Incipient Fault Detection Based On Principal Component Analysis

Principal component analysis (PCA) finds wide application in chemical monitoring and fault isolation as a data-driven modeling method. Current studies show that PCA combined with data filters can effectively enhance the detectability for incipient faults. However, it is not clear how it works. Moving average (MA) and exponentially weighted moving average (EWMA) are most popular data filters. The combinations of PCA with these two filters are defined as MA-PCA and EWMA-PCA respectively, which are the research subjects in this paper.This paper firstly investigates the mechanism of the effect of MA filter and EWMA filter on the fault detection of PCA. The reason that MA-PCA and EWMA-PCA can detect incipient faults is analyzed respectively. This paper also derives the critical fault magnitude of MA-PCA and EWMA-PCA. For MA-PCA, the relationship among critical fault magnitude, detection latency and window length is analyzed, based on which a novel method for selecting proper window length is also proposed to avoid the subjectivity. For EWMA-PCA, the relationship among critical fault magnitude, detection latency and forgetting factor is analyzed, based on which the forgetting factor is determined. Then a comparison of the detectability between MA-PCA and EWMA-PCA is proposed. It is proved that MA-PCA has a smaller critical fault magnitude and lower fault detection rate than EWMA-PCA with the same detection latency. Finally MA-PCA is applied to a numerical simulation and the Tennessee Eastman Process (TEP), and is compared with EWMA-PCA and conventional PCA monitoring method. The monitoring results illustrate the above conclusions.