Fault classification based on improved evidence theory and multiple neural network fusion

Combination results of the evidence theory will be out of accord when the evidences highly conflict with the real condition, some improved methods can solve this problem, but the convergence speed is rather slow. Furthermore, the combination of these improved methods will diverge when the evidences are consistent, thus limiting the application of evidence theory in condition monitoring system. In view of this, a novel evidence combination approach based on evidence confidence is proposed, and a multiple neural network fusion model is constructed for fault classification on the basis of this improved evidence combination method. A case of gear fault diagnosis using the proposed model is studied. The fault feature space is divided into several subspaces, and the corresponding sub-neural network classifiers are established. The output of these sub-neural network classifiers are used as the combination evidences. Finally, different faults are classified through combining the obtained evidences by the novel combination method. Comparing with the classification results of traditional evidence theory method, other representative improved methods and traditional neural network method, experiment results indicate the effectiveness of this improved evidence fusion method. ©2010 Journal of Mechanical Engineering.