← Back to Sessions

Approximate Inference of the EEW Event Parameters Using MCMC and Bayesian Networks

Session: Earthquake Early Warning Live in California! Current Status and Challenges I

Type: Oral

Date: 4/23/2021

Presentation Time: 10:15 AM Pacific

Description: 

An efficient earthquake early warning system (EEWS) requires accurate estimates of the seismic event parameters in order to translate them into reliable warning times. From the seismic risk reduction prospective, it is critical to perform uncertainty analysis of the event parameters derived from the first few stations which detected the P wave. This sparse data set, which provides warning in the region with the largest ground shaking, tends to exhibit the largest uncertainties in the magnitude and location estimates. At the same time, increasing the data set by incorporating additional stations will increase the blind zone around the epicenter. The use of alternative methods to compute the event parameters allows us to maximize the entropy and place additional constraints on the density functions. We employ the Bayesian networks to propagate the updates and compute the distributions associated with the nodes. The presence of large networks or non-Gaussian distributions suggests the use of the approximate inference methods. Markov chain Monte Carlo method offers an efficient tool to sample from the probability space in order to construct posterior distributions of the required event parameters. We illustrate our approach using a set of records in terms of accelerations and velocities from M4.8 earthquake occurred 19km NNE of Victoria, British Columbia on December 29th, 2015. Some of these records were obtained from the dedicated EEWS stations operating since 2009.

Presenting Author: Anton G. Zaicenco

Student Presenter: No

Authors