Dual-modal feature pyramid fusion and distribution alignment for cross-domain few-shot fault diagnosis
TL;DR
A dual-modal feature pyramid fusion and distribution alignment framework that introduces KL-divergence loss to supervise feature alignment, thereby stabilizing fine-tuning and suppressing overfitting and a training protocol combining meta-learning pre-training and targeted fine-tuning is developed to address extreme data scarcity.
Abstract
Meta-learning-based few-shot cross-domain fault diagnosis faces significant challenges in cross-speed scenarios due to extreme data scarcity and speed-induced distribution shifts. To address these issues, this paper proposes a dual-modal feature pyramid fusion and distribution alignment (DA) framework. First, a hybrid feature extraction architecture integrating time–frequency images and temporal signals is designed. By combining Swin Transformer, multi-scale ResNet, and GRU, a pyramidal tri-branch network extracts complementary features from time and frequency domains, significantly enhancing the characterization of complex fault patterns. Building on this, pseudo-feature prototypes are constructed via a dual-clustering strategy using Isolation Forest and K-means. It constrains distribution shifts and thereby improves cross-domain generalization. Finally, a training protocol combining meta-learning pre-training and targeted fine-tuning is developed to address extreme data scarcity. It introduces KL-divergence loss to supervise feature alignment, thereby stabilizing fine-tuning and suppressing overfitting. Single-shot cross-domain experiments on CWRU, JNU, and PU datasets achieve diagnostic accuracies of 99.32%, 95.52%, and 93.62%, respectively, demonstrating superior performance over state-of-the-art meta-learning methods.