Rolling Bearing Fault Diagnosis Considering Long-Term Dependence and Time-Frequency Feature Fusion

Most of the current research models on rolling bearing fault diagnosis can effectively improve the reliability of bearings, but there are still some shortcomings. To address the issues of limited long-term dependency and sparse feature representation in existing rolling bearing fault diagnosis models, this paper proposes a novel method that incorporates long-term dependency and time-frequency feature fusion. The proposed method is based on a cross-deployed basic block with identity mapping and down-sampling, enabling the capture of long-term dependencies in bearing vibration time series signals. Furthermore, the use of skip connections facilitates effective information flow, allowing the network to integrate local dependencies, long-term dependencies, and time-frequency features at multiple scales. Experimental results on the Case Western Reserve University bearing dataset, which includes ten fault types, demonstrate that the proposed model achieves a detection accuracy of 86.8%. This performance surpasses that of the One-Dimensional Convolutional Neural Network (CNN1D) by 9% and the Multi-Layer Perceptron (MLP) by 16.3%. Moreover, the accuracy improvement is attributed to the incorporation of long-term dependency and time-frequency feature fusion. This research offers valuable insights for the intelligent diagnosis and predictive maintenance of bearings.