Abstract: Bearing fault diagnosis plays a crucial role in preventing equipment failures, enhancing production efficiency, and reducing maintenance costs. Conventional diagnostic approaches are often dependent on extensive labeled datasets, a requirement that is frequently unmet in practice due to data scarcity. While unsupervised learning techniques alleviate the need for labeled data, they typically yield suboptimal diagnostic performance. To overcome these limitations, this paper introduces a semi-supervised few-shot bearing fault diagnosis method that integrates an improved ReMixMatch framework with physical feature fusion. The proposed approach begins by processing raw vibration signals via the Hilbert transform to derive their envelope. A set of physical features is subsequently extracted from these envelope signals. The most discriminative features are then selected based on an analysis of their correlation and strength. The network architecture is further enhanced by incorporating a multi-scale residual module and a channel attention mechanism. These components collectively improve multi-scale feature representation and enable more efficient weighting of features across channels. Building upon this optimized architecture, we develop the 1D-ViMix network, which introduces modifications to the weak and strong augmentation modules within the standard ReMixMatch framework. These modifications lead to a significant improvement in diagnostic accuracy. Experimental results demonstrate the outstanding performance of our method in few-shot scenarios, achieving an average diagnostic accuracy of 95.3% and an average F1-score of 95.2%. The proposed approach exhibits superior diagnostic accuracy and robustness compared to existing unsupervised and semi-supervised methods, highlighting its substantial potential for practical health monitoring of rolling bearings.