A machine-learning-based method of detecting and picking the first P-wave arrivals of acoustic emission events in laboratory experiments

Abstract

SUMMARY Detecting and picking the first P-wave arrivals of seismic events in seismograms is fundamental in observational seismology. Recently, several machine-learning-based algorithms have been developed to incorporate human expertise for picking P-wave arrival times automatically. One shortcoming of these models is that they pick arrival times at individual seismic stations separately, which need to be sorted and associated to identify the seismic event. Also, most of them rely on existence of P-wave arrivals in the seismograms to be picked. Here, we developed a machine-learning-based seismic event detection and P-wave arrival time picking method called MultiNet and applied it to acoustic emission (AE) waveform data recorded in laboratory experiments. The MultiNet uses 2-D waveform images from multichannel AE recordings as the input to a convolutional neural network (CNN) to detect whether there is an AE event in an image and, if so, uses a fully convolutional neural network (FCN) to pick the P-wave arrival time at each channel in the image. We tested the MultiNet using 550 known AE events recorded during syn-deformational phase transformation from olivine to spinel in Mg2GeO4 (an analogue to Mg2SiO4) in a high-pressure experiment. Waveform data of 50 events were used to train the neural networks and the rest of data were used to validate the method. At the optimal image length and detection threshold, the CNN was able to detect all 500 known events plus 48 more events missed previously. Overall, 98.7 per cent of P-wave arrival times picked by the FCN were within 0.5 $\mu$s from the manually picked times. The average picking errors at different channels range from 0.01 ± 0.05 to −0.06 ± 0.22 $\mu$s. Our method greatly reduces the amount of human labour in picking P-wave arrival times for event location and source moment tensor inversion. It can be easily adapted to process continuous waveform data of a seismic network for earthquake detection and location in real time.