Goal-oriented system identification for safe operation of engineering systems

The goal of this paper is to use existing results and methods to propose a new type of closed-loop system configuration that maximizes safety and robustness of operation. The plant is assumed to be unknown in the beginning of the process. Two competing system identification methods (based on neural networks and Kalman filters) are combined, exploiting the advantages of each. The neural networks based system identification is used for fault detection, while a linear time-varying ARMA model, estimated in real time by a Kalman filter, is used for controller design so that the plant output follows a predetermined trajectory (setpoint)