Weld crack detection and quantification using laser thermography and deep learning algorithms

Abstract

Welding is used in many infrastructures because it has high usability, convenience, and excellent performance compared to other joining methods. However, the weld is susceptible to repeated or accumulated loads, resulting in cracks. Early diagnosis is required before cracks become larger. There are various non-destructive methods of inspecting weld cracks, but they require a lot of human resources or are less accurate. Therefore, a new inspection method that overcomes these problems was developed in this study. The objective of this study is to detect and quantify weld cracks using laser thermography. The proposed method consists of visualization of laser heating, augmentation of crack data, and detection and quantification of cracks. In the laboratory test, the proposed method has an AP of 95.1%, length error of 1.049 mm, and width error of 0.144 mm. In Yeongjong grand bridge field test, all 6 weld cracks are detected, and length error is 0.9032 mm, and width error is 0.0640 mm.