<p>In real industrial environments, the operating conditions of bearings are complex and changeable, and only a few training samples are available. This limits the ability of existing fault diagnosis methods to extract discriminative features and reduces their generalization. To address these challenges, this paper proposes an intelligent fault diagnosis framework called MERSN-ISwinT, which is based on a multi-scale efficient residual shrinkage network (MERSN) and an improved swin transformer (ISwinT). First, the original vibration signals are converted into time-frequency maps using continuous wavelet transform to enrich fault-related information. Then, MERSN, with its multi-scale dense residual structure, is used to extract deeper critical features. Furthermore, ISwinT is further designed to capture global dependencies. It applies a sliding window mechanism to fuse cross-window features step by step and uses a hierarchical structure to gradually expand the receptive field. By combining MERSN and ISwinT, the MERSN-ISwinT integrates local and global features to obtain more discriminative and high-quality fault representations. Experiments on two bearing datasets show that MERSN-ISwinT has both stability and strong generalization. The method achieves more than 99 % accuracy under multiple cross-working conditions. Even with strong noise interference, the accuracy remains at 95.69 %. With only a small number of samples, the accuracy still reaches 98.12 % and 91.25 % on the two datasets. These results outperform state-of-the-art methods and show the potential of the proposed method for industrial applications.</p>