Neural Network Learning: Crustal State Estimation Method from Time-Series Data

Predicting earthquake activity is challenging because of its complex mechanism. Theoretical models help us to understand some aspects; however, for practical purpose, it is difficult to predict earthquake activity using only models. Therefore, we usually utilize observational data by either fitting parameters to models or making models or indicators directly from the data. In this study, we chose the latter. We tried to reveal nonlinear relationships between observed data. We adopted machine learning, using three-layered neural network, as a method for extracting non-linear features. We applied this method to time-series data obtained from a multi-channel observatory before and after an earthquake. Thus, we obtained a trajectory of feature quantities that exhibited cyclic behavior, returning to the same state as before. We expect that the trajectory will be useful for judging whether aftershocks will occur, or not, by the feature quantities not returning, or returning, to their initial states.