The traditional fault diagnosis method often relies on the complex signal processing procedures and experts’ rich experience. It requires precise signal segmentation, which is a tedious process and is not conducive to the field use. In this paper, the deep belief network (DBN) method optimized by particle swarm optimization (PSO) is used to directly extract features from the original power data, and the restricted Boltzmann machine (RBM) is employed to fit the data features layer by layer, achieving the data dimension reduction at the same time. Then, extreme learning machine (ELM) is used to classify each state, thereby improving the diagnosis speed. The results show the accuracy reaches 96%, which is a 4% improvement, and the required time is significantly reduced, when compared to support vector machine (SVM) optimized by PSO.