Energy-Efficient Sensing

• Energy-efficient sensing is the design and optimization of sensor systems that minimize power use while ensuring accuracy, low latency, and robust data transmission.
• It employs methods such as guided processing, analog joint source-channel coding, and prediction-based techniques to achieve energy savings up to 1500× in some implementations.
• This approach is vital for sustainable applications ranging from wearable devices to large-scale networks, enabling battery-free and energy-harvesting deployments.

Energy-efficient sensing refers to the theory, design, and deployment of sensor systems, protocols, and architectures that minimize energy expenditure for data acquisition, processing, and communication, while meeting requirements on accuracy, latency, or information throughput. Recent research spans embedded deep learning, analog/digital co-design, integrated sensing and communications (ISAC), spectrum sensing, and hardware architectures from wireless sensor networks (WSNs) to emerging battery-free and energy-harvesting scenarios.

1. Theoretical Foundations and Resource Trade-offs

Classical models of energy-efficient sensing treat sensor resources—sensing, processing, and communication—as joint or competing consumers of limited energy budgets. In canonical parallel Gaussian source settings, minimizing mean-squared distortion subject to separate or joint constraints on sensing and communication energy admits closed-form resource allocation via "greedy plus water-filling" strategies under variance and cost orderings. Key results (Liu et al., 2012):

• Fractional Sensing: For each source $i$, the measured fraction $\theta_i$ is chosen according to sensing cost $\epsilon_i$ and variance $\sigma^2_i$, optimally trading sensing omissions (increased distortion) against rate/bandwidth allocation post-sensing.
• Joint Budget Optimization: When a single energy budget $B$ covers both sensing and transmission, only the "best" source–channel pairs (largest $\sigma^2$, lowest $\epsilon_i$, lowest channel noise $N_i$) are sensed and transmitted, others are dropped ($\theta_i=0$). For large $B$, full activation becomes optimal; for low $\theta_i$0, only low-cost sources are used.
• Bandwidth/Energy Asymptotics: The structure transitions between LCF– (cost-first) and ESF– (variance-first) regimes as $\theta_i$1 increases; maximum savings occur in the intermediate regime.

This approach delivers explicit guidelines for joint allocation and quantifies the distortion savings from explicitly accounting for per-sample acquisition energy, not just communication (Liu et al., 2012).

2. Embedded and Adaptive Sensing Architectures

Next-generation designs use cross-layer energy controls, guided processing, and intelligent system partitioning:

2.1. Guided Processing Beyond Duty-Cycling

Rather than naïve duty-cycling, "guided-processing" (Le et al., 2017) breaks the sensing pipeline into cascaded or graph-based chains. At each inexpensive stage, data frames are screened via lightweight features, and only information-rich instances are passed on to higher-cost, higher-accuracy detectors. Posterior probabilities are compared to dynamically optimized thresholds, rigorously balancing error risks (miss, false alarm) against per-stage energy:

$\theta_i$2

This results in up to $\theta_i$3 lower miss rate and $\theta_i$4 lower false-alarm rate compared to duty-cycling at the same energy, and generalizes to tree/DAG structures for multi-modal and multi-sensor networks (Le et al., 2017).

2.2. Prediction-Based and Co-Designed Sensing

In wireless sensor nodes, merging asynchronous wake-up receivers (WuR) and model-based sensing (MBS) achieves persistent, ultra-low-power sampling (Raza et al., 2016). By aggressively offloading predictions to sub-microcontroller devices (e.g., a PIC running derivative-based prediction) and relegating the main MCU/radio to deep-sleep, node average power drops below $\theta_i$5W—enabling indoor energy-neutrality.

