NFFNet: A Deep Learning Framework for Fault Diagnosis Using Classifier-Guided Graph Attention Networks and Blind Deconvolution

Ensuring the safety and reliability of train transmission systems is crucial for efficient rail transit operations. These systems are prone to faults that can lead to service disruptions or accidents, making effective fault diagnosis essential for operational safety. However, existing methods often fail to extract features effectively in noisy environments and neglect interactions between components, such as vibration propagation from motors to axle boxes. To address these limitations, this paper proposes NFFNet, a Graph Attention Network that incorporates component spatial relationships for fault diagnosis in train transmission systems. NFFNet employs a Blind Deconvolution (BD) technique combined with a Deep Residual Shrinkage Network (DRSN) for robust and adaptive feature extraction under noisy conditions. It further integrates spatial relationships within a graph attention network to model interactions between components, enhancing diagnostic accuracy. By recovering latent fault signals obscured by noise and dynamically modeling interactions among adjacent components, NFFNet achieves more reliable and precise fault diagnosis. Experimental results demonstrate that NFFNet significantly improves accuracy and robustness in high-noise environments compared to state-of-the-art methods.