Unsupervised bearing fault diagnosis and noise-robust data cleaning using a Swin transformer-based dual-module framework

Abstract

Bearings are essential components in high-speed train bogies, playing a crucial role in ensuring safe and reliable operation. In recent years, data-driven intelligent fault diagnosis methods have demonstrated impressive performance. However, most existing approaches heavily depend on large volumes of high-quality labeled data, which are difficult to obtain in real-world industrial environments. To address these challenges, this paper presents a novel Swin Transformer-based parallel framework for unsupervised bearing fault diagnosis and intelligent data cleaning. The proposed method integrates hierarchical feature extraction with a multi-task optimization strategy. By leveraging shifted window attention, the model performs joint vibration signal denoising and fault detection, while a dynamic weighting mechanism adaptively balances reconstruction and noise classification losses. Experimental evaluations on the Paderborn University and shaft crack datasets—under varying noise levels and load conditions—demonstrate the framework’s effectiveness. The model achieves a diagnostic accuracy of 98.6%, outperforming conventional autoencoder-based methods by 9.2%, and maintains over 98% robustness to noise (σ = 0.8), all without reliance on labeled data. The results suggest that dual-module architecture significantly enhances unsupervised fault diagnosis accuracy, and the diagnostic performance is highly sensitive to load and noise intensity, showing a nonlinear relationship.