PCA fused NN approach for drill wear prediction in drilling mild steel specimen

Transcription

1 PCA fused NN aroach for drill wear rediction in drilling mild steel secimen S.S. Panda Deartment of Mechanical Engineering, Indian Institute of Technology Patna, Bihar-8000, India, Tel No.: (M); Fax No.: S.S. Mahaatra Deartment of Mechanical Engineering, National Institute of Technology Rourkela,Orissa , India, Tel No.: (O); Fax No.: , India, Abstract: The resent aer describes use of rincial comonents for drill wear rediction. It also makes a comarative analysis in using large sensor based technique in redicting drill wear. In order to reduce the redundancy of the network, rincial comonent has been fused with artificial neural network (ANN) for rediction of drill wear. Large numbers of exeriments have been conducted and sensor signals have been acquired using data acquisition system. Cutting force, torque, vibrations along with other rocess arameters such as sindle seed, feed rate, drill diameter, chi thickness and surface roughness have been used as indicative arameters for characterizing the rogressive wear of drill. Princial comonent of these inut arameters has been derived thereafter and has been used to redict the flank wear using BPNN Keywords-comonent: Neuron; sensor integration; signal analysis; design of exeriment; flank wear; BPNN;PCA I. INTRODUCTION The reason for acquiring the drill wear state information is to enhance the redictive caability to allow the machine oerator to schedule tool change or regrind just in time to avoid under use or overuse of tools, avoid shutdown of machines due to damage, and to minimize scra or rework. Drill wear also affects the ability of the hole cutting system to satisfy secified erformance characteristics, such as hole roundness, centering, burr formation at drill exit, and surface finish. Wearing action in the tool is an inevitable henomenon leading to the loss of dimensional accuracy and ossible damage to the work iece. Hence, online rediction of cutting tool wear becomes an imortant issue for today s manufacturing industries. Many early works on deterioration and failure of drill have been reorted in literature. Altogether different tyes of drill wear can be recognized as outer corner wear, flank wear, margin wear, crater wear, chisel wear and chiing at the li as reorted by Kanai and Kanda []. Some of the revious works (Chungchoo and Saini [], Haili et al. []) reorted that average width of crater and maximum deth of flank wear could be used as to assess tool failure. It is relatively common to measure only the thrust or axial comonent of force for TCM in drilling as suggested by Thangaraj and Wright [4]. The sindle and feed motor current are also used as inuts to the TCM system by (Ramamurthi and Hough [5]; Liu and Chen [6] used eight indices of thrust force and torque as inut arameters for drill wear monitoring. Stiffness and daming roerties of /09/$ IEEE the drilling system has no effect on wear as suggested by Abu-Mahfouz [7]. Vibration monitoring techniques alied to the detection of drill failure have been reorted by several investigators (Wang et al. [8]; Ko and Cho [9]; Li et al. [0]; Liu and Wu []). Liu and Anantharaman [] used nine features reresenting drill condition as inut to multi layer ercetron and found 00% accuracy for correct classification of drill wear. Sanjay et al. [] used drill diameter, feed, cutting seed, time, force and torque as inuts and flank wear was estimated using different structures of artificial neural network (ANN). Since wear of drill related to number of direct and indirect factors and hence can be considered as dimension reduction roblem (Princial comonent analysis), a dimension reduction technique to identify the wear on drill. Rummel [4] used factor analysis using a mathematical model to study the human ability. Detailed descrition of factor analysis can be found in the relevant literature such as Kleinbaum et al. [5] aroach. Fisher [6] had develoed a linear classification algorithm to transform n dimensions attern to dimensions. Hong and ang [7] have used a method for constructing a classifier on the otimal discernment lane by using minimal distance criterion for multi classification roblems. II. BACK PROPAGATION NEURAL NETWORK The basic structure of a BPNN having inut, hidden and outut layers has been considered in the resent work. The inut layer receives information from external sources, and asses this information to the network for rocessing. The hidden layer receives information from the inut layer, and does the information rocessing. The outut layer receives rocessed information from the network, and sends the results to an external recetor. The inut signals are modified by the interconnection weights, known as weight factor w ji, which reresents the interconnection of i th node of the first layer to j th node of the second layer. The sum of modified signals (total activation) is then modified by a sigmoid transfer function. Batch mode tye of suervised learning has been used in the resent case. During training, the redicted outut has been comared with the desired outut, and the mean square error has been calculated. If the mean square error is more than a rescribed limiting value, it is back roagated from outut to inut, and weights are further Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOG ROURKELA. Downloaded on November 7, 009 at :5 from IEEE Xlore. Restrictions aly.

