11:30 〜 11:45
[SSS04-10] Deep Learning Technique for Temporal Site-to-Site Seismic Predictions using short-interval Fourier Amplitude Spectra
キーワード：strong ground motion, earthquake early warning, deep learning techniques, Fourier spectra prediction, ground motion prediction equation
In this work, advance seismic amplitude of impending ground shaking at target sites is predicted using Fourier spectra and deep learning, a supervised learning technique. Past strong ground motions could provide enough seismic activity data for a region’s earthquake early warning system. Historical sensor data, such as from surface and borehole accelerometers, could be coupled with modern deep learning techniques to provide a good estimate of real-time wave-propagation from a front-site to a target-site. During a strong motion event, the ongoing seismic waveform could be obtained from front-site stations which could enable future predictions in target stations if the relationship is mapped between both sites. Deep learning models have the capability of mapping nonlinear relationships between features and targets, which is explored in this study.
Wave-field prediction in the form of acceleration time-series is possible because current communication speed exceeds seismic wave velocities. Observation sites that are closer to the hypocenter than target sites provide early seismic activity waveform. Since the frequency content of seismic waves vary with propagation distance in addition to the media it passes through, therefore similarity between acceleration at front-site and target site is a challenge to map. The temporal change of the (short-time) amplitude Fourier spectra at each site could in-exchange provide the real-time response mapping between multiple affected sites. This process is similar to frequency-dependent site amplification prediction. To map these relationships and estimate future short-time seismic wave amplitude, deep learning methods are implemented.
The supervised AI algorithms used in this study are Deep Neural Networks (DNN) and Convolutional Neural Networks (CNN). The structures of these neural networks allow for non-linear mapping between inputs and outputs of training tasks. These deep learning algorithms can handle big series data; however, the data management and filtering are necessary for importing meaningful features to neural networks. A new proposed pipeline and deep learning model is created, where hybrid CNN-DNN models attain features of Fourier spectra of multiple sites and multiple components in the form of a temporal 2D array. The representation of this model is presented in Figure 1.
Firstly, historical time-series data, from multiple strong motion events, for front-field sites and far-field sites are gathered and filtered. Secondly, filtering is performed with Discrete Wavelet Transform (DWT) decomposition, and DWT ranks that contain previously selected predominant frequencies are attained. This process functions as a facilitator for better model training. Thirdly, fast Fourier transform (FFT) with a short time interval is then applied to the decomposed time-series data, and time-series intervals of site 1 (Input) and site 2 (Target) are paired. Multiple input sites are also possible. Finally, for each DWT rank, a deep learning model is created, trained with multiple historical data, and tested with blind datasets. This process is swift and fairly accurate in real-time because inference of trained models is not computationally expensive. The designed tool temporal output is shown in Figure 2.
This algorithm is implemented on three separate case-studies, which included sites in Miyagi and Iwate. A study is done on the variation of parameters set by this method. The three main parameters are the short time interval length, the stride length from one interval to its subsequent window, and the decimation factor of the time-series data. In all three case-studies, the DWT ranks {3, 4, 5, 6, 7} were chosen, as they included the various frequency bands necessary for site response evaluation. The deep learning models are modified to optimize predicted Fourier amplitude spectra at target sites. The available. The available safety time is provided, and the results are evaluated in multiple forms.
Wave-field prediction in the form of acceleration time-series is possible because current communication speed exceeds seismic wave velocities. Observation sites that are closer to the hypocenter than target sites provide early seismic activity waveform. Since the frequency content of seismic waves vary with propagation distance in addition to the media it passes through, therefore similarity between acceleration at front-site and target site is a challenge to map. The temporal change of the (short-time) amplitude Fourier spectra at each site could in-exchange provide the real-time response mapping between multiple affected sites. This process is similar to frequency-dependent site amplification prediction. To map these relationships and estimate future short-time seismic wave amplitude, deep learning methods are implemented.
The supervised AI algorithms used in this study are Deep Neural Networks (DNN) and Convolutional Neural Networks (CNN). The structures of these neural networks allow for non-linear mapping between inputs and outputs of training tasks. These deep learning algorithms can handle big series data; however, the data management and filtering are necessary for importing meaningful features to neural networks. A new proposed pipeline and deep learning model is created, where hybrid CNN-DNN models attain features of Fourier spectra of multiple sites and multiple components in the form of a temporal 2D array. The representation of this model is presented in Figure 1.
Firstly, historical time-series data, from multiple strong motion events, for front-field sites and far-field sites are gathered and filtered. Secondly, filtering is performed with Discrete Wavelet Transform (DWT) decomposition, and DWT ranks that contain previously selected predominant frequencies are attained. This process functions as a facilitator for better model training. Thirdly, fast Fourier transform (FFT) with a short time interval is then applied to the decomposed time-series data, and time-series intervals of site 1 (Input) and site 2 (Target) are paired. Multiple input sites are also possible. Finally, for each DWT rank, a deep learning model is created, trained with multiple historical data, and tested with blind datasets. This process is swift and fairly accurate in real-time because inference of trained models is not computationally expensive. The designed tool temporal output is shown in Figure 2.
This algorithm is implemented on three separate case-studies, which included sites in Miyagi and Iwate. A study is done on the variation of parameters set by this method. The three main parameters are the short time interval length, the stride length from one interval to its subsequent window, and the decimation factor of the time-series data. In all three case-studies, the DWT ranks {3, 4, 5, 6, 7} were chosen, as they included the various frequency bands necessary for site response evaluation. The deep learning models are modified to optimize predicted Fourier amplitude spectra at target sites. The available. The available safety time is provided, and the results are evaluated in multiple forms.