Kernel ridge regression estimation of high-frequency signals for nuclear fusion diagnostics

Abstract

Estimating missing signals in plasma measurements is a significant challenge, particularly in fusion plasma devices where high-frequency temporal evolution signals provide valuable insight into the behavior of plasma. Indeed, occasional signal absence can impact to misfunction of other active diagnostics, leading to misinterpretation of the plasma state and inaccurate management of related control systems that affect the overall performance of the device. Therefore, the replacement of synthetic and accurate signals is a crucial aspect for system performance and reliability in fusion diagnostics, particularly when original signals are missing or erroneous. Precise prediction of high-frequency signals refers to time-series synthetic data that are modelled at an extremely fine scale. Due to the large number of samples, high frequency observations generally contain a large amount of data and most forecasting techniques based on dimensionality reduction or feature selection are not applicable as valuable information is lost before training periods. Besides, many real-time signals such as those in ITER, are non-stationary, meaning that their spectral content changes with time. In the current work a novel algorithm is presented based on kernel ridge regression (KRR) confidence machine. It provides a balance between computational efficiency and high accuracy across the entire frequency spectrum.