Experimental Investigation to Improve Inspection Accuracy of Magnetic Field Imaging-Based NDT Using Deep Neural Network

• Published in: Russian journal of nondestructive testing Vol. 58; no. 8; pp. 732 - 744
• Main Authors: Park, Seung-Kyu, Kim, Jaemin, Park, Duck-Gun, Jo, Minho, Lee, Jinyi, Lee, Jonghwan
• Format: Journal Article
• Language: English
• Published: Moscow Pleiades Publishing 01-08-2022; Springer Nature B.V
• Subjects: Accuracy; Alternating current; Artificial neural networks; Characterization and Evaluation of Materials; Chemistry and Materials Science; Defects; Direct current; Electromagnetic Methods; Flux density; Inspection; Magnetic flux; Materials Science; Neural networks; Nondestructive testing; Structural Materials; Training; transformed images; defect detection; non-destructive testing (NDT); deep neural network; magnetic image
• ISSN: 1061-8309; 1608-3385
• DOI: 10.1134/S1061830922080101

A deep neural network is expected to be a useful tool to improve the accuracy of defect detection in Non-Destructive Testing (NDT). In this article, a deep neural network-based technique to improve the defect detection accuracy of an advanced NDT using imaging the distribution of time-varying magnetic flux density (hereinafter, magnetic image) was investigated. Although deep neural networks require training on large amounts of data to achieve high performance, it is not easy to obtain large amounts of useful training data from many magnetic image-based NDT sites. So, we explored a way to improve the defect detection accuracy even with a limited amount of training data by mapping the widely scattered defect information into a specific region. In this article, a deep neural network for magnetic image-based NDT was trained using transformed images in which the alternating current (AC) components of the magnetic image signal were preserved and the direct current (DC) offset values were matched to a single reference value. Here, the defect information is mainly contained in the AC components. Experiments demonstrated that the deep neural network trained using transformed images significantly improved defect detection accuracy compared to the conventional deep neural network trained on images without transformation.