← Back to Gallery

A Hybrid Deep Learning Framework for Denoising Distributed Acoustic Sensing Data

Distributed Acoustic Sensing (DAS) is a powerful tool in seismic monitoring, enabling high-resolution strain detection over long distances with fiber-optic cables. This cost-effective technology has diverse applications in seismology, but DAS recordings are often affected by erratic and horizontal noise, compromising their effectiveness in seismic analyses.

To address these challenges, we propose a hybrid denoising framework that combines traditional signal processing techniques with advanced deep learning. The core of the framework builds on Deep Image Prior (DIP), an unsupervised deep learning method that excels at mitigating random noise. However, DIP struggles with complex, overlapping noise types due to limitations in network design and hyperparameter sensitivity. To overcome these issues, we integrate DIP with a Multi-Head Attention Regression Network (MH-ARN), which incorporates fully connected layers and multi-head self-attention blocks to capture significant details in DAS data.

Our approach starts with dynamic patch-based processing using Non-Local Means (NLM) for high-frequency noise suppression. A variance-based patch selection targets signal-rich regions for training. Horizontal noise is mitigated through a frequency-wavenumber dip (FK-dip) filter applied after the first iteration, with iterative MH-ARN refinements ensuring effective noise suppression.

Applications to synthetic and real-world DAS datasets from the FORGE geothermal field show that our framework significantly improves the signal-to-noise ratio while preserving key details. It outperforms conventional DIP and integrated denoising methods, which combine operators like bandpass, structure-oriented median, and FK-dip filters. This multi-step process enhances signal quality and overcomes the limitations of traditional DIP methods, offering a robust solution for complex noise in DAS applications.

Session: Fiber-optic Sensing Applications in Seismology [Poster]

Type: Poster

Room: Exhibit Hall

Date: 4/15/2025

Presentation Time: 08:00 AM (local time)

Presenting Author: Yapo Abole Serge Innocent Oboue

Student Presenter: No

Invited Presentation: 

Poster Number: 59

Additional Authors