Multimode process monitoring based on fault dependent variable selection and moving window-negative log likelihood probability

Abstract Multimode process monitoring has attracted much attention in academia and industry in past decades. Generally, all the measured variables are involved to monitor a process. However, irrelevant variables may degrade the monitoring performance due to over-fitting, and increase the online computational complexity excessively. In order to monitor the faults possibly affecting the operational safety and product quality, it is important to select appropriate monitored variables that are closely related to these faults. This paper explores the problem of multimode process monitoring with variable selection. A novel algorithm based on the Kullback-Leibler divergence is proposed for variable selection in multimode processes, which effectively selects the most informative variables about the concerned faults. Then, the one-step Viterbi algorithm with low computational complexity is developed to implement online mode identification, which utilizes both spatial characteristics and temporal correlations to identify operating modes accurately. By introducing the moving window technique, a new detection index, i.e. moving window-negative log likelihood probability (MWNLLP), is proposed to capture the dependency of samples and further improve the detection performance for the concerned faults. A numerical example and the Tennessee Eastman process (TEP) are adopted to demonstrate the effectiveness of the proposed method. Specifically, MWNLLP can effectively detect the fault 3 in TEP with a detection rate over 95%.