Abstract: Aiming at the problems of various transformer faults and insufficient fault samples, a transformer fault judgment method based on acoustic signal feature fusion was proposed. A large number of normal sound signals were extracted from the normal transformer to find out the characteristics of the transformer. In different spatial domains such as time domain and frequency domain, feature extraction was carried out on the collected normal transformer acoustic signals, and the feature set that was most consistent with the transformer acoustic signals was found by the maximum-correlating-minimum-redundancy algorithm, and then the weighted entropy principal component method was used to obtain the fusion index for the feature set. The fusion index was compared with the normal signal and fault simulation signal of 220 kV and 110 kV transformer of a substation under the condition of noise interference. The results show that the fusion index obtained by this method can distinguish clearly between normal and simulation fault, and has good recognition effect, which verifies the feasibility of this method.