Design and implementation of deep learning-based framework for multi-class fault diagnosis in complex chemical process systems

Remigius Nnadozie Ewuzie, Shivaneswar Gunasekaran, Zainal Ahmad, Norazwan Md Nor,
Design and implementation of deep learning-based framework for multi-class fault diagnosis in complex chemical process systems,
Engineering Applications of Artificial Intelligence,
Volume 162, Part D,
2025,
112630,
ISSN 0952-1976,
https://doi.org/10.1016/j.engappai.2025.112630.
(https://www.sciencedirect.com/science/article/pii/S0952197625026612)
Abstract: Fault diagnosis in modern chemical plants is increasingly challenging due to process complexity, nonlinearity, and high-risk operations, where undetected faults can cause severe safety and economic consequences. Conventional machine learning (ML) models suffer from reliance on handcrafted features, poor generalization in high-dimensional spaces, and limited labeled data, resulting in reduced diagnostic performance. To overcome these challenges, we propose a scalable deep learning (DL) framework for multi-class fault diagnosis in chemical processes. The framework employs convolutional neural networks (CNN), autoencoders (AE), and long short-term memory (LSTM) networks to automatically extract spatial and temporal features from multivariate process data. Validated on the Tennessee Eastman process (TEP) benchmark, CNN achieved the highest diagnostic performance, with 88 % accuracy, 91 % precision, and 89 % F1-score. AE also performed strongly, with 85 % accuracy and 82 % F1-score, while LSTM achieved 71 % accuracy and 75 % F1-score, all outperforming conventional machine learning models, which scored between 48 % and 52 % accuracy. Superior AUC scores (micro-average: 1.00; macro-average: 0.99) confirm the framework's robustness, including in detecting overlapping and rare faults. The proposed two-phase approach, offline training and real-time monitoring, offers a practical solution for improving fault diagnosis accuracy, adaptability, and early warning capabilities in industrial processes.
Keywords: Process monitoring; Machine learning; Deep learning; Multi-class classification; Performance comparison; Fault diagnosis