Recurrent Autoencoder for Fault Detection in a Polymerization Reactor Train

A fault detection system based on a recurrent autoencoder is proposed, suitable for continuous processes with complex nonlinear dynamics. The system consists of three modules: 1) a recurrent autoencoder; 2) a fault state detection module; and 3) an interpretation module. The training strategy only requires process data obtained under normal operating conditions. The fault detection module analyzes the total reconstruction error of the measured variables obtained by the recurrent autoencoder. The interpretability module examines the reconstruction errors of the individual variables and highlights variables potentially responsible for the detected anomaly. The developed system is applied to a computationally simulated Styrene‐Butadiene rubber latex production process. The latex is synthesized in a train of continuous stirred tank reactors that may be affected by various types of faults. Simulation results show a satisfactory performance of the proposed fault detection system. Failures whose data patterns differ significantly from those of normal operation are effectively detected. In contrast, some subtle faults only cause minor deviations in the data patterns and are therefore difficult to detect rapidly, leading to delays in detection. In most cases, the interpretability module can correctly identify the variable more strongly associated with the detected fault.