- Email: [email protected]
Artificial neural network based tool condition monitoring in micro mechanical peck drilling using thrust force signals
Accepted Manuscript Title: Artificial neural network based tool condition monitoring in micro mechanical peck drilling using thrust force signals Author: K. Patra A.K. Jha T. Szalay J. Ranjan L. Monostori PII: DOI: Reference:
S0141-6359(16)30447-0 http://dx.doi.org/doi:10.1016/j.precisioneng.2016.12.011 PRE 6505
To appear in:
Precision Engineering
Received date: Revised date: Accepted date:
27-11-2014 26-6-2016 30-12-2016
Please cite this article as: Patra K, Jha AK, Szalay T, Ranjan J, Monostori L.Artificial neural network based tool condition monitoring in micro mechanical peck drilling using thrust force signals.Precision Engineering http://dx.doi.org/10.1016/j.precisioneng.2016.12.011 This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
Artificial neural network based tool condition monitoring in micro mechanical peck drilling using thrust force signals K. Patra1*, A. K. Jha1, T. Szalay2, J. Ranjan1 and L. Monostori2 1
Mechanical Engineering department, Indian Institute of Technology Patna, Bihta, Patna-801103, India Department of Manufacturing Science and Engineering, Budapest University of Technology and Economics, H-1111, Budapest, Hungary 2
*Corresponding author: Email: [email protected], ph. +91-612-2552012
Highlights
Hole numbers represent drill wear stages well.
Prediction of hole number gives tool wear status of the micro-drill.
Higher feed and lower cutting speed values improve tool life. Artificial neural network accurately predicts hole number from thrust force. The prediction values are in good agreement with experimental values.
Abstract Micro scale machining process monitoring is one of the key issues in highly precision manufacturing. Monitoring of machining operation not only reduces the need of expert operators but also reduces the chances of unexpected tool breakage which may damage the work piece. In the present study, the tool wear of the micro drill and thrust force have been studied during the peck drilling operation of AISI P20 tool steel workpiece. Variations of tool wear with drilled hole number at different cutting conditions were investigated. Similarly, the variations of thrust force during different steps of peck drilling were investigated with the increasing number of holes at different feed and cutting speed values. Artificial neural network (ANN) model was developed to fuse thrust force, cutting speed, spindle speed and feed parameters to predict the drilled hole number. It has been shown that the error of hole number prediction using a neural network model is less than that using a regression model. The prediction of drilled hole number for new test data using ANN model is also in good agreement to experimentally obtained drilled hole number. Keywords: Micro-drilling, peck drilling, tool breakage, thrust force, tool wear, artificial neural network, regression analysis
1. Introduction In the modern era, there has been a continuing trend towards miniaturization for better performance, higher quality and less expensive products in various industries such as aerospace, biomedical, electronics, communication and automotive [1]. Successful miniaturization of products requires high precision micromachining methods for producing micro-features including micro-holes [2]. In general, laser beam machining, ultrasonic machining, electro-chemical machining, electrical discharge machining and mechanical micro-drilling are used to produce micro-holes. Among these methods, mechanical micro-drilling is gaining wide acceptability due to its capability to produce good quality micro-holes in most of the engineering materials in a short processing time [3]. Mechanical micro-drilling can be applied for producing micro-holes in various mechanical, electronic, optical, ornamentals and micro-fluidic device components made of metal/alloys, composites and plastics. However, tool wear of the micro-drill during the micro-drilling operations is one of the critical issues, since it considerably affects the quality of the machined parts. That’s why monitoring of tool wear during the machining operation and prediction of micro-drill breakage is very important to maintain high hole quality and improve the breakage life of the drill and efficiency of the micro-drilling process [4-5].
The monitoring and prediction of tool wear and breakage in macro-machining are mainly done indirectly through thrust force, torque, current, spindle power, vibration, acoustic emission measurements [6-11] and through machined surface texture measurement using machine vision [12]. Attempts have also been made in micro-domain machining to apply indirect methods for tool wear and breakage monitoring. Kim et al. [3] applied thrust force monitoring to increase tool life before drill failure in deep-micro-hole drilling. Kondo et al. [13] also applied thrust force signals to monitor burr and pre-failure phase in microdrilling of stainless steel. Apart from thrust force signal, torque signal was also monitored to prevent the tool breakage in micro-drilling [14-15]. However, it was shown that magnitude of torque was very small compared to thrust force and special non-contact measuring equipment would be required for torque signal for its effectiveness in micro-tool wear monitoring applications. Input current signatures which were successfully applied for tool breakage [16] and wear detection [17] in macro-drilling operations were shown to be less effective than the spindle motor input impedance signatures for detection of drill breakage in micro-drilling [18]. However, Ogedengbe [19] showed the feasibility of tool condition monitoring of micro-milling using time-domain and frequency-domain features of current signatures. Workpiece vibration [20] and spindle motor vibration [21] signals were also shown to be effective for classification of tool wear in micro-milling and classification of tool breakage in micro-drilling, respectively. However, vibration signals in micro-scale cutting can be influenced by external noise due to its lower bandwidth [22]. Acoustic emission (AE) signals, which were less affected by external noise due to their higher bandwidth, were demonstrated tool condition with a high degree of confidence in micromachining [22-24]. Among the aforementioned various sensing methods, cutting force provides the most effective method for monitoring tool condition in a micro-machining process due to its higher signal to noise ratio and its best representation of the state of the machining process [22]. For effective monitoring of tool condition in micro-machining, it is essential to apply a viable, reliable and accurate modeling technique which can establish highly nonlinear relations of cutting conditions, process signals and tool states. Artificial neural network (ANN) based modeling approaches were considered to be one of the most viable, reliable and attractive approaches for tool condition monitoring in machining processes [25-28]. Even though there are plenty of examples of the applications of ANN to relate cutting forces with tool wear in macro-scale machining, these models can not be used directly in micro-scale tool condition monitoring as cutting forces are influenced by tool size effects at low feed micro-cutting [29-31]. Therefore, ANN models for micro-machining tool condition monitoring should be trained with signals obtained during micro-cutting operations. Tansel et al. [32-33] predicted the wear of micro-end mill from cutting force measurement and artificial neural network models. Malekian et al. [22] applied multi-sensors signals including cutting force, vibration, acoustic emission to monitoring tool wear in micro-milling based on ANN and fuzzy algorithms. Hsieh et al. [20] applied back propagation neural
network to relate spindle vibration features to tool wear in micro-milling process. Fu et al. [18] used an ANN model to recognize the waveform of the motor impedance to identify drill breakage. Beruvides et al. [34] applied neural network and neuro-fuzzy systems to correlate the cutting force and vibration signal features with the number of drilled holes. As the number of drilled holes can represent different stages of tool wear and tool breakage for a particular cutting condition [35], correlating sensor features with the number of drilled holes will actually indicate the condition of the micro-drill. However, Beruvides et al. [34] analyzed the cutting force and vibration signals only for five numbers of drilled holes in each of the three different cutting conditions. As no correlation between tool wear and hole number was established in their work, it is not clear how mere five numbers of drilled hole can represent all the stages of drill wear in a cutting condition. For effective tool condition monitoring for micro-drilling with ANN modeling technique, it is imperative to use process signal data for sufficient number of drilled holes covering the maximum portions of the drill wear states. Considerable numbers of drilled holes are produced to obtain significant amount of tool wear under each cutting condition. The main goal of this work is to develop an ANN model to predict the number of drilled holes using thrust force and cutting parameters of a micro mechanical peck drilling process. By predictiong the hole number, one can easily determine the condition of the tool and remaining tool life at that cutting condition. For some machining applications where micro-drills are reused after some interval, the proposed monitoring system can predict numbers of holes which were drilled prior to the present operation and the remaining tool life in any cutting condition. The performance of the ANN model is compared to that of regression model which is developed from the same experimental data sets used by the ANN model. Finally the ANN model is applied to new test data to predict the tool condition successfully, as a validation of the proposed micro-drill condition monitoring technique.
