Adaptive Canonical Correlation Analysis Method Based on Forgetting Factor for Fault Detection

In this paper, an adaptive canonical correlation analysis method is proposed for fault detection in time-varying processes. Firstly, a designed forgetting factor is used to update the canonical correlation analysis (CCA) model that builded with initial normal process data. Then, Mahalanobis distance is introduced as a classifier to distinguish whether data changes are caused by system modal changes or system abnormalities. In this way, the new model can not only be updated according to the system modality in real time, but also can accurately response to the occurrence of faults. Compared with traditional CCA-based methods, the proposed new method has the following two advantages: 1) it has a wider range of application scenarios since it can adapt to slow changes in the system or changes in operating points; and 2) it has a smaller amount of calculation because it only performs a simple data classification rather than require complex iterative operations on the threshold. The effectiveness of the new method is verified in a simulated superheated steam spray water temperature reduction process.