2 modified till the error or number of iteration are within a rescribed limit. Mean square error, E for attern is defined as = n E ( Di O i ) () i= where, Di is the target outut, and O i is the comuted outut for the i th attern. Weight change at any iteration t,isgiven by Δ Wt () = ηe () t+ α ΔWt ( ) () where η is learning rate, andα momentum arameter. Entire exerimental data set is divided into training set, testing set and validation set. The error on the testing set is monitored during the training rocess. The testing error normally decreases during the initial hase of training, as does the training set error. However, when the testing error starts increasing for a secified number of iterations, the training is stoed; and the weights at the minimum value of the testing error are returned. With the trained network, unseen data set (validation set) are verified and ercentage of variation of redicted outut (flank wear) with resect to the actual outut is thus evaluated. III. PRINCIPAL COMPONENT ANALSIS Princial comonent analysis is a statistical technique which generated uncorrelated linear combinations of original variables and account for the total variance and original data so that adequate information of original data can be extracted. In rincile, each of the rincial comonents is a linear combination of the original values for the variables; PC = c + c + c c ( axis Z ) PC PC = c = c... + c + c + c + c c c ( axis Z ( axis Z () Where, c a,b is the comonent score coefficient for variable b on PC axis Z a,and b is the score for variable b. Princial Comonent Analysis transforms a multivariate set of variables (,,..., ) to new variables (Z, Z,..., Z ), which are uncorrelated. The first rincial comonent consists of a rincial comonent coefficient (α i ) for each variable () such that there is maximal variance in the calculated score for each case (n); the factor score for each case is calculated as α. + α. + + α i. i + α. where i is the centered value for the i th ( i -meanforthei th variable). The first rincial comonent axis for the raw variables is the new axis Z. The second rincial comonent consists of the next set of rincial comonent coefficients (α i ) such that there is a maximal remaining variance in the calculated score for each case (n), and there is no correlation of the second rincial comonent with the first. Z is the second ) ) rincial comonent axis for. Further sets of rincial comonents (third, fourth, etc) can be calculated until no statistical significance can be attributed to that comonent (e.g. by χ test) or all m rincial comonents are comuted. IV. EXPERIMENTAL SET-UP In the resent work, a radial drilling machine (Batliboi Limited, BR68 model) has been used for the drilling oeration. Uncoated HSS drills with different diameters have been used for drilling holes in mild steel work iece at different cutting conditions. Different sensory devices such as dynamometer and vibration analyzer are used for sensing thrust force signals, torque signals and vibration signals in the exeriments is shown in fig.. Radial drilling machine Job fixture Comuter Sindle Drill Work iece Inut arameters (sensor signals and cutting condition) Neural Network model Performance analysis Dynamometer Vibration analyser Charge amlifier Data acquisition system Outut arameter (drill wear) Drill wear Prediction system Figure Schematic diagram of the exerimental set-u HSS drill bits with different diameters have been used for drilling in mild steel work-iece at different cutting conditions in dry state. Chis are collected during each cutting condition and average thickness is measured. Similarly, after each cutting oeration surface roughness of drill hole is measured with hel of ocket surface roughness tester (make Mahr) with maximum roughness deth 0. μm to 5. μm and maximum stylus force 5.0 mn. In all the drilling oerations erformed in the resent work, no coolant has been used. Root mean square (RMS) values of thrust force and torque signal are recorded through a iezo- electric dynamometer (Kistler, 97). Signals from the dynamometer were assed through low ass filter, amlified through charge amlifier (Kistler, tye 505 model), and stored in the comuter through a Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOG ROURKELA. Downloaded on November 7, 009 at :5 from IEEE Xlore. Restrictions aly.

