Artificial Intelligence-Based Hole Quality Prediction in Micro-Drilling Using Multiple Sensors

The prevalence of micro-holes is widespread in mechanical, electronic, optical, ornaments, micro-fluidic devices, etc. However, monitoring and detection tool wear and tool breakage are imperative to achieve improved hole quality and high productivity in micro-drilling. The various multi-sensor signals are used to monitor the condition of the tool. In this work, the vibration signals and cutting force signals have been applied individually as well as in combination to determine their effectiveness for tool-condition monitoring applications. Moreover, they have been used to determine the best strategies for tool-condition monitoring by prediction of hole quality during micro-drilling operations with 0.4 mm micro-drills. Furthermore, this work also developed an adaptive neuro fuzzy inference system (ANFIS) model using different time domains and wavelet packet features of these sensor signals for the prediction of the hole quality. The best prediction of hole quality was obtained by a combination of different sensor features in wavelet domain of vibration signal. The model’s predicted results were found to exert a good agreement with the experimental results.


Introduction
Due to the emergence of the miniaturization of products in aerospace, biomedical, communication, electronics and automotive industries, the low cost of such miniaturized products with their best-of-class quality makes them suitable for extensive use of micro-machining processes. The adopted micro-machining processes include microturning [1,2], micromilling [3][4][5][6], microdrilling [7,8]. Micromachining is generally defined as the machining process that produces miniature component or feature in of the range of 1µm to 999 µm [9]. Several non-conventional methods such as micro electric discharge machining (µEDM) [10][11][12], laser micromachining [13,14], ultrasonic [15][16][17], electron the rules of the fuzzy model and the final ANFIS model of the system is obtained. Malekian et al. [25] used features of acoustics emission, force and acceleration sensors, cutting parameters and the tool edge radius value as input to the neuro-fuzzy algorithm to estimate tool conditions (good, average, bad). Beruvides et al. [48] used ANFIS for predicting tool wear in micro-drilling.
For an effective ANFIS-based tool condition monitoring system, it is necessary to use data of process signals for a number of drilled holes up to the drill breakage for a variety of cutting conditions. In this work, the vibration signals and cutting-force signals have been applied individually and in combination to determine their effectiveness for tool condition monitoring applications in terms of hole quality and also to determine best strategies for tool condition monitoring. The time domain and wavelet features of these signals have been used in ANFIS models to predict the hole quality. Furthermore, the hole quality prediction by wavelet features of vibration signal was compared with the same by the time domain features.
The purpose of this article is to predict the quality of the holes obtained in micro-drilling through monitoring the condition of the tool and detecting the tool breakage by thrust force, torque and vibration signals using artificial intelligence.

Materials and Method
Micro-drilling experiments were conducted on CNC micro machining centre (Model DT110, Mikrotools Pte Ltd, Singapore) as shown in Figure 1. A triaxial accelerometer and a Kistler Minidyn 9256C2 dynamometer (Aachen, Germany) were used to record the vibration signal, force signal and torque signal. The sampling frequency of 10,000 Hz was set in NIPXI 1052 data acquisition system (National Instruments, Austin, TX, USA) for obtaining vibration signals. The force and torque were acquired by charge amplifier 5070A (Kistler, Aachen, Germany) and dynoware software with a sampling rate of 1000 samples/s. Solid carbide drills (manufactured by Seco Tools GmBH, Fagersta, Sweden) with diameter of 0.4 mm were applied on austenitic stainless steel workpiece to make blind hole of 1.0 mm depth. The experimental conditions are shown in Table 1. As the feed value is very small (in micrometer range) in microdrilling, high spindle speed generally is used to reduce the machining time for hole making. Anand et al. [49] used very high cutting speed (15-40 m/min corresponding to 10,000-25,000 rpm) and showed that the specific thrust force value does not vary much by increasing the cutting speed from the recommended cutting speed range of the manufacturer. Therefore, this work used cutting speed range of 12.8-24 m/min (10,186 rpm to 19,099 rpm). There is also another reason for using higher cutting speed. The higher cutting speed (spindle rpm) was shown to accelerate tool wear [50]. Therefore, by using higher cutting speed, we could get increased cutting force and vibration with less number of drilled hole. This helps us to develop the model will less number of experiments.

