A Compact CNN-Transformer Model for Robust Fault Diagnosis in Large-Scale Chemical Processes

Abstract

Ensuring the safe and stable operation of large-scale chemical processes requires accurate fault detection and diagnosis under nonlinear dynamics and strong variable interactions. This study investigates deep learning-based fault diagnosis for the Tennessee Eastman Process (TEP), focusing on whether performance improvements beyond near-saturation Long Short-Term Memory (LSTM) baselines can be achieved. A standardised TEP dataset with 52 measured and manipulated variables is used, excluding the non-detectable fault cases (IDV 3, 9, and 15), resulting in a 19-class classification problem. A hybrid Convolutional Neural Network–Transformer (CNN-Transformer) architecture is proposed in which one-dimensional convolutional layers capture local cross-variable correlations, while a Transformer encoder models long-range temporal dependencies through self-attention. To ensure fair comparison, both the proposed model and a strong LSTM baseline are trained and evaluated under identical preprocessing, optimization, and evaluation protocols. The CNN-Transformer achieves an overall classification accuracy of 99.92%, marginally outperforming the LSTM baseline (99.86%). Although the numerical improvement is slight, the proposed model consistently yields higher macro-averaged F1-scores and reduced fault-wise misclassification, indicating enhanced robustness in challenging fault scenarios. The key contribution of this work is demonstrating that combining convolutional feature extraction with attention-based temporal modelling provides consistent class-level robustness beyond near-saturated recurrent architectures, while maintaining a compact structure suitable for practical deployment.