| dc.description.abstract | Abstract. Few-shot fault diagnosis presents significant challenges due to sample scarcity and the difficulty in extracting discriminative features. To address these issues, this article proposes a dual-branch integrated network (DBIN), a novel computational framework for few-shot learning. The framework introduces two key innovations: First, in the realm of metric learning, an adaptive memory-enhanced Mahalanobis prototypical network (AMEMPN) is designed. It leverages Mahalanobis distance to model feature covariance and employs an adaptive memory mechanism, significantly enhancing prototype accuracy and robustness. Second, a multigranularity fusion mechanism is established, which synergistically integrates the global class representations from the AMEMPN branch with the local topological structures captured by a relation-aware radial basis function (RBF) kernel graph neural network (RA-RKGNN) branch. This dual-branch architecture allows for a more comprehensive feature representation. Experimental validation on multiple benchmark datasets demonstrates that DBIN achieves superior diagnostic performance, exhibiting remarkable generalization capability under extremely limited sample conditions. This work provides novel insights for few-shot learning, offering significant theoretical and practical value for intelligent fault diagnosis. | |
