.png){kind=icon}
.png){kind=icon}
Return
F-DiffWave: a fault-aware wavelet-structured diffusion framework for robust fault diagnosis under variable operating conditions
.png){kind=icon}
PRE
AI
DOI：10.1088/1361-6501/ae2cb3.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Multi-source domain generalization (MDG) is a promising paradigm for fault diagnosis of rotating machinery under variable operating conditions. However, existing MDG methods are often limited by data imbalance and insufficient feature discriminability. Generative models can rebalance data, but generative adversarial networks (GANs) risk mode collapse and diffusion models are computationally expensive and slow to sample. To overcome these challenges, we propose F-DiffWave, a fault-aware wavelet-structured diffusion framework that integrates generative modeling and discriminative learning for robust domain generalization. F-DiffWave leverages discrete wavelet transform to decompose time-frequency representations into multi-scale sub-bands, reducing generative complexity and enhancing frequency-aware feature learning. A large-step diffusion GAN is employed to accelerate high-quality sample synthesis, guided by reconstruction and perceptual losses, while an enhanced U-Net architecture preserves critical frequency components. In parallel, a self-supervised contrastive loss increases intra-class compactness and inter-class separability across domains. Experimental results on two gearbox datasets validate that F-DiffWave achieves superior diagnostic accuracy and generalization performance compared to state-of-the-art MDG approaches.
Keywords:
multi-source domain generalization
fault diagnosis
diffusion models
wavelet transform
generative adversarial networks
Journal
IF:
3.4
Papers:
2.6K
Citations:
2.3W