Identification of vibrating structures and fault detection using neural networks

An investigation is undertaken to ascertain the suitability of network based nonparametric regression for multivariate nonlinear system identification and fault detection. A network that makes use of the theory of autoregressive models and functional approximation is proposed. A feature of this network is the use of different basis functions in each hidden layer. Observation of the network weights shows which basis functions dominate, thereby revealing information about the physical system. During this analysis step, only the dominant functions are retained to reduce error variance. In a further refinement step, several sigmoid functions are added to the network to generate a smooth stepwise approximation to the part of the mapping unexplained by the dominant basis functions. By implementing the network, it is possible to generate a structural model for a 3 kW induction motor. The network exhibits a 100% success rate in the detection of 1.8 A armature current variations.