High-quality labeled data serves as the cornerstone for achieving precise fault diagnosis. However, accurately annotating sample data comes with prohibitively high costs. To alleviate the burden of annotation costs and reduce the overall application cost of fault diagnosis models, this paper presents a hybrid supervised model based on sample quality assessment for fault diagnosis in rotating machinery, leveraging both labeled and unlabeled data to overcome the challenges of limited labeled data and high annotation costs. The approach begins with training a source model on a small set of labeled data to generate pseudo-labels for unlabeled data. Through Uncertainty and Cluster Sampling, high-quality pseudo-labeled data is selected for supervised training alongside genuine labeled data. The model employs adversarial loss between outputs of the source and classification models for unsupervised training, enhancing the classification model’s performance. Experimental results on bearing datasets demonstrate the method’s superior fault diagnosis capabilities compared to existing techniques, offering a cost-effective solution for fault diagnosis with limited labeled data.