A meta-transfer learning prediction method with few-shot data for the remaining useful life of rolling bearing

Rolling bearings are essential components of rotating machinery. It is crucial to predict and manage the health of rolling bearings. This article proposes a meta transfer learning-based remaining useful life (RUL) prediction approach with few-shot data for rolling bearing. First, multiple subtasks under variable operating conditions are constructed. A subtask and cross-subtask-based gradient optimization model is employed to extract degradation knowledge adaptively. The batch feature norm differences method is presented to reduce the impact of negative transfer and poor transfer performance. Interdomain transferable features are obtained by minimizing the difference in the number of feature paradigms between the source and target domains. Therefore, the Meta-SGD transfer learning approach realizes the RUL prediction under few-shot data and variable operating conditions. Two cases validate the effectiveness of the presented method.