Abstract
Distributed optical fiber acoustic sensing (DAS) technology is increasingly vital for vertical seismic profile (VSP) exploration due to its high spatial resolution and resilience to extreme conditions. However, DAS seismic data is often compromised by complex noise types, including random, optical abnormal, fading, and horizontal noise. These overlapping noise profiles challenge traditional denoising methods. This paper introduces ADNet, an attention-guided denoising convolutional neural network designed to effectively improve the signal-to-noise ratio (SNR) of DAS data. By integrating a sparse block (SB), feature enhancement block (FEB), attention block (AB), and reconstruction block (RB), ADNet extracts global features and hidden noise information while preserving weak seismic signals. Results from synthetic and field data demonstrate that ADNet outperforms conventional methods and standard deep learning networks in both signal retrieval and noise suppression.
Introduction: The DAS Noise Challenge
Distributed Acoustic Sensing (DAS) has revolutionized seismic exploration by converting standard fiber-optic cables into dense arrays of high-resolution sensors. Despite these advantages, DAS data is notoriously susceptible to a "cocktail" of complex noise. Unlike traditional geophones, DAS sensors encounter fading noise and persistent horizontal bands that often overlap with actual seismic signals in the frequency domain.
Traditional methods like band-pass filtering or wavelet transforms often struggle to separate these signals without significant loss of data fidelity. Existing deep learning models, such as DnCNN, frequently fail to capture the global context required to identify long-range noise patterns or preserve the subtle reflected waves necessary for accurate subsurface imaging.
Methods: The ADNet Architecture
To address these limitations, the researchers developed ADNet, a 17-layer architecture optimized for seismic denoising. The network is structured into four functional blocks:
- Sparse Block (SB): Utilizing a combination of standard and dilated convolutions, the SB expands the network's receptive field. This allows the model to identify broad, sprawling noise structures, such as horizontal bands, without a massive increase in parameters.
- Feature Enhancement Block (FEB): This block fuses global and local features through concatenation, ensuring the network retains original context as it processes deeper layers.
- Attention Block (AB): The AB acts as the network's "brain," weighing different features to focus on hidden noise information buried within complex seismic backgrounds.
- Reconstruction Block (RB): ADNet utilizes residual learning. Instead of mapping noisy data directly to clean data, the network predicts the noise itself. The final denoised image is produced by subtracting the predicted noise from the original input.
Training and Evaluation
Due to the lack of standard "clean" DAS datasets, the authors constructed a robust training set using elastic wave equations to create 5,000 pure seismic patches from 10 complex 2-D geological models. Realistic noise—including random, fading, and optical abnormal noise extracted from actual field records—was then injected into these patches. The model was evaluated against synthetic records and a field DAS record from the Tarim Basin in Xinjiang, China.
Results and Discussion
ADNet demonstrated superior performance compared to Variational Mode Decomposition (VMD), standard band-pass filtering, and the DnCNN model:
- Signal-to-Noise Ratio (SNR): On synthetic data, ADNet achieved an SNR of 15.33 dB, significantly higher than DnCNN (11.95 dB) and traditional filtering (2.82 dB).
- Structural Similarity (SSIM): ADNet maintained a higher SSIM of 0.786, indicating superior preservation of original seismic waveforms compared to DnCNN’s 0.623.
- Field Performance: In real-world tests, ADNet effectively removed heavy horizontal noise, revealing critical "up-going" reflected waves that other methods missed.
Conclusion
ADNet provides a powerful solution for the simultaneous suppression of multiple DAS noise types. By combining dilated convolutions with an attention mechanism, the network achieves high-fidelity denoising while remaining computationally efficient. While "ringing noise" remains a challenge due to its similarity to weak reflected signals, future refinements in training set diversity and architecture are expected to further enhance the fidelity of DAS data processing.
Summary updated