Current methods for bearing fault diagnosis under complex conditions face limitations in capturing temporal signal correlations, multi-dimensional features, and adaptive feature extraction. This research introduced a sophisticated method that integrates the Gramian Angular Field-Markov Transition Field (GAF-MTF) encoding method with an enhanced Squeeze-and-Excitation ResNet 50 (SE-ResNet 50) model to effectively address the issue at hand. The GAF-MTF method fuses static (GASF), dynamic (GADF), and probabilistic transition (MTF) features to convert 1D timing signals to 2D images, preserving temporal correlations and enhancing fault representation. These images are processed by SE-ResNet 50, which employs channel attention mechanisms to dynamically prioritize critical features and enhance stability. Experiments on the CWRU dataset achieved 99.87% accuracy, with validation on the Jiangnan University dataset yielding 99.05%, demonstrating great generalization ability. Additionally, we utilized t-SNE to reduce feature dimensions and analyzed the role of every residual layer. The framework provides reliable fault diagnosis under variable conditions, with future work targeting computational efficiency and lightweight architectures for broader industrial deployment.
