Feature extraction networks for dull tool monitoring

Abstract

Automatic feature extraction is a need in many current applications, including the monitoring of industrial tools. Currently available approaches suffer from a number of shortcomings. The Kohonen self -organizing neural network (SOW) has the potential to act as a feature extractor, but we find it benefits from several modifications. The purpose of these modifications is to cause feature variations to be aligned with the SO” indices so that the indices themselves can be used as measures of the features. The modified S O W is applied to the dull tool monitoring problem, and it is shown that the new algorithm extracts and characterizes useful features of the data. 1. MOTIVATION FOR THIS WORK Monitoring the evolution of machine tool performance is an area in which manufacturing industries are very interested. Great expense is potentially involved either in damaging a part by using a dull tool or in attempting to avoid this damage by replacing tools prematurely. Automatic on-line evaluation is thus a strong desire, but current systems to this end are based on features chosen by human observation of the data. While such systems have had success on some applications [1.2], they are inherently biased toward the use of features which humans can readily perceive and model. Figure 1 shows short-time Fourier transforms (spectrograms) of vibration patterns from holes drilled using both sharp and dull drills. It is easy to see that, for the dull tool, the main carrier frequency hi= decreased and that a 9.7 kHz resonance has become significant in amplitude. However, these features do not capture the full complexity of the evolution of the dulling process. The desired classification system would provide more infomation than is available using these hand-selected features. 2. TRADITIONAL FEATURE EXTRACTION Ideally, we would l i e to analyze the drilling vibration patterns by applying an automatic feature extractor to them, i.e. finding new variations which are present in the data set instead of being forced to pre-determine which variations we are interested in. However, the bulk of current automatic feature extraction is based on clustering approaches developed not for vector the-sequences but instead for single vectors (e.g. [4]). The abilily of clusters to express variation over a continuum is limited, because they introduce artificial boundaries and because they contain no sense of ordering. For example, a clustering method applied to the spectrograms of Figure 1 might map all main resonances below TI(semnda) Figure 1: Spectrograms of vibration patterns from sharp and dull tools. 11.5 kHz to cluster A. resonances above 12 kHz to cluster B, and resonances from 11.5-12 kHz to cluster C. Thus a tone which moved smoothly through the frequency specl” might be represented unnaturally by a series of abrupt jumps between disjoint and unordered clusters. One commonly used approach to feature extraction which can encode features over a continuum is principal component analysis (PCA)[4]. This approach involves taking a high-dimensional data set and extracting from it the directions of maximum variance. The “features” which are then extracted are the projections of the input data along these principal directions. A restriction of the PCA approach,though, is the fact that it is looking for features which are linear combinations of the input data vector elements. Figure 2 shows a region of some of the principal components trained on the drill spectrograms of Figure 1. The feature which the components are attempting to express is the frequency of the main carrier resonance, which varies from 11.2 kHz to around 12.5 kHz. However, these spectral vectors are related in a nonlinear fashion to the locations of the main resonant frequency, an as a result PCA is not able to express this feature 0-7803-2431 4/95 $4.00

DOI: 10.1109/ICASSP.1995.479704

5 Figures and Tables

Cite this paper

@inproceedings{Owsley1995FeatureEN, title={Feature extraction networks for dull tool monitoring}, author={Lane M. D. Owsley and Les E. Atlas and Gary D. Bernard}, booktitle={ICASSP}, year={1995} }