All projects

KAERI 2022 - 2023

AI anomaly detection for the KOMAC linac

An LSTM autoencoder for early detection of anomalies in quadrupole-magnet operation.

My contribution
Applied AI anomaly detection to KOMAC accelerator operating data as part of the intelligent-operation research program.
Operational-data study
Detected early anomaly signatures in five of ten historical voltage-fault cases, 30.7-286.6 minutes before the recorded fault signal-off.

Challenge

KOMAC (Korea Multi-purpose Accelerator Complex) operates a 100-MeV proton linear accelerator with Drift Tube LINAC (DTL) structures. The Drift Tube Quadrupole (DTQ) magnets focus the beam, but prolonged usage causes enamel wire corrosion and performance degradation. Existing threshold-based alarm systems are reactive, often leading to unnecessary shutdowns.

DTQ Failure
Fig. DTQ magnet structure (left) and failed DTQ after enamel wire corrosion (right)

Approach: LSTM Autoencoder

LSTM Autoencoder
Fig. LSTM Autoencoder architecture for DTQ magnet anomaly detection

Deep Learning Model

  • LSTM Autoencoder for time-series anomaly detection
  • 6.5 years of historical operational data (2017-2023)
  • Resistance-based analysis (voltage/current ratio)
  • MSE loss threshold of 0.125 for anomaly classification

Early anomaly detection

  • Learn normal operational patterns through reconstruction
  • Detect subtle anomalies before threshold-based alarms trigger
  • Investigate potential support for preventive maintenance

Results

Anomaly Detection
Fig. Anomaly score over time: normal cases vs detected anomaly cases
Metric Value
Reported overall accuracy 92%
Voltage-fault cases detected 5 of 10
Early detection range (detected cases) 30.7 - 286.6 min
Mean early detection (detected cases) 166.4 min
False alarms in normal operating periods 2 of 76 (2.6%)


These results come from a retrospective analysis of historical operating data. Early detection time is measured from the anomaly-score threshold crossing to the recorded voltage-fault signal-off, not to an independently predicted physical breakdown.

The study supports the potential for preventive maintenance; reductions in operational downtime or maintenance costs were not measured.

Publication

SCIE

  • D.H. Kim, H.S. Kim, H.J. Kwon, Y. Yu, G. Lee, “Deep learning-based anomaly detection in Drift Tube Quadrupole operation for the KOMAC LINAC,” Journal of Nuclear Science and Technology, 62(8), 709-714, 2025. DOI

Research Project

(2021-2023) Establishment of Intelligent Operation Platform for HANARO and Research Facilities, KAERI.