Genetic evolving of dynamic neural networks with application to process fault diagnosis

The robustness issue in model-based diagnosis of process faults is addressed by means of artificial neural networks. The symptoms are generated by using observer schemes with dynamic neural nets. Their design is based on a hierarchical genetic algorithm, extended back-propagation method and multiobjective optimisation. The evolutionary search of genetic type is used to find the optimal architecture of the dynamic networks. Static networks are then used to classify the symptoms. Application to a laboratory process illustrates the approach. It refers to component and instrument fault detection and isolation in a three-tank system.