Distributed Acoustic Sensing (DAS-N2N)

• DAS-N2N is a fiber-optic sensing paradigm that uses dense, spatially-distributed measurements to monitor environmental, seismic, acoustic, and anthropogenic activities.
• It employs advanced signal processing techniques—such as phase unwrapping, ML-based Noise2Noise denoising, and spectral analysis—to improve signal clarity and spatial resolution.
• The system integrates distributed machine learning via federated and meta-learning approaches, enabling real-time, privacy-preserving inference across multiple sensor nodes.

Distributed Acoustic Sensing Node-to-Node (DAS-N2N) is an advanced paradigm in fiber-optic sensing that leverages dense, spatially-distributed measurements along optical fibers to enable high-resolution, scalable, and often collaborative monitoring of environmental, seismic, acoustic, and anthropogenic phenomena. DAS-N2N encompasses both canonical single-cable DAS architectures and more recent developments including distributed ML-based denoising, federated and meta-learning across sensor nodes, and dual-fiber bistatic localization. These advances have positioned DAS not only as a cost-effective distributed array for conventional sensing, but as a foundation for intelligent, real-time, and privacy-preserving sensing at continental scale.

1. Physical Principles and Architecture

DAS exploits Rayleigh backscatter within optical fibers: a pulsed or modulated coherent light source is launched into the fiber; scattered returns are coherently detected and multiplexed by arrival time to yield phase measurements at meter-scale increments ("channels") along tens to hundreds of kilometers (Xenaki et al., 25 Feb 2025, Rading et al., 17 Jul 2025). Localized mechanical strain $s(z,t)$—caused by seismic, acoustic, or anthropogenic disturbances—modulates the optical phase over a gauge length $L_g$, with the measured differential phase shift

$$\Delta\phi(z,t) = \frac{4\pi n}{\lambda} \int_0^{L_g} s(z', t) \, dz'$$

where $n$ is the fiber refractive index and $\lambda$ is the laser wavelength (Ramirez-Torres et al., 15 Sep 2025).

Interrogators vary: phase-sensitive OTDR (φ-OTDR) is standard, offering $\sim$1--10 m spatial resolution, kHz data rates, and gauge lengths tunable to optimize SNR versus resolution. For node-to-node or dual-fiber operation, two fibers are laid in parallel, each end equipped with transmitters and receivers to support bistatic TDOA localization (Grythe et al., 24 Sep 2025).

Table: Representative DAS Acquisition Parameters

Parameter	Symbol	Typical Value / Range
Gauge length	$L_g$	10–100 m
Channel spacing	$\Delta z$	1–10 m
Pulse width	$T_p$	50–500 ns
Sampling rate	$f_s$	500–5000 Hz

2. Signal Processing, Denoising, and Spectral Analysis

After acquisition, DAS data (differential phase $\Delta\phi(z,t)$) undergoes several processing steps:

• Phase unwrapping to recover continuous phase (removing $2\pi$ modulo effects);
• Calibration/scaling to convert to strain or acceleration units, factoring in system and environmental calibration parameters;
• Noise removal, including trend subtraction, bandpass filtering, and—critically—data-driven denoising methods.

DAS-N2N advances signal denoising using a weakly supervised neural denoising paradigm: by splicing two fibers in a loop, recordings from each fiber yield paired, independent noisy copies of the same underlying signal. A neural network (lightweight U-Net, 47k parameters) is trained to map one noisy instance to the other, suppressing incoherent noise while preserving coherent signals. This Noise2Noise (N2N) loss is mathematically justified: for zero-mean, independent noise, minimizing $\mathbb{E}[\|f_\theta(x_1) - x_2\|^2]$ recovers the supervised clean-target optimum (Lapins et al., 2023).

DAS-N2N denoising outperforms classical Butterworth bandpass, Wiener, and self-supervised blind-spot methods (e.g., jDAS) in SNR improvement—yielding up to $2$–$3\times$ SNR gain over raw, and $1.5\times$ over bandpass, while processing 30 s × 985-channel files in $<$1 s on commodity GPUs (Lapins et al., 2023).

