Webex: believe everyone is no stranger to the plant, but how much understanding of the daily maintenance and testing of a nuclear reactor?Lei feng net understands, for nuclear power plant safety checks on a regular basis as people to the hospital regularly examined, is very necessary.The traditional manual testing efficiency is low, not only in some small steel surface crack can be difficult to find.The gap once the testers, will leaks of radioactive substances into the water or air, to cause human life risk.In the era of AI, therefore, urgent need to find new ways to replace the traditional detection.
For nuclear power plant, regular check to before have accident or problem become serious, find cracks or find other problems.Crack detection in power plant, however, is not so easy, because the nuclear reactor are underwater, testers can't directly to the test, only by examining the video camera frame by frame to carefully check the metal surface.
Mohammad Jahanshahi (Purdue University, similarly hereinafter), Purdue University, professor of the department of civil engineering.He put forward a better way, deep study and use of GPU accelerationMachine learningTo realize automatic plant crack detection.On May 8 to 11, 2017 GTC held in silicon valley, he'll talk about how to realize the automation of power plant and other infrastructure.Lei feng network (public) : lei feng network) will be on the scene at the first time for the conference.
"In a nuclear power plant, even a small crack causes radioactive leak," Jahanshahi said, "it can be spread and lead to a nuclear accident."Cracks in the cost of is very big also.Worsening after radioactive tritium underground pipeline leak into the groundwater, Jahanshahi said in Vermont Yankee Nuclear Power Plant (Vermont Yankee Nuclear Power Plant), 2010 accident resulting in loss of up to $700 million.He also added that the 1996 Connecticut Milestone plant due to valve leakage accident, spent $254 million.
The ageing of the nuclear power plant
Jahanshahi foreseen in this moment arrived.According to the world nuclear industry status report, nearly 15% of the global nuclear energy equipment running time is more than their default 40 years of life, in the United States, more than a third of the equipment.Several countries, including the United States authorized power plant life to 60 years.
With the ageing of the nuclear power plant, their parts become more susceptible to cracks caused by heat, pressure and corrosive chemicals or other problems.In just the past decade, the world at least a dozen plant reported the crack problem.
Jahanshahi says, one reason is that the plant problems lack of detection.He is in a recent issue of journal of computer-aided civil and infrastructure engineering in published the results of his research.
The problem of too much, too little to prevent
Jahanshahi with purdue university doctoral student Fu - Chen Chen, a cooperative development of automation system, will detect equipment problem before problems become worse.
Buildings, like people, if you find an early "symptoms", can avoid "sick".
Actually Jahanshahi and Chen is not the first to eat crab, previously there are other methods to test the crack.But like other design is used to check the detecting method of the single frame in the video, often miss some subtle gap, but also it is difficult to distinguish some anomalies, such as solder joints and scratches.
With AI detect cracks in nuclear power plant
Purdue university system called CRAQ (crack recognition and quantification), namely fracture identification and quantification of information fusion of multiple video frame to find texture change in the steel surface cracks may be produced.The system can be seen under different lighting conditions and different angles of gaps in the video.
Researchers using machine learning technique developed their original system, and now they're building deep learning algorithm to improve the accuracy.Teams use CUDA parallel computing platform, with thousands of video frame detection algorithm to train it.Pascal architecture is based on the nvidia Titan X and GeForce GTX 1070 GPU, and cuDNN.
Jahanshahi hope deep learning ways to improve America's infrastructure.He said: "with computer GPU computing ability promotion, we can make use of computer vision, image processing, and deep learning to solve the problem."