A NEURAL NETWORK APPROACH FOR EARTHQUAKE EARLY WARNING SYSTEMS

Abstract

A neural network based approach for earthquake source parameter estimation has been tested The approach which is largely based on the new technology for Earthquake Early Warning —PreSEIS. The system was tested using a simulated array of strong motion sensors capable of real-time streaming. At regular time intervals after the first sensor is triggered the source information from the P —wave is fed to a two layer feed forward neural network (TIFF) for the estimation of epicentre and magnitude. The system was trained and tested using 779 earthquake scenarios (5.5<M<8) in the Himalayan belt. The system was found so fast that after 5 seconds of first sensor trigger the sources parameter estimates were found with much less error. Again it was found at much higher time windows after 10 seconds errors got considerably reduced. From time windows 3 seconds to 10 seconds the errors in latitude have come down from 0.22° to 0.08°.After 15 seconds system predicted to an error of +/- 0.06° (i.e., to an error offsets of 8 km), whereas for longitude the errors came down from 0.64° to 0.14° and then to 0.14°. For the second network the errors in magnitude have come down to as much as +/- 0.21 units. The analyses showed decreased uncertainty in the estimates over the time, which revealed a trade-off between the reliability of the estimates and the remaining warning time available