FS-PTL: A unified few-shot partial transfer learning framework for partial cross-domain fault diagnosis under limited data scenarios

Abstract

Traditional supervised learning-based fault-diagnosis models often encounter performance degradation when data distribution shifts occur. Although unsupervised transfer learning can address such issues, most existing methods face challenges arising from partial cross-domain diagnostic scenarios with limited training data. Therefore, this study introduces a unified few-shot partial-transfer learning framework, specifically designed to address the limitations of data scarcity and partial cross-domain diagnosis applicability. Our framework innovatively takes ridge regression-based feature reconstruction as a nexus to integrate episodic learning with an episodic pretext task and weighted feature alignment, thereby enhancing model adaptability across varying working conditions with minimal data. Specifically, the episodic pretext task enables the learned features with generalization abilities in a self-supervised manner to mitigate meta-overfitting. Weighted feature alignment is performed at the reconstructed feature level, allowing partial transfer with a significantly increased number of features, while further reducing overfitting. Experiments conducted on two distinct datasets revealed that the proposed method outperforms existing state-of-the-art approaches, demonstrating superior transfer performance and robustness under the conditions of limited fault samples.