<p>The implementation of deep learning in fault diagnosis is challenging due to the small sample, since the deficiency of labeled fault data constrains the model’s efficacy. To address this challenge, a new similarity filtering-based pseudo-label learning approach (SFPL) is proposed based on dual information and sample attention mechanism. SFPL utilizes unlabeled data for pre-training through data similarity calculations. It also includes a sample attention mechanism that assigns weights to the samples to improve the efficiency of the model’s learning. Additionally, a data filtering mechanism based on cosine similarity is introduced to enhance the quality of pseudo-labels. These pseudo-labels are used to fine-tune the model for high-accuracy fault diagnosis. Validation experiments on two datasets show that the suggested method can achieve high accuracy and stability with only a few labeled samples.</p>