Expanding Wavelet-Transform-Based Neural Network Denoiser Performance Using Utah Regional Data
Description:
Seismic waveform data can be interpreted as a superposition of some target signal defined as being from a unique source of interest and noise generated from all other sources. Previous projects focused on the development of convolutional neural network (CNN) based denoisers have shown the viability of incorporating discrete wavelet transforms (DWTs) to improve performance on both constructed test and real-world data. The advantages of utilizing the DWT in the U-Net style architecture are that, when compared to conventional pooling functions, the wavelet transform functions allow for a higher degree of information retention between model layers. Evaluation of both models on test data showed that the MWCNN model (average cross-correlation, CC, value of 0.85) outperformed the CNN model (average CC value of 0.75) in its ability to recover the ground truth signal component with little amplitude distortion. Ultimately, the Multi-level Wavelet-based Convolutional Neural Network (MWCNN) denoiser outperformed conventional CNN-based denoisers when trained on the same single-station based training data set. Here, we utilize a data set of over 300,000 constructed seismograms using recordings from three-component stations spanning the entirety of the University of Utah Seismograph Stations network to improve the transferability of the MWCNN denoiser to broader regions of seismicity. Additionally, we seek to improve the performance of the MWCNN denoiser on real-world datasets by expanding the real-world data evaluation set beyond its prior single-station single-channel extent. Lastly, we demonstrate the ability of the MWCNN denoiser to both improve detection rates and increase signal-to-noise ratio (SNR) values for those detected events, while not distorting the target signal amplitude values, on continuous segments of data.
Session: Opportunities and Challenges for Machine Learning Applications in Seismology [Poster]
Type: Poster
Date: 4/19/2023
Presentation Time: 08:00 AM (local time)
Presenting Author: Louis A. Quinones
Student Presenter: No
Invited Presentation:
Authors
Louis Quinones Presenting Author Corresponding Author laquino@sandia.gov Sandia National Laboratories |
Rigobert Tibi rtibi@sandia.gov Sandia National Laboratories |
|
|
|
|
|
|
|
Expanding Wavelet-Transform-Based Neural Network Denoiser Performance Using Utah Regional Data
Category
Opportunities and Challenges for Machine Learning Applications in Seismology