When you book a flight for the holidays, you may expect some turbulence in the air, but not while taxiing to and from the gate.
However, just like roads, airfield pavements are susceptible to weathering and damage from heavy loads. Over time, asphalt surfaces can develop cracks, leading not just to a bumpy ride but to loose fragments of pavement that may damage an aircraft’s turbine, propeller blades, jet engine, or tires.
“Cracked pavement at airfields requires prompt maintenance,” said Professor Andrew Yu. “Otherwise, runways may eventually require full removal and replacement, which can be up to a hundred times more expensive than crack filling.”
In 2022, the Tennessee Department of Transportation (TDOT) Aeronautics Division evaluated the pavement at 70 Tennessee airports. The report estimated that repairing or replacing all the damaged pavement would cost upwards of $400 million over four years. However, more than half of the issues TDOT identified could be resolved with surface treatments and crack sealing, which represented less than five percent of the projected repair budget.
TDOT currently relies on human inspectors to catch asphalt damage early enough to employ those less expensive repairs. Unfortunately, humans cannot manually inspect every foot of pavement at every airport. Instead, they measure the damage to a small area, then use that data to estimate the total extent and cost of repairs, which makes accurate scheduling and budgeting nearly impossible.
Yu, an expert in facility maintenance, is in the middle of a $200,000 TDOT-funded project to improve the efficiency of airfield inspections and maintenance scheduling by creating a machine learning (ML) program that can identify damaged asphalt from aerial drone photos of airports.
“The University of Tennessee has a strong collaborative relationship with TDOT on infrastructure maintenance projects,” Yu said. “This foundation of trust facilitates seamless, meaningful collaboration and enhances project continuity.”
Creating a Tool—and a Library
When the project began last year, Yu brought on his former student Zefeng Lyu (PhD ’23), who is now an assistant professor at Youngstown State University in Ohio.
“Dr. Lyu has already contributed significantly to this work,” said Yu. “His expertise in infrastructure maintenance and computation greatly benefit this project.”
In addition to identifying and measuring the length of new cracks, the ML program will need to recognize areas that were previously repaired with tar but now need more attention. Because tar repairs are darker than asphalt, the contrast between intact tar and new cracks is harder to differentiate on an aerial photo.
Another difficulty Yu and Lyu face is training the algorithm to recognize alligator cracks—cracks numerous enough that tar sealing is insufficient, so the entire area must be removed and repaved. Teaching the ML program to not only recognize large groups of cracks, but to accurately assess how much pavement around them needs to be replaced, requires training data that did not yet exist.
“As in most ML tasks, data quality is the most important factor, which directly impacts model efficiency,” Yu explained. “We have been collecting drone images of airfield pavement, manually annotating new cracks, previously repaired cracks and alligator cracks, and using that library to train the ML model to recognize those unique patterns in future pavement assessments.”
As their library of high-quality, annotated data grows, Yu and Lyu continuously compare their model’s recommendations with those from human inspectors. They hope that their model can soon be utilized in departments of transportation (DOTs) in other states or at the federal level.
The team also recently made their dataset publicly available on the IEEE DataPort, allowing researchers across the globe to work on similar infrastructure maintenance projects without annotating their own libraries first.
“We want other DOTs to benefit from this scarce resource,” Yu said. “I look forward to further collaborations with TDOT and other agencies in which we can extend ML technology to other areas of infrastructure sustainability and improvement.”
Contact
Izzie Gall (865-974-7203, egall4@utk.edu)