Bearing fault diagnosis based on LBWO-VMD and Swin-CNN-GAM dual-branch network

Abstract To address the issues of insufficient feature extraction and reliance on empirical parameter selection in traditional bearing fault diagnosis methods, this paper proposes an intelligent diagnosis method based on parameter adaptive decomposition and spatiotemporal feature fusion. This method integrates Lévy-flight Enhanced Beluga Whale Optimization (LBWO), Gramian Angular Difference Field (GADF), and a Swin-CNN-GAM dual-branch network. First, the Lévy flight strategy is introduced to improve the Beluga Whale Optimization (BWO) algorithm, achieving adaptive optimization of the key parameters for Variational Mode Decomposition (VMD). Effective Intrinsic Mode Functions (IMFs) are then selected and reconstructed based on a weighted comprehensive index. Subsequently, the reconstructed one-dimensional time-series signals are encoded into two-dimensional feature images using GADF to preserve temporal dependencies and enrich feature dimensions. Finally, a parallel dual-branch fusion architecture comprising Swin Transformer and CNN-GAM is constructed to capture the global long-range dependencies and local texture details of the images, respectively. Fault classification is achieved after fusing these features via channel concatenation. Experimental results on the CWRU and SEU datasets demonstrate that the recognition accuracies reach 99.11% and 99.19% respectively, exhibiting excellent feature extraction capability and generalization performance.