1:45 PM - 3:15 PM
[SCG55-P16] Virtual seismic network using deep learning and preliminary study
Keywords:Virtual seismic network, Deep learning, Ground motion estimation
In the past, severe house damage in a small area caused by strong ground motion was reported, suggesting that seismic ground motions change drastically within a small area. Therefore, it is important for disaster prevention to understand the actual ground motions during earthquakes and the causes of house damage. However, it is difficult to conduct ultra-high-density seismic observations, and even if it were possible, it would not be possible to observe strong ground motions of past damaging earthquakes.
In this study, therefore, we aim to develop a virtual seismic network that can estimate seismic ground motions in places where no seismographs are installed by combining a segmented observation method and deep learning technology. As a preliminary study, we have estimated the waveform attributes (P-wave arrival time and amplitude) using the dense seismic observation data on Hachijojima Island. 6 out of 44 stations were used as the reference stations and the relationship between reference stations and other stations was learned. After learning, the P-wave arrival time and amplitude were estimated from the data of the reference stations. The network we used was a fully-connected neural network with 33 units in the input layer, 3 hidden layers of 32, 64 and 128 units respectively, and 2 units in the output layer. Estimation results showed that for the data used for training, the mean error and standard deviation for the arrival times were 0.03 and 0.17 seconds respectively, and -1% and 33% for the amplitudes. For the data not used for training, the mean error and standard deviation were -0.03 and 0.33 seconds for arrival times and -5% and 46% for amplitudes.
In this study, therefore, we aim to develop a virtual seismic network that can estimate seismic ground motions in places where no seismographs are installed by combining a segmented observation method and deep learning technology. As a preliminary study, we have estimated the waveform attributes (P-wave arrival time and amplitude) using the dense seismic observation data on Hachijojima Island. 6 out of 44 stations were used as the reference stations and the relationship between reference stations and other stations was learned. After learning, the P-wave arrival time and amplitude were estimated from the data of the reference stations. The network we used was a fully-connected neural network with 33 units in the input layer, 3 hidden layers of 32, 64 and 128 units respectively, and 2 units in the output layer. Estimation results showed that for the data used for training, the mean error and standard deviation for the arrival times were 0.03 and 0.17 seconds respectively, and -1% and 33% for the amplitudes. For the data not used for training, the mean error and standard deviation were -0.03 and 0.33 seconds for arrival times and -5% and 46% for amplitudes.