← Back to Sessions

A Novel Physics-guided Contrastive Learning Strategy for Seismic Signal Analysis

Description: 

A moving seismic source generates signatures that change both temporally and spectrally. When combined with the seismic propagation due to the laterally varying structure of the shallow crust, we can expect significant variations in the received seismic signal. We present a contrastive learning strategy that uses temporal similarity, derived from time-frequency coherence, to estimate and aggregate consistent seismic signal patterns for complex anthropogenic sources. By leveraging physics-guided frequency-based temporal analysis, our approach captures the non-stationary characteristics of seismic data more effectively than standard methods, and under multiple propagation conditions.
To evaluate our approach under varying noise levels and signal complexities, we use three seismic and acoustic datasets for a variety of ground vehicles: (1) MOD, a self-collected set of seven vehicles plus a dismounted human; (2) ACIDS, containing nine ground vehicles traveling at speeds from 5-40 km/h; and (3) NSIN, featuring eleven commercially available vehicles in three different environments under varying speeds, distances, and directions. Using our physics-augmented embeddings in a lightweight downstream classification model (using just a few neural network layers), we create classifiers that are robust to the inherent non-stationarity in seismic signals. Experimental results show that time-frequency coherence yields more accurate inter-sample similarity assignments and improves downstream tasks with up to 7% higher target classification accuracy while requiring fewer labeled samples than conventional contrastive learning frameworks. These findings demonstrate the strong potential of the Foundation Models for advanced seismic signal understanding under varying noise and propagation conditions.

Session: Data-driven and Computational Characterization of Non-earthquake Seismoacoustic Sources - I

Type: Oral

Date: 4/16/2025

Presentation Time: 02:30 PM (local time)

Presenting Author: Denizhan

Student Presenter: Yes

Invited Presentation: 

Poster Number:

Authors