Chemical process fault detection using modified CVDA with memory mechanism

As an effetive dynamic data analytical tool, canonical variate dissimilarity analysis (CVDA) has been successfully applied to chemical process fault detection. However, the basic CVDA method adopts offline one-time modeling strategy and lacks the self-learning ability. That means, the CVDA model is built only based on the offline normal data and cannot learn the new fault knowledge during the online monitoring stage. To address this limitation, an enhanced CVDA modeling framework with memory mechanism (MCVDA) is proposed for monitoring chemical process faults more effectively. In the proposed method, the basic CVDA method is utilized to build the main monitoring model, and an auxiliary fault-memory CVDA model is designed by taking full advantage of the occurred fault data. Specifically, once a fault is detected, the corresponding fault data are analyzed for the development of the fault-memory CVDA model, where the fault-sensitive features are weighted for strengthening the detection of the same type of faults. Furthermore, the Bayesian inference technique is used to combine the monitoring results from the main CVDA model and the fault-memory CVDA model. When more fault data sets are available, the fault similarity judgement procedure is performed baased on the weight simialrity analysis to deterimine if the fault-memory CVDA model should be updated. The validation results on the TE process illustrate that the proposed MCVDA method has better fault detection performance than the basic CVDA method.