⚡ electrical engineering

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

This paper presents a Sagnac-assisted enhanced $\phi$ -OTDR architecture that mitigates polarization-induced fading and introduces a standardized benchmark framework, demonstrating that a dual-branch fusion model achieves superior event recognition performance (89.79% accuracy) compared to conventional methods on a 10-km distributed acoustic sensing system.

Original authors: Weiguang Wang, Fugen Wu, Hailing Wang, Xuechen Liang, Xiaobin Li, Ru Han, Tianchang Xie

Published 2026-06-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Weiguang Wang, Fugen Wu, Hailing Wang, Xuechen Liang, Xiaobin Li, Ru Han, Tianchang Xie

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to listen to a specific conversation in a very noisy, crowded room. Sometimes, the sound waves bounce off walls or get blocked by people, causing the voice to suddenly drop out or sound garbled. This is exactly the problem engineers face with a high-tech sensing technology called $\phi$ -OTDR, which uses fiber-optic cables to "listen" for vibrations (like footsteps, digging, or climbing) over long distances.

Here is a simple breakdown of what this paper does, using everyday analogies:

1. The Problem: The "Fading" Signal

The main technology, $\phi$ -OTDR, is great at pinpointing where a vibration happens along a long cable. However, it has a flaw: Polarization-Induced Fading.

The Analogy: Think of the fiber-optic cable as a long hallway. Sometimes, the "light" carrying the sound signal hits a weird angle or a dirty spot in the hallway, and the signal just disappears for a moment. It's like trying to hear a whisper through a wall that suddenly becomes soundproof for a split second. This causes the system to miss important events (like a thief climbing a fence) or to get confused by background noise.

2. The Solution: The "Sagnac Assistant"

To fix this, the authors added a second listening device called a Sagnac interferometer.

The Analogy: Imagine you are the main listener (the $\phi$ -OTDR), but you have a partner (the Sagnac) standing right next to you. When the main listener's signal fades out because of a "bad spot" in the hallway, the partner keeps hearing the sound clearly because they use a different method to catch the vibrations.
The Result: They combine these two streams of information. If one ear goes deaf for a second, the other ear is still working. This creates a "hybrid" system that is much harder to fool.

3. The Challenge: How to Compare Different Brains

The authors didn't just build the hardware; they realized that to prove this new system works, you need a fair way to test different computer programs (AI models) that interpret the sounds.

The Analogy: Imagine a cooking competition. You have a new, better oven (the Sagnac-assisted hardware). Now you want to see which chef (the AI algorithm) can make the best dish. But if Chef A uses fresh ingredients and Chef B uses frozen ones, the test isn't fair.
The Innovation: The authors created a "Standardized Benchmark." This is a strict set of rules for the competition. Everyone gets the exact same ingredients (data), the same prep time (preprocessing), and the same judging criteria. This ensures that if one chef wins, it's because they are better, not because they got lucky with the ingredients.

4. The Competition: Who Wins?

They tested four types of "chefs" (algorithms) to see which could best identify six different types of events (like walking, digging, knocking, or water flowing):

The Old School Chef: Uses simple, hand-written rules (Feature Engineering). Result: Struggled. It couldn't handle the complexity.
The Probabilistic Chef: Uses simple rules but adds a bit of "guessing" based on confidence. Result: Better, but still not great.
The Deep Learning Chef (Single-Branch): A smart AI that learns patterns on its own from one stream of data. Result: Very good.
The Deep Learning Chef (Dual-Branch Fusion): A smart AI that looks at both the main signal and the assistant signal simultaneously and combines them. Result: The Winner.

The Winner's Stats:
The "Dual-Branch Fusion" chef achieved 89.79% accuracy. More importantly, it made very few mistakes in two critical areas:

Nuisance Alarms: It didn't cry "Wolf!" when there was just wind or rain (low false alarms).
Missed Detections: It didn't miss the actual threats (zero missed detections in the test).

5. The Secret Sauce: How You Mix the Ingredients

One of the paper's most interesting findings is about Channel Grouping.

The Analogy: You have 12 different microphones. To feed them into the "Dual-Branch" AI, you have to split them into two groups (Branch A and Branch B).
The Discovery: If you just split them in half (Microphones 1–6 vs. 7–12), the AI performs poorly. It's like putting all the bass instruments in one group and all the violins in the other; they don't complement each other well.
The Fix: The best performance came from a "scrambled" mix, where the two groups contained a balanced variety of microphones that complemented each other. The paper shows that how you organize the data is just as important as the AI model itself.

Summary

This paper does two main things:

Hardware: It builds a "super-listener" by pairing a standard fiber-optic sensor with a Sagnac assistant to prevent signal dropouts.
Software/Testing: It creates a fair, standardized rulebook to test AI models. It proves that the best way to use this new hardware is with a "Dual-Branch" AI that fuses both signals, and that the way you organize the data inputs is critical for success.

The result is a system that is more reliable, makes fewer false alarms, and is better suited for real-world security and infrastructure monitoring than previous methods.

