Deep Learning Probabilistic Regression for Onset Time Determination
Abstract
A software tool was developed to determine a seismic phase's arrival time, which is essential for subsequent processes, and can unburden analysts from manual onset time picking. The tool is a deep learning (DL) algorithm using a convolutional neural network to analyze the raw signal. The algorithm is a regression model that makes a probabilistic estimation. The onset times are a continuous interpolation between samples, so the algorithm could be more accurate than the discrete sample rate. The probabilistic output is achieved using an ensemble mask process. The variation in the output provides a measure of the uncertainty. The models were trained on International Data Center data. The final model had a mean absolute error of 0.574s over the test dataset, and by using the estimated STD of the probability as a measure of prediction confidence, the mean absolute error is reduced to 0.439s. Initially, the algorithm was trained and tested on P-waves but later extended to simultaneously identify the presence of P- and/or S-waves and their onset time.
Document Details
- Document Type
- Technical Report
- Publication Date
- Mar 29, 2023
- Accession Number
- AD1211554
Entities
People
- Artemii Novoselov
- Greg Beroza
- Jesse Williams
- John Pace
Organizations
- Stanford University