Identifying Process Status Changes via Integration of Independent Component Analysis and Support Vector Machine

Abstract

Observations from the in-control process consist of in-control signals and random noise. This paper assumes that the in-control signals switch to different signal types when the process status changes. In these cases, process data monitoring can be formulated as a pattern recognition task. Time series data pattern recognition is critical for statistical process control. Most studies have used raw time series data or extracted features from process measurement data as input vectors for time series data pattern recognition. This study improves identification by focusing on the essential patterns that drive a process. However, these essential patterns are not usually measurable or are corrupted by measurement noise if they are measurable. This paper proposes a novel approach using independent component analysis (ICA) and support vector machine (SVM) for time series data pattern recognition. The proposed method applies ICA to the measurement data to generate independent components (ICs). The ICs include important information contained in the original observations. The ICs then serve as the input vectors for the SVM model to identify the time-series data pattern. Extensive simulation studies indicate that the proposed identifiers perform better than using raw data as inputs.