12:00 PM - 12:15 PM
[SCG51-11] Improvement of single station method in earthquake early warning using machine learning technique
Keywords:earthquake early warning, single station method, convolutional neural network, noise discrimination, train-induced vibration, epicentral distance estimation
Earthquake early warning (EEW) algorithms used in the Shinkansen (high-speed rail in Japan) and the Japan Meteorological Agency systems are equipped with single-station method that estimates source parameters such as epicentral distance (Δ) and magnitude (M) from the data observed by itself. This method can issue an alert earlier than using multiple station data, although the estimation should be less accurate. The Shinkansen accelerometers deployed at rail-side locations observe tremor data of running trains more frequently than due to an earthquake, suggesting that the rail-side seismometers need to discriminate seismic waves (that is, P waves) from the train vibrations using short data immediately after the shaking begins (typically 1 – 2 sec). In this study, we test a machine learning technique to see if the performances of the single station method are improved.
To check the discrimination performance between train vibration and seismic wave using machine learning, we analyze approximately 600 train-induced accelerograms recorded at the Hino civil-engineering testing site of Railway Technical Research Institute in Hino, Tokyo, which is nearby the Chuo line, and about 20,000 K-NET waveforms of Mj ‹= 9.0 and 0 ‹= Δ ‹= 200 km. The conventional technique used in the Shinkansen system looks at ratio between higher- and lower-frequency amplitudes based on recursive filters because running trains generate strong signal in the higher-frequency range in terms of the comparison with seismic data. This study uses convolutional neural network (CNN) method (He et al., 2016) and classifies the data into micro tremor, train vibration and earthquake in advance for supervised learning. As a result, the CNN provides the classification accuracies of train vibration and earthquake with more than 99%. We conclude that the machine learning technique can drastically upgrade the performance because both accuracies are about 90% using the conventional method.
We also introduce the CNN into the epicentral distance estimation in the single station method in which the C-Δ method is conventionally employed (Iwata et al., 2015). The same K-NET dataset mentioned above is utilized. We test several types of layer architecture for the CNN and consequently find that the best estimation accuracy is 0.190 in root mean square of logarithm error between the observed and predicted epicentral distances while it is 0.312 with the C-Δ method. We conclude that the machine learning can achieve a remarkable improvement in accuracy of the single station method in EEW.
To check the discrimination performance between train vibration and seismic wave using machine learning, we analyze approximately 600 train-induced accelerograms recorded at the Hino civil-engineering testing site of Railway Technical Research Institute in Hino, Tokyo, which is nearby the Chuo line, and about 20,000 K-NET waveforms of Mj ‹= 9.0 and 0 ‹= Δ ‹= 200 km. The conventional technique used in the Shinkansen system looks at ratio between higher- and lower-frequency amplitudes based on recursive filters because running trains generate strong signal in the higher-frequency range in terms of the comparison with seismic data. This study uses convolutional neural network (CNN) method (He et al., 2016) and classifies the data into micro tremor, train vibration and earthquake in advance for supervised learning. As a result, the CNN provides the classification accuracies of train vibration and earthquake with more than 99%. We conclude that the machine learning technique can drastically upgrade the performance because both accuracies are about 90% using the conventional method.
We also introduce the CNN into the epicentral distance estimation in the single station method in which the C-Δ method is conventionally employed (Iwata et al., 2015). The same K-NET dataset mentioned above is utilized. We test several types of layer architecture for the CNN and consequently find that the best estimation accuracy is 0.190 in root mean square of logarithm error between the observed and predicted epicentral distances while it is 0.312 with the C-Δ method. We conclude that the machine learning can achieve a remarkable improvement in accuracy of the single station method in EEW.