Authors - Sai Jagnyaseni Rana, Trailokyanath Singh, Pallavi Joshi, Sudhansu Sekhar Routray Abstract - This paper presents a data-driven closed-loop (CL) identification and controller reconstruction framework for interacting multivariable processes, validated on the benchmark Wood-Berry (WB) distillation column. CL reaction curve data are employed to identify process dynamics without interrupting operation. The measured step responses of diagonal and interaction channels are modeled using secondorder plus time-delay (SOPTD) structures, whose parameters are estimated through a hybrid particle swarm optimization (PSO) and nonlinear least-square fitting (NLSF) refinement scheme. The identified models are reduced to first-order plus time-delay (FOPTD) form using Skogestad’s approximation and further refined for improved accuracy. Based on the optimized FOPTD models and measured CL responses, decentralized PID controller are reconstructed using both PSO and reinforcement learning (RL) via a proximal policy optimization (PPO) agent. Simulation studies demonstrate that while PSO achieved reliable controller recovery, the RL-based approach provides superior transient matching and reduced tracking error. The results validate the effectiveness of the proposed framework for CL identification and data- driven controller reconstruction in interacting multivariable systems.
Open Zoom