Uncertainty-aware few-shot gearbox fault diagnosis via multi-scale spatial-attention and transformer integrated evidential deep learning

Abstract To address the challenges of data scarcity and unreliable prediction confidence in wind turbine gearbox fault diagnosis, an uncertainty-aware diagnostic and intelligent early-warning framework is proposed, integrating dynamic data augmentation with an improved evidential deep learning (EDL) framework. The proposed framework adaptively combines CutMix and Mixup strategies to effectively enrich training samples and alleviate overfitting in extremely few-shot scenarios. A multi-branch, multi-scale convolutional network with coordinate attention and a Transformer Encoder is designed to jointly capture robust local and global representations. Based on EDL, the model achieves both high diagnostic accuracy and reliable uncertainty quantification. Furthermore, the quantified uncertainty is visualized through a three-level control-chart-based early warning mechanism, enabling proactive and hierarchical fault alerts. Experimental results show that the proposed method maintains high accuracy, provides trustworthy uncertainty estimation, and supports dynamic early-warning support in few-shot conditions of wind turbine gearbox, showing strong potential for practical industrial deployment.