A Hybrid Deep Learning and Statistical Approach for Fault Detection and Diagnosis in AGRU Systems: Integration with Aspen Plus and Explainable AI

Abstract

Fault detection and diagnosis are vital functions in process industries, supporting real-time monitoring and ensuring plant safety. However, the presence of complex, high-dimensional, and highly correlated process data presents a significant challenge for traditional monitoring systems. In natural gas processing, the acid gas removal unit is a critical subsystem responsible for removing acid gases such as hydrogen sulfide and carbon dioxide. Faults such as foaming, tray corrosion, cooler fouling, low gas flow, and reduced amine circulation are common in AGRUs but have received limited attention in the literature, highlighting a significant research gap. To address this, a comprehensive study was conducted by simulating the AGRU process and its associated faults using Aspen Plus Dynamics. This enabled the generation of realistic, multivariate time-series data under normal and faulty operating conditions. For the fault detection phase, two approaches were developed and comparatively evaluated: principal component analysis, which employs Hotelling's T² and squared prediction error metrics, and a long short-term memory autoencoder, which utilizes reconstruction error and leverages sequential learning. Subsequently, a fault diagnosis was performed using bidirectional recurrent neural networks, specifically Bi-GRU and Bi-LSTM models, which were trained to classify fault types based on their temporal signatures. The Bi-GRU model demonstrated superior performance with an accuracy of 99.8% and an F1-score of 99.6%, indicating its suitability for robust fault classification. To enhance model interpretability, shapley additive explanations were applied to identify critical input variables influencing model predictions, and the results were compared with those from local interpretable model-agnostic explanations.