Partial domain adaptation ProtoNet: A partial transfer meta-learning framework for cross-domain fault diagnosis with limited labeled data

Abstract

Cross-domain fault diagnosis is essential for rotating machinery operating under variable working conditions, where the distribution of target-domain data often differs from that of the training data. However, most existing domain adaptation methods rely on sufficient training samples and assume that the source and target domains share the same label space. These assumptions are difficult to satisfy in practical industrial scenarios, where only limited labeled data are available, and the target domain often contains only a subset of source-domain fault categories. To address this limited-data partial domain adaptation (PDA) problem, this study proposes a PDA ProtoNet framework for cross-domain fault diagnosis. PDAPN first employs episodic prototypical learning to construct stable and discriminative class representations from scarce labeled source-domain samples. Then, a maximum of cosine similarity-based local partial alignment mechanism is introduced to identify target-compatible source-domain anchors and reduce negative transfer from source-private classes. In addition, a greedy fine-tuning strategy is designed to progressively incorporate high-confidence pseudo-labeled target samples, thereby refining target-domain decision boundaries and improving cross-domain generalization. Experiments are conducted on a machinery comprehensive diagnostics simulator gearbox dataset and a Harbin Institute of Technology bearing dataset under limited-data partial label-space mismatch scenarios. The results show that PDAPN outperforms representative existing methods, including source-only learning, closed-set domain adaptation, and recent PDA fault diagnosis approaches, demonstrating its effectiveness and robustness for limited-data partial cross-domain fault diagnosis.