Full text loading...
-
Multimodal Contrastive Learning on Time Series: Application to Various Scenario for Seismic Monitoring on Mars
- Publisher: European Association of Geoscientists & Engineers
- Source: Conference Proceedings, 86th EAGE Annual Conference & Exhibition, Jun 2025, Volume 2025, p.1 - 5
Abstract
Summary
Advances in Machine Learning (ML) and embedded systems have enabled the development of smart sensors and autonomous devices capable of both data acquisition and in-situ analysis for time series data (e.g., in acoustic, seismic …). These systems allow continuous monitoring, selectively transmitting or storing only the most relevant information.
This paper focuses on self-supervised learning, particularly contrastive learning with 1D-CNN models, to process in-situ stream datasets under varying labeling constraints. Unlike traditional supervised or unsupervised models, self-supervised learning leverages automatically labeled pretext tasks to extract robust feature representations without manual annotation. We investigate the impact of contrastive learning in two continuous monitoring scenarios using a multimodal dataset from the InSight mission on Mars, featuring seismic and pressure data. The first scenario employs supervised learning for continuous classification, optimized through contrastive pretraining with multiple data augmentation methods. The second scenario, based on one-class classifiers, detects unexpected events using strong features extracted via contrastive learning. While illustrated with Mars data, these approaches are applicable to other continuous monitoring contexts.
© EAGE Publications BV