2. Experimental Procedure 2.1 Experimental set-up and cutting force data acquisition The micro-drilling experiments were performed on a high stiffness Hembrug Slantbed Microturn 50 as shown in Fig. 1(a). It was equipped with a GMN precision high revolution spindle (max. speed: 120 000 rpm, maximum power: 1.2 kW). The positioning accuracy of this machine tool is within 100 nanometer, and because of the granite bed, the damping and heat resistance is very good. The consequence of the machine structure, the drilling direction was horizontal that is not usual in drilling (Figure 1(b)). However, the very smooth and accurate movement in this machine tool may exclude that causing reason of tool breakage. Hartner high speed steel drills of 0.5 mm diameter as shown in figure 1(c) are used to produce blind holes in tool steel workpiece material (AISI P20 + S – C 0.4; Mn 1.5; Cr 1.9; Mo 0.2; S 0.05) which is regarded as well machinable mold steel for machine-cut plastic molds and zinc die casting
dies. The depth of the drill holes were 1.5 mm and peck drilling strategy (with 0.5 mm step size) was applied with air cooling during drilling. In the experiments, the cutting parameters were varied in the range suggested by the tool producer. For measuring the thrust force in the peck drilling process, a Kistler 9257B type 3-component dynamometer and a Kistler 5019B type charge amplifier were used. The data was transferred to a computer using National Instruments 6024E DAQ Card at a sampling rate of 1000 Hz. A self-developed program in LabVIEW was used to acquire and archive the thrust force data in each peck drilling. The drilling process and the tool wear were inspected using digital microscope (Dino-Lite AM3013T with DinoCapture 2.0). 2.2 Design of Experiments A full factorial design of experiments with two factors (cutting speed and feed) and three levels were considered in this work. Table 1 shows three levels of cutting speed (16, 23 and 30 m/min) and their corresponding spindle speed (in rpm), and three levels of feed (0.001, 0.002 and 0.003 mm/rev). We repeated the drilling 90 times using a single micro-drill in each of total nine combinations of cutting parameters (feed and speed) to obtain substantial amount of drill wear. The maximum cutting force and the average cutting force for each step of peck drilling were measured during each experiment. 2.3 Measurement of drill wear Due to scaling down of the different components of the micro-cutting tool and its complex geometry, tool wear measurement is very challenging. Unlike macro-drilling process, it is not convenient to take out the micro-drill bit from the machine after each drilling as small change in setting may influence the drill wear process. Moreover, it is also not advisable to interrupt the operation after each drilling. In this work, images of micro-drill edges were taken by digital microscope (Dino-Lite AM3013T with DinoCapture 2.0) after each 30 drilled holes for a particular cutting condition. Figure 2 shows the images of fresh and used drill bit after different number of drilled holes. Tool flank wear and chisel edge wear can be observed in these images. Drill wear values were measured from digital image processing. However, the adherence of chips on the cutting edges makes measurement of tool wear very challenging. From the measurements of flank wear and chisel edge wear, it has been observed that progression of chisel edge wear against drilled hole numbers are more significant compared to that of flank wear for all the cutting conditions considered in this work.
3. Experimental results
3.1 Variation of tool wear with hole number and cutting conditions Figure 3 shows progression of chisel edge wear of micro-drills with increasing number of drilled holes over entire cutting periods under two different cutting speeds (16 m/min and 30 m/min) for feed value of 0.001 mm/rev. The evolution of chisel wear was found to follow similar three-stage process as described by Imran et al. [35]. During stage 1, tool wears rapidly as represented by high slope. Tool wear increased steadily in stage 2 which was indicated by lower slope in the chisel wear curve than the stage 1. Stage 3 is represented by rapid increase of the slope in the wear curve and catastrophic tool failure. However, tool wear rate at higher speed (30 m/min) was higher due to rapid heating that resulted to the shifting of the tool wear stages, higher tool wear and finally failure at lower number of drilled holes. Effects of feed and cutting speed were shown in Fig. 4 and Fig. 5, respectively. Initially, tool wear at different feed and speed did not differ much. However with increasing drilled hole numbers, tool wear was more for low feed and high speed conditions. Higher tool wear values at low feed and high speed in micro-drilling are driven by ploughing effects and rapid heating, respectively [35]. 3.2 Variation of thrust force with hole number and cutting conditions A typical thrust force profile of the micro peck drilling is shown in Fig. 6. In the thrust force profile, average force and maximum force of each step are also shown. Standard deviation of thrust force in each step is also shown with error bar having equal spread above and below the average value of the thrust force. The fluctuation of thrust force data during step drilling may be contributed by the tool runout and microdrilling size effects such as tool edge radius effects and cutting edge rubbing with the workspiece or ploughing [36-37]. It can be seen that there is negligle fluctuaion of measurement data during no cutting time. Therefore, measurement error can be considered as negligible. The variations of average thrust forces with hole numbers for each step at different feed values and constant cutting speed (23 m/min) are shown in Fig. 7(a)-(c). Average thrust force for initial holes is higher for higher feed rate as chip load is higher. However, for low feed rate, with increase of number of holes, tool wear increases rapidly due to ploughing and rapid heating that contribute to increase of average thrust force. Therefore, at higher number of holes, difference of average cutting force of 0.003 mm/rev feed with that of other feed reduces and in few instances, average thrust force may be more for low feed conditions. Variations of maximum thrust force with hole numbers for each peck drilling step at different feeds and constant speed (23 m/min) are shown in Fig. 8 (a)-(c). Compared to average thrust force, maximum value of thrust force is more affected by some other factors such as chip entanglement on cutting edge, variation of drill characteristics and non homogeneous structure of workpiece at microlevel, etc. Therefore, the difference of the maximum force of 0.003 mm/rev feed rate from that of other feed rates is not much. Similarly, variations of average and maximum thrust force values with drilled hole
number for each peck drilling step at different cutting speed and constant feed (0.001 mm/rev) are shown in Fig. 9(a)-(c) and Fig. 10(a)-(c), respectively. Both average and maximum thrust force values during each step follow increasing trends with number of drilled holes. Average and maximum thrust force values also increase with increasing cutting speed due to increasing heating as shown by Fig. 9(a)-(c) and 10(a)-(c), respectively.
4. Prediction of tool condition using ANN model 4.1 Development of ANN model An artificial neural network (ANN) is a mathematical or computational model that is inspired by the structure and/or functional aspects of biological neural networks [38]. In most cases, an ANN is an adaptive system that changes its structure based on external or internal information that flows through the network during the learning phase. Modern neural networks are usually used to model complex relationships between inputs and outputs or to find patterns in data. The Matlab neural fitting tool is used in this work to develop the neural network. Figure 11 shows the structure of the artificial neural network which consists of three layers: input, hidden and output layers. There are nine input variables (cutting speed, feed rate, spindle speed, first step average thrust force, second step average thrust force, third step average thrust force, first step maximum thrust force, second step maximum thrust force, and third step maximum thrust force) to the input layer. ‘Hole number’ is the only output variable from the output layer. The neurons of one layer are interconnected to the neurons of the previous and successive layers through weighted links. The network has been trained with Levenberg-Marquardt back propagation algorithm [39]. Levenberg–Marquardt algorithm is a combination of two training processes: steepest descent algorithm and Gauss–Newton algorithm. Around the area with complex curvature, the Levenberg–Marquardt algorithm switches to the steepest descent algorithm, until the local curvature is proper to make a quadratic approximation; then it approximately becomes Gauss–Newton algorithm to speed up the convergence significantly [34].
4.2 Training, validation and testing of ANN model Dataset for training, validation and testing of the ANN model consists of 810 numbers of patterns from total 810 experiments from 9 cutting conditions (90 holes x 9 cutting conditions) as shown in Table 2. Each pattern has nine input parameters and one target output parameter as mentioned in the earlier subsection 4.1. As hole number can represent the tool wear for a cutting condition, target output parameter considered in this work is the hole number. Input and output parameters have been normalized between 0 and 1. Normalization among the data is based on the maximum and mimimum values of each
parameter in the whole dataset. Following equation is used to determine the normalize value (xnorm) of any input or output parameter.
xnorm = (x- xmin)/(xmax-xmin)
(1)
where, x is the value of any parameter; xmin = minimum value of that parameter in the dataset; xmax = maximum value of that parameter in the dataset. The data has been divided randomly into 70% for training, 15% for validation, and 15% for testing of the ANN model. Data in each dataset covered all 9 cutting conditions shown in Table 2. The training data set is used for computing the gradient and updating the network weights and biases. The validation data set is applied to monitor the training process. The error of validation data normally decreases during the initial phase of training, as does the training set error. However, when the network begins to overfit the data, the error on the validation set typically begins to rise. The network weights and biases are saved at the minimum of the validation set error. The test data set error is used to compare the performances of different architectures of the ANN model. Mean square errors (MSE) of different architectures of ANN model, i.e., with different numbers of hidden neurons are compared to determine the best ANN architecture for this given problem. Figure 12 (a)-(e) show variations of MSE of training, validation and testing data with respect to number of epoch for different numbers of hidden neurons. Table 3 shows the variation of MSE of validation data with number of hidden neurons. Maximum performance has been obtained when number of hidden neuron in hidden layer is 10 and minimum value of MSE of validation is 0.00037803 at epoch 15 as shown in Fig.12 (b). Hence, the ANN architecture of 9-10-1 (9 input neurons, 10 hidden neurons and 1 output neuron) has been selected as the best architecture among the architechtures considered in this work. The network outputs i.e., predicted hole numbers, and regression plots of training, validation and testing datasets have been generated and shown in Fig.13. The network output (Y) values are very close to the target (T) values in all training, validation and testing cases. The regression plots of these data are overlapping to Y=T line.
