| Paper | Title | Page |
|---|---|---|
| TUPOPT062 | A Data-Driven Anomaly Detection on SRF Cavities at the European XFEL | 1152 |
|
||
|
Funding: This work was supported by HamburgX grant LFF-HHX-03 to the Center for Data and Computing in Natural Sciences (CDCS) from the Hamburg Ministry of Science, Research, Equalities and Districts. The European XFEL is currently operating with hundreds of superconducting radio frequency cavities. To be able to minimize the downtimes, prevention of failures on the SRF cavities is crucial. In this paper, we propose an anomaly detection approach based on a neural network model to predict occurrences of breakdowns on the SRF cavities based on a model trained on historical data. We used our existing anomaly detection infrastructure to get a subset of the stored data labeled as faulty. We experimented with different training losses to maximally profit from the available data and trained a recurrent neural network that can predict a failure from a series of pulses. The proposed model is using a tailored architecture with recurrent neural units and takes into account the sequential nature of the problem which can generalize and predict a variety of failures that we have been experiencing in operation. |
||
| DOI • | reference for this paper ※ https://doi.org/10.18429/JACoW-IPAC2022-TUPOPT062 | |
| About • | Received ※ 17 May 2022 — Revised ※ 14 June 2022 — Accepted ※ 15 June 2022 — Issue date ※ 24 June 2022 | |
| Cite • | reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml) | |
| WEPOMS036 | Accelerating Linear Beam Dynamics Simulations for Machine Learning Applications | 2330 |
|
||
| Machine learning has proven to be a powerful tool with many applications in the field of accelerator physics. Training machine learning models is a highly iterative process that requires large numbers of samples. However, beam time is often limited and many of the available simulation frameworks are not optimized for fast computation. As a result, training complex models can be infeasible. In this contribution, we introduce Cheetah, a linear beam dynamics framework optimized for fast computations. We show that Cheetah outperforms existing simulation codes in terms of speed and furthermore demonstrate the application of Cheetah to a reinforcement-learning problem as well as the successful transfer of the Cheetah-trained model to the real world. We anticipate that Cheetah will allow for faster development of more capable machine learning solutions in the field, one day enabling the development of autonomous accelerators. | ||
| DOI • | reference for this paper ※ https://doi.org/10.18429/JACoW-IPAC2022-WEPOMS036 | |
| About • | Received ※ 07 June 2022 — Revised ※ 14 June 2022 — Accepted ※ 01 July 2022 — Issue date ※ 01 July 2022 | |
| Cite • | reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml) | |