Team of UB engineers earn first place at the 2021 International Competition for Structural Health Monitoring

Four different illustrations of the same building. Each is very abstract and appear to identify different deficiencies.

Modeling, developed by UB engineers, identify different deficiencies in buildings. 

By Peter Murphy

Published July 14, 2022

A team made up of recent University at Buffalo engineering alumnus Seyed Omid Sajedi and current PhD student Kareem Eltouny, advised by assistant professor in civil, structural and environmental engineering, Xiao Liang, won the 2021 International Project Competition for Structural Health Monitoring (IC-SHM). 

“We had to push the boundaries to achieve high-resolution, yet efficient, autonomous visual inspection. ”
Xiao Liang, Assistant Professor
Department of Civil, Structural and Environmental Engineering

According to IC-SHM, the challenge, organized by the Structures and Artificial Intelligence Lab at the University of Houston, tackles “the challenge [of structural health monitoring] by fostering and encouraging innovations in the structural health monitoring community.”

Effective structural health monitoring (SHM) can identify the location and severity of structural damage from different sources of information such as images and vibration records in a timely manner. Machine-learning models focused on SHM can provide decision-makers with valuable information regarding repair, maintenance and safety assessment of infrastructure.

Specifically, the 2021 competition was launched to advance the frontiers of computer vision-based civil infrastructure inspection and monitoring. Teams were presented with three projects - each having its own datasets and objectives. UB’s team participated in the second project: Computer vision-based post-earthquake inspections of buildings.

“This was a sensational experience. We had to push the boundaries to achieve high-resolution, yet efficient, autonomous visual inspection,” Liang says. “I believe that working together closely to solve the many problems and challenges was one of the main motivations we had for winning the competition.”

In this project, teams were given a dataset, QuakeCity, including a simulated high-definition unmanned aerial vehicle (UAV) captured images of buildings that have suffered earthquake-induced damage. The surface damage textures in the image are produced using physics-based graphics models. All the images and data associated with each project were generated using different types of modeling.

The objective was to identify the precise location of damage, components and corresponding damage states by analyzing the image data of some of the buildings affected by the earthquake. Teams had to develop models to analyze the image data as accurately as possible in order to identify deficiencies and their precise locations.

“There are five distinct tasks we had to solve. We were given a 1080p HD image and had to provide 1080p HD segmentation for each of those tasks,” Liang says. “One of the main challenges is that the model must operate on high-resolution images and provide high-resolution masks. Most state-of-the-art-deep learning models are designed for medium to low-resolution images and masks.”

Participants had to develop models to identify structural and non-structural components, including walls, columns, beams and windows, components damage states (light, moderate, severe or N/A), cracks, spall (broken, small pieces of stone or concrete) and exposed rebar.

Liang, Eltouny and Sajedi drew on previous experience to find success in this competition. Liang specializes in infrastructure systems and structural dynamics, Eltouny’s research focuses on data-driven structural health monitoring, and Sajedi worked to develop machine learning models for structural health monitoring while earning his PhD in structural engineering.

“The IC-SHM event is a unique opportunity for AI-based structural health monitoring research because independent teams can optimize their deep learning models and compete on benchmark datasets,” Liang says. “Our models are designed for both high precision and efficiency. We developed trainable resizers that can be attached to a deep segmentation model to provide high-resolution segmentation with a fractional increase in computational costs.”

IC-SHM received 53 submissions from 12 countries.