Abstract: In order to further study typical partial discharge types of transformers, a feature recognition method is proposed by combining CCMDA and deep learning. Firstly, the dimension-reduction is carried out on the original data to retain the key feature; then, the deep learning framework is introduced to carry out the identification of PD characteristic, which is carried out by setting up the residual structure of ResNet34. The results show that, compared with kernel principal component analysis (KPCA), the dimensionality reduction effect of correlation coefficient matrix is more significant. When the image size is 64×64, the recognition performance is the best. When the learning rate is 0.001, loss function is the smallest. The proposed method has the highest recognition accuracy, which is obviously better than that of convolutional neural networks (CNN) and support vector machine (SVM). The robustness of the proposed method can be enhanced by introducing wavelet transform.