Fault Diagnosis in Chemical Reactors with Data-Driven Methods

Abstract

This study investigates fault diagnosis, encompassing fault detection, isolation, and estimation, with experimental data in a continuous stirred-tank reactor (CSTR) for the liquid-phase catalytic oxidation of 3-picoline with hydrogen peroxide. Two key faults were examined: coolant inlet temperature spikes (fault 1) and 3-picoline feed concentration decreases (fault 2). Data-driven methods, including random forest (RF) and k-nearest neighbors (KNN), successfully detected, isolated, and estimated faults under nominal conditions. However, both data-driven and model-based residual generators were disrupted by a shift in the heat transfer coefficient (U). An isolation forest (IF) algorithm was used for anomaly detection and model recalibration, restoring model-based performance. Updated data sets enabled RF and KNN to adapt effectively, demonstrating their scalability and adaptability. Experimental results highlight the strengths of both methods, advocating for a combined framework for robust fault diagnosis.