Using machine learning to detect bridge health
1
There are more than 620,000 bridges in the United States, and more than 42,000, or 6.8% are considered structurally deficient and in “poor condition,” according to a 2025 Report Card from America’s Infrastructure from the American Society of Civil Engineers.
Inspecting these bridges can be time-consuming and expensive, as they require specialty equipment and pose a safety challenge for those conducting the inspection.
In an effort to streamline the process and gather accurate data at a faster rate about the bridge health, PhD student Chengwei Wang is conducting research in the Structural Health Monitoring Laboratory under the direction of UIC Distinguished Professor Farhad Ansari. Wang will be using fiber optic sensors and a generalized machine learning approach to detect and quantify cracks.
“We know there are a lot of cracks on bridges. That’s one of the most common issues for bridges, especially for concrete bridges. When a crack gets too wide and propagates too deep, that can cause problems,” Wang said. “What we are trying to do is detect the crack and quantify the crack using sensors, which is going to be more efficient than a visual inspection that is subjective. Sometimes, some cracks were simply overlooked during inspection.”
To accomplish their goal, the researchers use a single distributed optical fiber sensor to detect the locations and quantify their size on entire sections of bridges. This information will allow real-time and archival data, as well as assess the condition of the bridges and understand how they develop and grow, and what threat they pose. They also use sensors on a number of laboratory bridge models, including cable-stayed and suspension structures, in the lab at UIC.
The sensors measure the damage at every location along the length of bridges. When there is a crack in the bridge, the sensor is elongated at that location and signifies its location.
The team is taking the research further with a machine learning algorithm that will save time by detecting the cracks.
“We came up with a machine learning method, so we don’t have to use our eyes to check the collected data. Without it, we have to pull the data, look through it, and say, ‘Okay, here is a strain jump, and here is a strain change, and that’s a possible crack location.’ The machine learning method detects strain changes automatically,” Wang said.
What makes this research even more unique is that the team trained the algorithm so it can be used in a general way on different kinds of bridges, rather than having to train it for individual bridges.
“With a generalized model, we don’t have to investigate a bridge only using our eyes anymore. We can send it to a specific project and let the algorithm do the work. One training can fit all,” Wang said.
This research has the potential to detect cracks before they grow, allowing municipalities to address the issues before the bridges fall into disrepair or become dangerous to the millions of motorists who use the bridges daily.