An improved real-time contrasts control chart using novelty detection and variable importance

Fault detection and isolation are important tasks in statistical process control. A real-time contrasts (RTC) control chart converts the statistical process-monitoring problem to the real-time classification problem, thus outperforming traditional monitoring techniques. An RTC assigns a class to reference data and the other class to a window of real-time contrasts. However, RTC control charts often fail to detect abnormal states when both normal and abnormal data exist together in the window. To enable more rapid detection of an improved RTC control chart, this paper proposes a multivariate process monitoring system with an improved RTC control chart. Although previous RTC control charts proposed by other studies outperform the original RTC chart, it is still difficult to detect an abnormal state when normal and abnormal data exist together. To overcome this problem, this paper proposes an RTC control chart using novelty detection and variable importance with random forests. Novelty detection and variable importance were used so that fault can be detected when the control limit could not be exceeded despite the abnormal state. The proposed method extracts representative data in the sliding window and adds the extracted data to the window to quickly detect the abnormal state. Experiments demonstrate the proposed method to outperform the original RTC chart.