Temporal-semantic alignment and adaptive calibration network for industrial zero-shot fault diagnosis

Abstract

Fault diagnosis is critical for mitigating potential risks and ensuring the operational safety of dynamic industrial systems. Zero-shot learning (ZSL) provides a promising solution for diagnosing unseen faults without labeled training samples. However, its application to industrial processes remains challenging because process measurements exhibit complex temporal dependencies, while cross-modal mappings may produce highly concentrated predictions around a small number of unseen-class prototypes. To address these challenges, this study proposes a temporal-semantic alignment and adaptive calibration network (TSAC-Net) for industrial zero-shot fault diagnosis. TSAC-Net employs a bidirectional long short-term memory network with temporal mean pooling to extract sequence-level dynamic representations from multivariate process windows. These temporal representations and fault attribute vectors are projected into a shared latent space and explicitly aligned through compatibility and attribute-prediction objectives, while SCE-based reconstruction regularization is incorporated to preserve process- and semantics-related information during cross-modal mapping. Furthermore, a transductive adaptive calibration strategy estimates prototype-wise similarity bias from the unlabeled evaluation batch and corrects prototype scores using a calibration coefficient selected from pseudo-unseen validation classes. Comprehensive experimental evaluations conducted on both the Tennessee Eastman (TE) process and the real-world Cranfield Multiphase Flow Facility (CMFF) dataset demonstrate the effectiveness of the proposed framework. The results show that the integration of dynamic temporal representation, temporal-semantic alignment, and prototype-score calibration improves average diagnostic performance and effectively alleviates prototype-level prediction concentration.