Empirical reduction chain:

• Baseline WSAN reporting: $\theta_i$6mW per node
• Aggressive DPM: $\theta_i$7mW (lifetime $\theta_i$8)
• Radio WuR: $\theta_i$9 mW ($\epsilon_i$0 gain)
• Model-based prediction/transmission: $\epsilon_i$1 µW ($\epsilon_i$2 gain)

This approach is scalable (hundreds of nodes), agnostic to sensor type, and robust to rare/aperiodic event traffic (Raza et al., 2016).

2.3. Hyperdimensional and Intelligent Control

HyperSense (Yun et al., 2024) demonstrates a hardware/software loop for real-time, sparse-event energy-efficient sensing:

• Always-on, low-precision ADC and HDC-based inference detect "object present".
• High-precision ADCs are gated only on predicted events, reducing redundant high-energy sampling.
• FPGA-based implementations achieve up to $\epsilon_i$3 system energy savings relative to always-on baselines, with up to $\epsilon_i$4 throughput gains over edge GPU deep learning.

This leverages hypervector similarity for fast, noise-tolerant classification, closing the loop between inference confidence and ADC activation (Yun et al., 2024).

3. Analog Joint Source-Channel Coding and Biocompatible Sensing

Analog architectures bypass expensive digitalization and computation, encoding sensor signals into physical quantities for minimal energy and maximal density:

• Rectangular and MOSFET-based AJSCC: Tier-1 "dumb" analog nodes perform direct Shannon-mapping in hardware (FM-mapped voltage, MOSFET I–V curves), flattening cost and power per node to $\epsilon_i$5W (Sadhu et al., 2017, Sadhu et al., 2020, Sadhu et al., 2019).
• Adaptivity and Biodegradability: Substrate sensors are fully passive (biodegradable FETs), powered by ambient (solar/piezo/microbial) energy at $\epsilon_i$6W per channel, enabling continuous, maintenance-free, dense deployments.
• Smart Processing: All demodulation, decoding, and ML-based adaptation of quantization parameters (e.g., via KDE/KLD estimation of source PDFs) are performed at Tier-2 digital CHs (Sadhu et al., 2020).

End-to-end SNR and MSE predictions agree with empirical deployments, confirming $\epsilon_i$7 savings over digital (Sadhu et al., 2017, Sadhu et al., 2019).

4. Compressive Sensing and Data-Driven Compression

Data-driven compressive sensing frameworks optimize both measurement matrices and dictionaries by directly training on the intended physiological (or general) signals, exploiting individual and structural sparsity (Xu et al., 2016).

• Co-training optimization formalizes a convex relaxation:

$\epsilon_i$8

• Experimentally: At 10$\epsilon_i$9 compression,
  • Isometry $\sigma^2_i$0 reduces by $\sigma^2_i$1 over random matrices,
  • Reconstruction SNR improves +15 dB over model-driven approaches,
  • $\sigma^2_i$290% segment-wise transmission energy savings.

Critical for battery-limited and always-on wearable/medical scenarios (Xu et al., 2016).

5. Energy-Efficient Sensing in Integrated Sensing and Communication Systems (ISAC)

6G and ISAC research frames energy-efficient sensing in the context of MIMO, OFDM, and beamforming architectures co-serving communications and radar (Nguyen et al., 12 Sep 2025, Hu et al., 6 Aug 2025, Wu et al., 2023, Ma et al., 2023):

• Multi-objective Optimization: System-level utility balances communication SE and estimation CRB—often via a weighted sum or subject to explicit constraints:

$\sigma^2_i$3

Solvers combine Dinkelbach's algorithm and SCA for tractable, near-real-time solutions (Nguyen et al., 12 Sep 2025).

• Key Trade-offs:
  • Increasing communication thresholds (SE) depletes power available to sensing beams, degrading angular/range accuracy as measured by CRB ($\sigma^2_i$416\% overall EE drop when SE goes from 4 to 8 bps/Hz at high $\sigma^2_i$5) (Nguyen et al., 12 Sep 2025).
  • Hybrid analog–digital MIMO/beamfocusing architectures, while circuit- and hardware-efficient, incur inevitable losses in estimation accuracy compared to fully digital, especially for range/distance CRB (Hu et al., 6 Aug 2025).

