A fault diagnosis method based on the combination of neural network and Dempster/Shafer (D-S) evidence theory is proposed. Feature level and decision level fusion of fault signal was realized， which was applied in research of the composite fault diagnosis of bearing. The output results of back propagation (BP), radial-based function (RBF), general regression neural network (GRNN) three kinds of neural networks were used as three body of evidence. Four kinds of compound fault characteristics of rolling bearing were regarded as system identification framework. Clustering coefficient was introduced as the weight distribution, and the basic probability assignment was recalculated. The D-S evidence theory is improved to improve the accuracy of the composite fault diagnosis of bearing.