A Quality-related Fault Detection Method Based on Weighted Mutual Information

Abstract

Quality-related fault detection has become a research hotspot in recent years, and its goal is to maximize the alarm rate for quality-related faults in process monitoring and minimize the alarm rate for faults that are irrelevant or self-adjustable. The traditional principal component analysis(PCA) and partial least squares(PLS) method use covariance to extract the principal component. As a second-order statistic, covariance can only extract Gaussian information without considering that the data may contain higher-order non-Gaussian information, so its statistical alarm rate is not the most accurate. To solve above problems, this paper proposes a quality-related fault detection method based on weighted mutual information. The first step is to extract the set of process variables that contain the most information about the quality variables through Bayesian fusion mutual information, and then use the fusion mutual information for the extracted process variables to eliminate some process variables. For the remaining process variables, use PLS algorithm based on maximum mutual information to extract principal components for statistical modeling. Finally, it is applied to Tennessee Eastman process(TEP) to simulate the feasibility and effectiveness of the proposed method.