Optimal transmit designs are eigenmode- and scenario-dependent; sensing-dominated (loose SE, tight CRB) regions favor shortest "on" durations and isotropic transmission, while comm-dominated regions require full-rank, water-filling on comm eigenmodes, with $\sigma^2_i$6 optimally scheduled (Wu et al., 2023).

• Waveform and Matched Filtering Energy Optimization: Discrete nonlinear FM (DNLFM) with time-frequency matched windows achieves constant-modulus (maximal PA efficiency) and arbitrary frequency windowing, yielding up to $\sigma^2_i$7 dB SNR and $\sigma^2_i$8 dB sidelobe suppression, minimizing wasted transmit energy—critical for multi-target, power-limited ISAC (Ma et al., 2023).

6. Large-Scale Architectural and Protocol Enhancements

Wide-area and battery-constrained networks benefit from adaptive sampling, hardware-aware orchestration, and protocol-level innovation:

• Energy-aware network functions (SCF, SECF, SAF) enforce closed-loop energy budgeting, selectively activating only essential sensors, adjusting sampling rates, sleep schedules, or offloading strategies, achieving $\sigma^2_i$945\% total energy savings in city-scale vehicular or V2X scenarios (Conceicao et al., 2024).
• Hierarchical and duty-cycled MAC: In vehicular, building or IoT deployments, duty-cycled wake/sleep, slotted random access, and collision/backoff analysis guide gateway and node schedules, enabling empirical energy savings of $B$0–$B$1 while bounding worst-case delay (Mishra et al., 2024).
• Cooperative spectrum sensing in cognitive WSNs: Energy-efficient routing with adaptive clustering (uniform vs non-uniform) and MST-based inter-cluster aggregation ensures maximal data delivery and minimizes per-bit energy, extending network lifetime by $B$2–$B$3 over LEACH (Surampudi et al., 2017).
• Advanced spectrum sensing: Optimization of sensing period and duration jointly minimizes energy, balancing the cost of spectrum "listening" against benefit of avoided retransmissions/collisions, solved via simple integer search (Cao et al., 2014).

7. Battery-Free and Energy-Harvesting Sensing

Ultra-low-power and battery-free paradigms exploit ambient energy:

• RF Energy Harvesting with REHSense: Purely passive receivers using harvested Wi-Fi energy as both sensor and power source achieve $B$4 accuracy on multiple fine-grained context tasks (respiration, activity, gesture) with $B$5 lower power than conventional Wi-Fi CSI approaches, harvesting up to $B$6 mW under standard routers (Ni et al., 2024).
• Sub-$B$7W analog systems: Dense analog networks leverage FET-based AJSCC, with scalability, biocompatibility, and long-term persistence—suitable for underwater, remote, and environmental monitoring, where maintenance or battery replacement is infeasible (Sadhu et al., 2019).

---

Conclusion

Energy-efficient sensing is realized through multi-scale optimization—algorithmic, architectural, analog/digital system, and protocol-level—with a diversity of mechanisms: guided and prediction-based processing, compressive and analog coding, co-designed hardware/software control, beamforming-aware joint optimization, and energy-aware orchestration at network scale. Explicit modeling of trade-offs among accuracy, latency, and multi-domain energy enables stable operation, minimizes waste, and unlocks sustainable, autonomous system lifetimes across applications from wearables to large-scale, battery-free, and ISAC-enabled networks (Nguyen et al., 12 Sep 2025, Xu et al., 2016, Raza et al., 2016, Le et al., 2017, Conceicao et al., 2024, Yun et al., 2024, Ni et al., 2024, Wu et al., 2023, Hu et al., 6 Aug 2025, Ma et al., 2023, Sadhu et al., 2017, Sadhu et al., 2020).