5:15 PM - 6:30 PM
[SCG52-P02] Toward development of dense low cost citizen seismic network: Performance comparison of different types of neural network architectures
Keywords:Neural network, Machine Learning, Seismic signal, Seismometer, Strong motion, CNN
We have developed a community based low cost seismic network in Yokohama, Japan, called Citizen Seismic Network (CSN), to monitor local scale strong motion which is closely linked to community’s life. Each sensor unit composed of 12 bit MEMS accelerometer and Raspberry Pi.
Since the units are supposed to be installed under high-noise condition such as inside of house where spiky noise made by human activities often misinterpreted as seismic signal, it is difficult to discriminate the seismic signals from other noises. In such condition, applying conventional detection method using amplitude ratio (e.g. STA/LTA) is problematic. To overcome the issue, we employed an artificial neural network that utilizes pattern recognition to retrieve seismic signals. In this study, we compared the 3 different types of neural networks, namely, 3 layered shallow network (ANN), convolutional neural network, and recurrent neural network. For the training and validation, we prepared numerical data set produced by seismic records obtained from K-net adding CSN sensor noise (~90000 for training and ~40000 for validation). Since it is possible the amount of the data is not enough for CNN and RNN, we also tested transfer learning using pre-trained model open to public. As preliminary results, we found that using pre-trained ANN on the Raspberry Pi in real time and making final decision using deep neural network pre-trained by transfer learning on server is the most effective.
Since the units are supposed to be installed under high-noise condition such as inside of house where spiky noise made by human activities often misinterpreted as seismic signal, it is difficult to discriminate the seismic signals from other noises. In such condition, applying conventional detection method using amplitude ratio (e.g. STA/LTA) is problematic. To overcome the issue, we employed an artificial neural network that utilizes pattern recognition to retrieve seismic signals. In this study, we compared the 3 different types of neural networks, namely, 3 layered shallow network (ANN), convolutional neural network, and recurrent neural network. For the training and validation, we prepared numerical data set produced by seismic records obtained from K-net adding CSN sensor noise (~90000 for training and ~40000 for validation). Since it is possible the amount of the data is not enough for CNN and RNN, we also tested transfer learning using pre-trained model open to public. As preliminary results, we found that using pre-trained ANN on the Raspberry Pi in real time and making final decision using deep neural network pre-trained by transfer learning on server is the most effective.