Feature extraction for machine learning to detect floating scrap during stamping using accelerometer

Abstract

This study entails an examination and comparison of features observed in an accelerometer signal obtained from floating-scrap detection during stamping with those obtained using the “center-of-gravity” method, which is conventionally used for anomaly detection via machine learning. The samples are detected using the Mahalanobis–Taguchi system. If using the center-of-gravity method, the unit space that contains normal samples without scraps cannot be separated from the error samples. Based on an estimated threshold for detecting all the error samples, the false-positive rate of the abovementioned method is 0.9 %. In this study, a suitable threshold that allows the system to detect 100 % of the error samples (and no normal samples) is estimated using six features. Detection using the six suggested features is more effective than that using only three features associated with the downward journey of the press slide. Features selected from two different events (i.e., the downward and upward journeys of the press slide) may result in more effective detections than features selected from only one event (i.e., the downward journey). To confirm the effect of tool wear, six experiments based on normal samples are conducted after all error samples are created. The machine-learning method with the suggested features is more insensitive to tool wear than the center-of-gravity method, as the features correspond closely to the size and deformation of a foreign object. Nevertheless, the unit space should be updated as tool wear progresses; otherwise, false positives may occur even when the suggested features are used.