In real industrial scenarios, due to the difficulty in obtaining fault data and the continuous emergence of new fine-grained fault modes under variable operating conditions, the demand for small-sample cross-domain fault diagnosis under variable operating conditions has gradually come to the fore. However, in the context of large-scale fine-grained fault classification, the current small-sample cross-domain fault diagnosis methods have shortcomings such as weak feature screening ability, poor model generalization, and difficulty in identifying new categories. To this end, this paper proposes a small-sample cross-domain fault diagnosis method based on attribute-guided fine-grained feature comparison learning, integrating supervised contrastive learning and multi-attribute learning on the basis of multi-task learning method, guiding the fault diagnosis model to efficiently differentiate the known fine-grained fault features in the source domain through the attribute learning process, and realizing the accurate identification of unknown new faults in the target domain based on the extracted semantic information on cross-category attributes and the fine-tuning of fewer samples. In this paper, by comparing the performance of each fault diagnosis method on the sample-small, multi-classification bearing fault dataset, we verify that the proposed method has superior cross-domain diagnosis performance in small-sample fine-grained scenarios.