4.3 Comparison of the ANN model with a regression model The number of drilled holes has been predicted using regression analysis for comparison to the performance of the ANN model prediction. Relation of output variable with input variables is established using the following regression equation (2):
(2)
where y p is the predicted value for n number of input variable x and target variable y. Coefficients b are obtained by minimizing sum square error of predicted (y p) and target value y for a set of m input-output pair. In this work, values of m and n are 810 and 9, respectively. The inputs and output parameters of regression model were same as ANN modeling parameters. The regression model fitting was done using Matlab statistical toolbox. The mean square error (MSE) and R2 values obtained from regression model are 0.00061 and 0.992, respectively. Table 4 shows the comparison between ANN and regression models in terms of MSE and R2 values. Figure 14 shows the plot between the experimental values and predicted values using both regression model and ANN model. From this plot, it can be seen that the ANN model predicted values are closer to the target values compared to those of the regression model.
4.4 Validation of the developed ANN model at new cutting conditions To validate the applicability of the developed ANN model, the trained network has been applied to test data from two new cutting conditions with same tool-workpiece material and same machining environment as discussed in section 2. For cutting condition 1 (CC#1), selected cutting speed and feed were 23 m/min and 0.0025 mm/rev, respectively. Whereas, selected cutting speed and feed were 35 m/min and 0.0035 mm/rev, respectively, in cutting condition 2 (CC#2). The peck drilling of three steps were repeated 90 times using a single micro-drill in each cutting condition to obtain substantial amount of drill wear. The average thrust force and maximum thrust force for each step of peck drilling were measured during each experiment. The normalization of cutting parameters and sensor signal features of both cutting conditions (CC#1 and CC#2) were normalized with respect to the earlier data from 9 cutting conditions reported in section 2. As the values of the cutting parameters in CC#1 are within the values of those considered in the earlier 9 cutting conditions, the normalized values of input and target values of CC#1 lie within 1. However, the values of the cutting parameters considered in CC#2 are outside the ranges of the earlier 9 cutting conditions. So the values of the normalized cutting parameters and the target parameter are obtained beyond 1. The simulation of the trained ANN model for these two new cutting conditions were performed in Simulink block driagram environment in Matlab neural network toolbox. The values of mean error of hole number prediction in CC#1 and CC#2 are 5.7% and 3.2%, respectively. Figure 15 shows the plot between the normalized predicted hole numbers and the normalized target hole numbers for all data sets, i.e., 810 data from 9 cutting conditions, 90 data from CC#1 and 90 data from CC#2. From this plot, it can be seen that the ANN model predicted values are more closer to the target values when data is taken from the same dataset used for training the neural
network than those of the data taken from completely new cutting conditions, i.e., CC#1 and CC#2. However, the mean prediction error values in these two new cutting conditions may be acceptable considering the high nonlinearity of the cutting force signals due to size effects in micro-drilling process. Regression model has also been applied for prediction of hole numbers of two new cutting conditions, i.e., CC#1 and CC#2. Figure 16 shows the plot between the normalized predicted hole numbers using regression model and the normalized target hole numbers for all data sets, i.e., 810 data from 9 cutting conditions, 90 data from CC#1 and 90 data from CC#2. Table 5 shows the comparison of the prediction performances of ANN model and regression model for CC#1 and CC#2 cutting conditions. The values of mean error of hole number prediction using regression model in CC#1 and CC#2 are 6.3% and 4.6%, respectively. It has been also found the error of hole number prediction for new cutting conditions using a ANN model is less than that using a regression model.
5. Conclusion In the present study, tool wear behavior of micro-drill with drilled hole numbers and cutting conditions of a micro mechanical peck drilling process were investigated. Later on an ANN model was developed to predict the number of drilled holes using thrust force and cutting conditions as input parameters. The performance of the ANN model was compared to that of regression model which was developed from the same experimental data sets used by the ANN model. Further the ANN model was applied to new cutting conditions for validating its tool condition predition performance. From this present study, the following conclusions were drawn: The progress of chisel wear in micro peck drilling of tool steel follows the similar three stages of tool wear as reported earlier. Stage 1 represents rapid tool wear and is followed by near uniform wear rate in stage 2. Finally, rapid tool wear and tool breakage is observed in stage 3. However, this study shows the shifting of each stage with change in cutting condition. The identification of the onset of stage 3 at different cutting condition should help to avoid breakage of tool during the cutting process. Results suggest that higher feed and lower cutting speed values should be undertaken for improving tool life by reduce ploughing effect and thermal effect, respectively. Experimental results show that hole number can represent drill wear stages well. The costly and process-interfering drill wear measurement after each drill hole can be avoided by the method of predicting hole number. Prediction of hole number for different cutting conditions using thrust force data indirectly gives tool wear status of the micro-drill.
The prediction results show that artificial neural network (ANN) is one of the accurate and viable techniques for predicting drilled hole number from the average thrust force, maximum thrust force and cutting conditions used in mechanical micro peck drilling. It has been also found the error of hole number prediction using a neural network model is less than that using a regression model. The prediction values of the ANN model with new experimental data were also in good agreement. The mean prediction error lies in the range of 3-6%, which can be considered to be acceptable considering the uncertainty and size effects in the microdrilling process. Hence, the proposed methodology can be successfully applied to tool condition monitoring of micro-drilling process at wide ranges of cutting conditions.
Acknowledgement This article is a part of the bilateral research project “Multi-sensors based intelligent tool condition monitoring in mechanical micro-machining” supported jointly by project no. INT/HUN/P-01/2012 and project number TÉT_10-1-2011-0233 funded by DST, Govt. of India and NIH, Govt. of Hungary, respectively.
References 1. I. S. Kang, J. S. Kim, M. C. Kang, K. Y. Lee, Tool condition and machined surface monitoring for micro-lens array fabrication in mechanical machining, Journal of materials processing technology, 201, 585–589, 2008. 2. J-S. Moon, H-S. Yoon, G-B. Lee, S-H Ahn, Effect of backstitch tool path on micro-drilling of printed circuit board, Precision Engineering 38, 691–696, 2014. 3. D. W. Kim, Y. S. Lee, M. S. Park, C. N. Chu, Tool life improvement by peck drilling and thrust force monitoring during deep-micro-hole drilling of steel, International Journal of Machine Tools & Manufacture, 49, 246–255, 2009. 4. J. Chae, S. S. Park, T. Freiheit, Investigation of micro-cutting operations, International Journal of Machine Tools & Manufacture, 46, 313–332, 2006. 5. H. Watanabe, H. Tsuzaka, M. Masuda, Microdrilling of printed circuit boards (PCBs) – Influence of radial run-out of microdrills on hole quality, Precision Engineering 32, 329–335, 2008. 6. K. Jemielniak, Commercial tool condition monitoring systems, International Journal of Advanced Manufacturing Technology, 15, 711–721, 1999. 7. H. Y. Kim, J. H. Ahn, S. H. Kim, S. Takata, Real-time drill wear estimation based on spindle motor power, Journal of Materials Processing Technology, 124, 267–273, 2002. 8. Y. T. Oh, W. T. Kwon, C. N. Chu, Drilling torque control using spindle motor current and its effect on tool wear, International Journal of Advanced Manufacturing Technology, 24, 327–334, 2004. 9. S. Garg, K. Patra, S. K. Pal, D. Chakraborty, Effect of different basis functions on a radial basis function network in prediction of drill flank wear from motor current signals, Soft Computing, 12 (8), 777-787, 2008. 10. K. Patra, S. K. Pal, K. Bhattacharyya, Fuzzy radial basis function (FRBF) network based tool condition monitoring system using vibration signals, Machining Science and Technology, 14 (2), 280-300, 2010. 11. K. Patra, Acoustic emission based tool condition monitoring system in drilling, Proceedings of the World Congress on Engineering, 3, 6-8, 2011. 12. S. Dutta, A. Datta, N. Das Chakladar, S. K. Pal, S. Mukhopadhyay, R. Sen, Detection of tool condition from the turned surface images using an accurate grey level co-occurrence technique, Precision Engineering, 36, 458–466, 2012. 13. E. Kondo, R. Kamo, H. Murakami, Monitoring of burr and prefailure phase caused by tool wear in micro-drilling operations using thrust force signals, Journal of Advanced Mechanical Design, Systems, and Manufacturing, 6(6), 885-897, 2012. 14. T. Terabayashi, Y. Daikoh, Y. Maeda, M. Masuda, Microdrilling of printed circuit boards (3rd report) - non-contact measurement of small cutting torque sign deformation of thin-walled elastic cylinder, Journal of The Japan Society of Precision Engineering, 67(6), 927–931, 2001. 15. T. Terabayashi, Y. Daikoh, Y. Maeda, M. Masuda, Micro-drilling of printed circuit boards (4th report) - a breakage prevention drilling using a quick stop mechanism of drill feed, Journal of the Japan Society of Precision Engineering, 68(4), 556–560, 2002.
