Damage classification and segmentation in extended shear tab connection using convolutional neural networks and transfer learning

The joint is one of the most critical parts of a building structure. In steel buildings extended shear tab (EST) connection is becoming an attractive alternative for light to moderate end shear connections due to its simplicity and economy. Detection and prediction of cracks in connections perform an important job in the maintenance of steel structures. Currently, the structural inspections of such joints are conducted by manually which is a tedious and expensive. It is essential to detect damages in joints in order to ensure structural safety. In this study a two-phase convolutional neural network (CNN) using transfer learning is presented for detection and segmentation of cracks and damages in component parts of an EST connection. Different pretrained networks using transfer learning, including AlexNet, GoogLeNet, ResNet-101 and VGG-16 are considered for crack and damage detection. The undamaged, cracked and damaged images of component parts of sixteen EST connections were generated through the finite-element simulation which were used to develop the CNN model. Segmentation and detection results show that VGG-16 model appears to give the best results with 100% precision and accuracy, followed by ResNet-101, and AlexNet, and finally GoogLeNet gives the least performance among the four methods selected. This study will be resourceful for quick reference for those who are working in structural health monitoring field.