Laboratory characterization of seismicty and development of a deep learning framework to denoise seismic data in the field

Evani, Sai Kalyan

Permalink

https://hdl.handle.net/2142/117527 Copy

Description

Title

Laboratory characterization of seismicty and development of a deep learning framework to denoise seismic data in the field

Author(s)

Evani, Sai Kalyan

Issue Date

2022-08-30

Director of Research (if dissertation) or Advisor (if thesis)

Popovics, John

Doctoral Committee Chair(s)

Popovics, John

Committee Member(s)

• Reis, Henrique
• Makhnenko, Roman
• Williams-Stroud, Sherilyn

Department of Study

Civil & Environmental Eng

Discipline

Civil Engineering

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• induced seismicity
• acoustic emission
• X-ray CT
• frictional slipping
• NDT
• machine learning
• deep learning
• signal processing
• denoising

Language

eng

Abstract

Subsurface injection encompasses a wide range of applications, for example injection of carbon dioxide is performed to reduce the impact of increasing levels of carbon dioxide contributing to climate change and injection of wastewater from oil and natural gas production is performed to avoid pollution of land and surface water bodies and to protect public health. Increases in seismic activity near injection sites have been observed and attributed to fluid injection process in several projects. Most of these induced events (especially in the case of carbon capture and storage) are micro-seismic in nature (magnitude<1.0 on the Richter scale) and are not typically felt by people on the surface. However, there have been documented instances of felt and damaging induced seismic events at injection sites in Oklahoma, northern California, South Korea, and Switzerland. Understanding the underlying mechanisms that give rise to these behaviors, and characterizing the factors that affect induced seismic events, is important to ensure public safety and improve public acceptance of activities that require subsurface fluid injection. This study aims to (1) better understand induced seismicity in faulted/fractured subsurface rock formations by designing and applying a carefully controlled laboratory-scale experimental test series that overcomes some of the drawbacks of previous studies, and (2) develop a signal denoising scheme to extract useful information from noisy induced seismic events recorded in the field to facilitate a better understanding of source mechanisms. To address the first objective, a series of experiments are conducted on cemented sand specimens, instead of rock cores. Test results demonstrate that cemented sand specimens can be used as a mechanical alias for rock cores at a laboratory scale. A protocol to manufacture cemented sand specimens with faults at predetermined orientations (30o and 45o relative to the cross section) is presented and the geometry of both surfaces on either side of the fault is characterized using X-ray CT. The roughness of the fault surface is quantified by deconstructing the surface geometry across different length scales and computing the roughness for each scale. Test specimens are tested in (1) core flooding, and (2) triaxial configurations to study slipping at the small and large scales respectively. The extent of fault gouging in a specimen is quantified by defining a cumulative gouging parameter using X-ray CT measurements before and after testing. Acoustic emission (AE) events emanating from the specimens in both configurations are monitored. An artificial intelligence-based deep event detection criterion is presented to better identify low energy AE events. Contribution of different frequency ranges to the overall amplitude of the AE signal is quantified by defining spectral ratio of frequency bins. It is demonstrated that the orientation of fault relative to the major principal stress direction affects the slip characteristics. Shear stress accumulated along the surface is used to quantify the extent of slipping in specimens tested in the core flooding configuration. With increase in the extent of slipping, the accumulated shear stress increases. The average accumulated shear stress for specimens with a 45° nominal fault orientation is more than twice of the accumulated shear stress for specimens with 30° nominal fault orientation. In addition to the fault orientation, it is also shown that the fault geometry affects slip characteristics. For specimens with the same nominal orientation, an increase in surface roughness at a scale close to the scale of slipping results in a decrease in the stress accumulated along the surface indicating that the extent of slipping reduces with increase in roughness. It is demonstrated that frictional slipping at a fault can occur at different length scales. The same fault can slip either at the small or large scales, depending on the deviatoric stress applied on the specimen. The specimens with a preexisting fault tested in the triaxial configuration slipped at a small scale for low levels of axial load (<20 kN, approximately 1900 psi deviatoric stress). At high levels of axial load (>80 kN, approximately 14,100 psi deviatoric stress), the same fault started slipping at the larger scale. It is also demonstrated that with an increase in the scale of slipping, the spectral ratio of (1) low frequency bins increases, and (2) high frequency bins reduces. In other words, AE generated by slipping at the small scale is dominated by high frequency components, whereas that from slipping at a large scale is dominated by low frequency components. It is shown that locations on the fault surface that exhibit a sudden change in surface normal are most susceptible to gouging/damage during frictional slipping. Also, the cumulative gouging observed on fault surfaces increases with increase in scale of slipping. To address the second objective, a deep learning framework: DDN-Net is developed and deployed to denoise low signal to noise ratio (SNR) micro-seismic events recorded at the Illinois Basin Decatur project injection site. The network is trained using only noise data from each geophone in the monitoring array. It is shown that p- and s-wave arrivals become more distinct when signal data are denoised using DDN-Net compared to bandpass filtering; on an average deep denoise data had a 19 dB higher SNR compared to bandpass filtered data; and the deep denoising micro-seismic data using DNN-Net results in a 25% reduction in uncertainty of the hypocenter location compared to traditional bandpass filtering.

Graduation Semester

2022-12

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/117527

Copyright and License Information

2022 Sai Kalyan Evani. All rights reserved.

Owning Collections