A FREQUENCY-BASED SITE FORECASTING METHOD USING DEEP LEARNING
Document Type
Book
Publication Date
2024
Abstract
Recently deep learning has attracted increased attention in predicting future ground motions from historical data. Such attention could boost the efficiency and speed of earthquake early warning (EEW) systems. Past developments have proven that deep learning has the potential to predict complicated systems accurately in temporal and spatial applications. However, there is yet to be an accurate method for forecasting ground motions due to earthquakes that also retain frequency characteristics. A multi-component technique is proposed for site spectra prediction (frequency-based site forecasting) of target-sites using a novel data-driven approach to sequence learning with deep neural networks. The novelty of this approach lies in using real-time observations from front-sites near the epicenters and rapidly estimating target-site spectra with high accuracy. Station sites that are close to hypocenters of strong-motion earthquakes (front-sites for strong-motion generation areas) can provide real-time spectra of the impending wave that will affect more distant sites (target-sites). Once the influence of earthquakes is recorded at front-sites and target-sites, a relationship between both sites’ spectra can be developed with a data-driven model. A deep learning technique can thus be used to explore features in input spectra from front-sites and map a relationship with their equivalent target-site’s spectra. For the real-time implementation of deep learning models, the real-time flowing waveform at sites can be obtained and transformed into ascending spectra. The development of spectra through time provides temporal spectral series that could be assessed by sequential learning methods, such as recurrent neural networks. The aim is to discover a deep learning model that can map the relation between the spectral series of front-sites and target-sites. For this EEW technique, it is vital to map the earliest relation possible, where series at front-sites are initiated from P-wave arrival while series at target-sites are initiated from S-wave arrival. Since useful engineering spectra can be in the form of Pseudo-Velocity Spectra, both are explored in this study. For the implementation process, two different case studies were chosen to evaluate the deep learning method. The evaluation is performed through quantitative analysis against ground motion prediction equations (GMPE), and additionally the estimated warning time and shaking intensity.
Recommended Citation
Torky, Ahmed A. PhD and OHNO, Susumu Ph.D., "A FREQUENCY-BASED SITE FORECASTING METHOD USING DEEP LEARNING" (2024). Civil Engineering. 245.
https://buescholar.bue.edu.eg/civil_eng/245