Electromagnetic Non-Destructive Testing of the Graphite Core of Advanced Gas-Cooled Reactors.

Abstract

The UK's advanced gas cooled reactors use graphite as a moderator and structural component. The graphite core degrades over time, and this is a concern to both controlling reactivity and the safety of the core. To inspect the graphite core, inductance spectroscopy is used as the soft-field imaging modality. This thesis considers the reconstruction of the conductivity depth profile and the detection of subsurface cracks. Within the thesis, multiple iterative inversion algorithms are compared for reconstructing the conductivity depth profile; of the iterative algorithms, second order methods are shown to be the best performing in different cases considering the prior estimate and noise level. Two machine learning algorithms are compared for direct inversion: multivariable polynomial regression and a convolutional neural network. The results show that the two algorithms have a comparable performance. A finite element model is used to generate data to train machine learning algorithms and in the iterative inversion of eddy current data. The model must accurately represent the physical experiment; therefore, there must be a rigorous calibration procedure. A constrained optimisation algorithm is presented for calibration and inversion of the eddy current data. The mean signal-noise ratio after calibration was 29.47 dB for a 3D coil model and 28.96 dB for the filament; prior to optimisation, the signal-noise ratio was 1.48 dB and 4.98 dB, respectively. In the inversion of graphite data, it is shown that the iterative algorithm generally does not improve on the prior estimate because the residual norm is below the discrepancy, causing steps to be rejected. The restrictions on the step could be relaxed but this increases susceptibility to modelling error. Instead, improvements can be made by increasing modelling accuracy. Feature extraction and classification algorithms are investigated for detection of subsurface notches. The best balanced accuracy achieved on synthetic test data was 71.45% using principle component analysis and a neural network. It is not clear how this compares with existing techniques.

Date of Award	1 Aug 2023
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Paul Wright (Supervisor) & Anthony Peyton (Supervisor)