Considering that the existing transformer fault diagnosis methods are only for a single fault feature, it is difficult to make an accurate and comprehensive judgment on the actual situation of the power transformer. On the basis of multi-dimensional information fusion of power transformer, a fault diagnosis method combining the improved extreme learning machine and the improved D-S evidence theory is proposed. The output of the limit learning machine is optimized by a posteriori probability mapping, and the probabilities of different labels are obtained, and the improved evidence theory is used to fuse the probability distribution matrix. The superiority of this method is verified by comparing and analyzing the diagnosis methods before and after optimization. This method has higher fault identification accuracy, and the accuracy rate reaches 96.50%, and can accurately identify various faults of power transformers, which can provide decision-making basis for condition maintenance.