An effective data enhancement method of deep learning for small weld data defect identification

• Published in: Measurement : journal of the International Measurement Confederation Vol. 206; p. 112245
• Main Authors: Yang, Wei, Xiao, Yancai, Shen, Haikuo, Wang, Zhipeng
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.01.2023
• Subjects: Data enhancement; Deep learning; Generative Adversarial Net; Transfer learning; Weld defect identification; Deep learning; Generative Adversarial Net; Weld defect identification; Data enhancement; Transfer learning
• ISSN: 0263-2241; 1873-412X
• DOI: 10.1016/j.measurement.2022.112245

•Application of supervised generative adversarial network to small sample data expansion of X-ray weld.•Quantitative evaluation of the generated defect samples by using the concept score.•We provide a possible scheme to solve the optical film recognition of industrial weld based on deep learning. Welding has become one of the most important manufacturing technology. Due to the operating environment, workmanship and welding parameters, different welding defects will inevitably appear in the welding process. In order to effectively identify these defects, X-ray films based on nondestructive testing are usually used. Owing to the small number of X-ray film samples, this paper proposes an attention self supervised learning auxiliary classifier generative adversarial net (ASSL-ACGAN) algorithm to expand the samples to improve the defect identification of small sample data sets. In addition, the influence of data transformation preprocessing on the sample quality of ASSL-ACGAN is also studied. Finally, intelligent defects identification based on transfer learning on two data sets is carried out. Experimental results not only suggest that ASSL-ACGAN based data enhancement is superior to wasserstein GAN (WGAN), WGAN gradient penalty (WGAN-GP) and auxiliary classifier generative adversarial net (ACGAN), but also prove the identification accuracy of ASSL-ACGAN exceeds that on original data set, with an average of 2.79%. The paper provides a possible scheme for defect identification of small number samples.