|
|
|---|
Materials for MLE Days and Summer School 2022 at TUHH presented by Antonin Sulc, check my website http://sulcantonin.github.io
15.09.2022, 13:30 - 15:00 Uhr
For smooth operation of the European XFEL, all subsystems involved must function perfectly so that there is only minimal downtime due to various failures. Many of these failures can be indicated by anomalous patterns in the data. Detecting these anomalies becomes particularly problematic given the enormous data throughput we record. In this session, we show how commercially available trainable unsupervised and supervised models for detecting anomalies at radio-frequency cavities and beam positions, which can be applied directly to accelerator control, identify numerous difficulties we encounter at the European XFEL.
Für einen reibungslosen Betrieb des European XFEL müssen alle beteiligten Teilsysteme perfekt funktionieren, damit es nur zu minimalen Ausfallzeiten aufgrund von verschiedenen Störungen kommt. Viele dieser Ausfälle können durch anomale Muster in den Daten angezeigt werden. Die Erkennung dieser Anomalien wird angesichts des enormen Datendurchsatzes, den wir aufzeichnen, besonders problematisch. In dieser Sitzung zeigen wir, wie die handelsüblichen trainierbaren unüberwachten und überwachten Modelle für die Erkennung von Anomalien an Hochfrequenzkavitäten und Strahlpositionen, die direkt auf die Beschleunigersteuerung angewendet werden können, zahlreiche Schwierigkeiten identifizieren, auf die wir beim European XFEL stoßen.
13.09.2022, 10:30 - 11:10
pdf link (ipynb basics) (ipynb autoencoder) (ipynb one class loss) (ipynb semi-supervised anomaly loss) (ipynb anogan) (ipynb orbits) (ipynb cavity)
- Error on Slide 23 with wrong exponentiation of the loss Semi Supervised Anomaly Detection is fixed now, sorry for confusion*
*Identification of unexpected behavior of some complex systems can provide us with valuable information about an ongoing problem. While nowadays, we often face an abundance of data, it is getting more and more valuable to identify either novel patterns or anomalies without very little or any human intervention. In this talk, we are going to introduce important concepts of anomaly detection. We present a taxonomy of anomalies since not all anomalies are clearly bounded by their intuitive meaning of sudden variation of some dimensions. Then we further introduce some concepts from state-of-the-art anomaly detection which uses currently the most popular machine learning framework Pytorch and explain and show some practical examples of one-class, semi-supervised learning and adversarial losses to train models to identify anomalies. Within the course, we give a particular emphasis to the intuitive understanding and ease of implementation which will be presented in simple Jupyter Notebooks. *