Fault detection using multiscale recursive principal component analysis for chemical process systems

Abstract

Process monitoring is essential for maintaining operational safety and product quality in chemical industries. Although conventional fault detection techniques are widely used, their static nature often leads to high false alarm rates (FAR) and missed detection rates (MDR) under dynamic conditions. To address these limitations, this study proposes a Multiscale Recursive Principal Component Analysis (MSRPCA)-based fault detection framework that combines multiscale signal decomposition with the adaptive capabilities of Recursive PCA (RPCA). The MSRPCA approach isolates process variations across different frequency bands while continuously updating the Principal Component Analysis (PCA) model using a moving window mechanism. This enables real-time adaptability and enhanced noise resistance. The proposed method is validated using the Tennessee Eastman Process (TEP), a widely used benchmark for chemical process monitoring under a range of fault types, including step, drift, and random variation disturbances. Fault detection performance is quantitatively assessed using FAR and MDR metrics across 20 distinct fault scenarios. The results demonstrate that MSRPCA consistently outperforms traditional techniques, significantly reducing false alarms while improving fault detection accuracy. For instance, in Fault 16, the MDR in the Hotelling’s T² (T²) chart decreased from 70.5 % (PCA) to 10.5 % (MSRPCA), while the FAR in the Squared Prediction Error (SPE) chart dropped from 21.3 % to 0 %. These findings underscore the robustness and effectiveness of MSRPCA for real-time fault detection in complex, time-varying, and noisy industrial environments.