The machine learning aspects in building the global smartphone seismic network

Abstract

MyShake is a global crowdsourcing smartphone seismic network to monitor and detect earthquakes. After it got released to the public in 2016, we arrived at more than 300,000 downloads with more than 800 detected earthquakes globally within 2 years. Machine learning plays a critical role in MyShake that makes everything happen. In this presentation, I will present the details of how we use the artificial neural network to distinguish earthquakes from the human activity movements recorded on a single phone in real-time. This includes how we do the data acquisition, pre-processing the data, addressing imbalanced datasets, feature engineering/selection, and evaluating the model. I will also talk other machine learning aspects in the MyShake network including the convolutional neural network that we built on the server to further classify the whole waveforms to find that caused by earthquakes, the adversarial machine learning for securities of the system, dealing with the dynamically changing network, training customized model for each user etc. These machine-learning applications in MyShake illustrate the power of combining data science and geophysics and provide good examples of how do we better facilitate the interactions of the two fields.MyShake is a global crowdsourcing smartphone seismic network to monitor and detect earthquakes. After it got released to the public in 2016, we arrived at more than 300,000 downloads with more than 800 detected earthquakes globally within 2 years. Machine learning plays a critical role in MyShake that makes everything happen. In this presentation, I will present the details of how we use the artificial neural network to distinguish earthquakes from the human activity movements recorded on a single phone in real-time. This includes how we do the data acquisition, pre-processing the data, addressing imbalanced datasets, feature engineering/selection, and evaluating the model. I will also talk other machine learning aspects in the MyShake network including the convolutional neural network that we built on the server to further classify the whole waveforms to find that caused by earthquakes, the adversarial machine learning for securities of the system, dealing with the dynamically changing network, training ...