Applying the self-organization feature map (SOM) algorithm to AE-based tool wear monitoring in micro-cutting

Abstract

This study applies a self-organization feature map (SOM) neural network to acoustic emission (AE) signal-based tool wear monitoring for a micro-milling process. An experiment was set up to collect the signal during cutting for the system development and performance analysis. The AE signal generated on the workpiece was first transformed to the frequency domain by Fast Fourier transformation (FFT), followed by feature extraction processing using the SOM algorithm. The performance verification in this study adopts a learning vector quantification (LVQ) network to evaluate the effects of the SOM algorithm on the classification performance for tool wear monitoring. To investigate the improvement achieved by the SOM algorithms, this study also investigates cases applying only the LVQ classifier and based on the class mean scatter feature selection (CMSFS) criterion and LVQ. Results show that accurate classification of the tool wear can be obtained by properly selecting features closely related to the tool wear based on the CMSFS and frequency resolution of spectral features. However, the SOM algorithms provide a more reliable methodology of reducing the effect on the system performance contributed by noise or variations in the cutting system.

Introduction

The increasing demand for miniaturized and lightweight products have increased the demand for micro-machining technologies such as laser cutting, electrical discharge machining (EDM), electron beam (E-beam) machining, lithographie galvanoformung abformung (LIGA) processing, and mechanical cutting. This is particularly true for manufacturing devices or molds with machined features ranging in size from 10 μm to 1 mm. Compared to non-traditional processes, micro-mechanical cutting technology offers greater flexibility in manufacturing 3D feature, more precise machining, a lower limit in the material selection, and lower impact on the environment. However, micro-mechanical cutting involves more challenges than conventional size cutting. These challenges include easy wear and breakage due to weak tools, the requirement for a high speed spindle to provide sufficient cutting speed, the fixture design for small workpiece, and fewer choices of tool material. Finally, the mechanical model developed for a conventional cutting process cannot be used for micro-cutting.

A number of studies over the past few decades discuss tool condition monitoring on the conventional scale and have been reported in a number of review papers [1], [2], [3], [4], [5]. For the development of signal processing algorithms, neural network models is one of the methods to predict the tool wear condition [6]. Selecting the right sensor type is very important to obtain enough information for classification when developing a monitoring system. Researchers have used different sensors to monitor tool condition in conventional cutting, including force sensors [7], [8], vibration sensors [9], sound signal sensors [10], acoustic emission (AE) signal sensors [11], [12], [13], [14], [15], [16], [17], spindle current sensors [18], temperature sensor [19], and stress sensor [20]. However, the AE signals are the primary candidates for monitoring tool condition in micro-machining [21].

