Less is More: The Dilution Effect in Multi-Link Wireless Sensing

Wireless sensing approaches promise to transform smart infrastructures into privacy-preserving motion detectors, yet commercial adoption remains limited. A common assumption may explain this gap: that denser sensor deployments yield better accuracy. We tested this assumption with a 12-day naturalistic study using a 9-node ESP32-C3 mesh (72 sensing links) in a residential environment. Our results show that a single well-placed link outperformed the full 72-link mesh (AUC 0.541 vs. 0.489, Cohen’s $d$=0.86). Even a random link selection matched optimized selection ($p$=0.35). The benefit comes from avoiding multi-link fusion, not from choosing the right link. We attribute this to a “dilution effect”: links whose Fresnel zones miss activity regions contribute noise that overwhelms signal from informative links. In our deployment, strategic link placement mattered 2.7$\times$ more than classifier choice. We release 312 hours of labeled CSI data, firmware, and analysis code to enable validation across diverse environments.

💡 Research Summary

Wireless sensing based on Wi‑Fi promises privacy‑preserving motion detection for smart environments, yet commercial products remain absent despite a decade of algorithmic advances. The prevailing assumption—that denser sensor deployments automatically improve accuracy—has gone largely unchecked. In “Less is More: The Dilution Effect in Multi‑Link Wireless Sensing,” Khamaisi and Rodrigues challenge this assumption with a rigorous, real‑world experiment.

Physical Basis
The authors ground their hypothesis in Fresnel‑zone physics. A Wi‑Fi link can only sense a human when the person’s body intersects the link’s first Fresnel zone—an ellipsoidal region where path‑length differences are within half a wavelength. If the activity region does not intersect a link’s Fresnel zone, the mutual information between that link’s Channel State Information (CSI) and occupancy is essentially zero, regardless of classifier complexity. This leads to the “dilution effect”: when many links are fused by simple feature concatenation, the majority of “dead” links (those whose Fresnel zones miss the activity area) inject pure noise, overwhelming the informative signal from the few useful links.

Experimental Design
To isolate this effect, the authors built a 9‑node ESP32‑C3 mesh (each node $5) forming a complete directed graph of 72 unidirectional links. Nodes were placed in realistic household locations (power outlets, shelves) and classified as Entrance, Mid‑Room, or Dead‑Corner based on their Fresnel‑zone overlap with typical human trajectories. Over 12 days (January 2‑13 2026) they collected 312 hours of raw CSI (≈10⁷ packets) while simultaneously recording ground‑truth occupancy with ESP32‑CAM cameras and entry/exit event logs. Labels were assigned high confidence (65 %) when directly observed and medium confidence (35 %) when inferred, yielding a realistic class imbalance (203 h empty vs. 109 h occupied).

Feature Engineering & Models
CSI was processed in 10‑second sliding windows (≈50 packets). For each of the 52 OFDM sub‑carriers per link the authors computed three statistical descriptors: Normalized Baseline Variability Index (NBVI), amplitude variance, and Median Absolute Deviation (MAD). Features were Z‑score normalized per link. Five classifiers spanning linear to deep learning—logistic regression, SVM, random forest, a three‑layer MLP, and a small CNN—were trained on the concatenated feature vectors (792 dimensions for the full mesh).

Evaluation Protocol
Temporal leakage was avoided by a nested, day‑level cross‑validation with forward‑chaining: for day d the model was trained on days 1…d‑1 and tested on day d. Link selection (e.g., ranking by SNR) was performed exclusively on training data within each inner loop to prevent optimistic bias. Performance was measured by ROC AUC, complemented by paired t‑tests, Wilcoxon signed‑rank tests, Cohen’s d effect sizes, and bootstrap confidence intervals (10 000 resamples).

Key Findings

1. Single‑Link Superiority – The best individual link achieved AUC = 0.541, significantly higher than the 72‑link fused model (AUC = 0.489). The effect size was large (Cohen’s d = 0.86) and persisted in 7 of 10 folds.
2. Monotonic Degradation with More Links – Adding links consistently reduced AUC; the impact of link count on performance was 2.7 × larger than the choice of classifier.
3. Random vs. Optimized Selection – Randomly picking three links performed statistically indistinguishably from an optimized greedy selection (p = 0.35). Hence, the benefit stems from avoiding multi‑link fusion rather than from sophisticated link ranking.
4. Dilution Effect Confirmation – Although the median AUC of individual links was 0.633, the high‑dimensional concatenation diluted this information, confirming the theoretical prediction.
5. Practical Engineering – The authors detail low‑cost firmware optimizations (binary packet format, lock‑free ring buffers, backend‑orchestrated TDMA with a 720 ms cycle yielding 1.4 Hz full‑mesh sampling) that achieved >99 % uptime and <0.1 % packet loss over the entire deployment.

Contributions & Open Resources
The paper provides (i) empirical evidence that denser Wi‑Fi sensing meshes can hurt detection performance, (ii) a quantitative comparison showing deployment topology outweighs algorithmic sophistication, (iii) a publicly released dataset of 312 hours of labeled CSI, the ESP32‑C3 firmware, and the data‑collection pipeline, enabling reproducibility and cross‑environment validation.

Limitations & Future Work
The study is confined to a single living‑room layout; generalization to multi‑room houses, varied furniture arrangements, or different frequency bands remains to be tested. Ground‑truth labels, while carefully constructed, are not 100 % certain, and activity types (e.g., walking vs. sitting) were not distinguished. Future research should explore (a) adaptive link‑selection mechanisms that dynamically deactivate “dead” links, (b) ensemble methods that treat each link independently before decision fusion, and (c) on‑device inference to assess feasibility on ultra‑low‑power edge nodes.

Conclusion
Khamaisi and Rodrigues overturn the intuitive “more sensors = better performance” mantra for Wi‑Fi‑based motion detection. By linking physical Fresnel‑zone constraints to statistical learning outcomes, they reveal a dilution effect that degrades accuracy when many uninformative links are fused. Effective system design should therefore prioritize strategic placement of a few high‑quality links or employ architectures that mitigate the noise contribution of dead links. The open dataset and detailed hardware implementation provide a valuable foundation for the community to further investigate scalable, privacy‑preserving wireless sensing.