Proceedings of the
9th International Conference of Asian Society for Precision Engineering and Nanotechnology (ASPEN2022)
15 – 18 November 2022, Singapore
doi:10.3850/978-981-18-6021-8_OR-01-0279
Study on Wire and Arc Additive Manufacturing of Metals Using Image Processing and Reinforcement Learning
Department of Power Mechanical Engineering, National Tsing Hua University, No. 101, Sec. 2, Guangfu Rd., East Dist., Hsinchu City, 300044, Taiwan (R.O.C.)
ABSTRACT
Currently, metal additive manufacturing systems can melt metal materials using high-energy sources. The most common high-energy sources such as lasers and electron beams are considered the stable energy sources, but the biggest weakness of these sources is the costly and high energy consumption system. In this research, the wire and arc additive manufacturing (WAAM) is used which characterized by low cost. However, the imperfect energy control system makes this manufacturing technology have the following shortcomings: accuracy and surface irregularities. In this research, the monitoring vision system is set up to capture fused deposition images while the CNC platform is moving at speed 1.5 mm/s. About 1300 continuous images are synthesized by template matching, and then the full images under different feeding rates will be extracted contour information by using edge detection. The contour information will be used for training reinforcement learning model to obtain a good feeding rate strategy to improve the profile accuracy. Therefore, this research integrates computer vision, image processing, and reinforcement learning technologies to learn the best feeding rate parameter strategy. Then use this set of parameter strategy in the process to reduce the appearance change of the printed result. This research uses a parameter strategy decision-making method based on the Q-learning algorithm to efficiently find out the good feeding rate strategy that can reduce appearance change to 0.29 mm in few training episodes. This method does not need to spend a lot of experimental time to find good process parameters.
Keywords: Reinforcement Learning, Image Processing, Wire and Arc Additive manufacturing, Plasma Arc Welding, Q-learning.
Department of Power Mechanical Engineering, National Tsing Hua University, No. 101, Sec. 2, Guangfu Rd., East Dist., Hsinchu City, 300044, Taiwan (R.O.C.)
ABSTRACT
Currently, metal additive manufacturing systems can melt metal materials using high-energy sources. The most common high-energy sources such as lasers and electron beams are considered the stable energy sources, but the biggest weakness of these sources is the costly and high energy consumption system. In this research, the wire and arc additive manufacturing (WAAM) is used which characterized by low cost. However, the imperfect energy control system makes this manufacturing technology have the following shortcomings: accuracy and surface irregularities. In this research, the monitoring vision system is set up to capture fused deposition images while the CNC platform is moving at speed 1.5 mm/s. About 1300 continuous images are synthesized by template matching, and then the full images under different feeding rates will be extracted contour information by using edge detection. The contour information will be used for training reinforcement learning model to obtain a good feeding rate strategy to improve the profile accuracy. Therefore, this research integrates computer vision, image processing, and reinforcement learning technologies to learn the best feeding rate parameter strategy. Then use this set of parameter strategy in the process to reduce the appearance change of the printed result. This research uses a parameter strategy decision-making method based on the Q-learning algorithm to efficiently find out the good feeding rate strategy that can reduce appearance change to 0.29 mm in few training episodes. This method does not need to spend a lot of experimental time to find good process parameters.
Keywords: Reinforcement Learning, Image Processing, Wire and Arc Additive manufacturing, Plasma Arc Welding, Q-learning.
