Fault Diagnosis Application of Train Air Brake Systems Under Strong Noise using Enhanced Residual Network

In high-speed train, the intricate nature of air brake systems coupled with fluctuating operating environments often leads to raw data being heavily contaminated with noise, presenting substantial challenges for current deep learning-based fault diagnosis methods. To address these issues, this paper introduces a noise-resilient optimized residual network. This approach is composed of three main components: the Inverted Residual Block (IRB), Efficient Channel Attention (ECA), and a soft thresholding function (ST-func). The IRB establishes signal correlations, enhances inductive bias capability, and extracts richer local feature information by varying the input channel dimensions. ECA serves as side paths for residual basic blocks, improving the model's perception of relationships between different channels in the input signal features, thereby increasing attention to critical channels. Finally, features learned from the Global Average Pooling and ECA modules are input into the ST-func to eliminate irrelevant features. Experimental results demonstrate that, compared to existing mainstream deep learning fault diagnosis methods, the proposed approach offers significant advantages in terms of lightweight design and robustness in handling noisy fault diagnosis tasks.