Zero-sample fault identification (ZSFD) has gained significant attention due to its notable success. However, current methodologies face challenges related to fault attribute labeling and feature extraction, which hinder their generalization and robustness. Specifically, relying on expert knowledge for fault attribute labeling is both time-intensive and laborious. Additionally, existing approaches typically extract fault features with in a single projection space, disregarding valuable information from target data, thereby leading to suboptimal ZSFD performance. To overcome these challenges, we introduce an innovative zero-sample fault identification method based on transfer learning. Our approach begins by transferring a shared knowledge dictionary, automatically learned from labeled source data, to the target data, thereby reducing ZSFD’s reliance on detailed fault descriptions. Next, we introduce a novel multiclass space projection model to capture discriminative fault features. Finally, we employ a pseudolabel mechanism to uncover inter-class and intra-class information within the target domain. Experiments with the Tennessee-Eastman process and a real-world hot roll steel process showcase the effectiveness and superiority of the method.