A hyper-conditional cross-attention meta-diffusion method for few-shot remaining useful life prediction of rolling bearings

Abstract Bearings are essential components of rotating machinery, and the effective maintenance and health management are crucial to ensuring system reliability. In the context of bearing remaining useful life (RUL) prediction with few-shot samples and varying operating conditions, meta-learning methods often suffer from gradient vanishing and degraded prediction accuracy during training. Motivated by these challenges, this paper proposes a hyper-conditional cross-attention meta-diffusion method for RUL prediction of rolling bearings across domains with limited data. Firstly, a base-class dataset and a novel-class dataset are constructed. The former is used for the training of ideal weights, while the latter is employed for task adaptation training and weight inference in the diffusion model. Then, a task-adaptive meta-diffusion framework is established by introducing the diffusion mechanism into meta-learning, where forward noise injection and reverse denoising are employed to achieve adaptive model updating. In addition, a weight-space dispersive loss is incorporated during the denoising stage to enhance the model’s adaptability and generalization when facing diverse operating conditions. Finally, two case studies are implemented to illustrate the effectiveness and practicality of the proposed method.