*********************************
There is now a CONTENT FREEZE for Mercury while we switch to a new platform. It began on Friday, March 10 at 6pm and will end on Wednesday, March 15 at noon. No new content can be created during this time, but all material in the system as of the beginning of the freeze will be migrated to the new platform, including users and groups. Functionally the new site is identical to the old one. webteam@gatech.edu
*********************************
Title: Seismic Processing via Machine Learning for Event Detection and Phase Picking
Committee:
Dr. James McClellan, ECE, Chair , Advisor
Dr. Ghassan AlRegib, ECE
Dr. Zhigang Peng, EAS
Dr. David Anderson, ECE
Dr. Andrew Newman, EAS
Abstract:
The increasing volume of seismic data from long-term continuous monitoring motivates the development of algorithms based on convolutional neural network (CNN) for faster and more accurate event detection and phase picking. However, existing CNN-based approaches depends on large training sets with millions of labeled samples which are only available for a limited number of well-studied regions. Moreover, inference of the large models from existing studies is only possible on cloud servers where multiple powerful GPUs are available. With the rapid development of seismic recording devices and connected sensor networks, lightweight yet accurate models are necessary for deploying neural network solutions on the edge devices. This thesis prototypes a feasible system for seismic event detection and phase picking on an embedded system with seismic sensors using a lightweight CNN model. A flexible CNN model capable of training on only tens of thousands of labeled samples is designed and validated on both small and large datasets. The deployment of the CNN model on continuous seismic waveform has resulted in newly detected events and accurate arrival times on multiple datasets from different geological regions. Transfer learning of the proposed CNN model is validated on small labeled datasets with and without fine-tuning. The CNN model is simplified based on the visualization of the internal CNN filter weights distribution. Quantization using lower-precision number of both data and filter weights are explored for future embedded devices with fixed-point number acceleration enabled.
