The way of signal feature extraction can directly affect the results of fault diagnosis, so a new method of combining feature vectors for effective fault mode identification was proposed to extract the signal features from the original signal that could maximize the characterization of the information it containsed. The energy of the effective IMF components of the original signal after empirical modal decomposition and the energy entropy of the signal were combined as feature vectors. Due to the absence of mechanical rotor fault diagnosis often occurs, the DPSO algorithm was proposed to optimize the BP neural network. The method was improved mainly by optimizing the initial weights and thresholds of the neural network. The results show that compared with the traditional BP neural network model, the iteration number of improved BP neural network model is significantly reduced, the training time is correspondingly shortened, the model training accuracy and the fault diagnosis correct rate are also improved.