Application of prototype contrastive learning with dynamic pseudo-labeling and attention mechanism for aircraft fault diagnosis

Abstract

To address the underutilization of unlabeled samples and the interference introduced by pseudo labels under imbalanced data distributions, a prototype contrastive learning framework integrating dynamic pseudo-labeling and an attention mechanism is proposed. The method aims to improve the effective use of unlabeled data while mitigating the adverse effects of pseudo-label noise. First, a confidence-based dynamic pseudo-label filtering strategy is introduced, in which the predictive uncertainty of the model on unlabeled samples is quantified to select high-confidence samples, thereby reducing the negative influence of low-confidence samples during training. Subsequently, a sample-level self-attention mechanism is designed to update class prototypes through similarity-based weighting among samples, enabling the prototypes to adaptively capture the distribution of unlabeled data. Finally, a time–frequency prototype contrastive loss is incorporated together with an entropy-guided metric strategy to balance the contribution of minority classes and suppress the impact of highly uncertain samples on model updates. Experiments conducted on aircraft engine casing bearing datasets demonstrate that, under semi-supervised conditions with limited labeled data and abundant unlabeled samples, the proposed method significantly improves the accuracy and robustness of minority classes. The results further highlight the critical role of the attention mechanism in similarity-based weighting of unlabeled features during prototype updating.