A Noise-Resilient Fault Diagnosis Method Based on Optimized Residual Networks

In industrial applications, the complexity of systems and the dynamic nature of operating environments often result in raw data being heavily contaminated by noise. This presents significant challenges for traditional deep learning-based fault diagnosis methods, which may struggle to handle such noisy data effectively. To address these challenges, this paper proposes a noise-resilient optimized residual network. The proposed network comprises three main modules: the inverted residual block (IRB), the efficient channel attention (ECA) mechanism, and a soft thresholding function (ST-func). The IRB enhances the model’s ability to capture signal correlations, improve inductive bias, and extract richer local features by varying the input channel dimensions. The ECA mechanism, designed as auxiliary paths for the residual blocks, improves the model’s ability to perceive relationships between different channels in the input features, thus focusing more on critical channels and boosting overall performance. Finally, the features extracted through Global Average Pooling and ECA are passed through the ST-func to remove irrelevant information and further refine the feature set. Experimental results demonstrate that, in comparison to existing mainstream deep learning fault diagnosis methods, the proposed approach offers significant advantages, including a lightweight design and superior robustness in handling noisy fault diagnosis tasks, making it highly suitable for practical industrial applications.