The emergence and dynamics of filamentary structures associated with
edge-localized modes (ELMs) inside tokamak plasmas during high-confinement mode
is regularly studied using Electron Cyclotron Emission Imaging (ECEI)
diagnostic systems. ECEI allows inference of electron temperature variations,
often across a poloidal cross-section. Previously, detailed analyses of
filamentary dynamics and classification of the precursors to ELM crashes have
been done manually. We present a machine-learning-based model, capable of
automatically identifying the position, spatial extent, and amplitude of ELM
filaments. The model is a deep convolutional neural network that has been
trained and optimized on an extensive set of manually labeled ECEI data from
the KSTAR tokamak. Once trained, the model achieves a 93.7% precision and
allows to robustly identify plasma filaments in unseen ECEI data. The trained
model is used to characterize ELM filament dynamics in a single H-mode plasma
shot. We identify quasi-periodic oscillations of the filaments' size, total
heat content, and radial velocity. The detailed dynamics of these quantities
appear strongly correlated with each other and appear qualitatively different
during the pre-crash and ELM crash phases.

本研究提出了一种基于深度卷积神经网络的机器学习模型，可自动识别位置、空间范围和振幅不同的 ECEI 电子回旋发射成像系统中的 ELM 暴露结构，该模型在 KSTAR 托卡马克上训练和优化，精度可达 93.7％，并用于表征 ELM 丝线动态行为，并发现在 ELM 前期和 ELM 崩溃阶段之间有明显区别。