[4L2-GS-13] AI application: Anomaly detection and maintenance

[4L2-GS-13-03] Anomaly detection of seismogram with autoencoder

〇Kanei Kaku¹, Yuki Miki¹, Yasuhiro Fujita¹, Keisuke Kiritoshi¹, Koji Ito¹ (1. NTT Communications Ltd)

Keywords:Anomaly Detection, Machine Learning, Deep Learning

It became possible to collect large scale of seismogram rapidly since the construction of Hi-net.
This shows mechanical failure are not just obvious breakdowns but also include the aging failure.
Obvious breakdowns were practicable to detect using previous methods. However, aging failure were discussed by researchers and specialists using its feature of frequency map. This make the delay of detecting failure. Not just this but because of its conclusion depending on discussion, it was hard to evaluate its correctness.
In this paper, we propose the autoencoder method to predect sensor's anomaly behavior.
Proposing method is to visualize the abstractness of sensor's anomaly behavior and evaluate it.
Since the aging phenomenon need years data for observation, we built high-speed data processing system for serving big data.
We experimented on real data for the valuation of proposed method.From this result, the proposed method shows its capability for detecting anomaly seismogram.