SagnacAssisted Enhanced OTDR for Distributed Acoustic Sensing: A Standardized Benchmark and Engineering Evaluation Framework

Abstract: Phase-sensitive optical time-domain reflectometry ($φ$-OTDR) is widely used in large-scale distributed acoustic sensing (DAS) because it provides distributed spatiotemporal monitoring over long sensing distances. Its field performance can still deteriorate because of polarization-induced fading (PIF), local signal degradation, and strong environmental interference. This study develops a Sagnac-assisted enhanced $φ$-OTDR sensing architecture and a standardized benchmark framework for engineering-oriented DAS event recognition. The Sagnac interferometer provides a continuous phase response that supplements fading-prone observations in the $φ$-OTDR channel, and heterogeneous signal alignment is achieved using a cross-correlation procedure implemented on an FPGA platform. The benchmark protocol compares conventional feature-engineering methods, probabilistic shallow classifiers, single-branch deep models, and dual-branch fusion models under consistent data partitioning, preprocessing, and metric definitions. Experiments on a 10-km sensing fiber with six representative acoustic event classes show that the dual-branch fusion model provides the most favorable trade-off among the evaluated methods, reaching 89.79\% accuracy, 89.83\% macro-F1, and a nuisance alarm rate of 5.00\% on the balanced test set. The results also show that channel grouping strongly affects dual-branch evaluation, indicating that deployment-oriented conclusions should be based on accuracy, macro-F1, nuisance alarm rate, false negative rate, and latency rather than accuracy alone. This work provides a physically motivated enhancement strategy for $φ$-OTDR-based DAS and a reproducible benchmark protocol for future fusion-oriented sensing research. The implementation and scripts for reproducing the DAS event-recognition experiments are publicly available at https://github.com/wawa-abc/das.

• The paper introduces a standardized benchmark for Sagnac-assisted c5-OTDR recognition, comparing shallow, probabilistic, single-branch deep, and fusion models under consistent data splits, preprocessing, and deployment metrics.
• The dual-branch fusion CNN achieves 89.79% accuracy, 89.83% macro-F1, 5.00% nuisance alarm rate, zero false negatives, and 12.79 ms latency on six-class events measured over 10 km of G.652.D fiber.
• Channel organization is a critical engineering variable: tested groupings changed fusion accuracy from 51.25% to 78.75%, while the reported physical hybrid benefit remains provisional because experiments used a unified-channel surrogate rather than separately recorded Sagnac and c5-OTDR data.

Motivation and scope

Phase-sensitive optical time-domain reflectometry ($\phi$-OTDR) is among the most widely deployed distributed acoustic sensing (DAS) modalities, but single-configuration systems remain vulnerable to polarization-induced fading (PIF), local signal degradation, and strong environmental interference. The paper under review addresses this problem on two fronts simultaneously: a physical-layer enhancement that augments $\phi$-OTDR with an auxiliary Sagnac interferometer channel, and an evaluation-layer contribution consisting of a standardized, reproducible benchmark for comparing recognition routes under identical data partitioning, preprocessing, and metric definitions (2606.05754). The authors are explicit that the contribution is not a new classifier; the dual-branch fusion CNN is presented as one representative route within the benchmark rather than as the central methodological novelty.

Sagnac-assisted architecture

The hybrid front-end preserves $\phi$-OTDR's role in spatially resolved localization via the round-trip mapping $z = c\tau / 2n_{eff}$, while the Sagnac interferometer supplies a continuously demodulated phase response that is less susceptible to local fading. The Sagnac output intensity follows $$I(t) = I_1 + I_2 + 2\sqrt{I_1 I_2}\cos[\Delta\phi(t)]$$, with phase recovery via a $3\times3$ coupler and orthogonal demodulation. Heterogeneous signal alignment is achieved by maximizing a normalized fragment-level cross-correlation between the Sagnac phase sequence and the $\phi$-OTDR spatiotemporal response, implemented on an FPGA platform for real-time synchronization.

A notable methodological honesty appears here: because the benchmark dataset is stored as a unified 12-channel matrix rather than as physically separated Sagnac and $\phi$-OTDR files, the experimental evaluation of fusion operates through controlled channel grouping over that matrix. The authors state plainly that this constitutes a surrogate evaluation protocol for the hybrid concept, not a strict file-level comparison of two independently stored sensing modalities.