3 data acquisition system (Advantech, PCL 88 HG, 6 channel analog to digital (A/D) converter in 6 bit digital time-discrete, and 00 khz samling rate). Two iezo electric accelerometers (model Bruel & Kjaer, tye 496) have been used to cature vibration signals. One accelerometer has been attached on the to surface of the mild steel secimen to extract feed vibration and other on the side surface of the mild steel secimen to extract radial vibration. Signals from accelerometer were assed through vibration analyzer (Bruel & Kjaer, tye 560 D) in the frequency range 7 Hz-5.6 khz. RMS of maximum amlitude of vibration both in feed and radial direction were collected through Bruel & Kjaer ulse software (version 7) and is stored in the comuter through data acquisition system through data recorder (tye Bruel & Kjaer, Tye 770). The otical microscoes along with Carl-Zeiss software interfacing have been used to measure flank wear. The maximum flank wear has been used as the criterion to characterize the drill condition, and is obtained by measuring the wear at different oints on either of the cutting edges. V. RESULTS AND DISCUSSION Drilling oerations have been conducted over a wide a range of cutting condition. Sindle seed has been varied in the range 5 rm to 000 rm in four stes. Feed rate hasbeenvariedfrom0.to0.6mm/rev in four stes. High-seed steel (HSS) drills of four different diameters of (9mm, 0mm, mm and mm) have been used for drilling 64 numbers of through holes of 5mm deth for various combination of sindle seed, feed rate and drill diameter in mild steel lates. VI. WEAR PREDICTION B PCA FUSED NN In order to extract the rincial comonent, a test statistics of different variable of samle size of 64 have been studied. XLSTAT version 6 has been used for finding of rincial comonents. Table rovides the Eigen values for all nine rincial comonents, with the ercentage of the variance which they exlain and their cumulative ercentage of variation exlained. Cumulatively Eigen values one, two and three exlain 98.% of the total variance. Rest of the comonent exlains only about.89% of cumulative variance. Total rincial comonent axes have been considered from the resent work is gauged from the fig. as these three comonents exlain about 98.% of the total variance. Taking consideration of these three rincial comonent (PCA, PCA and PCA) which is having all the variables at different score coefficient. So these three rincial comonents is now acting as inut to neural network and outut is drill flank wear. Now network has been trained at different condition of learning rate, momentum coefficient and number of neuron in hidden layer. Large numbers of runs were given for selecting the best architecture and table shows only the best and the second best network in terms of the validation error and the number of iteration. Eigen value Comonent number Figure Screen lot The otimum network architecture in the resent case has been observed to be -- with η = 0. and α = 0.. Variation of mean square error in training and testing with number of iteration for the otimum network is shown in fig.. After the network has been trained, it has been validated with unknown data samle. Fig. 4 shows ercentage of error between actual value and the redicted value and it can be observed that resent neural network of -- with η = 0. and α = 0. redicts the results within ± 6.4 % error. Mean square error training curve testing curve PCA based BPNN (,η=0., α=0. ) Number of Iteration Figure Variation of mean square error with number of iteration of -- with η = 0. and α =0. This otimum network is later on used for classifying the new set of attern of low wear and high wear. Hence in the same exerimental environment another 64 test data ( data in low range of wear and data in high range of Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOG ROURKELA. Downloaded on November 7, 009 at :5 from IEEE Xlore. Restrictions aly.

4 Table Variance of Eigen values Value PC PC PC PC 4 PC 5 PC 6 PC 7 PC 8 PC 9 Eigen value Variance (%) Cumulative variance (%) Table Otimum network architectures for PCA based BPNN Architecture MSE training MSE testing Maximum Validation error (%) L-M-N η α Iteration Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOG ROURKELA. Downloaded on November 7, 009 at :5 from IEEE Xlore. Restrictions aly.

