Scholarlink

<p>This study applies a transformer-based conditional variational autoencoder (T-CVAE) model for anomaly detection in pulse waveform signals from the High Voltage Converter Modulator (HVCM) and Klystron at the Korea Multipurpose Accelerator Complex (KOMAC). Building upon prior work using CVAE models for anomaly detection in Spallation Neutron Source (SNS) accelerators, the T-CVAE model was tailored to the specific characteristics of KOMAC data by optimizing hyperparameters and leveraging transformer-based architectures for enhanced feature extraction. Experimental results demonstrate that the model effectively learns the distribution of normal signals, as validated through boxplots, the receiver operation characteristics (ROC) curve and kernel density estimation (KDE) analyses. Anomalies are detected through significant reconstruction loss differences between normal and abnormal signals. By reliably identifying pre-fault conditions, the proposed system offers a promising approach to improving operational reliability and minimizing unplanned downtime in KOMAC's proton linear accelerator.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Anomaly detection in KOMAC high-power systems using transformer-based conditional variational autoencoder

  • 期刊论文

  • 发表日期：

    2025年3月10日

查看全文
  • Gi-Hu Kim,
  • Hae-Seong Jeong,
  • Han-Sung Kim,
  • Hyeok-Jung Kwon,
  • Dong-Hwan Kim

摘要

This study applies a transformer-based conditional variational autoencoder (T-CVAE) model for anomaly detection in pulse waveform signals from the High Voltage Converter Modulator (HVCM) and Klystron at the Korea Multipurpose Accelerator Complex (KOMAC). Building upon prior work using CVAE models for anomaly detection in Spallation Neutron Source (SNS) accelerators, the T-CVAE model was tailored to the specific characteristics of KOMAC data by optimizing hyperparameters and leveraging transformer-based architectures for enhanced feature extraction. Experimental results demonstrate that the model effectively learns the distribution of normal signals, as validated through boxplots, the receiver operation characteristics (ROC) curve and kernel density estimation (KDE) analyses. Anomalies are detected through significant reconstruction loss differences between normal and abnormal signals. By reliably identifying pre-fault conditions, the proposed system offers a promising approach to improving operational reliability and minimizing unplanned downtime in KOMAC's proton linear accelerator.