A Data-Driven Approach to System Dynamics Modeling and Control Design

Abstract

This thesis presents an in-depth investigation on modeling and simulation of an optical fiber extrusion system and its controllers in production. With measured production data during the fiber drawing process, a long short-term memory (LSTM) neural network is architected, implemented, and trained to model the process dynamics of the fiber drawing plant. Training experiments were conducted to investigate the effect of several parameters on the model’s performance. Furthermore, statistical analysis models with assumed structures are employed as part of the black-box system identification process to model controllers in the production system, subject to noise and disturbances. With aforementioned components, a closed-loop simulation of the fiber extrusion system is then developed in MATLAB, proving the feasibility of simulating mechanical systems in production using learned models. The approach developed in this study is suitable for data-driven deployment of many kinds of manufacturing plants in production, which may have limited operational domains due to mechanical constraints. The simulation, once implemented in hardware, could potentially replace the laborious, iterative tuning process of the controllers, and serve as a design tool to optimize these controllers using a digital twin.