Spectral analysis (via short-time Fourier or f–k transforms) enables detection and discrimination of features such as microseisms, vessel signatures, and animal calls by their frequency and spatial signatures (Xenaki et al., 25 Feb 2025, Ramirez-Torres et al., 15 Sep 2025).

3. Distributed Machine Learning and Node Collaboration

Traditional DAS processing is limited by site-specific calibration and fixed signal-processing heuristics. The DAS-N2N paradigm generalizes by enabling multiple geographically distributed DAS nodes to collaborate via federated learning (FL) or meta-learning for robust multi-site inference (Zhong et al., 11 Jun 2025).

• Federated learning (FL): Each DAS node trains local models on its own labeled windows (e.g., for human activity classification) and shares only model weights (e.g., 18MB SR-Net) with a central server, not raw acoustic traces. Standard FedAvg aggregation is used; after $\sim$20–30 communication rounds, accuracy and F1 converge to $>90\%$ on cross-node tasks—recovering same-site accuracy despite climate, soil, and depth heterogeneity. Communication efficiency (orders of magnitude below raw data transfer) and privacy are direct outcomes (Zhong et al., 11 Jun 2025).
• Meta-learning (e.g., Reptile): Nodes without network connectivity or only small labeled datasets (<10 samples) can rapidly adapt pretrained models with few-shot fine-tuning. Meta-initialized SR-Nets require 1–2 minutes and $<$10 samples to achieve $>90\%$ accuracy on new deployments.
• These methods successfully neutralize cross-site generalization breakdowns common in independently trained models ($\sim$40% accuracy) (Zhong et al., 11 Jun 2025).

4. Multi-Functional Sensing: Environmental and Infrastructure Applications

DAS-N2N enables high-resolution, scalable monitoring in a broad set of domains:

• Seismic and Newtonian noise mitigation: DAS arrays match or exceed the performance of dense geophone/seismometer networks for ground-motion monitoring in the 3–20 Hz band, with coherence $\rho(f)>0.8$ and residual noise cancellation factors $\eta(20\,\mathrm{Hz})=0.11$—indistinguishable from conventional systems, but with far greater spatial coverage (Rading et al., 17 Jul 2025).
• Oceanography and environmental monitoring: Submarine DAS arrays, as on the OOI Regional Cabled Array, resolve source locations to tens of meters and angles to $\pm5^\circ$, discriminating whale calls and ship noise using combined phase and frequency-domain analysis (Xenaki et al., 25 Feb 2025).
• Maritime security: DAS-N2N repurposes submarine cables for vessel detection/localization. Advanced ML (XGBoost or compact NNs) achieves $>90\%$ vessel detection F1 and $141$ m mean absolute error for vessel distance estimation over 2,774 channels and 10-day deployments (Ramirez-Torres et al., 15 Sep 2025). AIS integration and rigorous ground-truthing confirm practicality for real-time, continuous surveillance.
• Smart infrastructure and active travel monitoring: Roadside/underground fibers support passive, privacy-preserving travel activity monitoring (walking/cycling) with scalable, FL/meta-learned neural architectures.

Table: Exemplary DAS-N2N Applications

Domain	Performance / Metrics	Reference
NN mitigation	$\rho>0.8$, $\eta(20\text{Hz})=0.11$	(Rading et al., 17 Jul 2025)
Vessel localization	F1=$0.91$, MAE=$141$ m	(Ramirez-Torres et al., 15 Sep 2025)
Travel monitoring	$>90\%$ accuracy (FL/meta), $64\times512$ windows	(Zhong et al., 11 Jun 2025)
Ocean fin-whale loc.	$~30-50$ m spatial, $\pm0.1$ km range	(Xenaki et al., 25 Feb 2025)

