TFAR-UDP: A time-frequency adaptive domain generalization framework for cross-condition fault diagnosis of transmission systems

To address the severe distribution shifts of transmission systems under cross-speed and cross-load conditions, as well as the inability of conventional convolutions to adaptively decouple time-frequency features, which renders models vulnerable to source domain overfitting amidst temporal impact scale variations and spectral energy shifts, thereby leading to the deterioration of generalization performance under unknown scenarios, an adaptive domain generalization framework named TFAR-UDP is proposed. Firstly, the time-frequency adaptive receptive field network (TFAR-Net) is designed, integrating time-frequency decoupled isotropic convolutions, adaptive multi-scale hollow convolutions and band-specific attention mechanisms to achieve joint modeling of fault impulse evolution, resonance band responses and multi-scale time-frequency textures. Secondly, the proposed unified dual-path (UDP) perturbation strategy, which utilizes exact feature distribution matching spatial-domain ordering statistical perturbations and Fourier frequency-domain amplitude spectral perturbations to expand the source domain sample space from both spatial statistical distribution and spectral distribution perspectives, adopts a homoscedastic uncertainty weighting mechanism to adaptively coordinate the dual-path optimization process. To validate the effectiveness of the proposed framework, a system-level fault diagnosis task comprising key components such as bearings, gears, rotors and motors is constructed, and cross-speed and cross-load generalization experiments are conducted on the self-built XUST-IDC dataset and the publicly available BJTU-RAO dataset. The results show that TFAR-UDP achieves average accuracy rates of 92.11 and 81.08% in cross-speed and cross-load experiments on the XUST-IDC dataset, respectively; on the BJTU-RAO dataset, 84.83 and 93.83%, respectively, outperforming mainstream domain generalization methods. The results validate the effectiveness and robustness of the proposed framework in fault diagnosis across operating conditions for transmission systems.