In this paper, the measured data of four kinds of faults in a substation of the Southwest power grid are used as the data set. A method based on ensemble learning (EM) is designed for the fault diagnosis of a high voltage direct-current (HVDC) system, one which can significantly improve the speed, accuracy, and robustness of fault diagnosis. First, data preprocessing for the four types of fault data is conducted. At the same time, the feature of fault data is extracted and trained. The fault data label is used to average the weight of the fault data set. Then the classification error of the current weak classifier for the weighted data set is calculated, as well as the weight of the current classifier in the strong classifier. Finally, the distribution of the weights of the training samples is updated to obtain a strong classifier. From the trained model, the fault types are identified in the different data sets. Compared with a back propagation neural network fault diagnosis model, the proposed method can achieve more than 89% diagnostic accuracy in multiple tests, with a low error rate and strong robustness. It is conducive to fault identification and rapid diagnosis in an HVDC system in operation.