5. Dual-Fiber and Node-to-Node Localization

The dual-bistatic optical forward transceiver (dual-fiber node-to-node, or "DAS-N2N bistatic") architecture uses parallel fibers and transmit/receive pairs at both ends to enable TDOA-based acoustic source localization (Grythe et al., 24 Sep 2025). Each receiver records modulations $z_1(t), z_2(t)$ from a remote acoustic source; the relative delay $\delta_\tau = \tau_1 - \tau_2$ is estimated via the cross-ambiguity function (CAF),

$$\Psi_{\rm cross}(\tau, \nu) = \int y_1(t) y_2^*(t+\tau) e^{j2\pi\nu t} dt$$

with maximization yielding the TDOA and (optionally) Doppler shift. The spatial resolution is limited by acoustic bandwidth $B_\beta$ and sampling—$\Delta_\beta = v_f/B_\beta$, with CRB for range estimation scaling as $1/(|\kappa|^2 N B_\beta^3)$, where $|\kappa|^2$ is the link loss and $N$ the number of samples (Grythe et al., 24 Sep 2025).

Principal factors limiting accuracy include frequency-dependent fiber coupling, multipath, ambient colored noise, clock synchronization, and SNR degradation due to fiber attenuation or suboptimal coding. Mitigations include optimal fiber selection, code design, robust CAF peak-interpolation algorithms, and GNSS-disciplined clocks. This approach allows integrated acoustic communication and ranging, extending the functionality of submarine communication infrastructure (Grythe et al., 24 Sep 2025).

6. Performance, Scalability, and Practical Considerations

DAS-N2N architectures consistently demonstrate:

• Orders-of-magnitude improved spatial coverage relative to conventional seismic arrays, e.g., O($10^3$) virtual sensors per fiber versus O($10^3$) individual geophones (Rading et al., 17 Jul 2025).
• Robust real-time denoising with neural architectures that process kilometer-scale, kHz-rate acquisitions in sub-second latencies (Lapins et al., 2023).
• Edge deployability and privacy: small-model inference on embedded or IoT hardware, with communication-efficient FL or meta-learning, supports continuous, privacy-preserving monitoring (Zhong et al., 11 Jun 2025).
• Spectral fidelity across 0.2–20 Hz or higher, with SNR governed by gauge length, pulse width, and channel spacing (Xenaki et al., 25 Feb 2025).
• Residual factors/metrics (e.g., $\eta(f)$ for NN cancellation; F1 and MAE for classification and localization) matching or exceeding traditional instrument performance.

Deployment limitations and open questions include fiber-coupling heterogeneity, gauge-length constraints for low-frequency detection, necessity for accurate strain→acceleration calibration (requiring slowness estimates and channel selection), and the generalizability of pretrained neural models across cable types and interrogators. Ongoing work seeks to improve SNR stability, address coherent narrowband noise, enable adaptive online calibration, and expand generalization to new domains and sensor layouts (Lapins et al., 2023, Zhong et al., 11 Jun 2025).

7. Outlook and Future Directions

DAS-N2N is evolving towards integrated, AI-enhanced sensing platforms combining large-scale physics knowledge, distributed learning, and robust signal extraction from complex noise environments:

• Generalization across diverse environments is being addressed by FL/meta-learning and robust channel calibration (Zhong et al., 11 Jun 2025).
• Multi-task fusion (e.g., combining vessel detection, speed, size, and classification with meteorological data) and advanced array processing are in development for enhanced maritime security and environmental monitoring (Ramirez-Torres et al., 15 Sep 2025).
• Edge ML implementation aims for sub-second response and improved autonomy in real-world deployments.
• Security/authentication mechanisms for hybrid communication-sensing systems are required for operational reliability in critical infrastructure protection (Grythe et al., 24 Sep 2025).
• Long-term stability and self-adaptation—including semi-supervised or self-supervised online learning—will underpin persistent DAS-N2N deployments spanning continents.

Release of benchmark datasets such as Marlinks-NS (Ramirez-Torres et al., 15 Sep 2025) and real-world cross-validated performance metrics is advancing the scientific foundation for next-generation distributed sensing arrays, situating DAS-N2N as a principal solution for dense, scalable, and intelligent environmental monitoring.