| 作成者 |
|
|
|
|
|
| 本文言語 |
|
| 会議情報 |
|
| 出版タイプ |
|
| アクセス権 |
|
| 関連DOI |
|
| 以下を参照 |
|
| 概要 |
This study explores the automation of welding inspection processes in shipyards by leveraging deep learning technologies to develop a system that aims to improve both efficiency and accuracy. We perfo...rmed 90 horizontal fillet weld tests in a laboratory environment and 96 tests in a field environment on a T-shaped joint specimen, obtaining output logs from the welding machine during the welding process. We investigated deep learning computational methods and data preprocessing, introducing input data standardization and grid search for hyperparameters. As a result, the adjusted coefficient of determination for estimating leg length and undercut depth was improved. Next, we examined the frame size and the amount of frame shift for separating the input data were investigated, and the optimal values were identified that met practical criteria in a laboratory environment.
In addition, we conducted field tests using a simple welding carriage commonly used in shipyards to validate the applicability of our system in real-world environments. The results showed that the adjusted coefficients of determination for leg length and undercut depth estimation were 0.69 and 0.45, respectively, demonstrating the potential of our approach to improve production efficiency through automated weld quality inspection. However, the study also identifies future challenges, including the need for more comprehensive training data, the incorporation of environmental data, and improvements in the estimation capabilities for various weld appearance features. This research serves as a step towards the automation of welding processes in the shipbuilding industry and provides directions for future research.続きを見る
|
Mendeley出力
