The transformer fault diagnosis technique based on a single detection method is difficult to identify the same type of defects of oil-paper insulation in detail, which cannot meet the requirements of power system on equipment operation reliability under the background of rapid development of deep offshore wind power. Therefore, an oil-paper insulation defect identification method based on information fusion of phase-resolved partial discharge (PRPD) spectrum and dissolved gas analysis (DGA) data was proposed. Six kinds of electrode models were designed and made to simulate the typical defects of surface discharge in transformers with different electric field inhomogeneity coefficients, and PRPD and DGA data were collected. Then convolutional neural network (CNN) and back propagation neural network (BPNN) were adopted to recognize the patterns of PRPD spectrum and DGA feature vector of six kinds of defects, respectively. Finally, the CNN-BPNN information fusion model based on D-S evidence theory was proposed to realize joint diagnosis based on PRPD spectrum and DGA data. The results show that the CNN-BPNN model based on the D-S evidence theory can effectively correct the wrong output of the single criterion model and reduce the uncertainty of the classification results. When the input dimensions of PRPD spectrum are 8×8, 16×16, and 32×32, the recognition accuracy of the model integrated with the DGA feature vector is 93.21%, 97.53%, and 99.17%, respectively, which is 4.81%, 2.78%, and 0.84% higher than that of PRPD single criterion model. The CNN-BPNN model can effectively integrate the electrical physical information and chemical product information of partial discharge, which not only improves the accuracy of defect identification, but also enhances the confidence of the output results, and reduces the data storage requirements, providing accurate, reliable, and lightweight defect identification methods for intelligent operation and maintenance of transformers.