Interpretable and robust fault diagnosis of rotating machinery in noisy environments via improved high-order spatial interactions network

According to the problems of the existing fault diagnosis (FD) model being affected by noise and lacking interpretability, this paper proposed an innovative FD model for rotating machinery, named noise critical layer adaptation (NCLA). By designing the module of noise robustness criticality (NRC), the model effectively focuses on layers most affected by noise, significantly improving classification accuracy and interpretability in noisy environments. Furthermore, this study designed an improved feature extraction framework based on the high-order spatial interactions with recursive gated wavelet convolution (WTConv) network, which enables the model to decompose and process signal components at different frequencies, enhancing its robustness and capability to capture fine-grained features. Unlike traditional models that rely on datasets with identical feature distributions, the proposed model was pre-trained on noise-free data and tested on noisy datasets, which aligns more closely with actual engineering applications. Experimental results of the two cases demonstrated that the model exhibits superior generalization and robustness across various noise conditions, outperforming conventional approaches. Additionally, by visualizing the impact of noise on critical layers, the proposed model addresses the limitations of the black box in deep learning methods.