5:15 PM - 6:45 PM
[STT37-P05] Predicting Peak Ground Acceleration for Earthquake Early Warning by Deep Learning Technologies
Keywords:Earthquake Early Warning, Peak Ground Acceleration, Artificial Intelligence, Ground Motion Prediction
Traditional earthquake monitoring systems conduct the monitoring process by checking the seismic waves from multiple seismograph stations and detecting an earthquake event based on certain abnormal phenomenons observed in the waveforms, such as the ratio of the short-term average and long-term average exceeding a pre-defined threshold. When an earthquake occurs, an early warning decision can be made using the observed waveforms from the first few seismograph stations and a warning message can be issued to remote sites at the very early stage of the event. However, the traditional early warning process may create a blind zone in the area close to the epicenter, where may face the most serious hit by the shaking waves. In this study, a deep learning model is developed to predict the peak ground motion at a certain site using the first few seconds of the on-site waveforms observed at the corresponding location. From previous studies in the literature, certain features in the initial P waves, such as the peak displacement of the waveforms, may imply the shaking intensity for the later S waves. The proposed deep learning model can extract more features from a short window of the initial P waves and infer whether the potential ground motion intensity may exceed a certain threshold for the later destructive S waves. An early warning message can be issued if the predicted ground motion may exceed a chosen threshold. From the experiments, the proposed scheme can achieve very high accuracy for issuing on-site early warning messages. The warnings can cover the blind zone for traditional early warning systems and may further prevent possible disasters that occur close to earthquake epicenters.