Transformer internal fault diagnosis usually needs to use dissolved gas in oil, but the information extraction, detection and analysis process is cumbersome, and the real-time performance is poor. Therefore, this paper proposes a method of transformer internal fault diagnosis which only needs electrical information. The frequency domain fault characteristics of short-circuit current and differential current are extracted by wavelet packet analysis. The fault characteristics of zero sequence current are represented by the maximum value. The information fusion technology is used to fuse all the fault features, and the BP neural network algorithm is used to diagnose the internal electrical fault types of transformer. Using Matlab / Simulink platform modeling and simulation analysis, the results show that the proposed internal electrical fault identification strategy of transformer has high accuracy and high reliability.