Gaussian process-based quasi-coherent noise suppression in magnetic confinement devices with superconductors

Abstract

Gaussian process (GP)-based technique suppressing quasi-coherent noises, i.e. structured noises, is developed which is more effective than conventional denoising techniques such as using frequency-domain filters. Superconducting devices like KSTAR, EAST, JT-60SA and ITER require separate sets of normal conducting magnetic coils inside the tokamak vacuum vessels to achieve a prompt control of fusion-grade plasmas in response to various fast and abrupt plasma activities such as vertical displacement events. Hence, these in-vessel control coils are typically operated with high-frequency switching power supplies which generate quasi-coherent noises. Semi-conductor based bolometers in KSTAR, for instance, are vulnerable to the quasi-coherent noise that makes a tomographic reconstruction for the 2D poloidal radiation map with the noise-contaminated signals flawed. By modeling the quasi-coherent properties of the noise as multivariate Gaussian distribution and generating the kernel function for the GP solely based on the measurements, the proposed method is able to suppress the noise whose performance is superior to the conventional filtering schemes. The method not only suggests an estimate of the denoised signal but also informs the consistent (with the measurements) uncertainty of the estimate at a level smaller than the standard deviation of the quasi-coherent noise. Performance of the method is confirmed with synthetic data containing the quasi-coherent noises, and it is applied to the measured data obtained by the KSTAR bolometers.