16. Y.S. Tarng, B.Y. Lee, Amplitude demodulation of the induction motor current for the tool breakage detection in drilling operations, Robotics and Computer Integrated Manufacturing, 15, 313–318, 1999. 17. K. Patra, S. K. Pal, K. Bhattacharyya, Artificial neural network based prediction of drill flank wear from motor current signals, Applied Soft Computing, 7, 929-935, 2007. 18. L. Fu, S.-F. Ling, C.-H. Tseng, On-line breakage monitoring of small drills with input impedance of driving motor, Mechanical Systems and Signal Processing, 21, 457–465, 2007. 19. T. I. Ogedengbe, R. Heinemann, S. Hinduja, Feasibility of tool condition monitoring on micromilling using current signals, Journal of Technology, 14(3), 161-172, 2011. 20. W-H. Hsieh, M-C. Lu, S-J. Chiou, Application of back propagation neural network for spindle vibration-based tool wear monitoring in micro-milling, International Journal of Advanced manufacturing Technology, 61, 53-61, 2012. 21. C-R. Huang, M-C. Lu, C-E. Lu, Y.-W. Hsu, Study of spindle vibration signals for tool breakage monitoring in micro-drilling, Proceedings of the 8th World Congress on Intelligent Control and Automation, Taipei, Taiwan, June 21-25, 2011. 22. M. Malekian, S. S. Park, M. B. G. Jun, Tool wear monitoring of micro-milling operations, Journal of Materials Processing Technology, 209, 4903–4914, 2009. 23. I. S. Kang, J. S. Kim, M. C. Kang, K. Y. Lee, Tool condition and machined surface monitoring for micro-lens array fabrication in mechanical machining, Journal of Materials Processing Technology, 201, 585–589, 2008. 24. K. A. Bourne, S. G. Kapoor, Process Monitoring During Micro-Drilling via Acoustic Emission, Ultrasonic Sound, and Spindle Load Sensors, ASME 2012 International Manufacturing Science and Engineering Conference, Notre Dame, Indiana, USA, June 4–8, 2012, pp. 781-790. 25. E. Jantunen, A summary of methods applied to tool condition monitoring in drilling, International Journal of Machine Tools and Manufacture, 42(9), 997–1010, 2002. 26. B. Sick, Online and indirect tool wear monitoring in turning with artificial neural networks: a review of more than a decade of research, Mechanical Systems and Signal Processing, 16 (4), 487–546, 2002. 27. P. P. Sam, A. S. Varadarajan, A multi-sensor fusion model based on artificial neural network to predict tool wear during hard turning, Journal of Engineering Manufacture, 226(5), 853-860, 2012. 28. J. Xu, K. Yamad, K. Seikiy, R. Tanaka, Y. Yamane, Effect of different features to drill-wear prediction with backpropagation neural network, Precision Engineering, 38, 791–798, 2014. 29. A. Aramcharoen, P. T. Mativenga, Size effect and tool geometry in micromilling of tool steel. Precision Engineering, 33, 402–407, 2009. 30. F.Vollertsen, D. Biermann, H.N. Hansen, I. S. Jawahir, K. Kuzman, Size Effects in Manufacturing of Metallic Components, CIRP Annals - Manufacturing Technology, 58, 566–587, 2009. 31. R. S. Anand, K. Patra, M. Steiner, Size Effect in Micro Drilling of Carbon Fiber Reinforced Plastic Composite, Production Engineering Research and Development, 8, 301-307, 2014.
32. I. Tansel, O. Rodriguez, M. Trujillo, E. Paz, W. Li, Micro-end-milling—I. wear and breakage, Int International Journal of Machine Tools and Manufacture, 38, 1419– 1436, 1998. 33. I. Tansel, T. Arkan, W. Y. Bao, N. Mahendrakar, B. Shisler, D. Smith, M. McCool, Tool wear estimation in micro-machining part II: neural-network-based periodic inspector for non-metals. International Journal of Machine Tools and Manufacture, 40, 609–620, 2000. 34. G. Beruvides, R. Quiza, R. D. Toro, R. E. Haber, Sensoring systems and signal analysis to monitor tool wear in microdrilling operations on a sintered tungsten-copper composite, Sensors and Actuators A: Physical, 199, 163-175, 2013. 35. M. Imran, P. T. Mativenga, P. J. Withers, Assessment of machining performance using the wear map approach in micro-drilling, International Journal of Advanced manufacturing Technology, 59, 119-126, 2012. 36. R. S. Anand, K. Patra, Modeling and simulation of mechanical micro-machining—a review, Machining Science and Technology, 18 (3), 323-347, 2014. 37. RS Anand, K Patra, Extracting Specific Cutting Force Coefficients in Micro Drilling with Tool Edge Radius Effects, Applied Mechanics and Materials, 799, 256-260, 2015. 38. L. V. Fausett, Fundamentals of neural networks, 1st edition. Prentice Hall, 1994. 39. B. M. Wilamowski, S. Iplikci, O. Kaynak, M. O. Efe, An algorithm for fast convergence in training neural networks, International Joint conference on Neural Networks (IJCNN’01), Washington DC. 1778-1782, 2001.
(a) The Hembrug machine tool
(b) The drilling environment with the sensors
Hartner HSS-E-PM drill d1 = 0.5 mm, l2 = 3.4 mm, d2 = 1 mm
(c)The tool and its important parameters Figure 1. Experimental setup and tool details
(a)
(b)
(c)
(d)
Figure 2. Images of 0.5 mm drill bit (a) before first hole, (b) after 30 holes, (c) after 60 holes, and (d) after 90 holes; for cutting speed: 16 m/min; spindle speed:10185.9 rpm; and feed: 0.001 mm/rev.
Figure 3. Chisel wear v/s hole number (feed: 0.001 mm/rev; cutting speed: 16 and 30 m/min)
Figure 4. Variations of chisel wear with number of drilled holes for different feed (at constant cutting speed of 30 m/min)
Figure 5. Variations of chisel wear with number of drilled holes for different speed (at constant feed of 0.001 mm/rev)
Figure 6. A typical profile of thrust force for different steps of a micro mechanical peck drilling ( drill diameter: 0.5 mm; feed: 0.002 mm/rev; cutting speed: 23 m/min).
(a) First step
(b) Second step
(c) Third step Figure 7. Variations of average thrust force with number of drilled holes for different steps of peck drilling (feed: 0.001, 0.002, and 0.003 mm/rev; cutting speed: 23 m/min).
(a) First step
(b) Second step
(c) Third step Figure 8. Variations of maximum thrust force with number of drilled holes for different steps of peck drilling (feed: 0.001, 0.002, and 0.003 mm/rev; cutting speed: 23 m/min).
(a). First step
(b) Second step
(c) Third step Figure 9. Variations of average thrust force with number of drilled holes for different steps of peck drilling (feed: 0.001 mm/rev; cutting speed: 16, 23, and 30 m/min).
(a). First step
(b) Second step
(c) Third step Figure 10. Variations of maximum thrust force with number of drilled holes for different steps of peck drilling (feed: 0.001 mm/rev; cutting speed: 16, 23, and 30 m/min).
Figure 11. Structure of ANN model with nine input and one output variables
(a)
(b)
. (c)
(d)
(e) Figure 12. MSE of ANN model with number of hidden neurons (a) 5, (b) 10, (c) 15, (d) 20 and (e) 25
Figure 13. Regression plot when number of hidden neuron is 10
Figure 14. Graph between predicted and target hole numbers (normalized) using ANN and Regression model.
Figure 15. ANN prediction of hole number (normalized) from training dataset and new testing dataset
Figure 16. Regression model prediction of hole number (normalized) from training dataset and new testing dataset
Table 1. Ranges of cutting speed, spindle speed and feed. Cutting Speed (m/min) Spindle Speed (rpm) Feed (mm/rev)
16,23,30 10185.9, 14642.3, 19098.6 0.001, 0.002, 0.003
Table 2. Cutting conditions with their corresponding hole numbers and chisel wear. Cutting condition 1 2 3 4 5 6 7 8 9
cutting speed (m/min) 16 16 16 23 23 23 30 30 30
spindle speed (rpm) 10185.9 10185.9 10185.9 14642.3 14642.3 14642.3 19098.6 19098.6 19098.6
feed (mm/rev) 0.001 0.002 0.003 0.001 0.002 0.003 0.001 0.002 0.003
hole no. 1-90 91-180 181-270 271-360 361-450 451-540 541-630 631-720 720-810
Table 3 MSE with variation of hidden neurons Number of hidden neurons
Mean square error of validation data 0.00048585 0.00037803 0.0004883 0.00053463 0.00069252
5 10 15 20 25
Table 4. Comparison between MSE and R2 for ANN and Regression model. Model ANN Regression
MSE 0.00040 0.00061
R2 0.995 0.992
Table 5. Comparison of ANN and Regression model prediction for new test data Model
Mean prediction error (%) CC#1
ANN Regression
5.7 6.3
CC#2 3.2 4.6
Max. chisel wear (mm) 0.020 0.020 0.015 0.022 0.023 0.021 0.029 0.019 0.018
S0141-6359(16)30447-0 http://dx.doi.org/doi:10.1016/j.precisioneng.2016.12.011 PRE 6505
To appear in:
Precision Engineering
Received date: Revised date: Accepted date:
27-11-2014 26-6-2016 30-12-2016
Please cite this article as: Patra K, Jha AK, Szalay T, Ranjan J, Monostori L.Artificial neural network based tool condition monitoring in micro mechanical peck drilling using thrust force signals.Precision Engineering http://dx.doi.org/10.1016/j.precisioneng.2016.12.011 This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
Artificial neural network based tool condition monitoring in micro mechanical peck drilling using thrust force signals K. Patra1*, A. K. Jha1, T. Szalay2, J. Ranjan1 and L. Monostori2 1
Mechanical Engineering department, Indian Institute of Technology Patna, Bihta, Patna-801103, India Department of Manufacturing Science and Engineering, Budapest University of Technology and Economics, H-1111, Budapest, Hungary 2
*Corresponding author: Email: [email protected], ph. +91-612-2552012
Highlights
Hole numbers represent drill wear stages well.
