This Work Package deals with the analysis of plant data and neutron current and flux modelling, through advanced signal processing and machine learning techniques. In particular, signal processing methods include Fourier and Hilbert transformations, multi-resolution wavelet analysis, while machine learning focuses on deep learning and deep neural networks (DNNs). Specific targets include the detection of abnormal fluctuations & their classification, handling of scarcity of in-core instrumentation, learning in selected reactor(s) and transferring it to others adaptation of the developed techniques towards their application to real data.
Different scenarios and simulated data have been generated in the first year of CORTEX, both in:
– frequency:
- 2 types of perturbations
- many locations and sensors
- large amounts of data
– and in time:
- 1 type of perturbation
- 4 locations and tenths of sensors.
Moreover, some data from real reactors were recently generated and provided to the members of Work Package 3 for analysis.
An analysis of all types of the simulated data was performed for detection of perturbation type and location, using novel DNNs with/without pre-training. Very promising results were obtained and already presented/accepted at two IEEE conferences. Current efforts concentrate on processing and analysing both simulated and real data.
CORTEX work presented at the World Congress in Computational Intelligence
Last July, our researchers from the University of Lincoln (UoL) attended the IEEE World Congress in Computational Intelligence (WCCI), held in Rio De Janeiro. Under the International Joined Conference in Neural Network (IJCNN) track, Stefanos Kollias and Francesco Calivà presented one of their latest publications, “A Deep Learning Approach to Anomaly Detection in Nuclear Reactors”. The work was produced in collaboration with the Chalmers University of Technology team. Top world-wide researchers in Computational Intelligence gathered to discuss about the latest developments in the field, covering topics such as biological neural networks, arti ficial neural computation, theoretical and real-word applied evolutionary computations. In their paper, the CORTEX team presented a deep learning framework able to unfold the induced neutron noise up to the reactor core spatial resolution.
The noise represented the reactor response to an absorber of variable strength in the frequency domain and was simulated using CORE-SIM. Furthermore, the paper proposed a Deep Convolutional Denoising Auto-Encoder, which is capable of fi ltering disturbing noise out of the signal, as well as reconstructing missing part of the signals.
Francesco Calivà presenting the CORTEX work at WCCI 2018
CORTEX work accepted for presentation at the Symposium Series on Computational Intelligence
The latest joint research from CORTEX groups at the University of Lincoln, Paul Scherrer Institute and Chalmers University of Technology was accepted for publication at the IEEE Symposium Series on Computational Intelligence, happening next November in Bengaluru, India.
SSCI is a unique opportunity where expert researchers and professionals gather and discuss the latest advances in computational intelligence. The accepted paper, which was co-authored by Fabio De Sousa Ribeiro, Francesco Calivà, Georgios Leontidis, Stefanos Kollias (UoL), Antonios Mylonakis, Christophe Demazière (Chalmers), Dionysios Chionis and Abdelhamid Dokhane (PSI), presented the first step towards a framework able to analyse reactor core signals simulated in both the frequency and time domains.
The Deep Learning based platform proved able to unfold neutron noise induced by an absorber of variable strength in the frequency domain (which was modelled using CORE-SIM) up to the reactor core resolution, irrespective of the presence of disturbing noise at various signal-to-noise ratios. With respect to signals in the time domain (which were simulated by using SIMULATE-3K) the proposed framework was proved capable of detecting the perturbation cause (e.g. the vibration of a fuel assembly), within one second of perturbation occurrence.
The UoL staff members look forward to presenting their work to an audience of experts. Meanwhile, they are advancing their research on the analysis of more simulated scenarios, towards monitoring of real data from nuclear power plants.
By Stefanos Kollias