München 2019 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
P: Fachverband Plasmaphysik
P 18: Postersitzung
P 18.106: Poster
Thursday, March 21, 2019, 16:30–18:30, Foyer Audimax
Convolutional LSTMs for Plasma State Classification — •Francisco Matos1, Vlado Menkovski2, Federico Felici3, Frank Jenko1, and The TCV Team3 — 1Max Planck Institute for Plasma Physics, Garching, Germany — 2Eindhoven University of Technology, Eindhoven, Netherlands — 3Swiss Plasma Center, Lausanne, Switzerland
During a tokamak discharge, the plasma can vary between different modes, Low and High confinement, with an additional intermediate state called Dithering. Furthermore, several events can happen during a discharge, namely ELMs when the plasma is in H mode. The state transitions and events in question are identifiable by a human expert post-shot by looking at features from several different diagnostic signals. Ideally, an approach should exist allowing to determine in real-time when these events occur. Convolutional neural networks (CNNs), typically used for image recognition, are ideal to automatically extract the data features necessary to determine when these events take place. However, CNNs do not keep track of temporal dependencies between different data points. As a result, they can make inconsistent predictions - for example, two successive transitions into the same state. Long Short-Term Memory Networks, or LSTMs, are a type of neural network designed specifically to keep track of temporal dependencies. By using convolutional layers for feature extraction and LSTM layers to keep track of temporal correlations, we train an automatic classifier to determine the plasma state. We use data from the TCV tokamak - specifically, photodiode, interferometer and diamagnetic loop signals.