Prediction of hole number gives tool wear status of the micro-drill.
Higher feed and lower cutting speed values improve tool life. Artificial neural network accurately predicts hole number from thrust force. The prediction values are in good agreement with experimental values.
Abstract Micro scale machining process monitoring is one of the key issues in highly precision manufacturing. Monitoring of machining operation not only reduces the need of expert operators but also reduces the chances of unexpected tool breakage which may damage the work piece. In the present study, the tool wear of the micro drill and thrust force have been studied during the peck drilling operation of AISI P20 tool steel workpiece. Variations of tool wear with drilled hole number at different cutting conditions were investigated. Similarly, the variations of thrust force during different steps of peck drilling were investigated with the increasing number of holes at different feed and cutting speed values. Artificial neural network (ANN) model was developed to fuse thrust force, cutting speed, spindle speed and feed parameters to predict the drilled hole number. It has been shown that the error of hole number prediction using a neural network model is less than that using a regression model. The prediction of drilled hole number for new test data using ANN model is also in good agreement to experimentally obtained drilled hole number. Keywords: Micro-drilling, peck drilling, tool breakage, thrust force, tool wear, artificial neural network, regression analysis
1. Introduction In the modern era, there has been a continuing trend towards miniaturization for better performance, higher quality and less expensive products in various industries such as aerospace, biomedical, electronics, communication and automotive [1]. Successful miniaturization of products requires high precision micromachining methods for producing micro-features including micro-holes [2]. In general, laser beam machining, ultrasonic machining, electro-chemical machining, electrical discharge machining and mechanical micro-drilling are used to produce micro-holes. Among these methods, mechanical micro-drilling is gaining wide acceptability due to its capability to produce good quality micro-holes in most of the engineering materials in a short processing time [3]. Mechanical micro-drilling can be applied for producing micro-holes in various mechanical, electronic, optical, ornamentals and micro-fluidic device components made of metal/alloys, composites and plastics. However, tool wear of the micro-drill during the micro-drilling operations is one of the critical issues, since it considerably affects the quality of the machined parts. That’s why monitoring of tool wear during the machining operation and prediction of micro-drill breakage is very important to maintain high hole quality and improve the breakage life of the drill and efficiency of the micro-drilling process [4-5].
The monitoring and prediction of tool wear and breakage in macro-machining are mainly done indirectly through thrust force, torque, current, spindle power, vibration, acoustic emission measurements [6-11] and through machined surface texture measurement using machine vision [12]. Attempts have also been made in micro-domain machining to apply indirect methods for tool wear and breakage monitoring. Kim et al. [3] applied thrust force monitoring to increase tool life before drill failure in deep-micro-hole drilling. Kondo et al. [13] also applied thrust force signals to monitor burr and pre-failure phase in microdrilling of stainless steel. Apart from thrust force signal, torque signal was also monitored to prevent the tool breakage in micro-drilling [14-15]. However, it was shown that magnitude of torque was very small compared to thrust force and special non-contact measuring equipment would be required for torque signal for its effectiveness in micro-tool wear monitoring applications. Input current signatures which were successfully applied for tool breakage [16] and wear detection [17] in macro-drilling operations were shown to be less effective than the spindle motor input impedance signatures for detection of drill breakage in micro-drilling [18]. However, Ogedengbe [19] showed the feasibility of tool condition monitoring of micro-milling using time-domain and frequency-domain features of current signatures. Workpiece vibration [20] and spindle motor vibration [21] signals were also shown to be effective for classification of tool wear in micro-milling and classification of tool breakage in micro-drilling, respectively. However, vibration signals in micro-scale cutting can be influenced by external noise due to its lower bandwidth [22]. Acoustic emission (AE) signals, which were less affected by external noise due to their higher bandwidth, were demonstrated tool condition with a high degree of confidence in micromachining [22-24]. Among the aforementioned various sensing methods, cutting force provides the most effective method for monitoring tool condition in a micro-machining process due to its higher signal to noise ratio and its best representation of the state of the machining process [22]. For effective monitoring of tool condition in micro-machining, it is essential to apply a viable, reliable and accurate modeling technique which can establish highly nonlinear relations of cutting conditions, process signals and tool states. Artificial neural network (ANN) based modeling approaches were considered to be one of the most viable, reliable and attractive approaches for tool condition monitoring in machining processes [25-28]. Even though there are plenty of examples of the applications of ANN to relate cutting forces with tool wear in macro-scale machining, these models can not be used directly in micro-scale tool condition monitoring as cutting forces are influenced by tool size effects at low feed micro-cutting [29-31]. Therefore, ANN models for micro-machining tool condition monitoring should be trained with signals obtained during micro-cutting operations. Tansel et al. [32-33] predicted the wear of micro-end mill from cutting force measurement and artificial neural network models. Malekian et al. [22] applied multi-sensors signals including cutting force, vibration, acoustic emission to monitoring tool wear in micro-milling based on ANN and fuzzy algorithms. Hsieh et al. [20] applied back propagation neural
network to relate spindle vibration features to tool wear in micro-milling process. Fu et al. [18] used an ANN model to recognize the waveform of the motor impedance to identify drill breakage. Beruvides et al. [34] applied neural network and neuro-fuzzy systems to correlate the cutting force and vibration signal features with the number of drilled holes. As the number of drilled holes can represent different stages of tool wear and tool breakage for a particular cutting condition [35], correlating sensor features with the number of drilled holes will actually indicate the condition of the micro-drill. However, Beruvides et al. [34] analyzed the cutting force and vibration signals only for five numbers of drilled holes in each of the three different cutting conditions. As no correlation between tool wear and hole number was established in their work, it is not clear how mere five numbers of drilled hole can represent all the stages of drill wear in a cutting condition. For effective tool condition monitoring for micro-drilling with ANN modeling technique, it is imperative to use process signal data for sufficient number of drilled holes covering the maximum portions of the drill wear states. Considerable numbers of drilled holes are produced to obtain significant amount of tool wear under each cutting condition. The main goal of this work is to develop an ANN model to predict the number of drilled holes using thrust force and cutting parameters of a micro mechanical peck drilling process. By predictiong the hole number, one can easily determine the condition of the tool and remaining tool life at that cutting condition. For some machining applications where micro-drills are reused after some interval, the proposed monitoring system can predict numbers of holes which were drilled prior to the present operation and the remaining tool life in any cutting condition. The performance of the ANN model is compared to that of regression model which is developed from the same experimental data sets used by the ANN model. Finally the ANN model is applied to new test data to predict the tool condition successfully, as a validation of the proposed micro-drill condition monitoring technique.
