Hyper-opinion dual branch evidential deep learning with distance perception method for open set fault diagnosis

Accurate fault diagnosis is critical to ensuring the reliable operation of equipment. However, traditional deep learning methods typically rely on a predefined set of known fault types, limiting their ability to identify unknown faults. Thus, this paper proposes an open set fault diagnosis method based on Hyper-opinion Dual Branch Evidential Deep Learning with Distance Perception (HDEDL-DP). Firstly, a dual-branch framework is proposed to capture sharp evidence supporting single proposition and vague evidence supporting composite propositions via hyper-opinion modeling. A projection mechanism converts the vague evidence into sharp evidence for model inference. Secondly, mahalanobis distance is used to provide distance perception. Finally, an OOD score integrating uncertainty estimation and distance perception is designed to identify unknown fault classes. Two case studies validate the effectiveness of HDEDL-DP.