[1L3-OS-17] 地震研究と人工知能

[1L3-OS-17-02] Efficient similar waveform search using deep hashing

〇Makoto Naoi¹, Shiro Hirano² (1. Kyoto University, 2. Ritsumeikan University)

Keywords:deep hashing, similar waveform search, earthquake detection

Similar waveform searching plays an important role in seismicity analysis. However, the high computing cost to calculate waveform cross-correlations is an obstacle when analyzing long, continuous waveform records with many stations. We developed a deep learning network to obtain 64-bit hash codes containing information on seismic waveforms and performed a similar waveform search for 16-channel, 10-MHz sampling continuous acoustic emission records obtained in a hydraulic fracturing experiment in the laboratory. The analysis using template hash codes of events cataloged with conventional techniques allowed us to detect 2.5 times more events. We also searched for events with similar waveforms without templates by calculating hamming distances among all windows cut from the continuous record. This calculation took only 15.5 hours under 120-thread parallelization. Our approach significantly reduces the required memory and calculation costs compared to a hashing approach based on random substitution, enabling similar waveform searching on a large-scale dataset.