Dissolved gas analysis (DGA) of insulating oil can provide an important basis for transformer fault diagnosis. To improve diagnosis accuracy, this paper presents a new transformer fault diagnosis method based on deep belief networks (DBN). By analyzing the relationship between the gases dissolved in transformer oil and fault types, the Non-code ratios of the gases are determined as the characterizing parameter of the DBN model. DBN adopts multi-layer and multi-dimension mapping to extract more detailed differences of fault types. In this process, the diagnosis parameters are pre-trained. A back-propagation algorithm adjusts them with the labels of the samples and optimizes the parameters. To verify the effect of the proposed method, the diagnostic DBN model is constructed and tested using various oil chromatographic datasets collected from the State Grid Corporation of China and previous publications. The performances of the DBN diagnosis model are analyzed by different characterizing parameters, different training datasets and sample datasets. In addition, the influence of discharge and overheating multiple faults on the diagnosis model is studied. The performance of the proposed approach is compared with that derived from support vector machine (SVM), back-propagation neural network (BPNN) and ratio methods respectively. The results show that the proposed method significantly improves the accuracy of power transformer fault diagnosis.