2. Experimental Procedure 2.1 Experimental set-up and cutting force data acquisition The micro-drilling experiments were performed on a high stiffness Hembrug Slantbed Microturn 50 as shown in Fig. 1(a). It was equipped with a GMN precision high revolution spindle (max. speed: 120 000 rpm, maximum power: 1.2 kW). The positioning accuracy of this machine tool is within 100 nanometer, and because of the granite bed, the damping and heat resistance is very good. The consequence of the machine structure, the drilling direction was horizontal that is not usual in drilling (Figure 1(b)). However, the very smooth and accurate movement in this machine tool may exclude that causing reason of tool breakage. Hartner high speed steel drills of 0.5 mm diameter as shown in figure 1(c) are used to produce blind holes in tool steel workpiece material (AISI P20 + S – C 0.4; Mn 1.5; Cr 1.9; Mo 0.2; S 0.05) which is regarded as well machinable mold steel for machine-cut plastic molds and zinc die casting
dies. The depth of the drill holes were 1.5 mm and peck drilling strategy (with 0.5 mm step size) was applied with air cooling during drilling. In the experiments, the cutting parameters were varied in the range suggested by the tool producer. For measuring the thrust force in the peck drilling process, a Kistler 9257B type 3-component dynamometer and a Kistler 5019B type charge amplifier were used. The data was transferred to a computer using National Instruments 6024E DAQ Card at a sampling rate of 1000 Hz. A self-developed program in LabVIEW was used to acquire and archive the thrust force data in each peck drilling. The drilling process and the tool wear were inspected using digital microscope (Dino-Lite AM3013T with DinoCapture 2.0). 2.2 Design of Experiments A full factorial design of experiments with two factors (cutting speed and feed) and three levels were considered in this work. Table 1 shows three levels of cutting speed (16, 23 and 30 m/min) and their corresponding spindle speed (in rpm), and three levels of feed (0.001, 0.002 and 0.003 mm/rev). We repeated the drilling 90 times using a single micro-drill in each of total nine combinations of cutting parameters (feed and speed) to obtain substantial amount of drill wear. The maximum cutting force and the average cutting force for each step of peck drilling were measured during each experiment. 2.3 Measurement of drill wear Due to scaling down of the different components of the micro-cutting tool and its complex geometry, tool wear measurement is very challenging. Unlike macro-drilling process, it is not convenient to take out the micro-drill bit from the machine after each drilling as small change in setting may influence the drill wear process. Moreover, it is also not advisable to interrupt the operation after each drilling. In this work, images of micro-drill edges were taken by digital microscope (Dino-Lite AM3013T with DinoCapture 2.0) after each 30 drilled holes for a particular cutting condition. Figure 2 shows the images of fresh and used drill bit after different number of drilled holes. Tool flank wear and chisel edge wear can be observed in these images. Drill wear values were measured from digital image processing. However, the adherence of chips on the cutting edges makes measurement of tool wear very challenging. From the measurements of flank wear and chisel edge wear, it has been observed that progression of chisel edge wear against drilled hole numbers are more significant compared to that of flank wear for all the cutting conditions considered in this work.
3. Experimental results
3.1 Variation of tool wear with hole number and cutting conditions Figure 3 shows progression of chisel edge wear of micro-drills with increasing number of drilled holes over entire cutting periods under two different cutting speeds (16 m/min and 30 m/min) for feed value of 0.001 mm/rev. The evolution of chisel wear was found to follow similar three-stage process as described by Imran et al. [35]. During stage 1, tool wears rapidly as represented by high slope. Tool wear increased steadily in stage 2 which was indicated by lower slope in the chisel wear curve than the stage 1. Stage 3 is represented by rapid increase of the slope in the wear curve and catastrophic tool failure. However, tool wear rate at higher speed (30 m/min) was higher due to rapid heating that resulted to the shifting of the tool wear stages, higher tool wear and finally failure at lower number of drilled holes. Effects of feed and cutting speed were shown in Fig. 4 and Fig. 5, respectively. Initially, tool wear at different feed and speed did not differ much. However with increasing drilled hole numbers, tool wear was more for low feed and high speed conditions. Higher tool wear values at low feed and high speed in micro-drilling are driven by ploughing effects and rapid heating, respectively [35]. 3.2 Variation of thrust force with hole number and cutting conditions A typical thrust force profile of the micro peck drilling is shown in Fig. 6. In the thrust force profile, average force and maximum force of each step are also shown. Standard deviation of thrust force in each step is also shown with error bar having equal spread above and below the average value of the thrust force. The fluctuation of thrust force data during step drilling may be contributed by the tool runout and microdrilling size effects such as tool edge radius effects and cutting edge rubbing with the workspiece or ploughing [36-37]. It can be seen that there is negligle fluctuaion of measurement data during no cutting time. Therefore, measurement error can be considered as negligible. The variations of average thrust forces with hole numbers for each step at different feed values and constant cutting speed (23 m/min) are shown in Fig. 7(a)-(c). Average thrust force for initial holes is higher for higher feed rate as chip load is higher. However, for low feed rate, with increase of number of holes, tool wear increases rapidly due to ploughing and rapid heating that contribute to increase of average thrust force. Therefore, at higher number of holes, difference of average cutting force of 0.003 mm/rev feed with that of other feed reduces and in few instances, average thrust force may be more for low feed conditions. Variations of maximum thrust force with hole numbers for each peck drilling step at different feeds and constant speed (23 m/min) are shown in Fig. 8 (a)-(c). Compared to average thrust force, maximum value of thrust force is more affected by some other factors such as chip entanglement on cutting edge, variation of drill characteristics and non homogeneous structure of workpiece at microlevel, etc. Therefore, the difference of the maximum force of 0.003 mm/rev feed rate from that of other feed rates is not much. Similarly, variations of average and maximum thrust force values with drilled hole
number for each peck drilling step at different cutting speed and constant feed (0.001 mm/rev) are shown in Fig. 9(a)-(c) and Fig. 10(a)-(c), respectively. Both average and maximum thrust force values during each step follow increasing trends with number of drilled holes. Average and maximum thrust force values also increase with increasing cutting speed due to increasing heating as shown by Fig. 9(a)-(c) and 10(a)-(c), respectively.
4. Prediction of tool condition using ANN model 4.1 Development of ANN model An artificial neural network (ANN) is a mathematical or computational model that is inspired by the structure and/or functional aspects of biological neural networks [38]. In most cases, an ANN is an adaptive system that changes its structure based on external or internal information that flows through the network during the learning phase. Modern neural networks are usually used to model complex relationships between inputs and outputs or to find patterns in data. The Matlab neural fitting tool is used in this work to develop the neural network. Figure 11 shows the structure of the artificial neural network which consists of three layers: input, hidden and output layers. There are nine input variables (cutting speed, feed rate, spindle speed, first step average thrust force, second step average thrust force, third step average thrust force, first step maximum thrust force, second step maximum thrust force, and third step maximum thrust force) to the input layer. ‘Hole number’ is the only output variable from the output layer. The neurons of one layer are interconnected to the neurons of the previous and successive layers through weighted links. The network has been trained with Levenberg-Marquardt back propagation algorithm [39]. Levenberg–Marquardt algorithm is a combination of two training processes: steepest descent algorithm and Gauss–Newton algorithm. Around the area with complex curvature, the Levenberg–Marquardt algorithm switches to the steepest descent algorithm, until the local curvature is proper to make a quadratic approximation; then it approximately becomes Gauss–Newton algorithm to speed up the convergence significantly [34].
4.2 Training, validation and testing of ANN model Dataset for training, validation and testing of the ANN model consists of 810 numbers of patterns from total 810 experiments from 9 cutting conditions (90 holes x 9 cutting conditions) as shown in Table 2. Each pattern has nine input parameters and one target output parameter as mentioned in the earlier subsection 4.1. As hole number can represent the tool wear for a cutting condition, target output parameter considered in this work is the hole number. Input and output parameters have been normalized between 0 and 1. Normalization among the data is based on the maximum and mimimum values of each
parameter in the whole dataset. Following equation is used to determine the normalize value (xnorm) of any input or output parameter.
xnorm = (x- xmin)/(xmax-xmin)
(1)
where, x is the value of any parameter; xmin = minimum value of that parameter in the dataset; xmax = maximum value of that parameter in the dataset. The data has been divided randomly into 70% for training, 15% for validation, and 15% for testing of the ANN model. Data in each dataset covered all 9 cutting conditions shown in Table 2. The training data set is used for computing the gradient and updating the network weights and biases. The validation data set is applied to monitor the training process. The error of validation data normally decreases during the initial phase of training, as does the training set error. However, when the network begins to overfit the data, the error on the validation set typically begins to rise. The network weights and biases are saved at the minimum of the validation set error. The test data set error is used to compare the performances of different architectures of the ANN model. Mean square errors (MSE) of different architectures of ANN model, i.e., with different numbers of hidden neurons are compared to determine the best ANN architecture for this given problem. Figure 12 (a)-(e) show variations of MSE of training, validation and testing data with respect to number of epoch for different numbers of hidden neurons. Table 3 shows the variation of MSE of validation data with number of hidden neurons. Maximum performance has been obtained when number of hidden neuron in hidden layer is 10 and minimum value of MSE of validation is 0.00037803 at epoch 15 as shown in Fig.12 (b). Hence, the ANN architecture of 9-10-1 (9 input neurons, 10 hidden neurons and 1 output neuron) has been selected as the best architecture among the architechtures considered in this work. The network outputs i.e., predicted hole numbers, and regression plots of training, validation and testing datasets have been generated and shown in Fig.13. The network output (Y) values are very close to the target (T) values in all training, validation and testing cases. The regression plots of these data are overlapping to Y=T line.
