A novel multi-source sensor correlation adaptive fusion framework with uncertainty quantification for intelligent fault diagnosis

Intelligent fault diagnosis using multi-source sensor fusion holds significant promise but faces challenges related to reliability due to variations in signal quality across sensors and inconsistencies in fault features. To tackle these issues, a multi-source sensor correlation adaptive fusion (MSCAF) framework with uncertainty quantification is proposed to enhance identification trustworthiness for intelligent fault diagnosis. The proposed MSCAF integrates Dempster–Shafer theory with Dirichlet distribution to model sensor uncertainty and split multi-source sensors into high-confidence and low-confidence sensors based on the consistency of cross-sensor fault information. High-confidence sensors are given greater weight, ensuring more reliable fusion. Then, the reward and penalty functions are introduced to assess their correlation weights. Meanwhile, Convolutional and graph neural networks are employed to enhance feature extraction and output category probabilities, which can ensure robust fusion across varying diagnostic scenarios. This approach allows adaptive weighting, optimizes fusion reliability, and enables manual intervention for low-confidence sensors. Experimental results demonstrate that the proposed MSCAF achieves superior diagnostic performance compared to state-of-the-art methods, confirming its efficacy in extracting reliable features with uncertainty quantification for intelligent fault diagnosis.