Fault diagnosis of bearings plays a critical role in the predictive maintenance and health management of mechanical equipment. However, compound faults diagnosis of bearings remains a challenging task, because compound faults are coupled together by multiple single faults, which are difficult to decouple and identify. In addition, in industrial scenarios, collecting sufficient compound fault samples for model training is unrealistic. To address these issues, we propose a generative zero-shot compound fault diagnosis model based on semantic alignment. The model includes a semantic alignment module, feature extraction module, generative adversarial module, and classification module. First, we propose a feature extractor that combines channel and spatial attention mechanisms to extract fault features from the original vibration signals. Second, we designed a novel fault semantic construction method that combines interpretable priori fault semantics in time–frequency domain and hidden feature distribution alignment. Third, in the generative adversarial module, we combine supervised contrastive loss and comparator into the improved Wasserstein generative adversarial networks with gradient penalty (WGANs_GP) to learn the mapping relationship between fault semantics and fault features. Finally, we use the pseudocompound faults features generated by the generator to train a classifier for classifying ground-truth unseen compound fault. The effectiveness of the proposed method is verified on a self-built bearing test bench, and the results show that the proposed model can achieve classification accuracy of 89.25% for compound faults under the condition of only using single-fault samples for training.