With the trend of large-scale wind turbines, real-time fault detection and diagnosis of generators is of great significance. In this study, an EWT-BiCRNN fault diagnosis method that integrates empirical wavelet transform and multi-scale spatiotemporal features is proposed, based on convolutional neural network (CNN) and long short-term memory network (LSTM). First, in the preprocessing stage, the empirical wavelet transform method is used to extract the modal components containing the vibration characteristic frequencies. Then, a three-way parallel convolutional neural network is designed to extract multi-scale vibration fault space features. At the same time, two parallel bidirectional long short-term memory networks are integrated to obtain global information using temporal characteristics. Finally, the diagnostic model is output, using cross loss as the loss function, and Adam is used as an optimization method to complete generator fault diagnosis. The loss is only 0.0267, and the accuracy reaches 0.9991. Compared with traditional fault diagnosis algorithms, the EWT-BiCRNN in this paper has more full-scale spatial characteristics and mines timing information before and after the fault occurs, making this method more promising for industrial applications.