Most current intelligent fault diagnosis models, dependent on specific working condition data for training, cannot effectively diagnose faults in unknown working conditions without data. Zero-shot learning (ZSL) can identify samples unseen during the training phase and has now been applied to the field of fault diagnosis. However, effective methods for constructing fault semantics in unknown conditions for zero-shot fault diagnosis are still lacking. This paper proposes a hybrid semantic attribute-based zero-shot learning model (HSAZLM) to address this issue. A novel approach to constructing the semantic space is introduced by combining manually defined semantic attributes (SA) with non-semantic attributes (NSA) to form hybrid semantic attributes (HSA). This addresses the issue of insufficient semantic information due to limitations in expert knowledge when defining SA. In the hybrid semantic attribute construction module, a denoising residual convolutional autoencoder (DRCAE) is employed as a semantic learning model to acquire NSA containing high-dimensional abstract features, enhancing the model's generalization and robustness. To verify the effectiveness and superiority of the proposed HSAZLM, experiments were conducted on both a publicly available bearing dataset and a self-built bearing dataset. The average accuracy rates are 96.82% and 96.42% respectively, exceeding the current state-of-the-art methods.