Fault detection and isolation for dynamic processes using recursive principal component analysis (PCA) based on filtering of signals

A systematic procedure for the fault detection and isolation of dynamic systems is presented. The inputs of the process first pass through the dynamic filters which represent the process dynamics. Then, principal component analysis (PCA) is applied to the data matrix consisting of these filtered signals and the process outputs for fault detection. In case of a fault being detected, owing to an artificial linear relationship existing in the data matrix, the last principal component (LPC) is adopted for fault isolation. A recursive algorithm for PCA based on rank-one matrix update of the covariance is derived to compute the LPC on line. Patterns of the LPC are devised to isolate these faults, which include constant-bias and high-frequency noises originating from sensor measurement, errors resulting from input disturbance and change in the process gain. Furthermore, the magnitude of the fault can also be identified from the computed LPC. An illustrative example is used to verify the effectiveness of the proposed method. (c) 2007 Curtin University of Technology and John Wiley & Sons, Ltd.