Attention-Based Gated Recurrent Unit Approach for Fault Diagnosis in the Tennessee Eastman Process

Abstract

Fault diagnosis in chemical processing systems remains difficult because these operations are highly complex, nonlinear, and safety-critical; undetected abnormalities can escalate into major operational and economic losses. Traditional FDD techniques often have limited capability in handling high-dimensional sensor measurements and typically overlook time-dependent behavior, which reduces their diagnostic reliability. In this study, an attention-enhanced gated recurrent unit (Attention-GRU) model is introduced to improve process fault identification. The architecture couples GRU layers with an attention module that highlights the most relevant temporal segments, strengthening the learned feature representations. The derived context vector is then passed to a fully connected SoftMax classifier. Using the Tennessee Eastman process (TEP) as the evaluation benchmark, the method achieved average accuracy, precision, and F1-score values of about 90%, surpassing the standard GRU baseline, which recorded roughly 86%, 85%, and 85%, respectively. The findings indicate better diagnostic performance, adaptability, and earlier fault recognition, providing a more robust and reliable solution for industrial chemical process monitoring.