Conductivity profiling of complex objects using multi-frequency eddy current measurements and a neural network approach
H Tesfalem, A J Peyton, A D Fletcher, M Brown and J Hampton
In this paper, the authors employ multi-frequency eddy current data extracted from a modelled graphite brick making up the cores of the advanced gas-cooled reactors. This data is used to train a simple feed-forward neural network. The training data for this study was collected by simulating the previously measured physical properties of the graphite bricks and interpolating these data to get a range of datasets that allowed the neural network to be trained. During the training process, the authors employ a back-propagation approach, along with a Levenburg-Marquardt optimisation algorithm. Two simple neural network training approaches were tested: multi-layer and layer-by-layer training approaches. The multi-layer approach employs a five-layer neural network with a back-propagation method to train and estimate all conductivity unknowns at the same time, whereas the layer-by-layer training approach estimates the unknown conductivity profiles incrementally. The trained network was then tested using unseen simulated datasets. The results from this study show that the multi-layer approach gives a mean profile error of 7% of the tested cases, whereas the discrete training approach gives a mean profile error of 3% and tends to generalise the neural network solution when tested with some unseen simulated datasets. In addition, the training and solution time of the neural network was found to be significantly shorter compared with the conventional inversion algorithm that the authors studied previously.