Time Domain Features of Process Signals
The time domain features extracted from different process signals are mainly arithmetic mean, root mean square (RMS) value, kurtosis, and standard deviation [51]. Among these features, RMS value is the most common one that represents most of the signal energy. The time domain features extracted from vibration signal is root mean square (RMS), whereas, mean value of thrust force and mean value of torque are the two other time domain features. The variations of vibration, thrust force and torque features with respect to the hole numbers before the tool breakage of different cutting condition are shown in Figure 2a-c. From Figure 2a. it is clearly seen that at constant feed rate f r = 0.003 mm/rev and different spindle speeds n = 14,186 rpm (CC5) and n = 19,099 rpm (CC4), an increase in rms of vibration Z signals is observed with respect to lower spindle speed n = 10,186 rpm (CC1). A similar trend with constant spindle speed n = 14,186 rpm and different feed rates f r = 0.002 mm/rev (CC2) and f r = 0.003 mm/rev (CC5) is also observed. In addition, for feed rate f r = 0.001 mm/rev, spindle speed n = 19,099 rpm (CC3), the rms of vibration Z signals increase after 20 number holes and increase sharply after 30 number holes. For feed rate f r = 0.002 mm/rev, spindle speed n = 14.186 rpm (CC3), rms of vibration Z signals increase after 37 number hole and increase sharply after 73 number hole. These phenomena can be explained by the growth of flank wear micro drill. Figure 2b shows that the mean torque increases with increasing hole number. However, with increasing spindle speed from n = 10,186 rpm (CC1), n = 14,186 rpm (CC5) to n = 19,099 rpm (CC4) and constant feed feed rate f r = 0.003 mm/rev, we observe a decrease in mean torque. Moreover, an increase in feed rate from f r = 0.002 mm/rev (CC2) to f r = 0.003 mm/rev (CC5) and a constant spindle speed n = 14,186 rpm on the contrary leads to an increase in mean torque. Figure 2c shows that mean thrust force tends to increase with increasing hole, but with less intensity compared to mean torque. However, with increasing spindle speed from n = 10,186 rpm (CC1), n = 14,186 rpm (CC5) to n = 19,099 rpm (CC4) and constant feed feed rate f r = 0.003 mm/rev, we observe increase in mean thrust force. A similar trend at feed rate from f r = 0.002 mm / rev (CC2) to f r = 0.003 mm / rev (CC5) and a constant spindle speed n = 14,186 rpm leads to an increase in mean thrust force. With increasing feed rate from f r = 0.001 mm/rev (CC3) to f r = 0.003 mm/rev (CC4) and a constant spindle speed n = 19,099 rpm, the trend also increases in mean thrust force. It can be seen that amplitudes of these signals increase non-linearly with the number of holes drilled. With increase in hole number, tool wear increases that contributes to increase RMS and mean values of these signals [23,45] shown in the figures. It can also be seen that RMS values of vibration signals increase suddenly just before the breakage of the tool [45]. On the other hand, the magnitudes of torque and thrust signals do not show any sudden change prior to tool breakage [23].

