Speaker
Description
Identifying energetic particle driven modes in spectrograms is crucial in generating large databases. These databases are essential for advancing physics research and training machine learning models for real time control. However, to train supervised machine learning algorithms, large databases labelled by humans must be used, which is practically impossible taking into account the amount and the complexity of the data that is required. To this end, we present two unsupervised and semi-supervised machine learning algorithms for mode identification without any human intervention, which can be used to build large databases.
First, a method for labeling spectrograms using computer vision algorithms is introduced. The computer vision pipeline [1] consists of a sequence of image processing steps: directional filtering, ridge enhancement, thresholding, and connectivity filtering. This approach is well-suited for semi-supervised identification of time-frequency structures across different plasma discharges. We present some applications to discharges in the JET tokamak that are challenging from the point of view of the complexity.
Second, a new time signal coding algorithm accelerated using neural networks is applied to identify modes directly from raw signals [2]. This method has been employed to extract modes from 1,291 discharges of the TJ-II stellarator. Afterwards, a novel method based on the use of Mutual Information can be used to create features based on generalized correlations between plasma signals and mode structures. By combining time-frequency decompositions of Mirnov coil signals with other diagnostic measurements such as density, plasma energy, and current, it is possible to perform an in-depth profiling of the mode activity in the whole database using clustering analysis. As an illustrative result, modes associated with Alfvénic activity and those correlated with the plasma current have been successfully extracted from the large database.
These two approaches open the door to automatic and fast identification of MHD activity without any human intervention to create large databases for subsequent transfer learning of the pre-trained state-of-the-art deep learning techniques, such as YOLO or Detectron.
[1] E d D Zapata-Cornejo et al 2024 Plasma Phys. Control. Fusion 66 095016
[2] E d D Zapata-Cornejo et al 2024 Nucl. Fusion 64 126057
| Presentation type | Oral |
|---|
