A meta multi-scale spatial attention network for interpretable few-shot remaining useful life prediction

Abstract Remaining Useful Life (RUL) prediction faces the challenges of insufficient representation and a lack of interpretability, especially under few-shot and complex operating conditions. To address these issues, this paper proposes a Meta Multi-Scale Spatial Attention Network (MeMSAN). Based on Integrated Gradients (IG) attribution theory, a meta-learning collaborative optimization objective is constructed, integrating RUL prediction loss with attribution consistency, which promotes collaborative convergence and guides the model to learn fast cross-task adaptation capabilities while simultaneously quantifying the marginal contribution of each feature. Additionally, a Multi-Scale Spatial Attention (MSA) mechanism is designed to jointly model local degradation patterns and global dependencies within the feature domain, thereby extracting highly transferable degradation representations. Experimental results on aero-engine and bearing datasets demonstrate that the proposed MeMSAN outperforms several benchmark models. Furthermore, it provides meaningful feature contribution explanations, validating its effectiveness and application potential in few-shot scenarios.