System Identification and Machine Learning Model Construction for Reinforcement Learning Control Strategies Applied to LENS System

Abstract

Identifying and controlling of additive manufacturing processes has the potential to improve part quality during the build process. The melt pool size of direct energy deposition processes has been related to part quality. In this paper, we investigate the use of system identification tools to device closed-loop controllers that are capable of regulating the melt pool size during the build process. Based on the results of linear models, machine learning approaches are investigated with the goal to obtain higher fidelity models, capable of characterizing the nonlinearities existing in such processes. In addition, a reinforcement learning controller is proposed that can accommodate the nonlinear behavior and the initial uncertainty in the model. Experiments with a direct energy deposition setup show improved part geometry using the linear model and controller. Simulation results employing the developed reinforcement learning controller show promise in enhanced control performance.