Enhanced Fault Detection in Industrial Process via a Hellinger Distance-Based Data-Driven Scheme

Accurate fault detection in modern chemical processes ensures safety and efficiency. However, detecting faults in noisy environments remains a significant challenge. This paper introduces an effective fault detection (FD) strategy that combines a supervised learning approach using partial least squares (PLS) with fault detection charts based on Hellinger distance (HD) and exponentially weighted moving average (EWMA). The PLS-HD-EWMA strategy utilizes PLS to generate residuals, while HD is employed to compute the divergence between actual residuals and fault-free residuals. EWMA is then applied to evaluate the HD, providing sensitivity to small changes for effective fault detection. The proposed strategy's efficacy is demonstrated through simulations on a benchmark continuous stirred tank reactor process, evaluating three common fault scenarios. The monitoring performance indicates that the PLS-HD-EWMA approach outperforms all the other conventional methods, offering superior fault detection capabilities.