMM) and ts error compensaton by the probe based on Radal Bass Functon Neural Network (RBFNN) s advanced n ths study. The advantages and dsadvantages for touch-trgger probe of OMM system are dscussed. Major factors that nfluence the probe measurement precson are analyzed. The measurement error compensaton based on RBFNN s presented. At last the expermental system wth touch-trgger probe s put forward and the experment ndcated that, usng the touch-trgger probe makes onmachne measurement more automatc and effcent and by usng RBFNN for error compensaton make on-machne measurement more precse. Keywords: Error compensaton, on-machne measurement, radal bass functon neural network, touch-trgger probe INTRODUCTION In tradton, a work pece after manufactured wll be nspected for form and poston errors by the Coordnate Measurng Machnes (CMM) whch are stand-alone measurement nstruments and generally located separately from a Machnng Center (MC) (Caballero-Ruz, 2007). Ths s called offlne measurement. Nowadays, as the development of the sensor and the demand of effcency, another measurement method for conventonal manufacturng process called On-Machne Measurement (OMM) appears. The dfference between them depends on whether the machnng process and nspecton process are done on the same equpment (Kwon and Yongjn, 2005). The offlne measurement method by CMM s the most popular now and many research works are done about CMM (Erkan et al., 2011; Naf et al., 2011; Marek and Woznak, 2005). However ths ncreases the overall manufacturng cost and tme to obtan the fnal product and the bottleneck phenomenon may be caused by the product stagnaton due to the tme lag between the machnng and nspecton process. In addton t s hard to measure the complex, large-szed parts (Kwon and Yongjn, 2005). Furthermore, t nevtably reduces the measurement precson due to the secondary fxture error when transfer the work pece from machne center to CMM (Cho et al., 2004). To overcome these problems, the methodology of OMM s accepted. As for OMM, t also has 2 types. One s n-process gaugng and the other s n-cycle gaugng. In-process gaugng means machnng and gaugng at the same tme for whch probe s hardly to use because the cuttng tool and the probe can t work n concurrence. The n-cycle gaugng means gaugng s done only before or after the machnng process, however the work pece wll not be dsassembled from fxture. In ths study n-cycle gaugng wll be manly dscussed. There are some advantages to use touch probe n an OMM system. Frst a touch probe s a, relatvely nexpensve and easy-to-use. In a machne center, the moton of a probe treated as a cuttng tool, s controlled by G code or macro program. After the machnng process s fnshed, the touch probe replaced wth a cuttng tool starts the measurement n the normal drecton to the machned surface. Second as an accessory of certan machne center the probe s easy to be ntegrated wth Computer Numercal Control (CNC) system. OMM system wth touch-trgger probe can reduce the producton tme and cost, can be wdely used for automatng and speedng part processng, elmnate errors caused by secondary fxture. However for hgh-accuracy applcatons wth touch-trgger probe, t measures parts along the machne tool axes (Chattopadhyay, 2008), the measured data nevtably nclude the probng errors orgnated from the structure of the probe and the postonng errors orgnated from the naccurate axs moton of a machne tool. There are largely systematc errors nfluence the accuracy, such as pre-travel varaton n dfferent probng drectons, stylus length, dameter, measure speed, trgger force, etc. These errors should Correspondng Author: Xaomng Qan, Department of Mechancal Engneerng, Nanjng Unversty of Aeronautcs and Astronautcs, Nanjng, Jangsu, , PR Chna 909

2 be elmnated from the measured data to obtan the true machnng error. Caballero-Ruz, only researched the geometrcal error (Caballero-Ruz, 2007), Reddy, Tppa S. focused on the tool wear (Reddy and Reddy, 2011), Valno, G. dd some work for tool compensaton (Valno et al., 2009). In fact, all the factors nfluenced the accuracy should be taken nto account together. When the probe touches certan pont of a work pece that s loaded n the machnng center and get the poston of ths pont that s measurement pont. However, due to the exstence of probe errors, the measurement pont has devaton wth the actual pont. Although we don t know the exact poston of actual pont, we can use theory pont to substtute the actual pont. In order to get more precse result, we need elmnate the devaton between the measurement pont and theory pont. However, due to the many types of error exstng n the measurement process, t s mpossble to fnd a formalze way to express all the error. The Radal Bass Functon Neural Network (RBFNN) s a good method to map the measurement pont to actual pont (Hongtao Zhang, 2011). A method to mplement the OMM wth the touchtrgger probe and error compensaton based on RBFNN was advanced n ths study. The basc technque of probe measurement error modelng was researched. The basc technque of probe measurement error modelng was dscussed. At last the method was valdated n the expermental system composed of DIXI 50 machnng center, Fanuc 16 control system, Blum CNC P probe and PC. The experment ndcated that, usng the touch-trgger probe makes on-machne measurement more automatc and effcent. In addton, once the model has been traned and tested, t can be mplemented ether on-lne or off-lne to compensate for probe errors. Fg. 1: The probe error model Fg. 2: The error factor of probe s federate THE MAJOR ERRORS FOR THE PROBE PRECISION A generalzed probe error model s dscussed n Caballero-Ruz et al. (2007), n whch the nfluencng factors of probe are group nto the followng four categores: mpact force, probe rgdty, stylus rgdty and operatng envronment. Besdes these factors we also fnd there some other factors, such as the probe movement, the structure of probe, the work mode of probe. So the probe error model can be rearranged as Fg. 1. In Fg. 1, there are fve categores that consst of several factors, respectvely. Operatng envronment ncludes the envronment temperature and machnng temperature. Probe structure ncludes probe rgdty, stylus rgdty, stylus ball s radus, stylus length and 910 sngle emttng angle. Probe movement ncludes federate, drecton of probe to approach work pece and mpact force. Work pece to be measured ncludes the shape, materal of work pece. Probe work mode ncludes measurement path and the number of pont to measurement. Each factor could mpact the measurement precson more or less. For example, we know that hgh probe approach speed results n bgger forces transmtted through the probe system and so a bgger dstorton of the stylus. As far as federate of probe concerned, as show n Fg. 2, at the measurng tme, the measurng axs should move wth constant federate Fg. 2a. Measurng when the machne s acceleraton Fg. 2b or deceleraton Fg. 2c can declne the repeatablty or result n measurement error. And so calbraton and tool measurement should be done wth the same federate.

