Fault detection using canonical variate analysis

The system identification method canonical variate analysis (CVA) has attracted much attention from researchers for its ability to identify multivariable state-space models using experimental data. A model identified using CVA can use several methods for fault detection. Two standard methods are investigated in this paper: the first is based on Kalman filter residuals for the CVA model, the second on canonical variable residuals. In addition, a third method is proposed that uses the local approach for detecting changes in the canonical variable coefficients. The detection methods are evaluated using three simulation examples; the examples consider the effects of feedback control; process nonlinearities; and multivariable, serially correlated data. The simulations consider several types of common process faults, including sensor faults, load disturbances, and process changes. The simulation results indicate that the local approach provides a very sensitive method for detecting process changes that are difficult to detect using either the Kalman filter or canonical variable residuals.