Research Of Tomographic Reconstructions On Tokamak Line-integrated Signals Based On Deconvolutional Neural Networks

A number of diagnostic systems in Tokamak device are based on line-integrated measurement along its sight of view.The line-integrated signals are sparse due to the limited size of both Tokamak ports and the diagnostics.It is well needed to reconstruct the two-dimensional(2-D)plasma cross section based on the one-dimensional(1-D)line-integrated signals,which will be helpful to the further analysis including MHD activities,impurity transportations or turbulent transportations.The prevalent reconstruction methods from 1-D signals to 2-D image are based on iterations,which are accomplished through optimization by imposing additional constraint terms to retrieve the emission intensities on plasma cross sections.However,this kind of approaches usually needs hours to iterate,besides they often require a priori information or constraint terms.In this thesis we developed a reconstruction method based on deconvolutional neural networks(deCNN).By training in advance,the network will be capable of recover new data samples.Once the training are completed,the network can gain a large amount of reconstructed images in seconds.In this thesis we give a brief introduction about deCNN.The network then trained through simulation data that are generated from an analytic parameter model,which could provide ground truth for the assessment of reconstructions.After that we explored the suitable network architecture for the configurations similar to EAST soft-X ray systems.By replacing loss function from L2 to structure similarity(SSIM),the network shows improved reconstruction accuracy.In consideration of noisy environment of the diagnostics,we present the noise robustness test for this method.The network shows admissible results under 10%,15%and 20%noise corruptions.