Reinforcement Learning based Straightening in Roll Forming

Abstract Roll forming is a sheet metal forming process used in a wide range of manufacturing industries that has high material utilisation and high production rates. The raw material is incrementally bent into the desired shape by consecutive forming stands. Due to the spatial bending, different longitudinal strains develop along the profile cross-section, leading to defects such as horizontal and vertical bowing or twisting. As a result, it is often necessary to rework the parts using straighteners. Their adjustment is typically based on trial-and-error procedures that are manual work intensive and dependent on individual experience. To obtain a higher level of automation, this paper presents a novel data-driven approach to controlling a straightener. Reinforcement Learning is used to control the complex system by learning through interaction. To generate the data needed for learning, a simulation pipeline is implemented, focusing on a symmetric hat profile. The core of the approach is a model-free Soft Actor-Critic algorithm for determining the optimal position of the straightening device. The algorithm was able to predict an optimal position for a specific maladjustment in all test cases and for a varying misalignment in eight out of nine cases.