Multimode process monitoring based on conditionally independent Bayesian learning

Abstract

In multimode process monitoring, both mode identification and fault detection are important for system safety and reliability. In this paper, a new data-driven multimode process monitoring method called conditionally independent Bayesian learning (CIBL) is proposed. Considering the strong assumption of conditional independence in naive Bayes, orthogonal transformation is first applied to measured variables to improve the extent of conditional independence in different operating modes, without excessively changing the data features. Then Bayes-based mode identification is adopted in transformed data, and the Mahalanobis distance of the transformed measurement vector serves as the detection index. With this orthogonal transformation, the mode identification accuracy can be effectively improved compared with naive Bayes. In addition, the fault detection performance of the proposed method outperforms the traditional multimode process monitoring method mixture principal component analysis.