U-DS-Net: An uncertainty-aware multimodal interpretable network for axlebox bearing fault diagnosis in metro vehicles

Axlebox bearing fault signals from metro vehicles are characterized by weak fault features under severe noise, which limits the diagnostic accuracy of conventional unimodal methods and leaves their decisions insufficiently interpretable. To address these issues, this study proposes U-DS-Net, an uncertainty-aware evidence reasoning network for multimodal signals. First, triaxial vibration and acoustic signals are used as multimodal inputs. Multiscale convolutions extract fault features under different receptive fields, while squeeze-and-excitation (SE) channel attention suppresses redundant noise. Second, vibration, acoustic, and joint vibroacoustic representations are constructed as multiple evidence sources to capture complementary axlebox bearing fault information. Finally, a two-stage evidence fusion strategy is developed on the basis of uncertainty-aware reasoning to achieve reliable multimodal diagnostic fusion. U-DS-Net is evaluated on the BJTU dataset and the self-built MABFD dataset. The results show that the proposed method outperforms several state-of-the-art approaches and provides interpretable diagnosis by quantifying modal contributions, evidence conflicts, and predictive uncertainty. The proposed framework offers a reliable approach to axlebox bearing fault detection and the safe operation of metro vehicles.