3 Fg. 3: RBFNN archtecture THE RBFNN RBFNN s a famous feed-forward network that s used for approxmaton, classfcaton, regulaton and pattern recognton (Hongtao Zhang, 2011). RBFNN just has a hdden layer whose actvaton functons are calculated based on dstance of the nputs and center of the actvaton functons. The general scheme of the RBFNN s shown n Fg. 3. RBFNN ncludes three layers, namely nput layer, hdden layer and output layer. W s the weghted value vector between the hdden layer and output layer. The actvaton functon of th hdden neuron s defned by: 2 x u Z ( x) = exp( ) 2σ 2 x = (x 1, x 2,.,x n ) s the nput set u = The center of th actvaton functon σ = A parameter to control the smoothness of the actvaton functon x- u = The Eucldean dstance between the nputs and the functon center (1) Mappng functon of the common RBFN wth Gaussan actvaton functon and weghted lnear summaton n output neurons s gven by: m y= wz ( x) = 1 Z (x) : The actvaton functon of th node w : The weght of th node (2) In order to provde greater performance, The RBFN wth local lnear model n the hdden layer s used. In ths way the connecton weghts between the hdden layer unts and output unts of the RBFN are replaced by a local lnear model: 911 Fg. 4: An OMM system structure w wx wx wx = n n (3) w j s the j th weght of th node and j = 1,, n. For optmzaton a cost functon f(x), n ntalzaton some partcles are scattered n an n- dmenson space randomly. Poston of every partcle s dsplayed by x and velocty of them by w. In every teraton, f(x) s calculated so ndvdual best poston p and global best poston p j, are ndcated. The poston and velocty vectors are updated by: w ( t + 1) = χα ( w() t + c1ϕ( p() t x()) t + c2ϕ( pj() t x())) t (4) x ( t+ 1) = x ( t) + w( t+ 1) x & a : Real number, the parameter x controls the magntude of w a : Magntude of the old velocty c 1 & c 2 : Postve constant φ : Dstrbuted random number n [0, 1] CASE STUDY (5) The On-Machne Measurement (OMM) experment system structure s shown n Fg. 4. The probe s composed of Optcal Module Probe (OMP) and Optcal Machne Interface (OMI). OMP, located between the probe head and the shank, receves CNC control sgnals va servo and transmts probe sgnals. Communcaton between the probe and the OMI s done va the optcal transmsson system. RS-232 seral communcaton s used to transmt the measurement NC program to the CNC controller and receve the measured data for further analyss usng a personal computer. The experment system hardware s composed of DIXI 50 machnng center, Fanuc 16 control system, Blum CNC P probe, mcrometer, standardzed rng gauge and PC. The probe styl are made of steel

4 Fg. 5: Measurement system for standardzed rng gauge Table 1: The parameters used for RBFNN Parameter Value Populaton sze 40 χ 0.8 α 1 c 1 2 c 2 5 and carbon wth φ5 mm, length 50 mm. The dameter of standardzed rng gauge s φ10± mm. In ths case, by usng hgh precse mcrometer, the standardzed rng gauge s put on the machnng center, whose center s consstent wth the center of spndle, as shown n Fg. 5. Then make ths center pont as orgn of work pece coordnate system, so each pont n the standardzed rng gauge n the drecton of +X, -X, +Y, -Y can be pnponted by calculatng, called theory ponts. These ponts can be treated as ponts n output layer of RBFNN. And the ponts measured by probe are ponts n nput layer of RBFNN, called measurement ponts. By ths way, the margn between measurement ponts and theory ponts expresses the error of the whole on-machne measurement system. In addton, the ponts n the drecton of +X, -X, +Y, -Y are symmetry, so the ponts n the drecton of +X are nspected. The choce of samples s arbtrary and s better to cover the whole work pece. Usng RBFNN to tran the sample data and the net learns from tranng. The change rule of the weght and threshold are defned as Eq. (4) and (5). The actvaton Eq. (1) s used. The objectve functon s the Mean Square Error (MSE) between targets and the model outputs: n 1 MSE = ( y y ) n = 1, 2 Res. J. Appl. Sc. Eng. Technol., 5(3): , 2013 (6) y, : The ponts measured by the probe y : The RBFNN model output and n s the number of data 912 Table 2: The tranng data set n +X Measurement ponts Theory ponts X Y +X Y Table 3: The error compensaton data n +X Measurement ponts Theory ponts Compensated ponts X Y +X Y +X Y The parameters are shown n Table 1. The maxmal tranng tme s and MSE s The key problem les n the exact quantty