Deep Learning for AGT
| Date: | May 25, 2018 |
|---|---|
| Last Updated: | May 27, 2018 |
| Categories: | |
| Tags: |
Contents
Introduction
This project is for AGT (Advanced Geophysical Technology) and 2018 SEG (Society of Exploration Geophysicists) conference. We concentrate on the seismic raw data and use deep learning approaches to improve the performance in some special application.
Beat Tone Network
The first application is for improving the FWI of Beat Tone, see here to learn the background of this application:
When the basic signal frequency increases, the effectiveness of the Beat Tone method reduces significantly. To solve this problem, we may use the synthetic low frequency data to perform the supervised learning and let the deep network predict the low frequency data based on high frequency ones. We feed the normalized 1D raw data and beat tone result into the network and get the prediction for low frequency results. The project could get good results for FWI inversion.
Denoising
Viewing the results collected by shot-gather devices as image, we may be able to use the deep learning tools to denoise the data. Now we only consider the supervised learning approach, which requires us to add random white noise to the clean synthetic data and train the network to let it regress on the clean data.
Indeed, this approach is a simple attempt and not good enough, we may try to explore how to perform the unsupervised learning with structured noise, which is more valuable for real applications.
Project Page
See here to view the main page of this project. This project has not been published yet, so in the main page we only provides some background and theoretical knowledge. You could not see any information for the real applications and data here.