Anomaly detection in the fan system of a thermal power plant monitored by continuous and two-valued variables. (September 2020)

• Record Type: Journal Article
• Authors: Wang, Min; Zhou, Donghua; Chen, Maoyin; Wang, Yanwen
• Is Part Of: Control engineering practice. Volume 102(2020)
• Publication Date: 2020-09
• Subjects: Anomaly detection -- Two-valued variable -- Continuous variable -- MHNBM -- Power plant; Automatic control -- Periodicals; 629.89
• DOI: 10.1016/j.conengprac.2020.104522
• Languages: English
• ISSNs: 0967-0661

Abstract

Abstract: The current data-driven anomaly detection approaches for practical industrial processes are strongly based on continues variables, represented by PCA, PLS and their variants. However, there exist a large amount of two-valued variables in large-scale processes, whose values are stored as 0 or 1. For example, the No.1 generator set of Zhoushan power plant, China, has 17381 variables, in which 8822 are two-valued variables. Therefore, an interesting problem naturally arises: how to combine these abundant two-valued information to enhance the performance of anomaly detection? This paper considers this problem and develops a novel mixed hidden naive Bayesian model (MHNBM) for anomaly detection, which can make the best usage of continuous and two-valued variables simultaneously. This kind of model is different from the traditional detection methods, in which the two-valued variables are totally eliminated in the data preprocessing. MHNBM is developed within the probability framework, and can efficiently enhance the detection performance through combining the two-valued variables. The effectiveness of the proposed MHNBM is validated through the simulation data and the actual data of Zhoushan power plant, China. Highlights: A novel data-driven process monitoring model called mixed hidden native Bayesian model (MHNBM) is proposed for anomaly detection. In the proposed model, both continuous and two-valued variables are utilized for detection. The variable selection is performed for both continuous and two-valued variables. Through the simulation data and the actual data of a power plant, the effectiveness of the proposed MHNBM is validated.