The single signal source feature cannot accurately identify the type of partial discharge. In order to solve this problem, a partial discharge pattern recognition method for high-voltage motors based on improved BP neural network and D-S evidence is proposed. The pulse phase information, ultra-high frequency signal and ultrasonic signal of different types of partial discharges are collected, and the eigen parameters of different signals are extracted. Then, the improved back propagation neural network (BPNN) recognition model based on whale optimization algorithm (WOA) is constructed to identify the partial discharge types, respectively. The recognition results of each recognition model are used as the evidence body to be fused using the D-S evidence combination rule. Finally, the fusion result is decided. The research results show that there are differences and uncertainties in the accuracy of independently identifying various types of partial discharges based on three types of single signal sources, and the recognition rates are 83.3%, 90.0%, and 83.3%, respectively. However, the commonality and difference of the three types of signal sources can be integrated and complementary, and have their own advantages, which can solve the problem of uncertainty in fault diagnosis. On this basis, the overall recognition rate of high-voltage motor partial discharge types based on D-S evidence fusion has increased to 96.6%, and three signal sources have been realized. Compared with a single model, the proposed method can identify partial discharge patterns stably and accurately, and has higher accuracy and reliability, which verifies the effectiveness and correctness of the proposed method.