Statistically-Guided Dual-Domain Meta-Learning with Adaptive Multi-Prototype Aggregation for Distributed Fiber Optic Sensing

TL;DR

This work tackles cross-deployment DFOS recognition under scarce target labels by proposing DUPLE, a dual-domain prototype-based meta-learning framework. It fuses time- and frequency-domain cues via adaptive multi-prototype representations, guided by a Statistical Guidance Network that leverages 26 signal statistics to weigh domain reliability and prototype sensitivity, and integrates a query-aware attention mechanism to adapt prototype aggregation at inference. The method achieves state-of-the-art performance on two real-world cross-deployment DFOS benchmarks (OSDG1 and OSDG2) under Leave-One-Deployment-Out evaluation, with statistically significant improvements over strong baselines and robust performance under data scarcity. These results demonstrate improved generalization, stability, and real-time feasibility for DFOS activity recognition across diverse deployment environments, advancing practical perimeter security sensing.

Abstract

Distributed Fiber Optic Sensing (DFOS) is promising for long-range perimeter security, yet practical deployment faces three key obstacles: severe cross-deployment domain shift, scarce or unavailable labels at new sites, and limited within-class coverage even in source deployments. We propose DUPLE, a prototype-based meta-learning framework tailored for cross-deployment DFOS recognition. The core idea is to jointly exploit complementary time- and frequency-domain cues and adapt class representations to sample-specific statistics: (i) a dual-domain learner constructs multi-prototype class representations to cover intra-class heterogeneity; (ii) a lightweight statistical guidance mechanism estimates the reliability of each domain from raw signal statistics; and (iii) a query-adaptive aggregation strategy selects and combines the most relevant prototypes for each query. Extensive experiments on two real-world cross-deployment benchmarks demonstrate consistent improvements over strong deep learning and meta-learning baselines, achieving more accurate and stable recognition under label-scarce target deployments.

Figures (7)

• Figure 1: The figure shows the acquisition of climbing vibration signals under four different deployment methods. (a) is the acquisition under the wire mesh fence scenario, (b) is the acquisition under the barbed wire scenario, (c) is the acquisition under the wall scenario, and (d) is the acquisition under the hangnail net scenario. Due to differences in distance from the signal source, installation structure, or coupling conditions, the waveforms acquired under each deployment method show significant differences in amplitude, frequency distribution, and pulse characteristics.
• Figure 2: Overview of the proposed Dual-Domain Meta-Learning framework. The architecture first integrates temporal and spectral information via a Feature Extraction module, using parallel 1D-CNN and 2D-CNN. Based on these features, the system performs Dual-Domain Prototype Construction to capture intra-class diversity from the support set. In parallel, the SGN Module derives instance-specific guidance signals from the input's statistical characteristics. This pipeline culminates in the Collaborative Decision module, which leverages the SGN guidance and a query-aware attention mechanism to adaptively aggregate prototypes for the final classification.
• Figure 3: A schematic diagram of the Statistical Guidance Network (SGN) is shown. This module takes the statistical features of the input signal as input and obtains intermediate feature representations through a physical property analyzer and an environmental feature extractor composed of multi-layer MLPs, respectively. These two representations are then concatenated and mapped to the base guidance signal by a guidance signal generator. Furthermore, the statistical features are also input to a domain importance predictor and a prototype sensitivity predictor to estimate the sample's dependence weights in the time and frequency domains, as well as its sensitivity to prototype aggregation, respectively. Finally, all guidance signals are output to downstream modules for dynamic adjustment.
• Figure 4: A schematic diagram of the Collaborative Decision module. The time-domain and frequency-domain prototypes are processed through two aggregation branches based on queries. The query consists of a guiding signal and two scalars in the time and frequency domains, which are used to modulate the two branches. The aggregation results from both paths are fed into the relation modeling unit and fused to obtain the final decision.
• Figure 5: Sensitivity analysis of model performance with respect to support set size $K \in \{1, 3, 5, 10\}$. Top Row (OSDG1): (a-c) DUPLE consistently achieves the highest Accuracy, Precision, and F1-Score, demonstrating exceptional robustness even in the extreme low-data regime ($K=1$), while baselines like ProtoNet struggle significantly. Bottom Row (OSDG2): (d-f) DUPLE exhibits superior data efficiency, showing a rapid performance surge from $K=1$ to $K=3$ and establishing a comprehensive lead in all metrics, unlike baselines that plateau at lower performance levels.