Machine learning-based online parameter estimation for noise-resilient process measurements using state-dependent models

Abstract

Measurement noise remains a critical challenge in industrial process control, often leading to inaccurate estimations, actuator wear, and degraded control performance. Traditional data reconciliation filters often rely on fixed-parameter models and require large sets of input variables, limiting their adaptability to process variations. This study addresses these limitations by introducing a noise reduction framework that combines dynamic data reconciliation with online parameter estimation, using a nonlinear state-dependent parameter (SDP) modeling approach. The proposed framework adaptively updates model parameters based on past reconciled data, enhancing robustness and accuracy under dynamic and noisy operating conditions. The method was evaluated in two case studies. In an industrial debutanizer process, the framework significantly reduced the standard deviation of manipulated variables by up to 54 %, improving control smoothness and actuator stability. In a simulated benzene–toluene distillation column, it outperformed a Refined Instrumental Variable-based Kalman Filter (RIV-KF) by reducing measurement noise by 50 %, while maintaining reliable performance under process state changes (PSC), even when unmodeled inputs varied significantly. Furthermore, the proposed filter decreased benzene concentration variability by 17 % across trays and reduced reboiler energy consumption by approximately 0.1 million kilocalories over 3.5 h. These results demonstrate the practicality of using reconciled-data-based online model adaptation for improving both measurement reliability and control efficiency in complex industrial processes.