A two-stage reconstruction-based OSR–ZSD framework for generalized zero-shot fault diagnosis in industrial processes

Abstract

In industrial processes, some fault types lack sufficient training data, and test samples may originate from both seen and unseen classes, giving rise to a generalized zero-shot diagnosis (GZSD) scenario. Existing methods often suffer from severe class bias, incorrectly classifying unseen faults as seen ones and thus compromising diagnostic reliability in practical applications. To address this issue, this study proposes a two-stage reconstruction-based open-set recognition (OSR)-zero-shot diagnosis (ZSD) framework for GZSD. In the first stage, a semantic-conditioned reconstruction model is developed for OSR, which identifies seen-class samples and adaptively rejects unseen ones by evaluating semantic consistency. In the second stage, an attribute-coupled ZSD model maps feature representations into an attribute semantic space and infers fault categories for unseen samples based on semantic similarity, thereby achieving unified diagnosis for both seen and unseen classes. Experiments on the Tennessee–Eastman process dataset demonstrate that the proposed method achieves an average accuracy of 98.32% for seen classes and 91.51% for unseen classes, with a harmonic mean of 94.74%, outperforming existing GZSL approaches in both recognition performance and balance. The results indicate that the proposed framework provides an effective and practical solution for generalized zero-shot fault diagnosis in industrial processes.