Benchmark design principles

The framework rests on four design commitments: (1) fair comparison across conventional feature-engineering baselines, probabilistic shallow classifiers, single-branch deep models, and dual-branch fusion models under one train–test split and preprocessing pipeline; (2) treatment of input organization—specifically channel grouping—as a controlled benchmark factor rather than a hidden implementation detail; (3) evaluation against operational risk using nuisance alarm rate (NAR), false negative rate (FNR), and inference latency alongside accuracy and macro-F1, with NAR and FNR defined via a background-versus-threat aggregation of the confusion matrix; and (4) separation of physical enhancement from algorithmic evaluation so that neither conflates the other's contribution.

The task is six-class recognition over a 10-km Corning G.652.D fiber covering perimeter intrusion, pipeline excavation, and environmental interference scenarios. The balanced dataset contains 15,419 samples split 80:20 by independent event segments to avoid leakage. A long-tail imbalanced variant was also constructed, though—as discussed below—its numerical results were not finalized at the time of writing.

Channel grouping optimization

Because branch construction cannot rely on predefined physical splits, the authors systematically searched 24 candidate groupings of the 12 channels into two equal branches, screening by validation accuracy with macro-F1, NAR, and FNR recorded as diagnostics. This step yields one of the paper's strongest empirical claims: the default contiguous partition (channels 0–5 versus 6–11) is the worst-performing representative candidate, achieving only 51.25% accuracy with a 44.50% FNR, whereas the top-ranked cross-group combination reaches 78.75% accuracy with 2.00% FNR—a spread of more than 27 percentage points attributable purely to input organization. The implication is direct: reported fusion performance can be substantially overestimated or underestimated by arbitrary branch assignment, so grouping must be part of any credible fusion benchmark.

Benchmark results

The main comparison establishes a clear hierarchy across technical routes:

Route	Representative method	Accuracy (%)	Macro-F1 (%)	NAR (%)	FNR (%)	Latency (ms)
Shallow single-route	MPE + ZCR + SVM	44.37	43.35	47.50	10.75	0.0127
Probabilistic shallow fusion	Fusion Features + PSVM	56.46	57.40	32.50	3.25	0.0173
Deep single-branch	Branch B CNN	86.04	85.89	12.50	0.00	3.35
Deep fusion	Fusion CNN	89.79	89.83	5.00	0.00	12.79

Shallow handcrafted descriptors (STFT + SVM at 41.25% accuracy) prove inadequate for the intra-class variability of field DAS data. Probability-enhanced decision making (PSVM) improves confidence structure and reduces FNR to 3.25%, but remains far below deep routes. The dual-branch fusion CNN achieves the best joint trade-off: 89.79% accuracy, 89.83% macro-F1, 5.00% NAR, zero FNR, at a latency of roughly 12.8 ms per sample—acceptable for online monitoring despite being nearly four times slower than the single-branch CNN.

Class-wise analysis shows the fusion advantage is concentrated where it matters operationally: categories with stable signatures (background, knock) are already handled well by single-branch models, while environmentally sensitive classes (water, shake) benefit most from branch interaction. The value of fusion therefore lies in reducing vulnerability to locally degraded observations rather than uniformly lifting all classes.

Limitations and open questions

Several boundaries should be acknowledged. First, the dual-branch evaluation is a channel-grouping surrogate over a unified 12-channel matrix, not a demonstration with physically separated Sagnac and $\phi$-OTDR recordings; the claimed physical complementarity of the hybrid front-end is thus motivated but not directly validated end-to-end in the recognition experiments. Second, the imbalanced long-tail evaluation lacks a complete exported numerical summary across all four route categories—the paper discusses expected behavior qualitatively and defers final numbers to a later pipeline export, leaving the central deployment question (robustness under realistic class imbalance) empirically unresolved. Third, the search-stage ranking score is validation accuracy alone, which could in principle favor groupings that do not optimize the multi-metric objective, although full-evaluation metrics are reported for the retained candidates. Finally, the framework relies on supervised learning with fixed annotations, and the authors identify semi-supervised adaptation, edge-deployed lightweight models, and multi-parameter (vibration/temperature/strain) sensing as open directions constrained by current annotation quality, compute placement, and transmission-distance limits.

Conclusion

This work couples a physically motivated Sagnac-assisted enhancement of $\phi$-OTDR with a standardized, reproducible benchmark for DAS event recognition. Its most consequential findings are quantitative: deep fusion outperforms shallow baselines by over 45 percentage points in accuracy while cutting NAR from 47.50% to 5.00%, and channel grouping alone shifts fusion accuracy by more than 27 points, establishing input organization as a first-class benchmark factor. The framework's insistence on multi-metric reporting—accuracy, macro-F1, NAR, FNR, and latency jointly—offers a sounder basis for deployment-oriented conclusions than single-metric comparisons, provided its surrogate grouping protocol and pending long-tail results are kept in view when interpreting the reported numbers.