Abstract: Picking up the arrival time of seismic phases is one of the fundamental problems of seismology. This study introduces a method based on LSTM neural network for automatic P phase picking. We transformed the arrival time problem into the probability problem of time redefined labeling, and a 4-layer neural network was built and trained on the North Korean nuclear earthquake vertical waveforms. The P phase arrival time of subsequent events were picked up accurately and effectively. The method shows a certain adaptability with ambient noise. Random input samples test shows that the input waveform data should be better with longer than 10 s after the P phase to get stable results. As an artificial intelligence method, LSTM neural network provides a new solution for seismic phase picking-up.

Key words: LSTM neural network, P phase picking, Nuclear explosion earthquake