Harnessing data from benchmark testing for the development of spalling detection techniques using deep learning

Expediting post-disaster recovery is an essential step to enhance the rapidity and resiliency of infrastructure, especially bridges that act as nodes of transportation networks. Machine learning can reduce impeding factors such as reliance on human inspection by automating inspection processes. Damages in images obtained from unmanned aerial vehicles (UAVs) can be classified within minutes of certain disastrous events without human intervention. In this chapter, we propose convolutional neural network models to classify the extent of concrete damage (spalling) obtained from benchmark testing under seismic excitations. The image data for concrete bridge columns were obtained from the published database from the shaking table experiment of a quarter-scale two-span reinforced bridge conducted at the University of Nevada, Reno, in 2004. The image data are trained with multiple CNN models using various techniques such as transfer learning, data augmentation, with transfer learning, without data augmentation, and with a combination of data augmentation and transfer learning, to determine which model provides the most accurate results obtained with the ResNet50 model. As observed, the model underperformed due to the excessive number of augmentation features. Therefore the model is revised by training with lesser data augmentation features to identify which feature improves accuracy. As described throughout the chapter, three distinct testing/training scenarios were proposed to assess the model's efficacy. The ResNet 50 model with ImageNet weights for transfer learning and augmentation techniques (vertical and horizontal flips and shear range) yielded 100 percent accuracy during training and 97 percent during testing.