How AI Spots Problems in Nuclear Plants That Other Systems MissMay 3, 2017
Regular safety checks are as critical to nuclear power plants as doctors’ check-ups are for people. For power plants, regular inspections can spot cracks and other problems before they become serious or even cause an accident.
But detecting cracks in nuclear power plants isn’t easy. Reactors are underwater, so inspectors can’t examine them directly. Instead, power station staff scrutinize hours of video from inspection cameras to spot cracks in the plant’s metallic surfaces. These cracks can be dangerous, discharging radioactive material into the water or air.
Mohammad Jahanshahi, a civil engineering professor at Purdue University, has a better way. He’s automating crack detection in nuclear plants with GPU-accelerated deep learning and machine learning. He’ll talk how how he’s automating inspections of nuclear power plants and other infrastructure at the GPU Technology Conference, May 8-11, in Silicon Valley.
“Even a tiny crack in a nuclear power plant can cause radioactive materials to leak,” said Jahanshahi. “It can propagate and cause a nuclear accident.”
Cracks are also expensive. Damages reached $700 million in a 2010 accident at the Vermont Yankee Nuclear Power Plant, Jahanshahi said, after deteriorating underground pipes leaked radioactive tritium into groundwater. A 1996 accident caused by a leaking valve at the Millstone nuclear power station in Connecticut cost $254 million, he added.
Nuclear Plants Growing Old
Jahanshahi’s advance comes at a time when the world’s nuclear energy facilities are showing their age. Nearly 15 percent of plants worldwide are operating beyond their predicted lifespan of 40 years. In the U.S., more than a third are, according to the World Nuclear Industry Status Report. Several nations, including the U.S., are licensing plants for up to 60 years.
As nuclear stations age, their components become more vulnerable to cracks and other problems caused by heat, pressure and corrosive chemicals. In the past decade alone, at least a dozen power plants worldwide reported cracking problems.
“One reason plants are deteriorating is insufficient inspections,” said Jahanshahi, who published the results of his research in a recent issue of the journal Computer-Aided Civil and Infrastructure Engineering.
An Ounce of Prevention
The automated system, which Jahanshahi developed with Purdue doctoral student Fu-Chen Chen, would make it easier to inspect facilities before problems worsen, he said.
“Structures are like humans. If you spot disease symptoms early, you can prevent the person from getting sick,” Jahanshahi said.
Jahanshahi and Chen aren’t the first to try to streamline the tedious manual process of spotting cracks in nuclear reactors’ steel surfaces. Other methods, which are designed to examine individual frames within the inspection videos, often miss tiny cracks, which are hard to tell apart from other anomalies like welds and scratches.
AI Detects Cracks in Nuclear Power Plants
The Purdue system – called CRAQ, for crack recognition and quantification – avoids this pitfall by fusing information from multiple video frames to find changes in the texture of the steel that might indicate cracks. Video makes it possible to see cracks under different lighting conditions and at different angles.
The researchers developed their initial system using machine learning techniques, but they are now creating deep learning algorithms to improve accuracy. The team is training its algorithms of several thousand frames of inspection videos using the CUDA parallel computing platform, the Pascal architecture-based NVIDIA TITAN X and GeForce GTX 1070 GPUs and cuDNN.
Jahanshahi hopes the deep learning approach will improve the state of the U.S. infrastructure, which recently received a disappointing D+ grade from the American Society of Civil Engineers.
“With the increasing computational capabilities of computers with GPUs, we can now use computer vision, image processing and deep learning to tackle this problem,” he said.