Traditional bearing fault diagnosis methods often suffer from low accuracy and weak model generalization under varying working conditions due to diverse sample distributions, scarcity of fault samples, and limited feature extraction capabilities of some few-shot learning algorithms. To address these challenges, this paper proposes a novel method for variable condition bearing fault diagnosis that combines a squeeze-and-excitation residual network (SE-ResNet) with meta-transfer learning (MTL). One-dimensional bearing vibration signals collected under different working conditions are converted into time-frequency images using continuous wavelet transform (CWT), thereby transforming the bearing fault diagnosis task into an image recognition problem. A squeeze-and-excitation (SE) attention mechanism is introduced to construct an SE-ResNet backbone network model. This focuses on more effective feature channels, thereby enhancing feature extraction and representation capabilities. Leveraging the advantages of transfer learning (which provides robust initial deep network parameters) and meta-learning (which enables rapid adaptation), the model undergoes sequential pre-training and meta-transfer training. This process yields a high-precision meta-transfer network that can be fine-tuned with only a small number of samples, ultimately achieving accurate bearing fault diagnosis under variable working conditions. The proposed method is validated using two benchmark datasets and a bearing fault simulation test bench developed in the laboratory. Comparative analysis with other methods demonstrates that the proposed method exhibits higher recognition accuracy and superior generalization performance for bearing fault diagnosis under both few-shot and variable working conditions.