Enabling Rewards for Reinforcement Learning in Laser Beam Welding processes through Deep Learning

摘要

Self-optimizing robots and machines in future factories are an exciting next step towards an ever more efficient industry. To achieve this goal, robots used in production must gain an understanding of the quality of their behavior. Machine learning can help us move closer to this goal. In this paper, we provide insights on the feasibility of self-optimized laser welding robots and show how accurate quality analysis based on deep learning and smart computer vision algorithms provide a reliable input for quality evaluation and ultimately a scoring function. Furthermore, the suggested scoring function can capture the defining properties of a weld. In turn, the score can be used as feedback to define a reward for a reinforcement learning agent’s action, which then optimizes the robot’s behavior accordingly. Our experiments show that we can achieve very good accuracy and consistency when evaluating the quality of the weld with deep learning and statistical modeling. Finally, we provide a production-oriented learning architecture that considers the scoring component in a reinforcement learning pipeline.