Adaptive semi-supervised meta-learning with pseudo-label for multiple working condition fault diagnosis

To address diagnostic performance degradation in industrial systems arising from scarce fault samples, limited labeled data, and significant distribution shifts across operating conditions, we propose an adaptive semi-supervised meta-learning method with pseudo-label selection (ASML). First, data-driven variable relationships are integrated with expert prior knowledge to construct an entity–relationship knowledge graph, and a relational graph convolutional network is employed to learn structured feature representations with physical and semantic constraints. Second, a center-removal-based embedding regularization mechanism is incorporated to regularize the episode-wise metric geometry, thereby stabilizing prototype estimation and improving distance-based classification under cross-condition distribution shifts. Third, pseudo-label reliability is evaluated using instance-wise margins smoothed by an exponential moving average, and a task-wise dynamic threshold is employed to filter low-confidence predictions, reducing noisy pseudo-labels and error accumulation under cross-condition shifts. Experimental results on two industrial datasets demonstrate that the proposed ASML method significantly outperforms existing semi-supervised and meta-learning approaches in diagnostic accuracy and generalization under small-sample and multi-condition scenarios.