Improved Swin Transformer Integrated with GADF Method for Bearing Fault Detection

Detecting bearing faults is of crucial importance for ensuring the reliability and performance of modern industrial equipment. In this research, a method integrating the Gramian Angular Difference Field (GADF), an improved Swin Transformer, and a Convolutional Neural Network with Global Attention Module (CNN-GAM) is proposed. The proposed approach utilizes GADF to convert one-dimensional vibration signals into two-dimensional images, which not only preserves the time-varying characteristics of the signals but also enhances the feature representation. The improved Swin Transformer efficiently extracts multi-scale and multi-level features, reducing computational complexity and broadening the receptive field of the model. The CNN-GAM module then optimizes the feature representation, with the convolutional neural network capturing local features and the global attention module assigning precise attention weights to accentuate crucial information, thereby significantly enhancing the accuracy of the model. This method has been validated using the bearing dataset from Case Western Reserve University (CWRU). After comprehensive data cleaning and processing, the model exhibited remarkable performance. In the first round of training, the accuracy reached 88%, and within 25 iterations, the validation accuracy was stably maximized at 99%, while the training loss was as low as 0.01. Compared with existing models, this approach demonstrates greater robustness, requires fewer training rounds, and maintains excellent stability. Even under simulated complex working conditions, it can accurately classify a wider range of fault types with minimal misclassifications. It achieves high precision and high reliability, and thus holds significant theoretical value as well as promising potential for industrial applications.