A new domain robust one-class fault detection framework for large-scale chemical processes

Abstract

The scarcity of fault samples in chemical processes necessitates one-class fault detection methods. Existing approaches often focus on mono-domain data, ignoring variability under different operating conditions. In practice, factors such as noise drift and fluctuating operational conditions lead to significant differences in sample distributions, resulting in high false alarm rates when these methods are applied to multi-domain data. To address this, we propose the Domain Robust One-Class (DROC) fault detection framework. DROC leverages self-supervised Contrastive Learning to align normal samples from different domains into a shared, stable feature space. It also incorporates Deep Support Vector Data Description with Outlier Exposure to improve feature compactness and discrimination. Additionally, a Transformer-based feature extractor and a novel Soft Brownian Offset algorithm are designed to capture temporal and inter-variable dependencies, improving fault detection accuracy. Applications to the Tennessee Eastman Process and a real-world pitting corrosion monitoring system demonstrate DROC’s superiority in one-class fault detection.