Abstract: In recent years, neural network models based on impulse frequency response analysis (IFRA) have been proven effective in detecting stator winding faults. However, these models are generally characterized by weak robustness and poor noise resistance. The reason is that most of the models adopt simple neural network architecture and conventional IFRA generally use fast Fourier transform (FFT) to perform time-frequency transformation on transient signals, while FFT is not suitable for processing transient abrupt non-stationary signals. In this paper, the stator winding of a loose wound synchronous machine is taken as the detection object and continuous wavelet transform (CWT) instead of FFT is used to process the transient signal of IFRA, and based on one-dimensional convolutional neural networks (CNN) and bi-directional long short term memory (BiLSTM) networks, a CNN-BiLSTM model is constructed to detect the fault of the data transformed by CWT. The experimental results show that, compared with other models with the transformed single, the CNN-BiLSTM model with CWT processed frequency domain sequence as input has the best average accuracy of 99.01%. The noise contrast experiment shows that data transformed by CWT can enable the fault diagnosis model to be more robust and generalized.