Evaluation and Application of Machine Learning Techniques to Data Conditioning Problems in Microseismic Data
- Author(s): Nava, Michael J
- Advisor(s): Rector, James W
- et al.
Hydraulic fracturing has evolved dramatically over the past decades. A number of new techniques have emerged in order to maximize production from organic-rich shale. For example, multistage fracturing, dynamically varying pumping parameters, horizontal drilling and finely-tuned perforation shots have all led to incremental improvements in the industry. With these engineering advancements, so too has the ability to monitor microseismic fractures expanded. An added benefit, or potentially an unintended consequence, of this new era of high frequency, high precision acoustic monitoring equipment is the generation of large scale digital data. With any real data set, there will inevitably be data conditioning problems that exist. Whether missing values, corrupt data, or poor experimental design and execution, there will be some constraint or obstacle that inhibits the cultivation of knowledge and insights.
The objective of this dissertation is to identify and understand where those limitations exist, to understand the genesis of those constraints - whether they arise from some physical limitation or from common data recording issues - and then apply an interdisciplinary approach to overcome those limitations.
To this end, we identify limitations caused by a typical, cost-effective microseismic monitoring geometry and pivot to understand and characterize microseismic events through spectral analysis. We build features that provide insight into the nature of microseismicity present in the data, which would otherwise elude us. Next, we incorporate information that is typically lost in the presence of high amplitude resonance and leverage this newly found data to identify specific microseismic attributes to make marked improvements on event location estimates. Through the inclusion of head waves and the use of inversion techniques, we reduce the uncertainty of microseismic event locations significantly. This is a fundamental step toward understanding the behavior of hydraulic fractures far beneath the surface of the earth.
Next, we turn to data science to continue to overcome data quality issues present in the data from a hydraulic fracturing project in the Marcellus shale. Specifically, machine learning and deep learning methodologies are applied to the data in order to recover meaningful information. The benefits of this are twofold. First, this work provides a data-driven approach to imputation through various learning methods. Second, it provides an understanding of the limitations and computational time required for various learning methods. This information will aid in the decision making of engineers who desire a more accurate solution or an accurate solution that can be used in real-time analysis.
Finally, we culminate the dissertation with an exploration into the ability to leverage ensemble learning methods to overcome poorly conditioned data sets with the objective of improving automated analysis steps. Specifically, we create an extensible computational paradigm that enables the automatic picking of waveform first arrivals. This is typically an arduous, time-consuming analysis step that suffers from inconsistent picks based on subjective assessment. Moving away from a human-in-the-loop system enables more transparency and reproducibility. Additionally, the total time for end-to-end analysis of first arrivals is dramatically decreased. Given the extensibility of this framework, expanding the use of the system to include full waveform classification is an appropriate next step.