In this data explosion epoch, data-driven structural health monitoring (SHM) and rapid damage assessment after natural hazards have become of great interest in civil engineering research. This report introduces deep-learning (DL) approaches and their application to structural engineering, such as post-disaster structural reconnaissance and vision-based SHM. Using DL in vision-based SHM is a relatively new research direction in civil engineering. As researchers begin to apply these concepts to structural engineering concerns, two critical issues remain to be addressed: (1) the lack of a uniform automated detection principle or framework based on domain knowledge; and (2) the lack of benchmark datasets with well-labeled large amounts of data.
To address the first issue, an automated and hierarchical framework has been proposed: the PHI-Net or Ø-Net for the PEER Hub Image-Net (https://doi.org/10.55461/PHIN01152018). This framework consists of eight basic benchmark detection tasks based on current domain knowledge and past reconnaissance experience. The second area of concern is based on the Ø-Net framework; a large number of structural images was collected, preprocessed, and labeled to form an open-source online large-scale multi-attribute image dataset, namely, the Ø-Net dataset. At the time of this writing, this dataset contains 36,413 images with multiple labels.
This report introduces herein three deep convolutional neuronal networks (CNN): VGG-16, VGG-19, and ResNet-50. The architecture design and network properties, etc., are described and discussed. For benchmarking purposes, a series of computer experiments are conducted. Multiple factors are considered in comparison studies under a fair setting of hyper-parameters and training approaches, i.e., using affine data augmentation (ADA) and transfer learning (TL). All experimental results are reported and discussed, which provide benchmark and reference values for future studies by other researchers developing new algorithms. These results reveal the great potential of using DL in vision-based SHM.
Finally, the first image-based challenge in structural engineering was held by the Pacific Earthquake Engineering Research (PEER) Center during the Fall of 2018. This challenge, designated as the Ø-Net Challenge, served as a pre-event prior to the open sourcing of the Ø-Net dataset and attracted worldwide attention and participation from researchers from around the globe.
Full List of PEER Reports: click here.
