A neural network based approach for machine fault diagnosis

In this paper a novel approach to classify the state of a machine based on vibration measurements and the use of dynamic neural network is presented. Some comparisons are made between the presented method, the linear classifier, the third-order nonlinear classifier, Markov model based classifier and the recurrent backpropagation network. The proposed classifier can be considered as a spatiotemporal neural network. Spatiotemporal neural networks are used to transform input patterns into timevarying class number output codes. Usually, time is assumed to move forward in small discrete steps. The recurrent backpropagation network and the Spatiotemporal Pattern Recognizer Neural Network (SPRAIN) are other examples of spatiotemporal neural networks. The spatiotemporal neural network can effectively store much more information than most other types of neural networks with same number of nodes. The presented approach is suited to machine maintenance for two reasons: Firstly, it can model temporal relations. For example, it can describe the development of the symptoms of faults. Secondly, it can efficiently handle large amounts of data. Because the spectral signatures of faults may change significantly depending on environmental, operating and physical conditions, the amount of training information is very large.