Data-driven soft sensor for online composition estimation with adaptive-window regression and QP-constrained kalman estimator.

Robust online composition estimation is crucial for sustaining energy efficiency in industrial distillation processes, especially under dynamically drifting and unobserved conditions. Classical model-based estimators often encounter limitations in adapting to nonlinearities, signal drift, and measurement noise, while pure data-driven techniques fail to generalize to unseen trends and process drifts, resulting in suboptimal servo and regulatory performances. To overcome these limitations, this study proposes a hybrid data-driven soft sensor framework that integrates a Quadratic Program-based Constrained Kalman Estimator (QP-CKE) with Piecewise Linear Regression (PLR), along with an Adaptive Window (AW) extension. The AW-PLR dynamically adjusts regression window length based on a Cumulative Sum (CUSUM) derived nonlinearity index, enabling better sensitivity to transients and non-stationary process behaviour. The proposed soft sensors are validated for an ethanol-water mixture separation in a pilot-scale laboratory distillation column. Performance is benchmarked against Extended Kalman Estimator (EKE), static QP-CKE, and a Support Vector Machine (SVM)-based regression model. Quantitative evaluation shows that the proposed AW-PLR based QP-CKE achieves approximately 35 % lower RMSE, 22 % higher SNR, and 25 % faster computation time, while also offering noise resilience and process interpretability compared to its counterparts. The proposed hybrid soft sensors demonstrate enhanced adaptability, computational efficiency, and robustness, supporting their suitability for integration into real-time soft sensing and control frameworks.