Wavelet Packet Features of Vibration Signals
As mean and RMS values of time domain features are determined by considering the whole signals, these are likely to be affected by the noise and disturbance present in the drilling process. To eliminate the noise and disturbances, many attempts have been made earlier to extract features in the frequency domain and the time-frequency domain [48,52]. Among different time-frequency domain signal processing methods, the most common is the wavelet packet transform which can overcome the poor time localization of a fast Fourier transform (FFT) and loss of important information due to not consideration of the detail part in case of a discrete wavelet transform (DWT). In this work, the wavelet packet transform is performed to decompose the vibration signal (up to 5 kHz) into 8 frequency bands (8 wavelet packets) up to level 3 [53]. These packets are denoted as packet (3,0), packet (3,1), packet (3,2), packet (3,3), packet (3,4), packet (3,5), packet (3,6) and packet (3,7) corresponding to frequency range of (0-0.625 kHz), (0.625-1. Wavelet packet coefficients are estimated using Daubechies 8 wavelet function. Wavelet packet feature (C ij ) which is the RMS value of the wavelet coefficients for a packet is determined to relate the condition of the tool at different cutting conditions. Here, i is the decomposition level and j stands for the packet number at the ith level. Figure 3a,b shows the time signal comparison of the first hole and 79th hole of the cutting condition (CC2). It can be observed that the amplitude of vibration for the 79th hole is very high in comparison with the same for the 1st hole as tool wear becomes severe for later case. This leads to tool failure after the 79th hole. The observation is in the similar line to that of trends reported earlier [45]. Figure 3c,d present all the 8 wavelet packets of 1st hole and the 79th hole (before the tool breakage), respectively. It can be observed that in packet (3,4), (3,5) and (3,7) the amplitude variation was very high compared with the 1st hole packet; however, the other packet's amplitude variation is comparably very low. Each wavelet packet carries information on the vibration signal in terms of wavelet coefficients. But some of the packets also contain noise or less important information depending on the process dynamics [50]. Figure 4a-e shows variations of wavelet coefficients with hole number for all 5 cutting conditions. In all 5 plots, 3 or 4 frequency bands such as C34, C35, C37 have non-linearly increasing trends as compared to other wavelet packets of the first hole to till brakage hole. The less amplitude packets contain noise and are not affected by the tool condition. Features that show increasing trends have been taken as input of soft computing techniques (i.e., ANFIS).

Hole Quality
The predictability and effectiveness of the two models are based on the hole quality. Typically, the hole roundness error is used as a measure of hole quality in drilling [54]. To determine the roundness error, maximum circles are inscribed inside and outside the hole profile and their radii are defined as outer radius and inner radius of the hole. The difference of outer radius and inner radius of the hole is called roundness error [55]. For each cutting condition, the images of all drilled holes before the drill breakage were captured by a high-resolution microscope, and its outer and inner radius of each hole are determined by digimizer software tool as shown in Figure 5a-c for the 1st, 40th and 79th holes of cutting condition CC2. The hole roundness error is calculated for individual hole and plotted against hole number for different cutting conditions as shown in Figure 5d  The roundness error value increases non-linearly with the drilled hole numbers. The nonlinear increase of roundness error with increasing hole number is related to the increase of cutting forces and vibration magnitudes. Increase in tool deflection due to change in cutting forces and vibration amplitude increases the roundness error [40]. Furthermore, it can be observed that the trend of roundness error is similar to that of the vibration with respect to the drilled hole number.

