Zero-shot condition monitoring of industrial systems via semantic manifold alignment under large semantic gaps

Abstract

Condition monitoring of modern industrial systems requires reliable fault identification from process variables and vibration signals, yet recognizing previously unobserved faults remains a major challenge when only limited categories are available during training. Zero-shot fault diagnosis (ZSFD) aims to recognize unseen industrial faults by leveraging semantic attribute descriptions, yet its generalization remains challenging when the semantic gap (SG) between seen and unseen classes is large. Existing ZSFD approaches largely follow a decoupled pipeline, where attribute prediction is optimized independently of the subsequent prototype matching, often resulting in biased projections and degraded diagnosis. To address this limitation, a framework termed semantic joint alignment network (SJA-Net) is proposed, which couples attribute prediction with semantic alignment through end-to-end joint optimization. Specifically, SJA-Net first strengthens attribute learning via a class-balanced prediction objective to reduce bias toward seen classes. The predicted attributes are then aligned to a topology-informed semantic embedding space to preserve inter-class relations during matching. Finally, a topology-aware contrastive objective enforces compact intra-class clustering around semantic anchors and enlarges inter-class margins, thereby improving robustness to large SGs and enhancing transfer to unseen fault categories. Extensive experiments on the Tennessee Eastman Process and the Case Western Reserve University rolling-element bearing vibration benchmark validate the effectiveness and robustness of SJA-Net for unseen-fault diagnosis in industrial condition monitoring applications.