A fault morphology-preserving and cross-source consistency learning method for domain generalization in bearing fault diagnosis

To address the cross-condition fault diagnosis challenges in rolling bearings, including the sensitivity of fault morphology to operating-condition interference, the limited stability of single-view representations, and the scarcity of target-condition samples in practical applications, this paper proposes a fault morphology-preserving and cross-source consistency learning network. The proposed method develops a unified framework from the joint perspective of fault morphology preservation and operating-condition disturbance suppression. Specifically, multiple complementary time–frequency structural representations are constructed and fused in a multi-scale collaborative manner to better preserve intrinsic fault morphology. A cross-source consistency learning mechanism is then introduced to reinforce shared fault structures across different operating conditions while suppressing condition-specific interference. In addition, a condition-adaptive receptive field modeling mechanism is designed to improve the characterization of anisotropy, scale drift, and local structural variations in fault textures. Extensive experiments on two widely used bearing datasets demonstrate that the proposed method consistently outperforms existing approaches, achieving an average diagnostic accuracy of over 99 %. Under Gaussian noise with an SNR of −5 dB, FMCSL-Net still achieves mean accuracies of 90.15 % and 90.14 % on the Ottawa and HUST datasets, respectively, demonstrating strong robustness and cross-condition generalization capability.