IMPROVING SAFETY OF A PILOT-PLANT REACTOR USING A MODEL-BASED FAULT-DETECTION AND IDENTIFICATION SCHEME

This work describes the experimental implementation of an automatic sc heme for the on-line detection and identification (FDI) of faults in t he sensors of an industrial scare pilot plant reactor under process co ntrol, where a pseudo zero-order exothermic chemical reaction is parti ally simulated. The main goals of this research are to enhance the saf ety of reactor operations and to demonstrate the potential of FDI for practical industrial applications. The automatic fault detection and i dentification method proposed here has two main steps: (1) the detecti on stage, which relies on a sequential statistical analysis of the pro cess parameters that are continuously estimated by means of a general regression software package (GREG) suitable for non-linear models; (2) the identification step, which is based on an Extended Kalman Filter (EKF) to provide values for the state variables estimates. These value s are compared to those given by the sensors thus enabling the identif ication of the faulty sensor. Moreover, this classification procedure ensures that automatic process control can still be carried on even in such a faulty situation. Despite the strong non-linearities and the h igh number of uncertainties, the proposed strategy exhibited very prom ising results concerning the detection and identification of the fault y sensors. Furthermore, it enabled a satisfactory controller performan ce for a reasonable period of time, when any of the sensors was disabl ed and control actions were solely based on state estimates. (C) 1998 Elsevier Science Ltd. All rights reserved.