ART-KOHONEN neural network for fault diagnosis of rotating machinery

In this paper, a new neural network (NN) for fault diagnosis of rotating machinery which synthesises the theory of adaptive resonance theory (ART) and the learning strategy of Kohonen neural network (KNN), is proposed. For NNs, as the new case occurs, the corresponding data should be added to their dataset for learning. However, the 'off-line' NNs are unable to adapt autonomously and must be retrained by applying the complete dataset including the new data. The ART networks can solve the plasticity-stability dilemma. In other words, they are able to carry out 'on-line' training without forgetting previously trained patterns (stable training); it can recode previously trained categories adaptive to changes in the environment and is self-organising. ART-KNN also holds these characteristics, and more suitable than original ART for fault diagnosis of machinery. In order to test the proposed network, the vibration signal is selected as raw inputs due to its simplicity, accuracy and efficiency. The results of the experiments confirm the performance of the proposed network through comparing with other NNs, such as the self-organising feature maps (SOFMs), learning vector quantisation (LVQ) and radial basis function (RBF) NNs under the same conditions. The diagnosis success rate for the ART-Kohonen network was 100%, while the rates of SOFM, LVQ and RBF networks were 93%, 93% and 89%, respectively. (C) 2003 Elsevier Ltd. All rights reserved.