Advancing fault diagnosis for ethyl benzene production plant: A machine learning and explainable AI approach

Fault diagnosis in chemical process plants is critical to ensuring operational safety, product quality, and economic performance. In complex units such as ethyl benzene production plant, early fault detection is challenging due to nonlinear process dynamics, strong variable interactions, and the scarcity of labelled fault data. To address this gap, this study presents a simulation-driven, explainable machine learning-based fault diagnosis system (FDS) capable of identifying multiple fault types without reliance on extensive historical plant data. A dynamic process simulation was conducted to generate datasets under both normal and faulty operating conditions. The system was first operated for 20 h in normal closed-loop mode with a 0.6-min sampling interval to form the baseline dataset. For each fault case, the simulation was reset and run for 5 h to allow stabilization, followed by 15 h of faulty operation, producing 2001 samples across 28 process variables per case. Fourteen representative faults, including sensor bias, actuator malfunctions, and process disturbances, were introduced. After feature selection, multiple classifiers were trained, and three high-performing yet diverse models, Ensemble Bagged Trees (tree-based), Support Vector Machine (kernel-based), and Fine KNN (distance-based), were integrated using a majority-voting mechanism to enhance robustness. EBT being best performing model achieved 99.98 % accuracy and an F1 score of 0.9988. Local Interpretable Model-agnostic Explanations (LIME) provided transparent insight into variable importance for each fault class, aiding interpretability. The novelty of this work lies in generating realistic labelled datasets entirely from dynamic simulation, combining heterogeneous algorithms for improved generalization, and embedding explainable AI (XAI) to create a scalable, transparent FDS framework applicable to other chemical processes with limited fault history.