4.3 Comparison of the ANN model with a regression model The number of drilled holes has been predicted using regression analysis for comparison to the performance of the ANN model prediction. Relation of output variable with input variables is established using the following regression equation (2):
(2)
where y p is the predicted value for n number of input variable x and target variable y. Coefficients b are obtained by minimizing sum square error of predicted (y p) and target value y for a set of m input-output pair. In this work, values of m and n are 810 and 9, respectively. The inputs and output parameters of regression model were same as ANN modeling parameters. The regression model fitting was done using Matlab statistical toolbox. The mean square error (MSE) and R2 values obtained from regression model are 0.00061 and 0.992, respectively. Table 4 shows the comparison between ANN and regression models in terms of MSE and R2 values. Figure 14 shows the plot between the experimental values and predicted values using both regression model and ANN model. From this plot, it can be seen that the ANN model predicted values are closer to the target values compared to those of the regression model.
4.4 Validation of the developed ANN model at new cutting conditions To validate the applicability of the developed ANN model, the trained network has been applied to test data from two new cutting conditions with same tool-workpiece material and same machining environment as discussed in section 2. For cutting condition 1 (CC#1), selected cutting speed and feed were 23 m/min and 0.0025 mm/rev, respectively. Whereas, selected cutting speed and feed were 35 m/min and 0.0035 mm/rev, respectively, in cutting condition 2 (CC#2). The peck drilling of three steps were repeated 90 times using a single micro-drill in each cutting condition to obtain substantial amount of drill wear. The average thrust force and maximum thrust force for each step of peck drilling were measured during each experiment. The normalization of cutting parameters and sensor signal features of both cutting conditions (CC#1 and CC#2) were normalized with respect to the earlier data from 9 cutting conditions reported in section 2. As the values of the cutting parameters in CC#1 are within the values of those considered in the earlier 9 cutting conditions, the normalized values of input and target values of CC#1 lie within 1. However, the values of the cutting parameters considered in CC#2 are outside the ranges of the earlier 9 cutting conditions. So the values of the normalized cutting parameters and the target parameter are obtained beyond 1. The simulation of the trained ANN model for these two new cutting conditions were performed in Simulink block driagram environment in Matlab neural network toolbox. The values of mean error of hole number prediction in CC#1 and CC#2 are 5.7% and 3.2%, respectively. Figure 15 shows the plot between the normalized predicted hole numbers and the normalized target hole numbers for all data sets, i.e., 810 data from 9 cutting conditions, 90 data from CC#1 and 90 data from CC#2. From this plot, it can be seen that the ANN model predicted values are more closer to the target values when data is taken from the same dataset used for training the neural
network than those of the data taken from completely new cutting conditions, i.e., CC#1 and CC#2. However, the mean prediction error values in these two new cutting conditions may be acceptable considering the high nonlinearity of the cutting force signals due to size effects in micro-drilling process. Regression model has also been applied for prediction of hole numbers of two new cutting conditions, i.e., CC#1 and CC#2. Figure 16 shows the plot between the normalized predicted hole numbers using regression model and the normalized target hole numbers for all data sets, i.e., 810 data from 9 cutting conditions, 90 data from CC#1 and 90 data from CC#2. Table 5 shows the comparison of the prediction performances of ANN model and regression model for CC#1 and CC#2 cutting conditions. The values of mean error of hole number prediction using regression model in CC#1 and CC#2 are 6.3% and 4.6%, respectively. It has been also found the error of hole number prediction for new cutting conditions using a ANN model is less than that using a regression model.
5. Conclusion In the present study, tool wear behavior of micro-drill with drilled hole numbers and cutting conditions of a micro mechanical peck drilling process were investigated. Later on an ANN model was developed to predict the number of drilled holes using thrust force and cutting conditions as input parameters. The performance of the ANN model was compared to that of regression model which was developed from the same experimental data sets used by the ANN model. Further the ANN model was applied to new cutting conditions for validating its tool condition predition performance. From this present study, the following conclusions were drawn: The progress of chisel wear in micro peck drilling of tool steel follows the similar three stages of tool wear as reported earlier. Stage 1 represents rapid tool wear and is followed by near uniform wear rate in stage 2. Finally, rapid tool wear and tool breakage is observed in stage 3. However, this study shows the shifting of each stage with change in cutting condition. The identification of the onset of stage 3 at different cutting condition should help to avoid breakage of tool during the cutting process. Results suggest that higher feed and lower cutting speed values should be undertaken for improving tool life by reduce ploughing effect and thermal effect, respectively. Experimental results show that hole number can represent drill wear stages well. The costly and process-interfering drill wear measurement after each drill hole can be avoided by the method of predicting hole number. Prediction of hole number for different cutting conditions using thrust force data indirectly gives tool wear status of the micro-drill.
The prediction results show that artificial neural network (ANN) is one of the accurate and viable techniques for predicting drilled hole number from the average thrust force, maximum thrust force and cutting conditions used in mechanical micro peck drilling. It has been also found the error of hole number prediction using a neural network model is less than that using a regression model. The prediction values of the ANN model with new experimental data were also in good agreement. The mean prediction error lies in the range of 3-6%, which can be considered to be acceptable considering the uncertainty and size effects in the microdrilling process. Hence, the proposed methodology can be successfully applied to tool condition monitoring of micro-drilling process at wide ranges of cutting conditions.
Acknowledgement This article is a part of the bilateral research project “Multi-sensors based intelligent tool condition monitoring in mechanical micro-machining” supported jointly by project no. INT/HUN/P-01/2012 and project number TÉT_10-1-2011-0233 funded by DST, Govt. of India and NIH, Govt. of Hungary, respectively.
References 1. I. S. Kang, J. S. Kim, M. C. Kang, K. Y. Lee, Tool condition and machined surface monitoring for micro-lens array fabrication in mechanical machining, Journal of materials processing technology, 201, 585–589, 2008. 2. J-S. Moon, H-S. Yoon, G-B. Lee, S-H Ahn, Effect of backstitch tool path on micro-drilling of printed circuit board, Precision Engineering 38, 691–696, 2014. 3. D. W. Kim, Y. S. Lee, M. S. Park, C. N. Chu, Tool life improvement by peck drilling and thrust force monitoring during deep-micro-hole drilling of steel, International Journal of Machine Tools & Manufacture, 49, 246–255, 2009. 4. J. Chae, S. S. Park, T. Freiheit, Investigation of micro-cutting operations, International Journal of Machine Tools & Manufacture, 46, 313–332, 2006. 5. H. Watanabe, H. Tsuzaka, M. Masuda, Microdrilling of printed circuit boards (PCBs) – Influence of radial run-out of microdrills on hole quality, Precision Engineering 32, 329–335, 2008. 6. K. Jemielniak, Commercial tool condition monitoring systems, International Journal of Advanced Manufacturing Technology, 15, 711–721, 1999. 7. H. Y. Kim, J. H. Ahn, S. H. Kim, S. Takata, Real-time drill wear estimation based on spindle motor power, Journal of Materials Processing Technology, 124, 267–273, 2002. 8. Y. T. Oh, W. T. Kwon, C. N. Chu, Drilling torque control using spindle motor current and its effect on tool wear, International Journal of Advanced Manufacturing Technology, 24, 327–334, 2004. 9. S. Garg, K. Patra, S. K. Pal, D. Chakraborty, Effect of different basis functions on a radial basis function network in prediction of drill flank wear from motor current signals, Soft Computing, 12 (8), 777-787, 2008. 10. K. Patra, S. K. Pal, K. Bhattacharyya, Fuzzy radial basis function (FRBF) network based tool condition monitoring system using vibration signals, Machining Science and Technology, 14 (2), 280-300, 2010. 11. K. Patra, Acoustic emission based tool condition monitoring system in drilling, Proceedings of the World Congress on Engineering, 3, 6-8, 2011. 12. S. Dutta, A. Datta, N. Das Chakladar, S. K. Pal, S. Mukhopadhyay, R. Sen, Detection of tool condition from the turned surface images using an accurate grey level co-occurrence technique, Precision Engineering, 36, 458–466, 2012. 13. E. Kondo, R. Kamo, H. Murakami, Monitoring of burr and prefailure phase caused by tool wear in micro-drilling operations using thrust force signals, Journal of Advanced Mechanical Design, Systems, and Manufacturing, 6(6), 885-897, 2012. 14. T. Terabayashi, Y. Daikoh, Y. Maeda, M. Masuda, Microdrilling of printed circuit boards (3rd report) - non-contact measurement of small cutting torque sign deformation of thin-walled elastic cylinder, Journal of The Japan Society of Precision Engineering, 67(6), 927–931, 2001. 15. T. Terabayashi, Y. Daikoh, Y. Maeda, M. Masuda, Micro-drilling of printed circuit boards (4th report) - a breakage prevention drilling using a quick stop mechanism of drill feed, Journal of the Japan Society of Precision Engineering, 68(4), 556–560, 2002.
