A Recommendation System Approach for Interference-Robust Sensor Subset Selection

Technical Summary: A Recommendation System Approach for Interference-Robust Sensor Subset Selection

Problem Statement
The paper addresses the challenge of efficient sensor subset selection in heterogeneous sensing environments, specifically for tracking moving targets (e.g., vehicles). In such settings, activating high-cost sensing modalities (like cameras) across all nodes is prohibitively expensive regarding computation, bandwidth, and battery life. The goal is to use low-cost, always-on sensors (acoustic) to recommend a small subset of nodes to activate their high-cost modalities.

Prior work in this domain relied on Received Signal Strength Indicator (RSSI) measurements from acoustic data. While RSSI is computationally cheap and compact, it aggregates all acoustic energy into a single scalar value. This aggregation makes RSSI-based approaches highly sensitive to environmental acoustic interference (e.g., wind, speech, passing vehicles), which degrades the accuracy of the recommended sensor subsets. The authors posit that while acoustic sensing cannot replace vision, richer acoustic representations could provide more robust recommendations without sacrificing the low-complexity nature required for real-time selective sensing.

Methodology
The authors propose a framework that reframes the sensor subset selection problem as a recommendation system task. Instead of performing explicit target localization or model-based posterior inference, the system learns a direct mapping from network-wide acoustic observations to the utility of candidate sensor subsets.

• Recommendation Paradigm: The network's acoustic state serves as the "context," and candidate sensor subsets serve as the "items" to be recommended.
• Two-Tower Architecture: The core of the method is a Two-Tower Multi-Layer Perceptron (MLP) architecture:
  • Context Tower: Encodes the time-varying acoustic state of the entire network.
  • Action (Candidate) Tower: Encodes the identity and geometry of a specific candidate subset of nodes.
  • Scoring: The embeddings from both towers are concatenated and combined with an element-wise product to capture compatibility. A prediction head outputs a scalar utility score. The subset with the highest score is selected.
• Feature Construction:
  • Context Vector: Unlike RSSI baselines, this vector aggregates node coordinates and frequency-band acoustic power features. The authors extract power in seven specific logarithmic frequency bands (ranging from 20 Hz to 6000 Hz), allowing the model to distinguish target signatures from interference.
  • Action Vector: Encodes the binary membership mask of the subset, normalized coordinates of selected nodes, and the subset size.
• Training Objective: The model is trained to minimize the mean-squared error between the predicted utility and a ground-truth utility function. The utility is defined based on the proximity of the selected subset to the target (specifically, a weighted sum of inverse distances to the closest nodes in the subset).

Key Contributions

1. Formulation: The paper formulates subset sensor activation as a recommendation problem where network observations define the context and candidate subsets define the recommendable items.
2. Architecture: It introduces a Two-Tower recommendation architecture that learns separate embeddings for network state and sensor subsets, enabling efficient real-time scoring of sensing actions.
3. Robustness via Frequency Bands: The authors demonstrate that utilizing frequency-band acoustic features significantly improves robustness to acoustic interference compared to scalar RSSI.
4. Real-World Validation: Through real-world outdoor deployments, the paper shows the method remains computationally lightweight (sub-millisecond to ~1 ms online computation) while outperforming prior RSSI-based approaches.

Experimental Results
The method was evaluated on two outdoor vehicle-tracking deployments using an IoT platform with NVIDIA Jetson Orin NX nodes:

1. Interference-Rich Deployment: A site with hilly terrain, buildings, and intermittent interference (speech, wind, construction).
   • Performance: The proposed Two-Tower model with frequency-band features achieved a mean closest-node containment accuracy of 98.39%.
   • Comparison: This represents a significant improvement over the RSSI-only Two-Tower model (80.44%) and model-based RSSI baselines (ranging from 71.33% to 77.03%). The frequency-band features allowed the model to distinguish the target's acoustic signature from interference sources where RSSI failed.
   • Computation: The total online computation time (including feature construction and inference) averaged 0.72 ms, well within the 200 ms sensing interval.
2. Open-Field Deployment: A larger, flatter site with fewer interference sources.
   • Performance: All methods performed well (exceeding 92% accuracy). The RSSI-only Two-Tower model slightly outperformed the frequency-band version (99.40% vs. 97.80%), suggesting that in benign acoustic conditions, the simpler RSSI approach is sufficient.
   • Computation: Computation times remained feasible, averaging roughly 1.08 ms for the full model.

Significance and Claims
The paper claims that its primary significance lies in demonstrating that recommendation-style architectures can be effectively adapted for efficient sensor-subset selection. By leveraging frequency-band features within a Two-Tower framework, the method achieves a substantial improvement in robustness against acoustic interference (up to ~20% accuracy gain in noisy environments) without compromising the low computational overhead required for real-time selective sensing. The authors emphasize that the approach does not aim to replace high-cost modalities like cameras but rather to activate them selectively and only when they are most likely to be useful, thereby optimizing resource usage in edge sensing networks.