Bearing Fault Diagnosis Using Dynamic–Static CNN Based on Multi-Sensor Fused Frequency Recursive Graph Under Complex Working Conditions

The operational state of rolling bearings in rotating machinery is critical to the reliability and operational safety of large-scale mechanical equipment. However, actual industrial scenarios often involve varying loads and rotational speeds, which can lead to changes in vibration characteristics and reduce the discriminability of fault-related features. Meanwhile, conventional one-dimensional vibration signals have limited capability in characterizing the complex relationships among frequency components and fault-induced dynamic variations. To address these issues, this paper proposes a multi-sensor fused frequency recursive graph (MSFRG) feature enhancement method, together with a parallel dynamic–static convolutional network embedded with an efficient channel attention (ECA) mechanism, termed ECA-PDSNet, for an end-to-end intelligent fault diagnosis scheme. The MSFRG enhances the representation of fault-related information through multi-sensor fusion, while the ECA-PDSNet simultaneously extracts stable static features and adaptive dynamic features, with the ECA module further emphasizing informative feature channels. Experimental results on the CWRU and YSU-219 bearing datasets demonstrate that the proposed method achieves classification accuracies of 100% and 99.5%, respectively, under single working conditions. When samples from multiple operating conditions are mixed for training and testing, the proposed method achieves accuracies of 99.75% on CWRU and 99.50% on YSU-219. Furthermore, in the CWRU cross-load experiments, the model trained under 0 HP achieves accuracies of 97.80%, 96.70%, and 88.50% under 1 HP, 2 HP, and 3 HP, respectively, with an average accuracy of 94.33%. These results demonstrate the effectiveness of the proposed method under fixed and mixed operating conditions, as well as its cross-condition generalization capability under varying loads.