I believe that everyone is familiar with nuclear power plants, but how much do you know about the routine maintenance and testing of nuclear reactors? Lei Feng.com understands that regular safety inspections of nuclear power plants are just as necessary for people to go to the hospital for physical examinations on a regular basis. Traditional manual inspections are not only inefficient, but also difficult to find for some fine steel surface gaps. Once these gaps are missed, they will leak radioactive substances into water or air, posing a life-threatening hazard to humans. Therefore, in the AI ​​era, there is an urgent need to find new ways to replace traditional detection.
For nuclear power plants, regular inspections are designed to find cracks or other problems before they cause an accident or the problem becomes serious. However, it is not so easy to detect cracks in nuclear power plants because the nuclear reactors are all underwater, and the inspectors cannot directly detect them. The metal surface can only be inspected frame by frame by the video captured by the inspection camera.
Mohammad Jahanshahi is a professor of civil engineering at Purdue University (the same below). He proposed a better way to use GPU to accelerate deep learning and machine learning to achieve automatic detection of cracks in nuclear power plants. On May 8-11, at GTC 2017 in Silicon Valley, he will talk about how he automates the inspection of nuclear power plants and other infrastructure. Lei Feng Network (Public No.: Lei Feng) will also arrive at the scene to report on the conference.
“In a nuclear power plant, even a small crack can cause radioactive material to leak,†Jahanshahi said. “It can spread and cause a nuclear accident.†The cost of the crack is also high. After the deteriorating underground pipeline leaked radioactive cesium into the groundwater, Jahanshahi said that at the Vermont Yankee Nuclear Power Plant, an accident in 2010 caused as much as $700 million in damage. He also added that the 1996 Milestone nuclear power plant in Connecticut suffered an accident caused by valve leakage, costing $ 254 million.
Aging of nuclear power plants
The foresight of Jahanshahi came at this moment. According to the world's nuclear industry status report, nearly 15% of global nuclear energy equipment runs longer than their default 40-year lifespan. In the United States, more than one-third of equipment is the same. Several countries, including the United States, have authorized power plants to last for 60 years.
As nuclear power plants age, their components become more susceptible to cracks or other problems caused by heat, pressure and corrosive chemicals. In the past decade alone, at least a dozen nuclear power plants around the world have reported crack problems.
Jahanshahi said that one of the reasons for the problem with the power station is the lack of detection. He published his findings in a recent issue of Computer-Aided Civil and Infrastructure Engineering.
Too many problems, too little prevention
The automation system developed by Jahanshahi in collaboration with Purdue University's Ph.D. student Fu-Chen Chen will detect equipment problems before the problem gets worse.
Buildings are like people. If you find "symptoms" early, you can avoid "illness."
In fact, Jahanshahi and Chen were not the first to eat crabs, and there were other ways to detect cracks. But like other methods designed to check a single frame in a video, you often miss out on some tiny gaps, and it's hard to distinguish between anomalies such as solder joints and scratches.
Using AI to detect cracks in nuclear power plants
Purdue's system is called CRAQ (crack recognition and quantification), which is the identification and quantification of cracks. The fusion information in multiple video frames is used to find the texture changes that may occur on the steel surface. This system can see cracks in the video under different lighting conditions and different angles.
Researchers used machine learning techniques to develop their original systems, and now they are building deep learning algorithms to improve accuracy. The team used the CUDA parallel computing platform to train its algorithms with thousands of frames of video. The Pascal architecture is based on the NVIDIA Titan X and GeForce GTX 1070 GPUs and cuDNN.
Jahanshahi hopes that deep learning methods can improve the state of infrastructure in the United States. He said: "As computer GPU computing power increases, we can use computer vision, image processing and deep learning to solve this problem."
Ceramic Housing
Ceramic Housing
YANGZHOU POSITIONING TECH CO., LTD. , https://www.pst-thyristor.com