Safer Bridges, Through Smarter AI
Stevens co-develops method that integrates relevant data from multiple sources, learning and analyzing with more predictive power
In 2007, an interstate bridge in Minneapolis collapsed into the Mississippi River, killing 13 and injuring more than 140.
In 2013, amid growing concerns about its deteriorating condition, New York’s 3-mile-long Tappan Zee Bridge spanning the Hudson River was fast-tracked for replacement — at a cost of $4 billion.
Isolated incidents? Maybe not: nearly 80,000 U.S. bridges should be replaced, according to the American Road & Transportation Builders Association’s latest assessment.
“Predicting bridge deterioration is challenging,” explains Stevens civil engineering professor Kaijian Liu, “because the deterioration of bridge conditions is affected by a multitude of factors.”
But many important data sources, such as local inspection reports, weather data and traffic data, are not used to assess bridge conditions and deterioration.
Now Liu, in collaboration with professor Nora El-Gohary from the University of Illinois at Urbana-Champaign, has developed a better idea: an AI that not only integrates multiple data streams related to bridge health, but analyzes them in an integrated manner using the new tools of machine learning.
More data, analyzed more deeply
One of the problems with current U.S. bridge assessment and maintenance is that it mostly relies on a single data point: the National Bridge Inventory (NBI), an annual Department of Transportation-mandated evaluation that classifies bridges into broad condition categories.
“In our view, NBI data are certainly important but themselves alone are not sufficient to accurately predict dangerous conditions or failures,” explains Liu. “Because NBI data do not capture information about bridge deficiencies and maintenance actions, which are detrimental to bridge conditions and deterioration.”
“Inspection reports contain a wealth of detailed information about the types, quantities, severities and causes of bridge deficiencies, as well as the types of performed maintenance and the applied maintenance materials.”
To improve assessments and predictions, Liu and El-Gohary drew in additional, multi-year data from sources including report notes from the Washington Department of Transportation and bridge traffic data including not only automobile traffic but also traffic from various classes of light and heavy trucks.
They also added publicly available data on temperature, precipitation and other weather conditions at bridge locations to the mix.
Then Liu and El-Gohary created and trained a new AI — of a type known as a recurrent neural network, or RNN — to learn about bridges’ evolving conditions over time, correlate that information with the multilayered data, and create predicted future condition assessments ranging from 1 (imminent failure predicted) to 9 (excellent).
During an early test of the new model’s predictive power, the duo assessed more than 2,600 bridges in Washington state. The result? The AI accurately flagged potential future issues in a bridge’s condition about 90% of the time, which was about 15% to 20% more accurately than existing methods would have done.
The method is also particularly accurate at forecasting the future condition of poor-condition (often aging) bridges, note the researchers.
“This is critical,” explains Liu. “Because we have so many fewer total cases of these ‘outlier’ bridges in the nation to study and train our prediction models on than average-condition bridges, the predictions about them are typically not as sound. But here they were very strong."
“And that’s important, because bridges trending toward the poorer end of the condition spectrum are obviously more likely to fail. You need to understand their deterioration more urgently than you might need to understand, for example, a newly built bridge in excellent condition.”
The pair found that their inclusion of state-level bridge deficiency and maintenance data of the type found in inspection reports, as well as of traffic and weather data, was critical in improving the performance of the model’s predictions.
“Bridges with similar geometric, structural and construction characteristics could have distinctive deterioration patterns, depending on the specific deficiency conditions of a bridge and the kind of maintenance it received — in addition to its ambient traffic and weather,” concludes Liu. “The NBI ratings alone do not sufficiently capture and differentiate the deterioration patterns of seemingly similar bridges.”
“This integrated data, from multiple sources, especially the inspection reports, make a real difference.”
Liu and El-Gohary’s research was reported in the Journal of Computing in Civil Engineering.