5:15 PM - 6:45 PM
[SCG53-P01] Enhancement of single station method for earthquake early warning systems using machine learning techniques
Keywords:machine learning, earthquake early warning, single station method, deep learning, convolutional neural network, gradient boosting decision trees
The earthquake early warning (EEW) systems of the Japan Meteorological Agency and Japanese high-speed rail (Shinkansen) operated by the JR companies employ the single station method that independently analyzes seismic records and issues an alert when necessary (Yamamoto & Tomori, 2013). Although this algorithm is less accurate in the estimation of seismic parameters such as epicenter location and earthquake magnitude than the multiple station method, it has the advantage of immediacy because there is no need to wait for seismic waves to arrive at other stations surrounding the source. In railways, there is a high compatibility with the use of single station method, as trains can be promptly halted if there is any risk or danger, and the operations can resume immediately upon safety confirmation. However, the need for enhanced performance, such as accuracy of alarms, has arisen due to various earthquakes. In this study, we applied recently developed machine learning techniques to the single station method and investigated their effectiveness.
For this problem, Noda et al. (2023) investigated using a Convolutional Neural Network (CNN), a typical deep learning technique. Additionally, Noda (2024, BSSA [in press]) utilized an automatic structure search (AutoSS) technique that is based on CNN. The AutoSS technique allows to select a model from numerous deep learning models that match the computational resources of seismometer, although models with higher accuracy generally come with increased computational loads. Thus, deep learning techniques tend to have higher computational loads during estimation, posing implementation challenges for the single station method on seismic instruments with limited computational resources. In this study, we evaluated Gradient Boosting Decision Trees (GBDT), known for relatively low computational demands. As a result, the estimated accuracy of seismic parameters obtained with GBDT was found to be nearly equivalent to that of CNN. Furthermore, GBDT exhibited an estimation execution time of approximately 3/100 on average compared to CNN. We conclude that the utilization of GBDT is effective.
The Shinkansen seismometers installed along the rails (referred to as rail-side seismometers) observe vibrations associated with the passage of trains. The single station method estimates seismic parameters using data as short as 1 second from the P-wave onset. Therefore, it is necessary to accurately classify from such short data whether the observed vibration is due to a passing train or an earthquake. In the method used in the current Shinkansen EEW system, the accuracy for this classification is about 90% (Iwata et al., 2015). Increasing this accuracy is directly related to reliable P-wave warnings during earthquakes, thereby improving the safety of running trains. Noda et al. (2022) reported that applying a CNN to this problem could improve the accuracy to about 97%. However, that study did not consider the time-series connection of the data, and the CNN model was trained directly on the amplitude data of each cut-out time window. Considering that, this study will apply a recurrent neural network method, such as LSTM (Long Short Time Memory), which can account for temporal recursion.
For this problem, Noda et al. (2023) investigated using a Convolutional Neural Network (CNN), a typical deep learning technique. Additionally, Noda (2024, BSSA [in press]) utilized an automatic structure search (AutoSS) technique that is based on CNN. The AutoSS technique allows to select a model from numerous deep learning models that match the computational resources of seismometer, although models with higher accuracy generally come with increased computational loads. Thus, deep learning techniques tend to have higher computational loads during estimation, posing implementation challenges for the single station method on seismic instruments with limited computational resources. In this study, we evaluated Gradient Boosting Decision Trees (GBDT), known for relatively low computational demands. As a result, the estimated accuracy of seismic parameters obtained with GBDT was found to be nearly equivalent to that of CNN. Furthermore, GBDT exhibited an estimation execution time of approximately 3/100 on average compared to CNN. We conclude that the utilization of GBDT is effective.
The Shinkansen seismometers installed along the rails (referred to as rail-side seismometers) observe vibrations associated with the passage of trains. The single station method estimates seismic parameters using data as short as 1 second from the P-wave onset. Therefore, it is necessary to accurately classify from such short data whether the observed vibration is due to a passing train or an earthquake. In the method used in the current Shinkansen EEW system, the accuracy for this classification is about 90% (Iwata et al., 2015). Increasing this accuracy is directly related to reliable P-wave warnings during earthquakes, thereby improving the safety of running trains. Noda et al. (2022) reported that applying a CNN to this problem could improve the accuracy to about 97%. However, that study did not consider the time-series connection of the data, and the CNN model was trained directly on the amplitude data of each cut-out time window. Considering that, this study will apply a recurrent neural network method, such as LSTM (Long Short Time Memory), which can account for temporal recursion.