Neural networks for pneumatic actuator fault detection

The suitability of artificial neural networks (ANNs) for detecting fault conditions in pneumatic control valve actuators is investigated. Specifically, the ability of a neural network to act as a predictor of correct valve behaviour is examined. Experimental results indicate that standard network architectures are unsuitable for temporal prediction of non-linear system behaviour. An original recurrent network architecture, designed specifically as a predictor and based on autoregressive models and functional approximation is therefore proposed. The performance of this network is evaluated using both measured data and data from simulations based on a mathematical model of the valve. Laboratory implementation of the fault detection system produced encouraging qualitative results, including high success rates for the detection of faults corresponding to valve Coulomb friction changes and input pressure offsets.