In an era of tight budgets, cities, provinces and states are always looking for ways to save money.
One way is to reduce routine inspections of infrastructure, including bridges.
But that has caused some to worry about the safety of North American bridges, especially since the 2007 collapse of a bridge over the Mississippi River in Minnesota. That tragedy killed 13 people.
The investigation that followed attributed the collapse to defects that safety inspectors missed. That, along with improvements of sensor technology and machine learning, has spurred researchers in their quest for quicker, cheaper and more reliable ways to evaluate the condition of bridge decks.
With the year not much more than half over, there have already been announcements of two new systems.
One of the systems is mounted on a small skid-steer robot. The other has the equipment mounted in a cart that can be towed along a roadway. But both make extensive use of machine learning technology to do the work.
Inspecting a bridge used to involve drilling through the road surface to check for signs of delamination and corrosion of rebar. Ground penetrating radar, or GPR, has simplified the work since the 1980s although sending teams of inspectors out to check bridges is still expensive and can require extended road closures. The upshot of all this is that many North American bridges are overdue for inspection.
It is estimated in the United States several thousand bridges fall into that category. In Canada, there might be several hundred.
A team at the University of Nevada, Reno, has built what they call the first fully autonomous robotic bridge inspector. It is made up of a GPR unit and electrical resistivity sensors to spot any corrosion of steel or deteriorating concrete.
Surface cracks are detected using an onboard camera. All this is mounted on a stubby-looking robot small enough that the whole system can shuttle back-and-forth along the side of the road without getting in the way of passing traffic.
A machine-learning algorithm converts the data in real time to produce a colour-coated map of the bridge with areas of weakness highlighted. The results are sent to human inspectors who can keep tabs on the robot as it goes about its business.
The system has been tested on four road bridges in Nevada, Montana, New Hampshire and Maine. In each case, the system proved to be faster and more accurate than human inspectors.
Spencer Gibb, a member of the research team, says the robot "takes the same amount of time to physically scan the bridge as a human inspector, but it processes the data in minutes instead of hours."
Meanwhile, at the University of Nebraska-Lincoln, Jinying Zhu and her team have devised an early-warning system for bridges that is based upon acoustics.
Her system is designed to produce a much faster and more accurate alternative to a conventional method of identifying delamination.
Delamination is a gradual separation of concrete layers that can subtly compromise the structural integrity of a bridge. Left unrepaired, it can lead to major problems.
Zhu’s system looks deceptively simple. It consists of a push-cart that drags several strings of brass balls along the concrete.
Those balls produce different acoustic frequencies depending on whether they’re being dragged over delaminated or pristine concrete.
Small, cheap microphones attached to the cart record the sound sending it to a laptop that collects and processes the signals.
While this is going on a specialized GPS is tracking the cart’s location down to a centimetre.
Correlating acoustic signals with the cart’s position, the system software can automatically generate a colour-coded map that shows the location and dimensions of the bridge’s defects.
Zhu says that bridge inspectors have traditionally looked for delamination by dragging chains across a bridge deck, listening for hollow spots and marking them with chalk or paint.
But that’s "a very slow process, and it’s subjective," Zhu says.
There is also something called and impact-echo test that involves tapping concrete with a steel ball and recording the acoustic signatures with a contact sensor.
That started Zhu wondering if she could meld the relative efficiency of the chain-drag approach with the accuracy of the impact-echo method.
After writing a machine learning algorithm that would help differentiate between the acoustic signals of delaminated and intact concrete, Zhu used microphones to record the sound generated by nearly a dozen different types of chain.
But no matter what she tried, the delamination-specific signals were being lost amid a sea of other acoustic waves that arose from the links clattering together.
That led her to replace chain links with metallic balls connected by nylon string. As she had hoped, the ball-and-string combination enhanced the desired signals while reducing the background noise.
Zhu and her colleagues have since tested the system on several bridges in Nebraska including two on a busy interstate highway. She found it takes her team about 20 minutes to assess a stretch of bridge that would have needed roughly two hours using the chain-drag method.
Next up will be more testing and adaptation that could improve the system’s speed.
"If we have a trailer behind a pickup or minivan, we can mount the system on the trailer," she says. "And even driving on a bridge, even though it might be at a slow speed, is still very fast compared to the conventional methods."
Central to both systems — one developed in Nevada and the other in Nebraska — are machine-learning algorithms.
Machine learning is a subset of artificial intelligence and gives computers the ability to learn without being explicitly programmed. It evolved from the study of pattern recognition. Thus, it can learn when it spots a pattern, even though it has not been programmed to spot that particular pattern.