Abstract: Aiming at the problems of slow convergence speed and large error in gear fault diagnosis based on traditional BP neural network, a gear fault diagnosis method combining EMD and BP neural network is proposed. Firstly, the basic principles of EMD and BP neural network are introduced. Then, the IMF components of gear time domain signals are extracted by EMD method , the energy characteristic parameters of fault signals in IMF components are calculated, and these energy characteristic parameters are used as input parameters of BP neural network for fault diagnosis. Sufficient sample data are collected on the gear transmission failure test-bed for experimental study. The results show that compared with the traditional BP neural network, the training error can be reduced from 0.01 to 0.001. In addition, the number of training iterations can be reduced to less than 10.