Research on Semisupervised Bearing Fault Diagnosis Combining Swin Transformer With Adaptive Pseudo-Labeling

Abstract Bearings are critical components of high-speed train bogies, acting as interfaces between rotating wheelsets and the bogie frame. This paper presents a semisupervised learning (SSL) approach for bearing fault diagnosis that addresses the challenge of limited labeled data in industrial settings. The method integrates Swin Transformer's feature extraction with an optimized pseudo-labeling strategy, achieving state-of-the-art performance while reducing dependence on labeled samples. A two-stage training process first builds baseline discriminative ability with minimal labeled data, then incorporates high-confidence pseudo-labels from unlabeled data through adaptive confidence thresholding based on prediction entropy. Experiments show 99.5% accuracy on the Paderborn dataset using only 30% labeled data—a 9% improvement over conventional methods. The system excels at identifying complex compound faults and remains robust to noise and data variability. This work also offers theoretical insights into data efficiency in semisupervised learning, defining optimal thresholds for labeled and unlabeled data use. The approach provides a cost-effective, scalable solution for industrial fault diagnosis while setting new benchmarks for semisupervised learning in mechanical fault detection.