Revealing noise: International research group with GRS participation uses artificial intelligence to detect damage in the reactor core at an early stage
The reactor core is the heart of a nuclear power plant. This is where energy is released by the fission of atomic nuclei. The prerequisite for this is that a neutron hits a fissile atomic nucleus of the nuclear fuel. The more neutrons there are in the reactor core, the higher the number of nuclear fissions and thus the reactor power.Since a great deal of energy is released in the reactor core during operation, many safety precautions are required. These include the measurement of the so-called neutron flux. This provides information about the number of neutrons (per cm2 and second) in the reactor core and thus about the current power of the reactor. In German pressurised water reactors, there are about 60 installed detectors permanently installed inside and outside the reactor pressure vessel that measure the neutron flux.
Scientists from Europe, the USA and Japan are currently developing new, intelligent methods in the CORTEX project with which conclusions can be drawn from the neutron flux measurement data about safety-relevant processes in the reactor core. Being part of the EU’s Horizon2020 research programme, the project is funded by the European Commission to the amount of around 5.5 million euros over a period of four years.
Neutron flux noise as information source
The measured values of the different neutron flux detectors are "noisy", i.e. they fluctuate around an average value with different amplitudes and frequencies. These fluctuations have different causes. These include, for example, vibrations of components in the reactor core caused by the coolant flow. An increase of the vibrations and a corresponding change of the neutron flux noise can also occur when certain defects occur. Therefore, the reactor's protection system automatically switches it off as soon as the vibration amplitudes become too large.
In the project, the researchers analyse the amplitudes and frequencies of the noise and derive information from this for the safety assessment of the reactor core. For this purpose, they develop various methods that can read and evaluate this noise signal. The methods are intended to provide early indications of an occurrence of possible disturbances and their location during the operation of reactors. This allows the plant operator to take appropriate countermeasures before any major damage can occur. For reactor safety research, the evaluation of neutron flux noise offers the possibility to develop a better physical understanding of its causes and possible influencing factors.
Training of neural networks
The researchers are using methods from the field of artificial intelligence to set up the new analysis system. Since the start of the project in autumn 2017, artificial neural networks have been developed. These networks are currently being trained to recognise certain patterns in the neutron flux noise. These are combinations of vibration frequencies and amplitudes, which - similar to a fingerprint - are specific to certain causes, such as a damaged fuel assembly.
The data used for this training, also known as machine learning, are currently derived from simulations. Using special software, the researchers simulate certain anomalies in the reactor core (e.g. fluctuations in coolant temperature, vibration of fuel assemblies) and the associated changes in the amplitudes and frequencies of neutron flux noise.
In a further step, the neural networks will be tested on the basis of data from two experimental reactors at Dresden Technical University and Lausanne University in Switzerland. The aim is to fundamentally validate the new method. To this end, the researchers will conduct numerous experiments in 2019.
In the final phase of the project, the efficiency of the new method will be demonstrated by applying it in various power reactors in Europe and the USA. For this purpose, the system will be trained on the specific patterns of the corresponding reactor types.
Tasks of GRS in CORTEX
In the project, GRS carries out neutron-physical simulations. Among other things, it uses the GRS-developed simulation code QUABOX-CUBBOX to calculate the neutron flux in the reactor core.
In addition, GRS supervises the work on the simulation of so-called fluid-structure interactions. This involves determining to what extent individual core components are influenced by the flow of the coolant and, for example, set in motion, and to what extent the movement of structures changes the behaviour of the flow. Since this interaction also influences neutron flux and neutron flux noise, an understanding of these relationships provides important information on possible mechanical causes of noise.
In order to train and apply the newly developed neural networks to special European or American reactors, GRS also coordinates the data exchange between the participating research institutions.
Find out more
Detailed information about CORTEX is provided by the project partners on the CORTEX website. In addition to the technical publications of the project, reports on project meetings and videos are also available there. A final documentation of the results will follow the completion of the project in autumn 2021.