Abstract: Objectives To address the low fault diagnosis accuracy of existing neural networks under unseen operating conditions, a single source domain generalization method based on causal learning is proposed. Methods First, source domain samples are transformed through data combination operations to construct an auxiliary domain that simulates the domain shift between the source and target domains. Then, a causal graph is constructed to represent the relations among input data, transformation factors, semantic features, and fault labels. Next, a counterfactual inference module is designed to infer the causal effects of different transformation factors on the diagnostic results, and correct the auxiliary domain features accordingly. Finally, domain alignment is achieved by minimizing the distance between source samples and auxiliary samples in the feature space, thereby enhancing diagnosis performance under unseen conditions. Results Experiments conducted on a self-built reducer test bench validate the effectiveness of the proposed method. The results show that the method achieves an accuracy of 90.34% without using any target domain samples, outperforming existing approaches. Conclusions The proposed approach can provide key technical support for fault diagnosis of industrial equipment under varying operating conditions and contribute to the advancement of intelligent manufacturing.