Open Collections

Abstract

Addressing faults in industrial systems is essential to ensure safe and efficient operations. Traditional fault classification methods, which depend heavily on labeled fault data, have limited applicability and struggle to address unseen faults. Current methods that aim to handle unseen faults often fail to distinguish between different fault types, lack explainability, or are restricted to processes with a single monitored variable. This work introduces a novel system capable of distinguishing between different faults and providing fault localization for improved explainability—all while relying solely on Normal Operating Conditions (NOC) data. At its core, the system employs a probabilistic recurrent neural network with a state-space-like structure to model the NOC process dynamics. The model is then used for fault detection and feature extraction. The latter is based on isolating the fault effect from the process measurements and applying manual feature extraction to the isolated component. Clustering these features reveals information about the underlying fault types by grouping similar observations. The process variables’ association with each cluster (discovered fault type) is ranked by analyzing each cluster through principal component analysis. The proposed approach is evaluated on the Tennessee Eastman Process (TEP), demonstrating a 0.89% average increase in detection accuracy and a 60% reduction in false positive alarms compared to state-of-the-art fault detection models. Furthermore, the system successfully differentiates between 12 distinct faults and provides fault localization rules that are consistent with both the literature and the TEP flow diagram.