With the expansion of the scale and complexity of railway communication systems, relay protection devices are prone to malfunction and missed faults in high-interference and multi-state overlapping environments. This study proposes a deep diagnosis method based on a multi-scale attention network, combined with convolution extraction, frequency domain enhancement and label correction mechanism, to improve the ability to identify key abnormal features in time series signals. Comparative experiments show that MSANet has an accuracy of 94.48%, an F1 score of 93.22% and a misdiagnosis rate of 3.02%, which is better than CNN-GRU, VAE-CNN and other methods. The anti-interference performance is stable, and the false alarm recovery time is shortened to 2.3 s. It still maintains an accuracy level of 88.6% under strong noise disturbance. This method has outstanding diagnostic adaptability and engineering deployment potential in complex signal environments.