Zhou, Hanting and Chen, Wenhe and Luo, Xinggang and Xia, Min (2026) TDAFDNet : A trustworthy data-augmented network for few shot fault diagnosis and noise interference. Applied Soft Computing, 189: 114545. ISSN 1568-4946
Full text not available from this repository.Abstract
The practical deployment of deep neural networks for intelligent fault diagnosis in industrial systems is often hindered by the scarcity of labeled fault data and distributional discrepancies between training and operational environments, especially under noisy conditions. To address these issues, this study proposes a trustworthy data-augmented fault diagnosis network (TDAFDNet) specifically designed for few-shot and noise-robust fault diagnosis. The framework integrates a multi-attention conditional variational autoencoder (MACVAE) to augment limited fault data, thereby enriching feature diversity and enhancing model generalization. A lightweight multi-channel depthwise separable residual network (MDS-ResNet) is introduced to achieve efficient and robust feature extraction. Furthermore, an adaptive attention-based risk-aware evidential loss function enables uncertainty-aware and risk-sensitive classification. To further improve interpretability and support risk-informed decision-making, a control chart-inspired uncertainty-based early warning mechanism is incorporated. Extensive experiments on benchmark rolling bearing datasets demonstrate that TDAFDNet outperforms existing state-of-the-art methods in diagnostic accuracy, noise robustness, and uncertainty estimation under data-scarce and noisy scenarios. These findings underscore the potential of the proposed approach for reliable and interpretable fault diagnosis in safety-critical engineering applications with limited and noisy labeled data.