STOCHASTIC APPROACHES TO DYNAMIC NEURAL NETWORK TRAINING. ACTUATOR FAULT DIAGNOSIS STUDY Krzysztof Patan Thomas Parisini Institute of Control and Computation Engineering, University of Zielona Góra, ul. Podgórna 50, 65-246 Zielona Góra, Poland K.Patan@issi.uz.zgora.pl Dept. of Electrical, Electronic and Computer Engineering DEEI-University of Trieste Via Valerio 10, 34127 Trieste, Italy parisini@univ.trieste.it Abstract: A paper deals with application of stochastic methods for dynamic neural network training. The considered network is composed of dynamic neurons, which contain inner feedbacks. This network can be used as a part of a fault diagnosis system to generate residuals. Up-to-date training algorithms, based on the classical back propagation, suffer from entrapment in local minima of an error function. Two stochastic algorithms are tested as training algorithms to overcome these difficulties. Efficiency of the proposed learning methods is checked using data recorded at Lublin Sugar Factory, Poland. Keywords: Actuators, dynamic modelling, fault detection, learning algorithms, neural network models. 1. INTRODUCTION The paper deals with application of stochastic methods for training dynamic neural networks. Such networks belong to the class of locally re- current globally feed-forward (Patan, 2000). They have an architecture that is similar to the feed- forward multi-layer perceptron and dynamic char- acteristics are included in their processing units. The networks under consideration are designed using the Dynamic Neuron Models (DNM). A single dynamic neuron consists of an adder mod- ule, linear dynamic system – Infinite Impulse Response (IIR) filter, and nonlinear activation module. When such neurons are connected into a multi-layer structure, a powerful approximating tool may be obtained. Taking into account that neuron by itself has dynamic characteristics, it is not required to introduce any global feedback to the network structure. In this way a simple This work was supported by the European Commision within the FP5 project DAMADICS architecture is obtained, what makes it relatively easier to elaborate a proper learning algorithm and contrary to recurrent networks to keep sta- bility of the neural model. Taking into consideration dynamic characteristics of the network, it is possible to apply it for mod- elling and identification of nonlinear systems. It is especially useful when there are no mathemati- cal models of the modelled system available, and analytical models and parameter-identification al- gorithms cannot be applied (Patton et al., 2000; Frank and K¨ oppen-Seliger, 1997). Based on the dynamic neural networks, the fault detection and isolation system for diagnosis of industrial pro- cesses can be designed (Patton et al., 2000; Chen and Patton, 1999). In recent few years this kind of neural networks was successfully applied in several fault diagnosis applications like fault detection and isolation in a two tank laboratory system or fault detection of the instrumental faults in the sugar evaporator (Patan and Korbicz, 2000; Ko- rbicz et al., 1999). In the present study, the dy- Copyright © 2002 IFAC 15th Triennial World Congress, Barcelona, Spain