This paper proposes the diagnostic method which is based on EMD-SVD feature extraction and random forest classifier in order to more accurately assess the transformer oil paper insulation aging stage. The paper Set up the experimental platform to collect air gap defect samples of different thermal aging stages partial discharge signals and obtained partial discharge signal characteristics after denoising processing and EMD-SVD feature extraction. The EMD-SVD feature are diagnosed by random forest classifier with the traditional BP neural network classifier and support vector machine respectively, and the results show that random forest classifier recognition result is superior to the traditional classifier. Compared with the traditional classifier, random forest classifier classification ability is better for EMD-SVD characteristics classification. The paper demonstrates that the EMD-SVD features combined with random forest classifier applied in oil paper insulation thermal aging phase recognition effect is better.