Real-time condition monitoring using neural networks

This paper describes a Pattern Recognition (PR) technique which uses Learning Vector Quantisation (LVQ). This method is adapted for practical application to solve problems in the area of condition monitoring and fault diagnosis where a number of fault signatures are involved. In these situations, the aim is health monitoring, including identification of deterioration of the healthy condition and identification of causes of the failure. In real-time situations, the problem of fault diagnosis may involve a large number of patterns and large sampling time, which affects the learning stage of neural networks. The study here aims at finding a fast learning model of neural networks for instances when a high number of patterns and numerous processing elements are involved. It begins searching for an appropriate solution. The study is extended to the enforcement learning models and considers LVQ as a network emerged from Competitive Learning model through enforcement training. Finally tests show an accuracy of 92.3% in the fault diagnostic capability of the technique.