Current meta-learning-based cross-domain few-shot fault diagnosis methods usually rely heavily on support samples for task-specific adaptation. As a result, they tend to overfit the source domain and are easily affected by errors caused by limited samples, which significantly weakens their performance in extreme few-shot settings. In addition, conventional Euclidean space modeling in meta-learning ignores the intrinsic geometric structure of data. To overcome these limitations, this paper proposes a cognition-aware meta-learning (CAML) framework for few-shot fault diagnosis across different machines and operating conditions. In CAML, a cognitive prototype construction method is designed to guide inner-loop adaptation more accurately while retaining generic meta-knowledge from the global task distribution. A cognition adaptive factor is further introduced into the loss to dynamically adjust the fitting intensity of the support set. Moreover, by using a projection distance-driven strategy, CAML achieves both geometry-aware feature embeddings and enhanced class discrimination. The proposed approach is validated on cross-bearing few-shot fault diagnosis tasks, and experimental comparisons demonstrate that it outperforms state-of-the-art few-shot diagnosis methods.

Meta learning,few-shot learning,cross-domain fault diagnosis,prototype,bearing