Greifswald 2024 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
P: Fachverband Plasmaphysik
P 25: Poster III
P 25.19: Poster
Thursday, February 29, 2024, 16:30–18:30, ELP 6: Foyer
Computer Vision Deep Learning-Based Shattered Pellet Injection (SPI) Shard Tracking at ASDEX Upgrade — •Johannes Illerhaus1,2, Paul Heinrich1,2, Mohammad Miah1,2, Gergely Papp1, Tobias Peherstorfer3, Wolfgang Treutterer1, Bernhard Sieglin1, Udo von Toussaint1, Hartmut Zohm1, Frank Jenko1, and the ASDEX Upgrade Team4 — 1Max-Planck-Institut für Plasmaphysik, Garching, Germany — 2Technische Universität München, Garching, Germany — 3Technische Universität Wien, Vienna, Austria — 4see the author list of U. Stroth et al. 2022 \textit{NF} 62 042006
A computer vision deep learning pipeline was constructed to automate the analysis of the more than 1000 videos created in lab experiments on the SPI test bench at ASDEX Upgrade. Our machine learning (ML) models provide highly accurate segmentation of the fragments shown in these videos. This allows for the labeling of the entire dataset, of which previously only 177 videos had been labeled using a pipeline based on traditional computer vision. The ML models eliminate the previously necessary human supervision, reduce the run time from months to a few hours and increase the accuracy and robustness of labeling. The shards are then tracked between frames with the goal of estimating their size and speed distributions. This enables using the experimental results to validate theoretical models predicting the right system setup and pellet attributes to produce the fragment distributions for optimal disruption mitigation. This will ultimately help inform design decisions for the ITER SPI, ITER's primary disruption mitigation system.
Keywords: Shattered Pellet Injection; Machine Learning; Computer Vision