In-situ estimation of welding quality and visualization of basis for the decision by combining welding monitoring and deep learning

2023 Fiscal Year Final Research Report

• Project/Area Number: 21K03806
• Research Category: Grant-in-Aid for Scientific Research (C)
• Allocation Type: Multi-year Fund
• Section: 一般
• Review Section: Basic Section 18020:Manufacturing and production engineering-related
• Research Institution: Osaka University
• Principal Investigator: Nomura Kazufumi 大阪大学, 大学院工学研究科, 准教授 (90397729)
• Project Period (FY): 2021-04-01 – 2024-03-31
• Keywords: アーク溶接 / モニタリング / 溶込み深さ / AI / CNN / 機械学習 / 判断根拠の可視化

Outline of Final Research Achievements

Robotic arc welding, a process used in various industrial fields, often encounters challenges unstable welding quality due to disturbances such as the gap fluctuation between the base materials and the misalignment of the wire target position. In response to these issues, we have developed a CNN-based machine learning model that estimates penetration depth from monitoring images of the welding phenomenon. Our model has been effective, but it also has its limitations, including being black-boxed and having low estimation accuracy in certain areas. In this study,we have visualized the basis for decision to identify important areas in the input image and clarify the physical relationship with the welding phenomenon. Moreover, we constructed an improved estimation model by shifting the input-output relationship.

Free Research Field

溶接，AI，検査・計測

Academic Significance and Societal Importance of the Research Achievements

昨今の機械学習は画像との相性が非常に高く溶接分野においても応用があり，画像から人の目でわかる特徴点を算出する操作を自動化したものが少なくない．本研究は，溶融池モニタリング結果を直接溶接品質と相関させ，重要な特徴量を可視化する手法の提案とその評価をしたものである．これは，人の目では一見わからない職人技能の可視化であり，真の自動化のためには必須であるといえる．一見してBlack Box化したAI応用技術であっても，物理的な溶接現象を反映したものであることが可視化され，更に入出力関係を吟味することで精度の高いモデルが構築できることを示すことができたことは学術的にも工学的にも非常に意義深いと言える．