The AE signals generated during the cutting process are primarily from the plastic deformation on the shear plane, flank wear land, and the friction on the tool/chip interface. For the conventional cutting, Dornfeld [11], [12] used AE signals to monitor tool wear in cutting processes. Jemielniak and Otman [13] developed a method to predict the tool failure condition based on the statistic analysis of AE signals. Kakade et al. [14] predicted the tool wear based on the analysis of AE signals. König et al. [15] found a close relationship between AE-RMS signals and tool chipping in the drilling process. Blum and Inasaki [16] observed better sensitivity of AE sensor for flank wear monitoring than the cutting force. Moriwaki and Tobito [17] studied the statistical values of AE signals to monitor the tool life of a coated tool. They concluded that the variance value is more sensitive to tool wear changes than the root mean square and mean value of the AE signal. After obtaining the signal generated in the cutting process, signal transformation, feature generation, and classifier design steps are necessary to map signal features to tool wear conditions. A number of feature generation and classifier designs have been proposed in the past few decades [22] including the backpropagation neural network (BPNN), radial basis function neural network (RBFNN), self-organizing feature map (SOM), adaptive resonance theory (ART) [23], linear discriminate analysis (LDA) [24], and fuzzy logic [25]. However, different methodologies have different features to meet different requirements. Some methods emphasize computing efficiency, while others focus on noise reduction capability or discriminate levels of the monitored events. The performance of applying certain algorithms is also affected by the characteristics of the collected signals and mechanical cutting conditions. Therefore, there is no algorithm that fits all cases well, and the optimum design of an algorithm depends on the domain knowledge obtained for monitored events and collected signals. Dornfeld [7] developed a tool condition monitoring system based on AE signals, cutting force, and spindle current. Dornfeld used an auto regressive series model and back propagation neural network in system development. Emel [26] obtained a good tool condition identification rate by adopting a backpropagation algorithm for signals with different cutting conditions. Hong [27], [28] used the wavelet transformation of cutting force signal as an input feature for a backpropagation neural network to identify the tool breakage. Elanayar and Shin [29], [30] measured the cutting force in cutting and developed a neural network for classifying the flank wear level. The SOM algorithm has been applied to the fault condition monitoring of gear boxes [31], wind turbines [32], engines [33], cutting tools [34], [35], [36], [37], [38], and welding [39]. Scheffer and Heyns [34], [35], [36] used the SOM algorithm to model the relationship between turning tool conditions and signals such as cutting force, AE signal, temperature, strain gauge. Their results show that the SOM neural network can effectively reduce the disturbance effect on monitoring the tool wear in turning processes. Silva [37] used an accelerometer for measuring vertical vibration, a microphone for recording sound emission and a strain gauged tool holder for force measurement in turning operation. The results showed that the computational architecture of self-organizing spiking neural network map has a greater potential to unveil embedded information in tool wear monitoring data sets and that faster learning occurs if compared to traditional sigmoidal neural networks. For the conventional milling process, Wang et al. [38] integrated the vision sensor, force sensor, along with SOM algorithms to detect the tool wear in cutting. In this study, the SOM-based feature processing will be developed to investigate its effect in reducing the signal variation effect caused by noise or the variation of cutting mechanism on tool wear monitoring in micro milling process.

Developing a tool wear monitoring system for micro-mechanical cutting involves more challenges than its conventional counterpart. Lee et al. reported AE signal applications in ultra-precision machining and drilling [21], [40]. Tansel et al. [41] studied tool breakage and tool wear monitoring in micro-cutting processes based on the AE signal and neural network technology. Jemielniak revealed a strong influence of tool wear on acoustic emission signal, providing acceptable results even while used separately [42]. The same study observed the cutting forces to be strongly disturbed by the resonance vibration of a table dynamometer. To study the classifier design of micro-tool monitoring, Zhu et al. [43], [44] adopted a discriminant feature selection approach to modeling of micro-milling tool conditions based on the cutting force using a hidden Markov model (HMM). Malekiana et al. [45] integrated accelerometers, force, acoustic emission sensors, and a vision system to monitor micro-milling operations. They also used a neuro-fuzzy method to train and determine the membership functions and rules of micro-end mills. Tansela et al. proposed a genetic tool monitor (GTM) [46] to identify the problems using an analytical model for micro-end-milling operations and genetic algorithm. They confirmed the capability of GTM to monitor micro-end-milling operations without any previous experience by estimating symmetrical wear and local damages at the cutting edges of a tool.

This study applies a self-organizing feature map (SOM) neural network to AE signal based tool wear monitoring for the micro-milling process. The reduction of the noise or system variation effect on the feature characteristics by applying the SOM algorithm in the feature process stage will be analyzed. The experiment in this study was designed to collect signals during cutting to aid system development and performance analysis. In the system development stage, the AE signals generated on the workpiece were first transformed to the frequency domain by fast Fourier transformation (FFT), followed by the feature extraction processing using the SOM algorithm. In the performance verification, learning vector quantification (LVQ) network was adopted as a classifier to evaluate the effect of the SOM algorithm on the classification performance for tool wear monitoring. To investigate the improvement of the SOM algorithms, this study also investigates cases applying only the LVQ classifier and both the feature selection based on the class mean scatter feature selection criterion (CMSFS) and LVQ.