According to a new study published in Seismological Research Letters, using seismic and acoustic data recorded 50 meters away from the reactor, it is possible to predict whether the reactor is on or off with up to 98 percent accuracy.
By applying several machine learning models to the data, researchers at Oak Ridge National Laboratory can predict when a reactor will be between turning on and off and estimate its power level with an accuracy rate of about 66 percent.
Chai Chengping, a geophysicist at Oak Ridge, the study's lead author, said the findings provide another tool for the international community to cooperate in verifying and monitoring the operation of nuclear reactors in a minimally invasive manner. "Nuclear reactors can be used for both benign and malignant activities. It would therefore be beneficial to the nuclear non-proliferation community to verify that nuclear reactors were operating as declared. ”
While seismic and acoustic data have long been used to monitor the structural properties of infrastructure such as earthquakes, buildings, and bridges, some researchers now use this data to look more closely at motion associated with industrial processes. In this case, Chai and his colleagues deployed seismic and acoustic sensors around oak ridge's high-throughput isotope reactor, a research reactor used to generate neutrons for physical, chemical, biological, engineering and materials science research.
The dynamic state of the reactor is a thermal process, with a cooling tower to dissipate heat. "We found that seismic acoustic sensors can record the mechanical characteristics of vibrating equipment such as fans and pumps in cooling towers, and their accuracy is enough to reveal operational problems," Chai said.
The researchers then compared some machine learning algorithms to find which one was best at estimating the reactor's power state from a particular seismic acoustic signal. The algorithms were trained to use only seismic data, acoustic data only, and both types of data and were collected for more than a year. They found that the combined data produced the best results.
Chai Chengping explains: "Seismic signals of different power levels present complex patterns that are difficult to analyze with traditional techniques. Machine learning methods are able to infer complex relationships between different reactor systems and their seismic-acoustic fingerprints and use them to predict power levels. ”
Chai and his colleagues found some interesting signals during their research, including the vibration of the noise pump when the reactor was off, and when the pump was replaced, the vibration disappeared.
Chai Chengping said linking seismic and acoustic features to different industrial activities and equipment is a long-term and challenging goal. For high-throughput isotope reactors, preliminary studies have shown that fans and pumps have different seismic acoustic fingerprints, and different fan speeds have their own unique characteristics.
"Some normal but less frequent activities, such as annual or occasional maintenance, need to be distinguished between seismic and acoustic data," Chai said. To better understand the relationship of these features to specific operations, we need to study the seismic and acoustic characteristics of instruments, as well as the background noise of various industrial facilities. ”