Open Access

Lv Wang, Junyu QiORCiD, Yi QinORCiD

• Zero-shot learning offers a promising approach to tackle the challenge of diagnosing compound faults even when training data for such faults is unavailable. However, the current zero-shot diagnosis methods have the following problems. First, existing methods typically only construct a single prototype based on prior knowledge, neglecting the integration of both theoretical prior knowledge and practical features. This oversight can easily result in zero-shot domain shift problems. Secondly, most zero-shot diagnosis methods focus on constructing the relationship between features and attributes, lacking pre-processing techniques for feature separation, which might lead to feature confusion. To tackle these problems, a contrastive cluster zero-shot (CCZS) model is proposed. First, a feature-guided prototype framework is constructed to alleviate the domain shift problem in zero-shot learning. This framework alleviates the domain shift problem of the single theoretical prototype by integrating actual feature clustering prototypes. Moreover, the constructed prototype is difficult to match with actual labels due to severe domain shift issues. To address issues, a prototype matching mechanism is developed. Additionally, a wavelet-based contrastive network is proposed to address the feature confusion problem. The proposed network can expand the distance between different prototype features and reduce the distance between similar prototype features, improving the separability of features. The experimental results on two bearing experiments validate the superiority of the CCZS model over other classic and advanced methods.