Improved swin transformer integrated with GADF method for bearing fault detection

Abstract Bearing fault detection is critical for ensuring the reliability and performance of modern industrial equipment. However, conventional methods often suffer from incomplete feature extraction for nonlinear vibration signals, with accuracy below 85%, while existing deep learning models face high computational complexity (e.g., the original Swin Transformer has up to 24 million parameters). To address these challenges, this study proposes a novel approach integrating Gramian Angular Difference Field (GADF), an improved Swin Transformer, and a Convolutional Neural Network with Global Attention Module (CNN-GAM), achieving enhanced feature representation, computational efficiency, and attention optimization. The proposed method introduces three key innovations: First, GADF converts 1D vibration signals into 2D time-frequency images, preserving 98% of time-domain features while enhancing feature representation. Second, the improved Swin Transformer employs a shifted window mechanism to reduce computational load by 42% while efficiently extracting multi-scale and multi-level features. Third, the CNN-GAM module combines local feature extraction via convolutional layers with a dual-channel attention mechanism, increasing the weight of critical features by 37%. Together, these innovations enable superior performance in industrial applications. Validated on the Case Western Reserve University (CWRU) bearing dataset, the GADF-Swin-CGAM model achieves 99% accuracy with only 25 training iterations—three times faster convergence than traditional methods—and a training loss as low as 0.01. Under complex operating conditions (1730 rpm), it attains 100% recognition rate for minor faults (0.53 mm size) with minimal misclassification. The model demonstrates exceptional robustness and stability, maintaining high precision even in noisy and variable industrial environments, while significantly reducing computational requirements compared to existing approaches. This work provides a high-precision, reliable solution for bearing fault diagnosis, with significant theoretical and practical implications. The integration of GADF for enhanced signal representation, an optimized Swin Transformer for efficient feature extraction, and CNN-GAM for attention-driven feature refinement offers a balanced approach that outperforms state-of-the-art methods in both accuracy and computational efficiency, making it particularly suitable for real-world industrial applications.