A new fault detection method based on an updatable hybrid model for hard-to-detect faults in nonstationary processes

Abstract

Fault detection is an effective means to guarantee the stable operation of industrial production. Fault signals are easily masked by nonstationary trends in the variables, which leads to hard-to-detect faults in nonstationary processes. In this paper, an updatable hybrid model for fault detection is proposed for the nonstationary characteristics and hard-to-detect faults of industrial processes. First, the stationary residuals of the nonstationary variables are combined with the stationary variables to form a combined matrix. Second, a monitoring model based on slow-feature-analysis-local-outlier-factor (SFA-LOF) is constructed, which extracts the slow features in the combined matrix and introduces a local outlier factor as the monitoring index. Third, the sensitive variables of faults that are hard to detect using SFA-LOF are screened, and refined models based on Kullback–Leibler divergence are constructed for hard-to-detect faults. Then, an updatable hybrid model based on the SFA-LOF model and the refined model is proposed. The hybrid model matches the detection models to the faults and is able to update the hybrid model by developing refined models. Finally, the Tennessee Eastman process is used to validate the effectiveness of the proposed fault detection framework.