of hdden layer s neurons. We frst use 1 neuron n hdden layer to tran the net. In the tranng process, when n the set tme and the MSE s less than 0.001, we thnk the tranng process fnshed and save the weght, threshold and quantty of hdden layer s neurons n a text fle. And ths process s so called error compensaton. Afterward, measure another data to test the net that has been traned. If the result can t be accepted, then the RBFNN wll be traned agan, otherwse t s a good tranng. Table 2 lsts some of data for tranng. In ths case, we all get 120 sample ponts n 4 drectons, ncludng 80 data for tranng and 40 data for test. Table 3 lsts some of the data for test the tranng. All these 120 sample ponts are sent to the PC from machne center by RS232 and could be stored n the database for later net tranng and testng. The transfer macro n FANUC control system s POPEN and PCLOS whch are orgnally used to send data to a

5 seres prnter. And we wrte a program to watch the seres port and get data once the data s sent from machne center. The method mentoned above can also be used to handle the drecton of -X, +Y, -Y and s smlar wth +X. Due to the structure of the probe and the characterstc of machne center, the error n drecton Z that s the spndle, can be gnored compared the error n other drectons. So n ths case we ddn t consder the error of drecton Z. The data lsted n Table 2 show that the method for error compensaton of touch-trgger probes s effectve. CONCLUSION Radal Bass Functon Neural Network (RBFNN) s a good knd of machne learnng methods. In ths study, an OMM system structure wth touch-trgger probe s put forward and then the advantages and dsadvantages for touch-trgger probe are dscussed. Major factors that nfluence the probe measurement precson are analyzed. A method for On-Machne Measurement (OMM) and ts error compensaton by the probe based on RBFNN s researched. The experment ndcated that, usng the touch-trgger probe makes onmachne measurement more automatc and effcent and by usng RBFNN for error compensaton make onmachne measurement more precse. In addton, the method can be used off-lne, only f there s a fle n whch all the data has correct format that could be read by the measurement system n ths study, an effectve hand trackng method s proposed based on hand occluson nformaton. Ths method s smple to operate by movng the user s hand freely for nteractve operatons. Some experments have been conducted to valdate that the proposed method can correctly predct the poston of the occluded parts and have better trackng accuracy. The future research wll address to get the nformaton of each fnger and hand gesture spottng for sgn recognton va sngle camera system. ACKNOWLEDGMENT Ths research s supported by Jangsu provnce scence and technology support plan project under project Research on Hgh-grade CNC system based on FPGA and ts applcaton n 5-axs machnng center (project no. SBE ). REFERENCES Caballero-Ruz, A., L. Ruz-Huerta, T. Badyk and E. Kussul, Geometrcal error analyss of a CNC mcro-machne tool. Mechatroncs., 17(4-5): Chattopadhyay, S., Study of accuracy of CNC machne tools. ASEE Annual Conference and Exposton. Cho, J.P., B.K. Mnb and S.J. Lee, Reducton of machnng errors of a three-axs machne tool by on - machne measurement and error compensaton system. J. Mater. Process. Tech., ( ): Erkan, T., R. Mayer and A. Woznak, Surface probng smulator for the evaluaton of CMM probe radus correcton software. Int. J. Adv. Manuf. Tech., 55(1-4): Kwon and Yongjn, Accuracy analyss of machnng process usng a spndle probe and a CMM to reduce the scrap. Am. Soc. Mech. Eng. Manuf. Eng. Dv., 16(1): Marek, D. and A. Woznak, CMM touch trgger probes testng usng a reference axs. Precs. Eng., 29(3): Naf, A., J.R.R. Mayer and A. Woznak, Novel CMM-based mplementaton of the mult-step method for the separaton of machne and probe errors. Precs. Eng., 35(2): Reddy, T. S. and C.E. Reddy, On-lne montorng of tool wear and surface roughness by Acoustc Emssons n CNC turnng. Int. J. Robot. Aut., 26(3): Valno, G., Y. Prado, J.C. Rco and B.J. Alvarez, 2009.Tool compensaton by means of touch trgger probes n CNC turnng IEEE 14th Internatonal Conference on Emergng Technologes & Factory Automaton, Span, pp: Zhang, H., J. Yang, Y. Zhang, J. Shen and C. Wang Measurement and compensaton for volumetrc postonng errors of CNC machne tools consderng thermal effect. Int. J. Adv. Manuf. Technol., 55(1):