Macroscopic–microscopic information fusion: a novel reliable bearing diagnosis method with physical embedding and uncertainty quantification

Lots of multimodal fusion models have been extensively applied in bearing fault diagnosis. However, they overlook the fusion uncertainty among different modalities and the fine-grained selection of fault information within a single modality. To this end, a novel reliable multimodal diagnosis framework named macroscopic–microscopic information fusion (MMIF) is developed. In MMIF, the microscopic physical embedding selection mechanism is first proposed to estimate the weight corresponding to different components of single modality, thereby enhancing the representation of intrinsic fault features. Subsequently, a new macroscopic uncertainty quantification scheme with D-S evidence theory is constructed to adaptively solve the uncertainty probability of each whole modality, achieving the fusion of multi-source monitoring signals. Besides, the joint optimization strategy with local weight regularization and global evidence calibration is designed to guarantee the effectiveness of the MMIF framework. Finally, the experimental results demonstrate on both vibration and current modalities that the proposed MMIF outperforms several state-of-the-art diagnosis methods. Source code is available at https://github.com/PHM-QQteam/MMIF .