Linear parameter-varying modeling and identification for condition-based monitoring of systems

Abstract

A linear parameter-varying (LPV) model and its new identification scheme are proposed for monitoring the status of a system. As the subsystem parameters are generally inaccessible, during the offline identification stage, emulators, which are transfer function blocks, are included at the measurement outputs to simulate different operating scenarios, including the nominal and abnormal ones. That is, instead of relying on the past or historical data, which may be unavailable, various operating scenarios are obtained from real-time simulation using emulators. The feature vector (coefficients of the system transfer function) is, in general, a nonlinear function of the emulator parameters and this nonlinear relationship is completely characterized by influence vectors which are the partial derivatives of the feature vector with respect to the emulator parameters. Influence vectors are identified off-line using linear least-squares fit to the emulated input–output data. The LPV model is comprised of (a) scheduling variables, and (b) emulator parameters to track respectively the variations in the operating points and to generate operational data. A Kalman filter is designed using the identified model of the system driven by the residual and is adapted to all the trajectories of the scheduling variables. The Kalman filter residual forms the backbone of the condition-monitoring and fault-diagnosis tasks. The proposed scheme is successfully evaluated on simulated systems as well as on a physical position control system.