13 May 2021 in IEEE Sensors Journal
This paper discusses an artificial neural inversion approach for non-destructive testing of the graphite moderator bricks that make up the cores of Advanced Gas-Cooled Reactors (AGR). The study employed fully connected feedforward neural networks consisting of four hidden layers and trained with backpropagation approach using Levenberg-Marquardt optimisation algorithm. The approach is based on multi-frequency (MF) eddy current (EC) measurements, and combinations of simulated and measured datasets from a laboratory sample and one of the operating reactor core, along with different regularisation parameters were used to train and test the networks. Various types of artificially generated errors were added to the data during training procedures, which in turn allowed for error tolerance and improved the generalisation of the ANNs to unseen test datasets. First, the ANN was tested using unseen simulated data, followed by the measurements collected from laboratory sample and one of the operating reactor core. The first test from the unseen simulated data showed mean profile error ranging between 1.30% and 8.20%, whereas the profiles estimated from reactor core measurements showed mean profile error ranging between 1.84% and 17.80% when compared with the resistivity measurements from trepanned graphite sample taken out of the reactor core. Further comparison of the network outputs against the profiles estimated using traditional iterative inversion algorithm indicates reasonable agreement between the two approaches with the exception of one case, but the solution time for the ANN was found to be over three orders of magnitude faster than the iterative inversion algorithm.
1 January 2022 in NDT & E International
1 February 2021 in Insight - Non-Destructive Testing and Condition Monitoring
The radial depth profile of the electrical conductivity of the graphite channels in the UK's advanced gas-cooled reactors (AGRs) can be reconstructed and estimated by solving a non-linear optimisation problem using the mutual inductance spectra of a set of coils. This process is slow, as it requires many iterations of a forward solver. Alternatively, a data-driven approach can be used to provide an initial estimate for the optimisation algorithm, reducing the amount of time it takes to solve the ill-posed inverse problem. Two data-driven approaches are compared: multi-variable polynomial regression (MVPR) and a convolutional neural network (CNN). The training data are generated using a finite element (FE) model and superimposed on a noise floor in the interval [20, 60] dB of the weakest amplitude point in the corresponding spectrum. A total of 5000 simulated datasets are generated for training. The results on smoothed test data show that the two models have a comparable mean percentage error norm of 17.8% for the convolutional neural network and 17.3% for multivariable polynomial regression. A further 500 unsmoothed profiles are tested in order to assess the performance of each algorithm on conductivity distributions where the conductivity of each layer is independent of another. The performance of both algorithms is then assessed on reactor-type test data. The results show that the two data-driven algorithms have a comparable performance when estimating the electrical conductivity depth profile of a typical reactor-type distribution, as well as vast deviations. More generally, it is thought that data-driven approaches for depth profiling of some electromagnetic quantity have the potential to be applied to other ill-posed inverse problems where speed is a priority.
18 April 2022 in IEEE Sensors Journal
We report on the use of a novel constrained optimisation algorithm for calibrating the finite element model in an eddy current inspection application. An accurate finite element forward model is often important in such eddy current applications for training neural networks or as part of an iterative solver. However, the subject of calibration of the model has not received much attention in the literature to date. We consider a multi-frequency, eddy current depth profiling application, which is important for non-destructive testing in the nuclear industry. In the optimisation algorithm, we use a Levenberg-Marquardt algorithm and a bisection search to ensure constraints are satisfied, coupled with a truncated Gradient and Hessian method. We calibrate two types of finite element models, one using a filament representation for the coils and the other using a full 3D approach. The results show the feasibility of using the constrained non-linear optimisation algorithm for tuning the finite element model parameters. The mean signal-noise ratio after tuning on a truncated spectrum was 29.47 dB for the 3D model and 28.96 dB for the filament; in contrast the mean SNR using the measured coil parameters was 1.48 dB and 4.98 dB for the two uncalibrated models, respectively. The results show that a filament model is competitive with a 3D model of the coils (within the nuclear graphite application); therefore, a computationally faster filament model can be used with minimal effect on accuracy.