EPS: An earthquake prediction system using Federated Learning

Abstract

Earthquake prediction is a challenging job and researchers are struggling to predict earthquakes since long time. The involvement of huge multidimensional data, data privacy considerations and transmission latency make it difficult to predict earthquakes in real time. In this research, an earthquake prediction system ‘EPS’ has been developed using python. It is unique among early earthquake warning systems as federated learning technique (FL) has been applied for incorporating the intelligence from cloud to the mobile edge devices. Features like onsite data collection, training of local data models and aggregation of these models for the generation of global data model makes FL capable enough to handle data privacy, data availability and resolve latency issues involved in earthquake prediction. The proposed system integrates multiple data sources using FL for generation of global data model, applies time series regression for earthquake prediction and transmits earthquake alerts to all connected users and agencies. Dataset from Western Himalayan zone has been collected from 1985 to 2020 considering five different aspects including seismicity, atmospheric temperature, atmospheric pressure, regional rainfall, and reservoir parameters using federated learning technique for experimentation. EPS has successfully predicted 97.96% of the earthquakes that have encountered in the region during past thirty-five years.