Seismic Acceleration Estimation Method at Arbitrary Position Using Observations and Machine Learning

Abstract

This study proposes a method of estimating the measurement data of nearby seismic stations by training an artificial neural network (ANN) through machine learning to understand the seismic acceleration time history at an arbitrary location where seismic acceleration time history is unknown. The ANN is trained using the observation data of 6 earthquakes at 10 ground seismic stations in Korea and 12 earthquakes at 212 underground seismic stations from the Korea Meteorological Administration. The location of the seismic station is assumed to be arbitrary in the untrained observation data to verify the validity of the trained ANN, and the measured and estimated data are compared. It is confirmed that the estimation accuracy of the ANN trained with the observation data of the underground seismic station is higher than that of the ANN trained with the observation data of the ground seismic station. The accuracy of the seismic acceleration estimation method proposed in this study is improved according to the level of learning data. It can also be applied as seismic acceleration to evaluate seismic damage or behavior of structures or facilities, even in places without seismic acceleration.