1:45 PM - 3:15 PM
[SCG55-P04] Development of Seismic Wave Detection System using Machine Learning for a Citizen Seismic Network
Keywords:Machine Learning, Deep Learning, Phase Pick, Raspberry Pi
Since Japan is one of the most seismically active country, various organizations have developed high-density seismic networks, which enable the immediate publication of seismic intensity at each location and earthquake early warning signals effectively. However, it is currently impossible to measure seismic intensity at a resolution finer than 20 km, which is the distance between seismographs, and earthquake early warning cannot deal with earthquakes occurring within an area of about 30 km. In addition, in recent years, high-rise buildings have been constructed rapidly in urban areas, and even within the same building, damage appears differently depending on the floor. In other words, for urban areas with a high population concentration, current seismic intensity measurements may not be sufficient for understanding and predicting damage.
Therefore, in this study, we propose a low-cost seismometer that can be installed in buildings and houses in addition to the existing seismic network. This seismic network is called CSN (Citizens Seismic Network) (Kim, 2016) and is being installed at several locations in Yokohama, Kanagawa Prefecture. Our low-cost seismometer consists of a MEMS accelerometer (Witmotion BWT901CL), which is relatively inexpensive and readily available, and a Raspberry Pi, a small computing device, to analyze the contentious seismic data obtained from the accelerometer.
In this presentation, we will focus on the construction of the algorithm to detect earthquake with the above mentioned seismometers. When the classical earthquake detection method, sta/lta ratio, was applied to sensor data, not only true earthquakes but also noises more than twice as large were misjudged as earthquakes. To solve this problem, we consider an approach using a machine learning model. The phase detection model is R2AU-Net, whose performance was evaluated by Nakamura et al. (JpGU 2023).
Since CSN has not accumulated a sufficient amount of data, we use the data obtained from K-Net, a strong-motion seismograph network operated by the National Research Institute for Earth Science and Disaster Prevention (NIED), for the training and validation. Since there is no phase-pick information in the K-Net data, we picked P- and S-wave onset using the ar_picker module in the ObsPy library and considered them as the ground truth. However, this resulted in low accuracy prediction in the evaluation. This was because the phase detection in the ar_picker module was not working well. Therefore, in order to build a model without using large-scale K-NET data, the first stage of training was performed using STEAD (Mousavi, 2019), large-scale global dataset of labeled earthquake. And then, we performed a transfer learning using small-scale K-NET data with manually picked P- and S-wave onset. The created model will be applied to actual CSN data to attempt to detect seismic phases and verify its performance. Furthermore, the possibility of the implementing the model on the Raspberry Pi will be discussed.
Therefore, in this study, we propose a low-cost seismometer that can be installed in buildings and houses in addition to the existing seismic network. This seismic network is called CSN (Citizens Seismic Network) (Kim, 2016) and is being installed at several locations in Yokohama, Kanagawa Prefecture. Our low-cost seismometer consists of a MEMS accelerometer (Witmotion BWT901CL), which is relatively inexpensive and readily available, and a Raspberry Pi, a small computing device, to analyze the contentious seismic data obtained from the accelerometer.
In this presentation, we will focus on the construction of the algorithm to detect earthquake with the above mentioned seismometers. When the classical earthquake detection method, sta/lta ratio, was applied to sensor data, not only true earthquakes but also noises more than twice as large were misjudged as earthquakes. To solve this problem, we consider an approach using a machine learning model. The phase detection model is R2AU-Net, whose performance was evaluated by Nakamura et al. (JpGU 2023).
Since CSN has not accumulated a sufficient amount of data, we use the data obtained from K-Net, a strong-motion seismograph network operated by the National Research Institute for Earth Science and Disaster Prevention (NIED), for the training and validation. Since there is no phase-pick information in the K-Net data, we picked P- and S-wave onset using the ar_picker module in the ObsPy library and considered them as the ground truth. However, this resulted in low accuracy prediction in the evaluation. This was because the phase detection in the ar_picker module was not working well. Therefore, in order to build a model without using large-scale K-NET data, the first stage of training was performed using STEAD (Mousavi, 2019), large-scale global dataset of labeled earthquake. And then, we performed a transfer learning using small-scale K-NET data with manually picked P- and S-wave onset. The created model will be applied to actual CSN data to attempt to detect seismic phases and verify its performance. Furthermore, the possibility of the implementing the model on the Raspberry Pi will be discussed.