Spatio-temporal graph convolutional network with domain generalization: A novel rotating machinery RUL prediction method in small samples

Deep learning-based methods for remaining useful life (RUL) prediction have gained significant attention in recent years. However, existing methods require large amounts of data for training, whereas acquiring full-life cycle data for machinery is both time-consuming and costly. To overcome this challenge, a novel spatio-temporal graph convolutional network with domain generalization (STGCN-DG) method is proposed in this paper. First, different degradation behavior modes of rotating machinery are identified and significant spatio-temporal dependence characteristics are extracted to mitigate the effect of insufficient sample size. To make full use of the data under different working conditions, this paper then develops a new domain generalization method to simultaneously minimize the marginal and conditional probability distributions of different domains, which guides the model to learn universal domain-generalized information. The proposed STGCN-DG method reflects the complex degradation characteristics of the machinery more comprehensively, and achieves accurate prediction of RUL in small sample scenarios. Extensive experiments are conducted to thoroughly evaluate the effectiveness and superiority of the proposed method, demonstrating its robust performance in small sample scenarios.