A frequency-prior guided dual-attention and gated fusion network for intelligent bearing fault diagnosis

Current deep learning methods for bearing fault diagnosis still suffer from insufficient exploitation of physical priors, difficulty in jointly modeling global and local features, and limited adaptability of fusion mechanisms. To address these issues, this study proposes frequency-prior attention and gated fusion network (FPAGF-Net), a deep learning framework that integrates frequency priors. First, the raw vibration signals are decomposed by variational mode decomposition into intrinsic mode functions with distinct center frequencies, which are incorporated into the network as explicit frequency priors. Subsequently, a dual-channel feature extractor is constructed, where a transformer encoder captures long-range dependencies and global patterns, while a temporal convolutional network models local impacts and transient details. To achieve effective feature integration, a frequency-prior-guided dual-attention mechanism (channel and temporal attention) is designed, together with a frequency-prior-driven gated fusion module that adaptively balances the importance of global and local features. Finally, the fused features are passed through a classifier to output the probability distribution of fault categories. Experimental results on the Case Western Reserve University and Machinery Failure Prevention Technology datasets demonstrate that FPAGF-Net outperforms comparison methods, with consistent improvements in both accuracy and robustness. Ablation studies further confirm the essential roles of frequency priors, dual attention, and gated fusion in the overall architecture. In summary, the proposed model effectively integrates signal-processing priors with deep learning capabilities, providing a novel and efficient solution for intelligent bearing fault diagnosis under complex operating conditions.