Prior-augmented measurement-signal fusion using a dual-branch cross-attention network for cross-domain bearing fault diagnosis

Abstract

Cross-machine bearing fault diagnosis is strongly affected by inconsistencies in vibration measurement conditions, including rotational speed, sampling frequency, and structural transmission paths. These factors cause speed-induced fault-frequency drift and cross-domain distribution discrepancies, making features learned from source-domain measurements unreliable in target-domain scenarios. Existing transfer learning (TL) methods are predominantly data-driven and insufficiently exploit mechanism-related prior information, which limits their interpretability and cross-machine generalization. To address these challenges, this paper proposes a prior-augmented dual-branch cross-attention network, termed PA-DCA Net, for cross-domain adaptive bearing fault diagnosis. First, a multi-scale S-transform with channel-weighted fusion is used to construct informative time-frequency representations from vibration signals. Meanwhile, a speed-normalized prior feature is introduced at the input-feature level to encode the relative rotational-speed discrepancy between the current operating condition and the source-domain reference condition. This prior feature is combined with eight conventional time-domain statistical features to form a statistical-prior feature vector. Second, an image-statistical dual-branch network is constructed, in which the image branch extracts deep time-frequency features and the statistical branch maps the statistical-prior vector into a high-dimensional representation. Multi-head cross-attention is then employed to achieve directed feature interaction between the two modalities. Third, a progressive TL framework integrating source-domain supervised pretraining, few-shot target-domain fine-tuning, CORAL, multi-kernel maximum mean discrepancy, and FixMatch-based consistency regularization is adopted. The proposed method is validated on five cross-domain tasks constructed from three public bearing datasets. PA-DCA Net achieves average accuracies of 97.10% and 96.92% on the Case Western Reserve University (CWRU)-to-Jiangnan University and CWRU-to-Huazhong University of Science and Technology cross-machine tasks, respectively, outperforming several representative transfer-learning baselines.