A unified approach to detection and isolation of parametric faults using a Kalman filter residual-based approach

A unified approach to detection and isolation of parametric faults in a physical system resulting from variations in the parameters of its constituting subsystems, termed herein as diagnostic parameters, is proposed here using Kalman filter residuals. Rather than use the feature vector made of the coefficients of the numerator and denominator of the system transfer function, which is known to be a nonlinear function of the diagnostic parameter variations, our proposed approach first shows and then exploits, for fault detection purposes, the fact that the Kalman filter residual is a multi-linear function of the deviations in the diagnostic parameters, i.e. the residual is separately linear in each parameter. A fault is then isolated using a Bayesian multiple composite hypotheses testing approach. A reliable map relating the diagnostic parameters to the residual is obtained off-line using fault emulators. The proposed unified scheme is successfully evaluated on both simulated data as well as on real data obtained from a benchmarked laboratory-scale coupled-tank system used to exemplify an industrial two-tank process.