Multiple detectors-based Bayesian inference and statistics analysis for incipient fault detection

To establish a secure and dependable operational setting for practical industrial processes, it is crucial to detect incipient faults promptly and accurately. In this work, a novel data-driven process monitoring approach is presented for incipient fault detection. First, inspired by the ensemble learning algorithm, multiple Mahalanobis distance (MD) detectors are trained, and corresponding detection indices for each sample are calculated. Second, with Bayesian inference, all the obtained indices are converted to probabilities for samples being faulty. Furthermore, the statistics of probabilities are analyzed by implementing sliding time window technique, ultimately achieving incipient fault detection. Finally, the effectiveness of the proposed strategy is demonstrated by a numerical example and the continuous stirred tank reactor benchmark process. The proposed approach inherits the advantages of ensemble learning, so it has the capability to improve the stability of detection performance. Meanwhile, due to the utilization of Bayesian inference, together with the sliding time window technique, it can better differentiate normal and faulty samples and improve the sensitivity to incipient faults.