AI-Based Algorithms for Real-Time Anomaly Detection in Ground Penetrating Radar Data

Abstract: Geophysical inversion is a well-established method for conducting non-invasive surveys of underground anomalies, such as buried infrastructure. Typically, inversion is performed using numerical methods: gradient descent and physics-based simulations are used to create a subsurface model that accurately explains the surface surveys. However, these methods are extremely computationally expensive, sometimes taking days or weeks to complete. Additionally, the inverse problems are highly underdetermined, and must be carefully regularized based on expert input. This creates a challenging bottleneck for field operators, who must accept lengthy delays between conducting surveys and starting work.

AI-based methods can help overcome these challenges. We will discuss advances in this direction based on work for the Grid Overhaul with Proactive, High-Speed Undergrounding for Reliability, Resilience, and Security (GOPHURRS) project. Our approach uses thousands of geophysical field surveys to train convolutional neural networks that can perform subsurface anomaly detection in near real-time. We will discuss the strengths and limitations of this AI-based approach, best practices for using physical data with AI, and our approach to optimizing our overall workflow.

Mathematics

Audience: researchers in the topic

PIMS-CORDS SFU Operations Research Seminar