Copyright C> IFAC Fault Detection, Supervision and Safety of Technical Processes, Washington, D.C., USA, 2003 IFAC c: 0 C> Publications www.elsevier.comllocatelifac DYNAMIC NEURAL NETWORKS FOR ACTUATOR FAULT DIAGNOSIS: APPLICATION TO THE DAMADICS BENCHMARK PROBLEM Krzysztof Patan .,1 Thomas Parisini·· • Institute of Control and Computation Engineering, University of Zielona Gora ul. Podgoma 50, 65-246 Zielona Gora, Poland, K.PatanDissi.uz.zgora.pl •• Dept. of Electrical, Electronic and Computer Engineering, DEEI-University of 7heste Via Valerio 10, 94127 7heste, Italy, parisiniDuniv.trieste.it Abstract : The paper presents results achieved during realization of the interna- tional project DAMADICS (Development and Application of Methods for Actuator Diagnosis in Industrial Control Systems). The proposed fault detection and isola- tion system is designed using a bank of dynamic neural networks. Each network is trained using a stochastic approximation method, which can be viewed as a fast alternative to back-propagation based algorithm. Simulation results are carried out using the real process data recorded at the Lublin Sugar Factory, Poland. Copyright © 2003 IFAC Keywords: Actuators, benchmark examples, fault diagnosis, dynamic models, neural networks, stochastic approximation, performance indexes. 1. INTRODUCTION Methods of FDI based on system identification and residual generation have been intensively studied for the last two decades. One of the most important classes of the FDI methods is neural modelling (Frank and Koppen-Seliger, 1997; Chen and ratton, 1999; Pat ton et al., 2000). Artificial neural networks can be applied to nonlinear sys- tems. They are useful when there are no math- ematical models of the diagnosed system, hence, analytical models and parameter-identification al- gorithms cannot be applied. One of the most inter- esting solutions of the dynamic system identifica- tion problem is the application of neural networks 1 This work was supported by the EU FP 5 project DAMADlCS 975 composed of Dynamic Neuron Models (DNM) (Ayoubi, 1994; Patan and Parisini, 2002b). Such neuron models consists of an adder module, a linear dynamic system - Infinite Impulse Re- sponse (HR) filter, and nonlinear activation mod- ule. Thus, the DNM activation depends on its actual inputs as well as inputs and outputs in pro- ceeding time. The relatively complex DNM allows one to design a neural network of a feed-forward multi-layer structure (Patan and Parisini, 2002a) . Derivation of the optimal neural network param- eters, however, is not a trivial problem. This pro- cess seems to be an optimization problem, which is intrinsically related to a very rich topology. The effectiveness of the gradient based algorithms in many cases is very poor, because it usually finds one of the local minima.