PSI - Issue 64

Jie Wang et al. / Procedia Structural Integrity 64 (2024) 1326–1333 Author name / Structural Integrity Procedia 00 (2019) 000–000

1327

2

1. INTRODUCTION Steel structures have excellent mechanical properties and are easy to assemble, making them widely used in large span buildings and bridges. However, steel box girders on bridges and crane beams in industrial plants are subjected to long-term dynamic loads, which may result in initiation of fatigue cracks. As the cracks propagate, they significantly endanger safety of the components and even the entire structure. Early detection, continuous monitoring and protective measures are crucial in ensuring structural integrity and safety. One method for detecting fatigue cracks is the regular manual inspection, but this approach is inefficient and prone to subjective errors. In construction of large-span bridges, structural health monitoring (SHM) systems are often installed. They utilize sensors such as fiber optic sensors, accelerometers, and vibration sensors to collect structural responses. Modal analysis or reliability analysis is conducted to assess structural damage or to predict the service life. However, it is generally complex for installation and operation. Since effects of structural damage are not significant at early stages, SHM is typically suited for detecting severe damage. Some other non-destructive detecting methods, such as ultrasonic by Halabe et al. (2001), thermal infrared by Eonid et al. (2021), magnetic particle, or millimeter wave, offer more reliable results than manual inspections. However, the higher costs and susceptibility to environment limit the widespread application. With the development of cameras and computer vision algorithms, many image-based detection methods have been proposed. Binarization processing, which has been applied to creep experiments by Masahiro et al. (1992) and cracks detection in steel billets by Hsu et al. (2016), often produces false results, which requires additional manual interpretation. Edge extraction operators like Sobel by Zhang et al. (2011) and Canny by Yeum et al. (2011) can highlight cracks in images, but they are significantly affected by factors such as algorithm parameters, lighting conditions, and image resolution. Machine learning and deep learning algorithms generally exhibit better robustness and higher accuracy. Alberto et al. (2016) designed color and texture descriptors for feature extraction and artificial neural networks (ANN) to detect corrosion and cracks on ships. The method struggled to effectively differentiate damage from complex backgrounds such as scratches or handwritten marks. Han et al. (2022) employed SLIC superpixel segmentation for downsampling input images, followed by the YOLOv3 model for selecting crack regions and multiple DeepLabv3+ models for joint decision-making of the crack shape. These methods have shown feasibility in detection of visible cracks with large scale in images. However, most fatigue cracks are invisible by naked eyes in static images. Since cracks open and close under dynamic loads, small-scale fatigue cracks can be detected by tracking the movement of the steel plate surface. One of the most direct ways is to use digital image correlation (DIC) techniques to obtain surface displacement fields. Al-Salih et al. (2021) proposed that location of crack tips could be determined by calculating convergence values of displacement. However, DIC generally requires speckle, which directly affects accuracy of the displacement field. In Kong et al. (2018), a video-based detection method was put forward, using Shi Tomas corner points and KLT sparse optical flow tracking to obtain motion of feature points. Feature points with significantly different motion were extracted to detect fatigue cracks. Feasibility of the method depended on the quantity of feature points detected by Shi-Tomas. However, it may be not feasible for steel structures, which are generally characterized with weak texture. In this paper, a method for detecting small-scale fatigue cracks using multiple frames of images was proposed. Motion of dense feature points across multiple frames was tracked to assess the surface displacement. Detection of the crack was based on the discontinuity in the displacement field around the crack. 2. Methodology This section outlines the general workflow of the proposed crack detection method. Only a consumer-grade digital camera was required, with no need for special treatment on the structure or environment. The method assumes: (1) the captured area approximates a plane, and (2) the camera is approximately perpendicular to the captured area and remains completely stationary. The approach is consisted of three steps: (1) dense matching of feature points, (2) assessment of the displacement field on the steel surface, and (3) detection of the crack.