Technical Summary: Sagnac-Assisted Enhanced $\phi$ -OTDR for Distributed Acoustic Sensing

Problem Statement
Phase-sensitive optical time-domain reflectometry ( $\phi$ -OTDR) is a dominant technology for large-scale distributed acoustic sensing (DAS), offering distributed spatiotemporal monitoring over long distances. However, its field performance is frequently compromised by polarization-induced fading (PIF), local signal degradation, and strong environmental interference. These physical limitations lead to unstable signal observations, resulting in missed detections and nuisance alarms, particularly under harsh field conditions. Furthermore, existing research often focuses on isolated structural improvements or device-level demonstrations, leaving a gap in system-level evaluation of hybrid architectures under realistic event-recognition scenarios. Additionally, the rapid advancement of AI-driven recognition algorithms lacks a standardized, reproducible benchmark that accounts for engineering constraints such as class imbalance, false alarm rates, and inference latency.

Methodology
The paper proposes a dual-pronged approach: a physical sensing enhancement and a standardized engineering evaluation framework.

Sagnac-Assisted Physical Architecture:
The authors developed a hybrid optical front-end that integrates a Sagnac interferometer with a conventional $\phi$ -OTDR system.
- Mechanism: The $\phi$ -OTDR channel retains its primary role in distributed spatiotemporal localization. The Sagnac interferometer, implemented with balanced polarization-maintaining fibers, provides a continuous, high-fidelity phase response that is less susceptible to local fading effects.
- Signal Processing: Heterogeneous signals from the two channels are aligned using a cross-correlation procedure implemented on an FPGA platform to establish a consistent temporal anchor. This creates a physically complementary arrangement where the Sagnac channel compensates for information loss in the $\phi$ -OTDR channel caused by PIF and environmental noise.
Standardized Benchmark Framework:
To evaluate this architecture and compare various recognition routes, the authors established a rigorous benchmark protocol using a 10-km sensing fiber dataset containing six acoustic event classes (background noise, manual digging, knocking, water flow, fence shaking, and walking).
- Input Organization: Since the dataset is stored as a unified 12-channel matrix rather than physically separated files, the benchmark treats "channel grouping" as a controlled variable. A search strategy was employed to identify optimal branch partitions for dual-branch fusion models, ensuring that performance gains are attributed to the fusion strategy rather than arbitrary data splitting.
- Model Categories: The framework compares four distinct technical routes under identical data partitioning and preprocessing:
  - Conventional feature-engineering methods (e.g., STFT + SVM).
  - Probabilistic shallow classifiers (e.g., Fusion Features + PSVM).
  - Single-branch deep models (Convolutional Neural Networks).
  - Dual-branch fusion deep models.
- Evaluation Metrics: Moving beyond simple accuracy, the benchmark emphasizes engineering-relevant metrics: macro-F1 score, nuisance alarm rate (NAR), false negative rate (FNR), and inference latency.

Key Results
Experiments conducted on the 10-km fiber dataset yielded the following findings:

Performance Hierarchy: A clear progression in performance was observed from shallow handcrafted features to deep fusion models. Conventional feature-engineering baselines (e.g., STFT + SVM) achieved low accuracy (~~41–44%). Probabilistic shallow fusion improved performance (~~56% accuracy) but remained inferior to deep learning approaches.
Optimal Route: The dual-branch fusion deep model achieved the most favorable trade-off, reaching 89.79% accuracy and 89.83% macro-F1 on the balanced test set. Crucially, it reduced the nuisance alarm rate to 5.00% and maintained a 0.00% false negative rate.
Impact of Channel Grouping: The study demonstrated that branch organization significantly affects fusion performance. The optimal grouping was not a simple contiguous split of channels; rather, cross-group combinations that preserved stronger complementarity between branches yielded superior results. This highlights that input organization is a critical factor in fusion-based DAS evaluation.
Engineering Trade-offs: While the fusion model exhibited higher inference latency (12.79 ms) compared to shallow baselines, the computational overhead was deemed acceptable for online monitoring. The model successfully balanced recognition quality with the suppression of nuisance alarms and missed events.

Significance and Claims
The paper positions its contributions as both a physical enhancement strategy and a methodological framework for the DAS community:

Physical Enhancement: The Sagnac-assisted architecture provides a physically motivated solution to mitigate polarization-induced fading, improving the observability of weak and low-frequency disturbances without relying solely on downstream algorithmic recovery.
Reproducible Benchmarking: The work establishes a standardized, reproducible benchmark protocol that shifts the focus from isolated laboratory accuracy to engineering-oriented reliability. By explicitly incorporating channel grouping optimization and multi-metric evaluation (NAR, FNR, latency), the framework allows for fair comparisons between diverse technical routes.
System-Level Validation: The authors claim that their results confirm the necessity of evaluating fusion-oriented DAS systems not just by classification accuracy, but by the joint optimization of recognition quality, alarm reliability, and deployment efficiency. The study asserts that the proposed dual-branch fusion approach offers a practically meaningful advantage for complex DAS event recognition in real-world monitoring scenarios.

The authors note that while the current benchmark relies on a unified 12-channel matrix as a surrogate for physically separated sensing files, the findings regarding the importance of input organization and the superiority of deep fusion routes provide a robust foundation for future research in hybrid sensing architectures.