Abstract: To address the non-stationary characteristics of rotating-rectifier open-circuit fault signals in three-stage generators, the difficulty in extracting effective fault features, and the limited interpretability of conventional deep learning methods, a fault diagnosis method, namely generalized S-transform-embedded convolutional neural network (GSTCNN) was proposed. By embedding the generalized S-transform kernel into the first layer of the convolutional neural network, a generalized S-transform convolutional layer with explicit physical time-frequency meaning was constructed. Channel-wise learnable frequency factors and window-width parameters are introduced to enable adaptive adjustment of fault-sensitive frequency bands and time-frequency resolution. On this basis, the generalized S-transform convolutional layer was integrated with a deep convolutional network to construct a fault diagnosis model with time-frequency interpretability. Experimental validation was conducted on multiple diode open-circuit faults of the rotating rectifier in a three-stage generator. The results show that under the 64-channel configuration, GSTCNN achieves an average diagnostic accuracy of 99.26% with a standard deviation of 0.47. Even under a low signal-to-noise ratio of -4 dB, the diagnostic accuracy remains as high as 97.86%. Furthermore, the learned frequency factors and window-width parameters reflect the key fault-related frequency bands and the model preference for time-frequency resolution. The above results verify the effectiveness of GSTCNN in feature extraction from non-stationary fault signals, and demonstrate that it can simultaneously achieve diagnostic accuracy, noise robustness, and model interpretability. This study provides a new method with high accuracy, strong noise robustness, and physical interpretability for open-circuit fault diagnosis of rotating rectifiers in three-stage generators.