16. Y.S. Tarng, B.Y. Lee, Amplitude demodulation of the induction motor current for the tool breakage detection in drilling operations, Robotics and Computer Integrated Manufacturing, 15, 313–318, 1999. 17. K. Patra, S. K. Pal, K. Bhattacharyya, Artificial neural network based prediction of drill flank wear from motor current signals, Applied Soft Computing, 7, 929-935, 2007. 18. L. Fu, S.-F. Ling, C.-H. Tseng, On-line breakage monitoring of small drills with input impedance of driving motor, Mechanical Systems and Signal Processing, 21, 457–465, 2007. 19. T. I. Ogedengbe, R. Heinemann, S. Hinduja, Feasibility of tool condition monitoring on micromilling using current signals, Journal of Technology, 14(3), 161-172, 2011. 20. W-H. Hsieh, M-C. Lu, S-J. Chiou, Application of back propagation neural network for spindle vibration-based tool wear monitoring in micro-milling, International Journal of Advanced manufacturing Technology, 61, 53-61, 2012. 21. C-R. Huang, M-C. Lu, C-E. Lu, Y.-W. Hsu, Study of spindle vibration signals for tool breakage monitoring in micro-drilling, Proceedings of the 8th World Congress on Intelligent Control and Automation, Taipei, Taiwan, June 21-25, 2011. 22. M. Malekian, S. S. Park, M. B. G. Jun, Tool wear monitoring of micro-milling operations, Journal of Materials Processing Technology, 209, 4903–4914, 2009. 23. I. S. Kang, J. S. Kim, M. C. Kang, K. Y. Lee, Tool condition and machined surface monitoring for micro-lens array fabrication in mechanical machining, Journal of Materials Processing Technology, 201, 585–589, 2008. 24. K. A. Bourne, S. G. Kapoor, Process Monitoring During Micro-Drilling via Acoustic Emission, Ultrasonic Sound, and Spindle Load Sensors, ASME 2012 International Manufacturing Science and Engineering Conference, Notre Dame, Indiana, USA, June 4–8, 2012, pp. 781-790. 25. E. Jantunen, A summary of methods applied to tool condition monitoring in drilling, International Journal of Machine Tools and Manufacture, 42(9), 997–1010, 2002. 26. B. Sick, Online and indirect tool wear monitoring in turning with artificial neural networks: a review of more than a decade of research, Mechanical Systems and Signal Processing, 16 (4), 487–546, 2002. 27. P. P. Sam, A. S. Varadarajan, A multi-sensor fusion model based on artificial neural network to predict tool wear during hard turning, Journal of Engineering Manufacture, 226(5), 853-860, 2012. 28. J. Xu, K. Yamad, K. Seikiy, R. Tanaka, Y. Yamane, Effect of different features to drill-wear prediction with backpropagation neural network, Precision Engineering, 38, 791–798, 2014. 29. A. Aramcharoen, P. T. Mativenga, Size effect and tool geometry in micromilling of tool steel. Precision Engineering, 33, 402–407, 2009. 30. F.Vollertsen, D. Biermann, H.N. Hansen, I. S. Jawahir, K. Kuzman, Size Effects in Manufacturing of Metallic Components, CIRP Annals - Manufacturing Technology, 58, 566–587, 2009. 31. R. S. Anand, K. Patra, M. Steiner, Size Effect in Micro Drilling of Carbon Fiber Reinforced Plastic Composite, Production Engineering Research and Development, 8, 301-307, 2014.
32. I. Tansel, O. Rodriguez, M. Trujillo, E. Paz, W. Li, Micro-end-milling—I. wear and breakage, Int International Journal of Machine Tools and Manufacture, 38, 1419– 1436, 1998. 33. I. Tansel, T. Arkan, W. Y. Bao, N. Mahendrakar, B. Shisler, D. Smith, M. McCool, Tool wear estimation in micro-machining part II: neural-network-based periodic inspector for non-metals. International Journal of Machine Tools and Manufacture, 40, 609–620, 2000. 34. G. Beruvides, R. Quiza, R. D. Toro, R. E. Haber, Sensoring systems and signal analysis to monitor tool wear in microdrilling operations on a sintered tungsten-copper composite, Sensors and Actuators A: Physical, 199, 163-175, 2013. 35. M. Imran, P. T. Mativenga, P. J. Withers, Assessment of machining performance using the wear map approach in micro-drilling, International Journal of Advanced manufacturing Technology, 59, 119-126, 2012. 36. R. S. Anand, K. Patra, Modeling and simulation of mechanical micro-machining—a review, Machining Science and Technology, 18 (3), 323-347, 2014. 37. RS Anand, K Patra, Extracting Specific Cutting Force Coefficients in Micro Drilling with Tool Edge Radius Effects, Applied Mechanics and Materials, 799, 256-260, 2015. 38. L. V. Fausett, Fundamentals of neural networks, 1st edition. Prentice Hall, 1994. 39. B. M. Wilamowski, S. Iplikci, O. Kaynak, M. O. Efe, An algorithm for fast convergence in training neural networks, International Joint conference on Neural Networks (IJCNN’01), Washington DC. 1778-1782, 2001.
(a) The Hembrug machine tool
(b) The drilling environment with the sensors
Hartner HSS-E-PM drill d1 = 0.5 mm, l2 = 3.4 mm, d2 = 1 mm
(c)The tool and its important parameters Figure 1. Experimental setup and tool details
(a)
(b)
(c)
(d)
Figure 2. Images of 0.5 mm drill bit (a) before first hole, (b) after 30 holes, (c) after 60 holes, and (d) after 90 holes; for cutting speed: 16 m/min; spindle speed:10185.9 rpm; and feed: 0.001 mm/rev.
Figure 3. Chisel wear v/s hole number (feed: 0.001 mm/rev; cutting speed: 16 and 30 m/min)
Figure 4. Variations of chisel wear with number of drilled holes for different feed (at constant cutting speed of 30 m/min)
Figure 5. Variations of chisel wear with number of drilled holes for different speed (at constant feed of 0.001 mm/rev)
Figure 6. A typical profile of thrust force for different steps of a micro mechanical peck drilling ( drill diameter: 0.5 mm; feed: 0.002 mm/rev; cutting speed: 23 m/min).
(a) First step
(b) Second step
(c) Third step Figure 7. Variations of average thrust force with number of drilled holes for different steps of peck drilling (feed: 0.001, 0.002, and 0.003 mm/rev; cutting speed: 23 m/min).
(a) First step
(b) Second step
(c) Third step Figure 8. Variations of maximum thrust force with number of drilled holes for different steps of peck drilling (feed: 0.001, 0.002, and 0.003 mm/rev; cutting speed: 23 m/min).
(a). First step
(b) Second step
(c) Third step Figure 9. Variations of average thrust force with number of drilled holes for different steps of peck drilling (feed: 0.001 mm/rev; cutting speed: 16, 23, and 30 m/min).
(a). First step
(b) Second step
(c) Third step Figure 10. Variations of maximum thrust force with number of drilled holes for different steps of peck drilling (feed: 0.001 mm/rev; cutting speed: 16, 23, and 30 m/min).
Figure 11. Structure of ANN model with nine input and one output variables
(a)
(b)
. (c)
(d)
(e) Figure 12. MSE of ANN model with number of hidden neurons (a) 5, (b) 10, (c) 15, (d) 20 and (e) 25
Figure 13. Regression plot when number of hidden neuron is 10
Figure 14. Graph between predicted and target hole numbers (normalized) using ANN and Regression model.
Figure 15. ANN prediction of hole number (normalized) from training dataset and new testing dataset
Figure 16. Regression model prediction of hole number (normalized) from training dataset and new testing dataset
Table 1. Ranges of cutting speed, spindle speed and feed. Cutting Speed (m/min) Spindle Speed (rpm) Feed (mm/rev)
16,23,30 10185.9, 14642.3, 19098.6 0.001, 0.002, 0.003
Table 2. Cutting conditions with their corresponding hole numbers and chisel wear. Cutting condition 1 2 3 4 5 6 7 8 9
cutting speed (m/min) 16 16 16 23 23 23 30 30 30
spindle speed (rpm) 10185.9 10185.9 10185.9 14642.3 14642.3 14642.3 19098.6 19098.6 19098.6
feed (mm/rev) 0.001 0.002 0.003 0.001 0.002 0.003 0.001 0.002 0.003
hole no. 1-90 91-180 181-270 271-360 361-450 451-540 541-630 631-720 720-810
Table 3 MSE with variation of hidden neurons Number of hidden neurons
Mean square error of validation data 0.00048585 0.00037803 0.0004883 0.00053463 0.00069252
5 10 15 20 25
Table 4. Comparison between MSE and R2 for ANN and Regression model. Model ANN Regression
MSE 0.00040 0.00061
R2 0.995 0.992
Table 5. Comparison of ANN and Regression model prediction for new test data Model
Mean prediction error (%) CC#1
ANN Regression
5.7 6.3
CC#2 3.2 4.6
Max. chisel wear (mm) 0.020 0.020 0.015 0.022 0.023 0.021 0.029 0.019 0.018