A novel identification scheme for physical systems with applications to system health monitoring

A novel identification scheme is proposed here for a wide class of nonlinear and highly complex physical systems, including manufacturing ones, which can be modelled using the linear parameter-varying (LPV) approach. The targeted applications include status monitoring, condition-based maintenance and fault diagnosis. In this approach, several operating points, selected over the entire system's operating regime, are used to model the system at each such point, by a linear model such that the set of all such local linear models forms the best approximation of the original nonlinear system. To ensure reliability and accuracy of both the proposed identification scheme and its applications, emulators are included at the output measurements to mimic likely operating scenarios resulting from variations in the subsystems such as the controller, sensor, actuators and plant. At each operating point, data collected from various emulator parameter-perturbed experiments is used to identify both the system and the influence vectors. The influence vector, a novel concept used here, is used to map the variations of the emulator parameters (and hence those of the subsystems) to the feature vector (transfer function coefficients). At the heart of the system's health monitoring setup used is a Kalman filter which uses its residual to detect a fault and the influence vector to isolate the faulty subsystem. The novel use of emulators and influence vectors is shown here to spawn various contributions, namely a generalization of the traditional LPV approach, a new multi-model LPV-based identification scheme, and hence a new health monitoring system based on it, that are both reliable, robust and accurate. The superiority of the performance of the proposed novel identification scheme over that of conventional schemes is shown both analytically and through a successful evaluation on both simulated and physical systems.