A Neural Network-Based Approach for On-site Earthquake Early Warning System

A Neural Network-Based Approach for On-site Earthquake Early Warning System

The on-site earthquake early warning system is under development for the area near the earthquake epicenter to provide information such as earthquake magnitude, the arrival time and the intensity of the strong shaking for free field as well as the structural response, etc. The real-time strong motion signals recorded from Taiwan Strong Motion Instrumentation Program(TSMIP) were used to train neural networks and the characteristics of the sensed earthquake accelerograms were learned. The neural networks provide a seismic profile of the arrival ground motion instantaneously after the shaking is felt at the sensors by analyzing the three components of the earthquake signals. Each data sample, consist of the first 1~10 second envelope of the complete earthquake accelerogram, was used as the input of the neural networks. The output of the neural networks provides estimates of the structural response and the time for emergency action. The neural network based algorithm is trained with 50149 accelerograms from 2505 earthquakes recorded in Taiwan. By producing informative warnings, the neural network based methodology has shown its potential to increase significantly the application of earthquake early warning system(EEWS) on hazard mitigation.