Efficient and reliable fault classification is beneficial to guide the dispatchers in finding and removing the fault quickly, thus restoring system power supply promptly, so the classification is of great significance for ensuring the safe and reliable operation of the system. A fault classification method based on deep belief network is proposed to overcome deficiencies of traditional fault classification, such as shallow intelligent methods' dependence on signal processing technology and artificial experience, and the lack of feature extraction and expression for complex power system. The raw data of each phase current- voltage and zero sequence current-voltage are taken as the network input, and the features of fault state are automatically learned and extracted from the original time-domain signals to realize the fault type identification. The simulation results of the IEEE 39-bus system and real fault cases of power grid show that the proposed fault type identification method has good capability of fault feature extraction. Besides, the proposed method keeps the original characteristics of data in the process of dimensionality reduction, and it is not affected by factors including transition resistance, fault time, fault location and load size. Therefore, it identifies fault types more accurately than other traditional artificial neural networks do.