Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Automated Anomaly Detection on European XFEL Klystrons

Antonin Sulc, Annika Eichler, Tim Wilksen

TL;DR

The paper addresses reducing downtime in European XFEL klystrons by detecting anomalies in high-power RF signals without labeled faults. It employs an unsupervised deep one-class classification using an LSTM encoder to map 205-feature sequential inputs to a latent center, minimizing $s(x) = \|f_\theta(x) - c\|_2$ for normal data and producing an anomaly score. Two real events (Mar 7, 2023 and Feb 1, 2024) demonstrate the method's ability to reveal precursors and distinct state clusters via anomaly scores and TSNE embeddings, enabling earlier diagnostics. This approach offers data-driven lead time for preventive actions, potentially increasing uptime and reducing costly downtime in the accelerator facility.

Abstract

High-power multi-beam klystrons represent a key component to amplify RF to generate the accelerating field of the superconducting radio frequency (SRF) cavities at European XFEL. Exchanging these high-power components takes time and effort, thus it is necessary to minimize maintenance and downtime and at the same time maximize the device's operation. In an attempt to explore the behavior of klystrons using machine learning, we completed a series of experiments on our klystrons to determine various operational modes and conduct feature extraction and dimensionality reduction to extract the most valuable information about a normal operation. To analyze recorded data we used state-of-the-art data-driven learning techniques and recognized the most promising components that might help us better understand klystron operational states and identify early on possible faults or anomalies.

Automated Anomaly Detection on European XFEL Klystrons

TL;DR

The paper addresses reducing downtime in European XFEL klystrons by detecting anomalies in high-power RF signals without labeled faults. It employs an unsupervised deep one-class classification using an LSTM encoder to map 205-feature sequential inputs to a latent center, minimizing for normal data and producing an anomaly score. Two real events (Mar 7, 2023 and Feb 1, 2024) demonstrate the method's ability to reveal precursors and distinct state clusters via anomaly scores and TSNE embeddings, enabling earlier diagnostics. This approach offers data-driven lead time for preventive actions, potentially increasing uptime and reducing costly downtime in the accelerator facility.

Abstract

High-power multi-beam klystrons represent a key component to amplify RF to generate the accelerating field of the superconducting radio frequency (SRF) cavities at European XFEL. Exchanging these high-power components takes time and effort, thus it is necessary to minimize maintenance and downtime and at the same time maximize the device's operation. In an attempt to explore the behavior of klystrons using machine learning, we completed a series of experiments on our klystrons to determine various operational modes and conduct feature extraction and dimensionality reduction to extract the most valuable information about a normal operation. To analyze recorded data we used state-of-the-art data-driven learning techniques and recognized the most promising components that might help us better understand klystron operational states and identify early on possible faults or anomalies.
Paper Structure (12 sections, 1 equation, 4 figures)

This paper contains 12 sections, 1 equation, 4 figures.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. One Class Loss
  5. Architecture
  6. Inputs
  7. Pre-processing
  8. Results
  9. Event on Mar. 7th 2023
  10. Event on Feb. 1st 2024
  11. Conclusion
  12. Acknowledgements

Figures (4)

  • Figure 1: Anomaly scores $s$ of the event that took place on Mar. 7th 2023 (top) and Feb. 1st 2024 (bottom) Top: Notice the bump around 9:13:16 which increases the $s$ until the klystron is shut down (after 9:23:45). Right: The anomaly score fluctuates quite significantly over an extended period (10:24:26 - 15:03:06). There is a peak around 14:30, which precedes severe disruption of observed signals.
  • Figure 2: The phase of the event that took place on Mar. 7th 2023. Each column is one waveform FD.RI of one pulse.
  • Figure 3: Amplitude and phase of FD.RI of the event from Feb. 1st 2024.
  • Figure 4: T-SNE Embedding shows a reduced projection of the network from $M$ dimensions onto 2D while distance is preserved. The colors of the left image encode different $s$ levels (red between $(0.04,0.06)$, blue between $(0,0.04)$ and green $(0.06,\infty)$. Similarly on the right figure, red color encodes $s$ above $0.007$.