Entropy-Oriented Semi-Supervised Dynamic Prototype Contrastive Learning for Rotating Machinery Fault Diagnosis

Abstract

Rotating machinery, a core component of mechatronic systems, plays a vital role in ensuring operational reliability and efficiency in aerospace, robotics, and manufacturing. However, challenges like high cost of data annotation and imbalanced data distributions hinder accurate fault diagnosis. Thus, this article proposed an entropy-oriented semisupervised dynamic prototype contrastive learning (ESDPCL) for rotating machinery fault diagnosis with limited labeled samples. First, a dynamic pseudolabeling selection policy (DPSP) is designed to filter high-confidence samples from imbalanced labeled data, thereby reducing the impact of low-confidence samples on model accuracy. Then, a sample self-attention mechanism calculates the correlation between high-confidence unlabeled samples and the initial prototypes, dynamically adjusting the prototype positions to enhance the network’s generalization under limited labeled data conditions. Finally, an entropy-oriented normalized time-frequency prototype contrastive loss is developed to optimize the feature distribution across both the time and frequency domains. This loss function also incorporates a normalization strategy to amplify the contribution of minority samples while utilizes entropy-guided Euclidean distance to suppress high-uncertainty features, addressing limitations in traditional contrastive loss designs. Experimental validation on the public and the self-made datasets demonstrates the accuracy of ESDPCL is 97.75% and 97.30%, which outperforms other methods with limited and imbalanced data.