Abstract: Rolling bearing fault diagnosis is a critical task in condition monitoring and intelligent maintenance of rotating machinery, as bearing failures may lead to unexpected downtime, safety risks, and significant economic losses. In recent years, deep learning–based diagnostic methods have shown strong capability in extracting discriminative features from complex vibration signals. However, most existing supervised approaches rely heavily on large-scale labeled datasets, which are difficult and costly to obtain in real industrial environments. Although unsupervised learning methods reduce the dependence on labeled data, their diagnostic accuracy and robustness are often insufficient for engineering applications. Therefore, developing an effective fault diagnosis method that can achieve high performance under limited labeled data conditions remains a challenging problem. To address this issue, in this study, a semi-supervised few-shot fault diagnosis method is proposed for rolling bearings based on physical feature fusion. The proposed method is developed upon an improved framework of semi-supervised learning with distribution matching and augmentation anchoring (ReMixMatch), integrates physical feature engineering with a physics-guided deep neural network architecture, forming a Semi-Supervised learning with physical feature fusion (1D–ViMix) model. The core objective is to fully exploit a small number of labeled samples together with a large amount of unlabeled data, while enhancing model interpretability through the incorporation of physical prior knowledge. Specifically, raw bearing vibration signals are initially processed via band-pass filtering and Hilbert transform to obtain envelope signals, which effectively suppress low-frequency background components and highlight fault-induced modulation characteristics. Based on the envelope signals and their envelope spectra, multiple physical features are extracted, including time-domain statistical indicators, envelope spectrum frequency-domain features, and segmented frequency-band energy features. To construct a compact and discriminative physical feature set, a systematic feature selection strategy is adopted. This strategy combines single-feature predictive performance evaluation using Macro–F1 scores, Pearson correlation analysis to identify redundant features, and a forward greedy search algorithm with correlation constraints, ensuring a balance between diagnostic performance, stability, and feature diversity. In terms of network design, a deep one-dimensional convolutional neural network is employed as the backbone to capture temporal characteristics of vibration signals. Multi-scale residual blocks are incorporated to enhance the representation of fault-related patterns across different time scales and frequency components. Furthermore, a physical prior regularization mechanism is introduced at the classification stage to guide the learning of classification weights. This regularization encourages the model to assign higher importance to strongly correlated physical features while suppressing weakly related ones, thereby improving consistency between model decisions and bearing fault mechanisms and enhancing interpretability. To improve the effectiveness of semi-supervised learning for vibration signals, the original ReMixMatch framework is further optimized. Pseudo-label generation is stabilized using exponential moving average smoothing, followed by distribution alignment and sharpening to mitigate class imbalance and improve label confidence. Additionally, signal-oriented weak and strong augmentation strategies, including noise injection, random masking, and amplitude scaling, are designed to effectively match the characteristics of mechanical vibration data. Consistency regularization and mixup are jointly applied to labeled and unlabeled samples, enabling more effective utilization of unlabeled data and improving model generalization under few-shot conditions. Experimental validation is conducted on the rolling bearing fault dataset from Paderborn University under few-shot learning scenarios, where only a limited number of labeled samples are available for each fault category. Extensive experiments, including multiple random trials, comparative studies with representative semi-supervised and small-sample learning methods, ablation tests, and interpretability analysis, are performed. The proposed 1D-ViMix model achieves an average diagnostic accuracy and F1-score of 95.4%, significantly outperforming the comparison methods. The results demonstrate that the integration of physical feature priors, multi-scale feature modeling, and improved semi-supervised learning strategies effectively enhances diagnostic accuracy, robustness, and interpretability. This study provides a practical and explainable solution for rolling bearing fault diagnosis under limited labeled data conditions in real industrial applications.