Copyright 0 IFAC Fault Detection, Supervision and Safety of Technical Processes, Washington, D.C., USA, 2003 IFAC COl> Publications www.elsevier.comllocatelifac FAULT DETECTION WITH OBSERVERS AND GENETIC PROGRAMMING: APPLICATION TO THE DAMADICS BENCHMARK PROBLEM Marcin Witczak * Ron J. Patton" Jozef Korbicz • • Institute of Control and Computation Engineering, University ofZielona G6ra, ul. Podgorna 50, 65-246 Zielona G6ra, Poland, e-mail: {M.Witczalc.J.Korbicz}@issi. uu.gora.pl .. Control and Intelligent Systems Engineering, Department of Engineering, University of Hull, Cottingham Road, East Yorkshire HU6 7RX, United Kingdom, e-mail: R.J.Patton@hull.ac. uk Abstract: This paper is focused on the problem of designing a fault diagnosis scheme for a class of non-linear systems. The one objective is to show how to employ a genetic program- ming technique to obtain state-space models of non-linear systems. Another objective is to employ a modified version of the well-known unknown input observer to form a non-linear deterministic observer for the purpose of residual generation. The final part of the paper shows how to use the proposed approach to tackle fault detection of the DAMADICS benchmark. Copyright © 2003 IFAC Keywords: fault diagnosis, non-linear systems, system identification, observers, genetic algorithms I. INTRODUCTION It is well known that there is an increasing demand for modem systems to become more effective and reliable. This real world's development pressure has transformed automatic control, initially perceived as the art of designing a satisfactory system. into the mo- dem science that it is today. The observed increasing complexity of modem systems necessitates the deve- lopment of new control and supervision techniques. To tackle this problem. it is obviously profitable to have all the knowledge concerning a system behaviour. Un- doubtedly, an adequate model of a system can be a tool providing such knowledge. Indeed. nowadays. advan- ced techniques for designing controllers are also based on models of systems. As most of industrial systems exhibit a non-linear behaviour, this has been the main reason for further development of non-linear system identification theory. Indeed, a few decades ago, non- linear system identification was a field of several ad- 1101 hoc approaches. each applicable only to a very re - stricted class of systems. With the advent of neu- ral networks. fuzzy models (Nelles. 2001), and mo- dem structure optimization techniques (Koza. 1992). a much wider class of systems can be handled. This implies the possibility of designing more reliable con- trol and supervision tools. Irrespective of the identification method used. there is always the problem of model uncertainty. i.e. the model-reality mismatch. Thus. the better the model used to represent a system behaviour. the better the chance of improving the reliability and performance in diagnosing faults. Unfortunately, disturbances as well as model uncertainty are inevitable in industrial systems. and hence there exists a pressure creating the need for robustness in fault diagnosis systems. This robustness requirement is usually achieved in the fault detection stage.