Aiming at the difficulty of partial discharge defect identification of ultra-high frequency (UHF) signals, a discharge identification method based on ensemble empirical mode decomposition (EEMD) and deep learning is proposed. Firstly, the original signal is decomposed by EEMD, and the time scale feature is used to optimize the unstable signal, and the modal aliasing is avoided by multiple set averaging to reduce the signal noise. Secondly, the convolutional neural network is used to train and learn the multi-class feature information of the discharge signal, and the reconstruction method is used to weight the output results, which avoids the deficiency of single feature recognition. The test results show that the ensemble empirical mode decomposition realizes the signal-to-noise separation of the detection signal, and the machine learning algorithm with multi-feature information realizes the classification and recognition of the discharge signal, which has high accuracy and reliability of defect judgment.