To tackle the challenges of sample scarcity and performance degradation inherent in traditional deep learning–based bearing fault diagnosis, this study proposes a novel fault diagnosis framework grounded in a multi-physical-field-coupled digital twin. First, a high-fidelity digital twin model is developed by comprehensively incorporating the bearing’s geometric configuration and thermo-solid coupling effects, enabling a more realistic representation of its operational behavior. The model is then employed to synthesize twin datasets corresponding to diverse fault modes, thereby mitigating the scarcity of real experimental samples. Second, to overcome the limitations of conventional deep learning architectures in multi-scale feature representation, a multi-scale convolutional neural network (MSCNN) is devised and jointly trained with a limited set of real samples to enhance fault recognition accuracy under small-sample conditions. Finally, extensive experiments are conducted on a bearing test bench platform to validate the proposed approach. Experimental results demonstrate that the proposed method substantially enhances diagnostic accuracy across diverse operating conditions, exhibits superior generalization and stability, and offers a promising solution for intelligent fault diagnosis in small-sample scenarios.