Partial discharge is an important cause of insulation breakdown of high-voltage electrical equipment, but also an important indicator of insulation deterioration, in view of the current switchgear partial discharge conventional detection methods have a small amount of detection information, poor timeliness, low diagnostic accuracy and other issues, this paper proposes a convolutional neural network detection method that can be integrated in mobile devices, and for the actual situation there is a problem of uneven discharge class samples, a fault sample generation method is proposed. The collected ultrasonic signal is de-denoised and pre-processed into a two-dimensional temporal spectrogram by short-term Fourier transform, and the partial discharge pattern is identified in the input convolutional neural network, and the adversarial network is used to generate the fault sample for the problem of uneven fault sample in the actual scene. The example experiments show that the accuracy rate of the proposed method in this paper reaches more than 97%, the computing power reaches 0.27 seconds under the condition of t710 computing power of the mobile terminal, and the error of the MSE generated data sample is lower than 0.067.