Model Construction Based on Real-Time System Identification Method (Proposal of a Model Construction Method for Anomaly Detection of Control Systems)

Abstract

This paper presents a real-time system identification method for the monitoring or anomaly detection of control systems. Real-time system identification technique is an extension of ordinary one to on-line, which is usually performed off-line due to its computational load. An application of real-time system identification enables us to construct numerical models reflecting on the instantaneous properties or behavior of control objects by using hardware acceleration. First, we explain that the ARX model, which has the simplest representation in system identification models, is suitable for hardware design and its implementation on field programmable logic devices shortens the calculation time of parameter estimation comparable to the servo cycles of control systems. Then, we discuss some difficulties derived from closed-loop system identification and the consequent ill-conditions in its calculation. As real-time system identification is supposed to be used for actual machine operations, their reference signals for control are sometimes set to constant values which are understandably not sufficient for system identification. This problem can be addressed by a modified estimation procedure considering reference signal conditions and a model with partially fixed values. Finally, the effectiveness of these designs is investigated by numerical simulations and experiments. They show us that both of accuracy and stability of parameter estimation are obtained.