Due to the complexity of the internal structure of the motor, there is a strong nonlinear relationship between the fault feature and the type of fault; also, the methods of asynchronous motor fault diagnosis are manual extraction of features, which requires a large number of prior knowledge, abundant signal processing theory and practical experience as support; at the same time, the amount of samples used in pattern recognition is too small, which may lead to overfitting. Therefore, a fault diagnosis method based on short-time Fourier transform (STFT) and convolutional, neural, networks (CNN) is proposed. The method uses a single vibration signal as a monitoring signal, and uses STFT to convert the signal into a time spectrum, and serves as a sample input for the network to supervise the training, which ensures the diversity of the sample, improves the robustness of the network and achieves accurate fault diagnosis. It is compared with the traditional motor fault diagnosis method and the stacked denoising autoencoder. The test results show that this method can effectively diagnose motor fault.