Aiming at the problems of low accuracy of bearing fault diagnosis due to different sample distribution and few fault samples under varying working conditions, and limited feature extraction and weak generalization ability of some small sample algorithms, a bearing fault diagnosis method by squeeze-excitation-resnet (SE-Resnet) and meta-transfer learning (MTL) under few-shot and variable working conditions is proposed. Firstly, the one-dimensional bearing vibration signal under different working conditions is converted into time-frequency image by continuous wavelet transform (CWT), thus transforming the bearing fault diagnosis into image recognition. Secondly, the squeeze-excitation attention mechanism is introduced to build a backbone network model of SE-Resnet to focus on more effective feature channels and enhance feature extraction capabilities. Then, taking advantage of the advantages of transfer learning that can provide good initial parameters of the deep network and rapid learning of meta-learning, pre-training and meta-transfer training are carried out in turn to obtain a high-precision meta-transfer network that can be achieved by fine-tuning with a small number of samples, thereby realizing bearing fault diagnosis under variable working conditions. Finally, the proposed method is verified by two benchmark datasets and a bearing fault simulation test bench built in our laboratory, and compared with other methods. The results show that the proposed method has higher recognition accuracy and generalization performance for bearing fault diagnosis under few-shot and variable working conditions.