← Back to Sessions

Deep Implicit Time Series Modeling for Earthquake Phase Picking on Edge Devices

Description: 

Earthquake monitoring and understanding seismic activities are pivotal in assessing seismic hazards and ensuring community resilience. One critical aspect of earthquake analysis involves accurate phase picking, where precise identification of seismic wave arrival times is fundamental for robust event characterization. Traditionally, seismic phase picking demands labor-intensive efforts and often struggles with noises, irregularities, and other complexities inherent in seismic data, until the emergence of deep learning-based automatic phase pickers. State-of-the-art models like EQTransformer and Phasenet have transformed this task, exhibiting remarkable precision in detecting and classifying seismic phases, significantly augmenting the efficiency of earthquake monitoring. However, these advanced models often rely on high-powered servers with multiple GPUs, limiting accessibility, especially for those reliant on mobile or edge devices (e.g. Raspberry Shakes) that lack significant computing capabilities. To address these challenges, we explore a novel deep-learning framework – times series modeling with a deep implicit model – on the seismic phase picking problem. The implicit model uses a state-driven approach, which distinguishes itself from its traditional, layered-based deep learning counterparts. It is capable of emulating classical feedforward-based neural networks while offering promise in compressing models without exhaustive retraining. We plan to first train the implicit model on the STEAD dataset. The trained model will then be compressed by utilizing the low-rank technique. An offline performance test on data collected by Raspberry Pi and Shake will then be performed to understand the effectiveness of model compression. We hope this approach advances seismic phase picking accuracy and pioneers a pathway to democratize deep learning for seismic monitoring on resource-constrained devices.

Session: Leveraging Cutting-Edge Cyberinfrastructure for Large Scale Data Analysis and Education [Poster Session]

Type: Poster

Date: 5/2/2024

Presentation Time: 08:00 AM (local time)

Presenting Author: Alicia

Student Presenter: Yes

Invited Presentation: 

Authors