Abstract: It is commonly accepted by many researchers that multiple evidence from different sources of different importance or reliability are not equally important when they are combined according to Dempster-Shafer theory, but it is seldom considered in the existent combination methods. A new method is presented to solve this problem, by which the considered evidence are first balanced according to the weighted average of all and then combined. The method is incorporated into a neural network classifier, which is based on Dempster-Shafer theory, to construct a weighted evidence network and the network is applied to mechanical equipment fault diagnosis problem in the followed experiments. The experiment results demonstrate the excellent performance of this network as compared to the improved RBF network; also the validity of the proposed method in improving the combination’s accuracy of multiple evidence is proved.

Key words: Combination rules, Evidence theory, Neural network, Pattern recognition