Estimation of Melting State at Groove Root in Pulsed MAG Welding Using Deep Learning

• Published in: QUARTERLY JOURNAL OF THE JAPAN WELDING SOCIETY Vol. 39; no. 4; pp. 322 - 333
• Main Authors: Denchi, WANG, Gewei, ZHANG, YAMANE, Satoshi
• Format: Journal Article
• Language: Japanese
• Published: Tokyo JAPAN WELDING SOCIETY 2021; Japan Science and Technology Agency
• Subjects: Armor; Ceramics; Deep learning; Estimation of weld pool penetration; Gas metal arc welding; GMA Welding; Grooves; Image classification; Machine learning; Melting; Metal active gas welding; Penetration depth; Pulsed current welding; Training; V groove; Weld pool image
• ISSN: 0288-4771; 2434-8252; 2434-8252
• DOI: 10.2207/qjjws.39.322

Robotic welding has been introduced in GMA welding to save labor. Estimation of the penetration depth of the molten pool is important to obtain good welding results in this welding. The authors used the image of the weld pool as input for deep learning to identify the welding state and estimate the melting state of the root of the groove. In this study, a single-sided downward-facing weld with a ceramic backing material is used. In order to apply deep learning, it is necessary to construct training data. A large amount of time is required to construct the training data. The relationship between the melting state of the root of the groove and that of the ceramic backing material was determined by basic experiments. The ceramic backing material melts when the brightness of the state in front of the weld pool increases. Using this feature, the state of the tip of the weld pool in the molten pool image was classified into three types. This facilitated the image classification. In addition, since only the tip ofthe molten part was targeted, the image was less susceptible to changes in the bevel shape. This was applied to the case where the gap varied from 4 mm to 6 mm. The effectiveness of the estimation of the melting state of the root of the groove was confirmed.