Multiple source-free domain adaptation via efficient-DDPM and adaptive fine-grained weight integration for imbalanced fault diagnosis
Abstract
Reliable operation of rotating machinery is pivotal to industrial production safety and efficiency. Constrained by data privacy concerns, source-free domain adaptation has emerged as a research focus in fault diagnosis, as it avoids the need to access source domain data. However, target domain data collected from real-world industrial scenarios often suffer from class imbalance and distribution shift, which severely degrade diagnostic performance. To address these challenges, this article proposes a novel multiple source-free domain adaptation method based on efficient denoising diffusion probabilistic model (Efficient-DDPM) and adaptive fine-grained weight integration for imbalanced fault diagnosis. Specifically, the Efficient-DDPM utilizes extremely few labeled target-domain fault samples to generate high-quality synthetic samples for minority fault classes. It incorporates an improved U-Net architecture and optimized time steps to enhance both sample quality and computational efficiency. Meanwhile, multiple pretrained source models are introduced to provide richer prior knowledge, and a dual-attention enhancement module is developed to learn intra-domain and inter-domain integration weights, achieving an optimal balance between sample specificity and domain consistency. In addition, a loss reweighting strategy guided by uncertainty quantification is constructed to mitigate the impact of low-quality pseudo-labels. Experiments on two bearing datasets verify that the proposed method achieves superior cross-domain diagnostic performance under constraints of data privacy and class imbalance, confirming its effectiveness and practicality.