A Transformer-based Deep Learning Approach For Microearthquake Detection And Its Application

Detecting every earthquake accurately and efficiently provides a significant foundation for seismic research.In recent years,there has been increasing interest in leveraging machine learning,especially deep learning,to improve microseismic detection performance.Although considerable research efforts have been made in this direction,most state-of-the-art solutions are based on convolutional neural network(CNN)structures.However,CNN-based models cannot fully capture the global and dynamic features of time-series data(i.e.seismic waves).To solve this,here we build a microseismic detection model based on the attention mechanism,which can accurately detect earthquakes from the noises and improve the catalog.Further,to explore the relationship between improved catalog and foreshock sequences,we use the advanced seismic phase picking method on the large dataset to widely detect new events and their arrival time,and then analyze the foreshock sequences before strong earthquakes.Specifically,(1)We propose a new deep learning method,Trans Quake,for seismic P-wave detection based on the frontier sequence model Transformer.Specifically,Trans Quake can utilize the STA/LTA algorithm to adapt the three-component structure of seismic waves as input and utilize the multi-head attention mechanism for interpretable model learning.Extensive evaluation results on the 2008 Wenchuan Mw7.9 earthquake aftershock dataset clearly show that Trans Quake can outperform the state-of-the-art algorithm baseline to achieve the best detection performance.At the same time,the experimental results also verify the interpretability of Trans Quake's understanding of the results,such as the attention distribution of seismic waves at different locations,and the correlation between coda waves and P waves helps to discriminate waveforms.(2)We use the advanced seismic phase picker EQTransformer,which based on the attention mechanism,to scan the continuous waveform data before the 14 strong earthquake events in the Chinese Mainland.Further,we utilize the machine learning method to evaluate each magnitude of newly detected events.We found that using an improved,highly complete earthquakes catalog helps us to discover unusually frequency changes before the strong earthquakes.Many of these foreshocks were so small in magnitude that they had been missed before.Our observations help to narrow the gap in the proportion of foreshocks between laboratory earthquakes and natural earthquakes and provide some indications for the arrival of strong earthquakes.