5 wear) were selected. The classification accuracy of the network was calculated by taking the number of correctly classified samle by the network and divided by the total number of samle into the test data is reresented as confusion matrix in table. Table shows that classification accuracy of the network on average is aroximately 90%. Predicted flank wear(μm) validation error +/- 6.5% error line PCA based BPNN (,η=0., α=0.) Actual flank wear(μm) Figure 4 Comarison of Predicted value with actual value of -- with η =0.andα =0. Table Confusion matrix for wear classification No of Low High samle wear wear Accuracy Low wear % High wear 96.87% % VII. CONCLUSION From the resent study following conclusions have been made. Generalized Conclusions: (i) Direct rocess arameters such as sindle seed, feed rate, and drill diameter, has definitive effect on rogressive flank wear of the drill. (ii) Wear on the drill (condition of the drill) affects the sensor signals such as thrust force, torque, vibration and hence these could be used in the drill wear rediction, (iii) Drill wear (condition of the drill) also affects the chi thickness as well as the surface roughness and hence thickness of the chi and the surface roughness could also be used as indicative arameters for rediction of drill condition. Derived Conclusions: (iv) Nine inut variable could be better reresented by three rincial comonent (v) PCA based BPNN architectures took aroximately 000, iteration to redict the drill flank wear within ± 6.5% error band. (vi) Network could able to classify low wear and high wear within an accuracy of 90%. REFERENCES [] Kanai, M. and Kanda,. (978), Statistical characteristic of drill wear and drill life for standardized erformance tests, Annals of CRIP, 7(), 6-66 [] Chungchoo, C. and Saini, D. (00), A comuter algorithm for flank and crater wear estimation in CNC turning oerations, Int. J. of Mach. Tools & Manuf., 4, [] Haili, W., Hua, S., Ming, C. and Dejin, H. (00), On-line tool breakage monitoring in turning, J. of Mater. Process. Technolo., 9, 7 4. [4] Thangaraj, A and Wright, PK (998), Comuter assisted rediction of drill failure using in-rocess measurements of thrust force, J. of Eng. Ind., 0,9 00 [5] Ramamurthi, K. and Hough, CL (99), Intelligent real-time redictive diagnostics for cutting tools and suervisory control of machining oerations, ASME, J. of Eng. for Ind., 5, [6] Liu, T. I. and Chen, W.. (988), Intelligent detection of drill wear, Mech. Syst. & Signal Process., (6), [7] Abu-Mahfouz, I. (00), Drilling wear detection and classification using vibration signals and artificial neural network, Int. J. Mach. tools & Manuf., 4, [8] Wong, S.V. and Hamouda, A.M.S. (00), Machinability data reresentation with artificial neural network, J. of Mater. Process. Technolo., 8, [9] Ko, T.J. and Cho. D.W. (994), Cutting state monitoring in milling by a neural network, Int. J. of Mach. & Tools Manuf., 4(5), 659 [0] Li, X., Dong, S. and Venuvinod, P. K. (000), Hybrid Learning for Tool Wear Monitoring, Int. J. of Adv. Manuf. Technolo., 6, 0 07 [] Liu, T.I. and Wu, S.M. (990), On-line detection of drill wear. J Eng. Ind.,, 99 0 [] Liu, T.I. and Anantharaman, K.S. (994), Intelligent classification and measurement of drill wear, ASME, J. Eng. Ind., 6, 9 [] Sanjay, C., Neema, M.L. and Chin, C.W. (005), Modeling of tool wear in drilling by statistical analysis and artificial neural network, J. of Mater. Process. Technolo., 70, [4] Rummel. R.J. (988), Alied factor analysis, Evanston, Northwestern University Press. [5] Kleinbaum, D., Kuer, L., and Vannelli, A. (986), Alied regression analysis and other multi variable methods, California: International J. of Production Research, 4, [6] Fisher, R. A. (96), The use of multile measurement in taxonomic roblems, Annals of Eugenis, 7, [7] Hong, Z. Q. and ang, J.. (99), Otimal discernment lane for a small number of samles and design method of classifier on the lane, Pattern Recognition, 4 (4), Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOG ROURKELA. Downloaded on November 7, 009 at :5 from IEEE Xlore. Restrictions aly.