PSI - Issue 70

Anjireddy Mummadi et al. / Procedia Structural Integrity 70 (2025) 417–423

418

1. Introduction Due to the advantages of bolted joints such as easy assembly, versatility, and durability in extreme temperatures, gained popularity in steel and metal structures. Bolted joints in steel structures play critical role in the safety of in steel infrastructures like buildings, bridges etc. However, bolt loosening poses the risk to integrity of the structure demanding careful monitoring and maintenance. Bolts are tended to loosen due to dynamic loads, temperature stresses, and improper installation. Detection of bolt loosening is essential and play crucial role in ensuring the safety and also to increase the intended life of steel infrastructure, as highlighted by Zahir M. et al. (2024). Sensor based methods involve deploying ultrasonic sensors, piezoelectric sensors and strain gauges to assess tension loads directly or by ultrasonic attenuation. Authors like Xu et al. (2019) researched on the challenges associated with sensor-based approaches such as installation, maintenance, and their sensitivity to environmental factors in bolt loosening detection. Alternatively, Xie et al. (2023) explored non-destructive methods, such as digital image correlation, to assess loosened bolts in wind turbines, emphasizing the limitations of sensor-based techniques, including restricted coverage and difficulties in data interpretation. Recent work by Li et al. (2023) showed how well deep learning algorithms worked in combining the methods for image processing like Hough Line transformation to increase the precision of bolt loosening detection. HLT and deep learning-based R CNN models were utilized to identify the corroded and loosened bolt edges in steel joints. By using deep learning techniques for bolt loosening detection, researchers have built on these advancements (Alapati M. et al., 2025; Burdette, 2022; Cha et al., 2016). Their results demonstrate the benefits of these approaches, such as increased cost effectiveness, reduced human involvement, and resilience to environmental factors. Liu et al. (2021) developed a vision based YOLOv 10-s deep learning model with a two-step segmentation method for high-precision bolt loosening detection in real-world environments. Similarly smith et al., (2014) studied the use of the YOLOv5 model to measure bolt loosening by calculating rotation angles. Even in extreme environmental conditions, these approaches have shown to be quite accurate and reliable. 2. Deep Learning Techniques in Object Detection Object detection plays an important place in computer vision and deep learning as it is much more accurate and effective. Conventional methods sometimes rely on comprehensive research strategies, which may be time consuming. Elective search improved this by breaking down images into meaningful regions and generating around 2,000 possible object locations, or "proposals," per image. These methods are studied using CNN-based models. Gradually, major developments like Fast R-CNN, Faster R-CNN, and Mask R-CNN have notably increased both the accuracy and speed of object detection methods. 2.1. Fast R CNN By resolving computational inefficiencies, Girshick (2015) developed Fast R CNN, which improved Region-based Convolutional Neural Networks (R CNN). Fast R CNN combines feature extraction and classification into a single framework rather than handling each region proposal independently. The model uses a Region of Interest (RoI) pooling layer to create fixed-size feature boxes after using a shared convolutional network to extract features. Following that, they are processed through fully connected layers for bounding box regression and classification. This lowers computation costs and greatly improves detection speed. 2.2. Faster R CNN Ren et al. (2015) optimised object identification efficiency by creating the Faster R CNN, which includes a Region Proposal Network (RPN). This method develops high-quality region recommendations using convolutional feature maps, eliminating the need for additional techniques. The strategy dramatically improves processing efficiency by combining convolutional layers from the RPN and detection network. Furthermore, Ren et al. (2015) improved the