<p>In order to solve the problems of insufficient actual working condition samples, a scarcity of labeled samples, and insufficient robustness and accuracy of deep learning in bearing fault diagnosis. This article proposes a few-shot learning method based on hypergraphs for bearing fault detection, which achieves bearing fault detection under unbalanced samples. Specifically, a hypergraph learning method applied to signal spectrograms has been proposed. Firstly, the signal is converted into a spectrogram to increase information density. Then, a hypergraph is constructed to model complex inter-sample relationships. Critically, unlike conventional graph models that are limited to capturing pairwise relationships, this approach uses hyperedges to connect multiple sample vertices simultaneously, enabling it to effectively model the high-order correlations among features of complex faults (e.g., concurrent faults). By performing label propagation on this richer structure, fully utilizing labeled data and significantly improving feature quality in the absence of samples. A few shot image classification method is used in hypergraph learning to solve the problem of sample imbalance by iteratively applying unlabeled data. Afterwards, only a small amount of fault data is needed to complete the model training under actual working conditions, achieving effective bearing fault detection. The experimental verification of the Case Western Reserve University (CWRU) dataset confirms the practicality of the proposed diagnostic method.</p>