Incipient fault detection for dynamic chemical processes based on enhanced CVDA integrated with probability information and fault-sensitive features. (June 2022)

• Record Type: Journal Article
• Authors: Deng, Xiaogang; Liu, Xiaoyue; Cao, Yuping; Cong, Lin; Li, Zhe
• Is Part Of: Journal of process control. Volume 114(2022)
• Page Start: 29
• Page End: 41
• Publication Date: 2022-06
• Subjects: Dynamic process -- Incipient fault detection -- Canonical variate dissimilarity analysis -- Probability information -- Fault-sensitive features; Process control -- Periodicals; Fabrication -- Contrôle -- Périodiques; Process control; Periodicals; Electronic journals; 660.281
• DOI: 10.1016/j.jprocont.2022.04.001
• Languages: English
• ISSNs: 0959-1524

Abstract

Abstract: Recently, canonical variate dissimilarity analysis (CVDA) has emerged as an efficient incipient fault monitoring method for dynamic chemical processes. Nevertheless, the basic CVDA method omits the mining of the probability information hidden in the canonical variate vectors and does not make full use of the fault-sensitive features provided by the prior fault data. Aiming at these two limitations, this paper presents an enhanced CVDA method, called fault-sensitive probability-related CVDA (FSPRCVDA), for better monitoring incipient faults. Firstly, the traditional CVDA is improved to the probability-related CVDA (PRCVDA) by applying the Kullback–Leibler divergence to measure the probability distribution changes of canonical variates. Secondly, fault-sensitive features are extracted by local Fisher discriminant analysis on the normal and prior fault data. The extracted fault-sensitive features are used to construct the auxiliary model for assisting the primary PRCVDA model, which leads to the holistic FSPRCVDA model. The case study on a continuous stirred tank reactor system shows that the proposed FSPRCVDA method has higher incipient fault detection rate than the basic CVDA method. Highlights: An improved CVDA modeling method is proposed by utilizing both normal and prior fault data. Kullback–Leibler divergence is applied to measure probability distribution changes of canonical variates. Fault-sensitive features are extracted by local Fisher discriminant analysis. Overall monitoring statistics are constructed by fusing the probability-related CVDA features and the fault-sensitive features.