Survey of neural network approach for induction machine on-line diagnosis

Fault detection and diagnosis is currently a very important problem in induction machine management. Both model-based method and expert systems have been suggested to solve the problem. Recently neural networks have been advocated as a possible technique to handle diagnostic tasks providing them with an effective improvement. Neural networks can be applied autonomously or can integrate with existing diagnostic tools. Several architectures for fault diagnosis are studied. In this paper the attention is focused to the multilayer perceptron and to the self organizing map networks which, with different features, seem best suited for induction machine diagnostic tasks. The application of the different neural architectures to specific problem by practical examples is discussed. In particular it will be shown that the synergy between the two mentioned neural architectures provides a global diagnosis approach: NNs specifically trained for dealing with certain tasks are the basic elements. A first level NN classifies the fault, then several second level fault specific NNs (FS-NNs) for stator short-circuits, bearing damages, rotor bar breakages etc. evaluate the fault severity.