Enhanced fault detection in the Tennessee Eastman Process using Mahalanobis distance based maximum likelihood estimation

Abstract Process fault detection is crucial for ensuring the reliability and efficiency of complex industrial systems. In this study, we evaluate and benchmark six algorithms, including ensemble learning light gradient boosting (LGB), deep learning long short-term memory (LSTM) neural networks, K-means clustering (KMC), principal component analysis (PCA), Mahalanobis–Taguchi system (MTS) and maximum likelihood estimation-based Mahalanobis distance (MLE-MD), for robust process monitoring. Using the Tennessee Eastman Process (TEP) simulation dataset, we conduct exploratory data analysis (EDA), which includes the use of control charts and statistical tests for Gaussianity. A comparative analysis reveals that the MLE-MD model significantly outperforms the other algorithms in overall fault detection, achieving the highest average fault detection score (74.1 %). The proposed MLE-MD framework integrates maximum likelihood estimation with Mahalanobis distance to explicitly model the full covariance structure of multivariate process data. The superiority of MLE-MD is attributed to its ability to account for the data’s covariance structure and distribution, offering robustness against non-spherical cluster shapes. Crucially, MLE-MD achieved a 79.4 % success rate on the challenging fault IDV19, which PCA failed to detect entirely. MTS did achieve better FDR scores in some faults including IDV19, but has failed in detecting IDV5 and IDV16, while also being expensive in terms of memory. In contrast, while MLE-MD excelled quantitatively, PCA provided enhanced interpretability by generating “fault signatures” through Q-contribution heatmaps. This work provides valuable quantitative insights for practitioners and researchers, validating MLE-MD as the most effective method among those tested for advancing robust fault detection and diagnosis (FDD) systems in industrial environments.