Unsupervised anomaly detection for earthquake detection on Korea high-speed trains using autoencoder-based deep...

Area of Science:

• Geophysics
• Artificial Intelligence
• Transportation Engineering

Background:

• High-speed trains are vulnerable to seismic events, necessitating reliable detection systems.
• Conventional earthquake detection methods like STA/LTA struggle with false positives in operational environments.

Purpose of the Study:

• To develop an advanced earthquake detection system for high-speed trains using unsupervised deep learning.
• To improve detection accuracy and reduce false alarms compared to existing methods.

Main Methods:

• Utilized autoencoder-based deep learning models for unsupervised anomaly detection.
• Trained models on normal high-speed train vibration data and validated with seismic data.
• Compared performance against the Short Time Average over Long Time Average (STA/LTA) model.

Main Results:

• The proposed deep learning model demonstrated superior earthquake detection capabilities for Peak Ground Acceleration (PGA) ≥ 0.07.
• Significantly reduced false earthquake detections during normal train operations.
• Accurately identified normal operational states, unlike the STA/LTA method.

Conclusions:

• The autoencoder-based deep learning approach offers a more reliable earthquake detection solution for high-speed rail.
• This method is effective even with limited seismic data, benefiting regions with low seismicity.
• Enhanced safety and operational integrity for high-speed trains in seismically active areas.