A reliable rolling bearing fault diagnosis method based on a time-frequency fusion network with frequency-band attention

Abstract

Reliable rolling bearing fault diagnosis remains challenging because fault information is distributed across transient impacts, temporal dependencies, and frequency-dependent energy patterns. To address this challenge, this study proposes TFBF-Net, a temporal–frequency fusion network with frequency-band structural attention. The temporal branch combines one-dimensional convolution, BiGRU, and a lightweight Transformer encoder to capture local impacts, bidirectional dependencies, and long-range temporal relations from raw vibration signals. The CWT branch preserves the ordered frequency-scale structure of time–frequency maps and generates band-specific attention weights by pooling only along the temporal dimension. The two compact branch features are concatenated and classified by a lightweight fusion head. Experiments on the Beijing Jiaotong University-Rail Autonomous Operations (BJTU-RAO), CWRU, SEU, and GDUPT datasets show that TFBF-Net achieves an accuracy of 88.45 ± 0.81% on GDUPT while maintaining moderate model complexity. These results indicate that the proposed method provides effective fault discrimination under the adopted strict splitting protocol.