Punching process monitoring using wavelet transform based feature extraction and semi-supervised clustering

Abstract Punching is a common sheet metal fabricating process for many industrial applications. On-line health condition monitoring of punching process is becoming more and more important in order to detect and correct process failures in time, and ensure the consistence of the product quality. Effective feature extraction of the process is critical for monitoring the punching process. In this work, piezoelectric strain sensors are used to measure the strain on press column surface which is the response of the press to the stamping force. A feature extraction approach based on the wavelet transform and the energy distribution of the reconstructed wavelet coefficients is proposed. The energy distribution is used as the process feature for similarity distance calculation for the process clustering. Semi-supervised clustering is applied to the process monitoring considering that many normal data sets are available while the failure data is difficult to obtain in practice. The proposed method is applied to the column strain signals from punching process and simulation failure data. The results show that combining the wavelet energy distribution of the strain signal and the semi-supervised clustering is an effective method for health condition monitoring and failure detection in punching processes.