Health prognosis of bearings based on transferable autoregressive recurrent adaptation with few-shot learning

Data-driven prognostic and health management technologies are instrumental in accurately monitoring the health of mechanical systems. However, the availability of few-shot source data under varying operating conditions limits their ability to predict health. Also, the global feature extraction process is susceptible to temporal semantic loss, resulting in reduced generalization of extracted degradation features. To address these challenges, a transferable autoregressive recurrent adaptation method is proposed for bearing health prognosis. In the enhancement of few-shot data, a novel sample generation module with attribute-assisted learning, combined with adversarial generation, is introduced to mine data that better matches the source sample distribution. Additionally, a deep autoregressive recurrent model is designed, incorporating a statistical mode to consider the degradation processes more comprehensively. To complement the semantic loss, a semantic attention module is developed, embedded into the basic model of meta learning. To validate the effectiveness of this approach, extensive bearing prognostics are conducted across six tasks. The results demonstrate the clear advantages of this proposed method in bearing prognosis, especially when dealing with limited bearing data.