Adaptive Neuro Fuzzy Inference System (ANFIS) Model for Time Domain Signals
As the hole roundness error depends on the tool condition and cutting conditions, prediction of hole quality through input process signals (vibration, thrust and torque) at any cutting condition can tell the cutting tool condition. The decision can be taken whether the tool is still acceptable or not based on the hole quality. An adaptive neuro-fuzzy inference system (ANFIS) has been introduced to relate input process signals with hole quality parameter. The Matlab ANFIS tool has been used to develop the neuro-fuzzy network. In ANFIS model, fuzzy inference system is superimposed to the neural network. The neural network layer has input, output and a hidden layer. The hidden layer is imposed by fuzzy inference system (i.e., membership function layer and rule-based layer) and an activation function. Features extracted from different sensors and the cutting conditions were used as the inputs to this ANFIS to determine the output parameter, i.e., hole quality. Figure 6a shows the inputs, membership values, rules and combined output architecture of ANFIS model. Here, input parameters X1, X2 and X3 are the mean thrust (Fz), mean torque (Mz) and RMS vibration (Az) from time domain signals. For training, data from sensor features, i.e., RMS value of vibration signal, mean value of thrust force and torque at 4 cutting conditions (CC1 to CC4 of 175 holes) have been employed. For testing, data with same sensor features from another cutting condition data (CC5 of 24 holes) have been used. Hole quality value could be assigned as an output. Based on hole-quality prediction, we identified the tool condition as 'acceptable' or 'not acceptable'. From Figure 5, it can be observed that roundness errors of the initial drilled holes at different cutting conditions remain in the range of 30-40 µm, similar to the errors reported by Anand and Patra [54] for microdrilling with fresh cutting tools. Though with increase in hole number the roundness error increases, the value remains within 70 µm for most of cutting conditions for most of the time. Thereafter, we observe a sudden increase of roundness error till the breakage of the microdrill. Therefore, hole roundness error has been considered 'acceptable' and 'not acceptable', corresponding to hole roundness less than 70µm, and greater than 70 µm, respectively. Normalization of both input and output parameters have been conducted and these parameters lie between 0 and 1. A trapezoidal membership function has been chosen for both input and output. Figure 6b define the acceptable and not acceptable hole quality with intersection regions. The presented results can help the operator or the control system to take proper tool changing decisions. Table 2 lists the combinations of the sensor features, taken for training and testing. The minimum mean training error was 0.0639 for the data set with combination of all force, torque and vibration signal. The corresponding testing error was found to be 0.07037. Figure 6c,d compares the normalized predicted hole quality with the actual hole quality (roundness error) using training and testing data, respectively, obtained by combining mean thrust (Fz), mean torque (Mz) and RMS vibration (Az). Figure 6d also shows the hole quality of acceptable and not acceptable values for the testing data against the hole number. The deviation of some of these data comes due to the non-linear feature of data. Nonlinear features of the data arise due to the uncertainty and complexity of the micromachining process. So we developed a model based on acceptable hole quality. Even though we found deviations of some of the data points, the mean testing error of the selected multi-sensor data fusion approach is low and acceptable. Among the employed strategies of hole-quality prediction, the best performance was given by fusion of different sensor signal; this is due to fact that it takes a wide range of signal bandwidth. At a wide range of bandwidth, the features extracted are more accurate and noise free.

ANFIS Model Using Wavelet Packet-Based Vibration Features
An ANFIS model has also been developed where the inputs were high sensitive wavelet features of Z direction vibration signal at different cutting conditions and the output was the hole quality. The structure of the wavelet packet feature based ANFIS is similar to that shown in Figure 6a. For training, data of 3 highly sensitive wavelet features (C34, C35, C37) of vibration signals for 4 cutting conditions of 176 holes have been considered. For testing, data with same wavelet features of 24 holes of another cutting condition's data (CC5) have been employed. In this approach also the normalized values between 0 and 1 are used for all input and output parameters. After training, the ANFIS was ready to estimate tool condition based on hole-quality prediction with the testing data. By ANFIS model, the mean training error and the testing error of vibration signal was found as 0.054371 and 0.080652. Table 3 presents the performance comparison of ANFIS models using RMS vibration features and wavelet packet features of vibration signal in Z direction. The errors of training and testing steps are less in a wavelet packet approach as compared to a time domain approach.

Conclusions
This work develops the indirect tool-monitoring techniques for monitoring tool condition and detecting tool breakage based on hole quality. For this purpose, features were extracted from the thrust force, torque and vibration signals. The following conclusions are drawn from this work:

•
Cutting force signals, i.e., thrust and torque have been studied in the time domain, and vibration signals from the workpiece are studied in the time domain as well as the time frequency domain. The values of the extracted features increase for a continuing drilled hole and become the maximum before breakage.

•
Prediction of drilled hole quality is better by considering all the appropriate features collectively rather than when they are considered individually. The results establish that the fusion of sensor signal features can produce higher accuracy of tool-condition prediction.
• Hole-quality prediction by ANFIS model reveals the best performance among the different monitoring strategies it is fusion of all the time signal feature. The wavelet packet-based approach is more accurate in prediction of hole quality as compared to time based approaches of vibration signals. Moreover, the different monitoring strategies are found to perform better in fusion of all sensor signal.