Feature extraction of waveform signals for stamping process monitoring and fault diagnosis.

Abstract

Sheet metal stamping is a very complex manufacturing process. The current quality control practice of using periodic inspection of the stamped parts cannot satisfy the high stamping production requirement, due to its low sampling frequency and incapability of root cause determination of process faults. Meanwhile, stamping tonnage waveform signals, which are measured on-line, contain rich information that can be related to both the product quality and process variables. However, little research has been done in the past to utilize on-line tonnage waveform signals for stamping process control. In this thesis, the emphasized research is how to fully utilize in-process tonnage waveform signals to monitor the conditions of stamping process variables through the integration of engineering knowledge, statistical multivariable analysis, and signal processing of wavelet analysis. Three fundamental issues have been studied. (1) Feature lossless data compression is proposed and implemented for the first time to efficiently collect tonnage signals for process monitoring and fault diagnosis. Compared with data compression using a denoising approach, the feature lossless data compression approach is more efficient, further removing signals irrelevant to process performance. The remaining data contain only those potential features for process monitoring and fault diagnosis. (2) A new hierarchical feature extraction method is developed for process diagnosis with variable interactions by using a two-level fractional factorial design of experiment (DOE). The unique characteristics of the proposed methodology are that it considers the interactive variables in the feature extraction of a tonnage waveform signal, which is useful for multiple stamping process fault diagnosis. (3) A novel methodology is presented to develop a diagnostic system by using waveform signals with limited prior fault samples, where the continuous in-process learning and diagnostic performance enhancement are emphasized through automatic feature extraction and optimal feature subset selection during the use of a diagnostic system in a manufacturing process. This research contributes to the science of using waveform signals for process monitoring and diagnosis. Though the stamping process is used as an application example in the thesis, the developed methodologies are generic and can be used in various manufacturing diagnostic applications where the waveform signals are available.