Dynamic Resource Optimization for Decentralized Estimation in Energy Harvesting IoT Networks

Claudio Battiloro, Paolo Di Lorenzo, Paolo Banelli, Sergio Barbarossa

TL;DR

This paper introduces dynamic resource allocation strategies for IoT networks where sensors must estimate time-varying signals using only the energy they harvest from their environment. By leveraging stochastic Lyapunov optimization, the authors develop algorithms that balance the "energy-accuracy tradeoff"—maximizing estimation precision while ensuring sensor batteries never run dry, all without needing prior knowledge of energy arrival or radio channel statistics.

What Problem Does This Paper Address?

In the burgeoning world of the Internet of Things (IoT), we often deploy "set-and-forget" sensor networks in hard-to-reach places. These sensors are increasingly equipped with Energy Harvesting (EH) capabilities (e.g., solar, wind, or vibration) to extend their lifespan indefinitely. However, EH introduces a massive headache for engineers: uncertainty.

Traditional "static" optimization assumes we know exactly how much energy we have and how much the radio channel will fluctuate. In reality, both energy arrivals and communication quality are highly dynamic. If a sensor transmits too much data to achieve high estimation accuracy, it might deplete its battery and go dark. If it saves too much energy, the "Fusion Center" (the central hub) receives poor-quality data, leading to inaccurate signal reconstruction. This paper tackles the challenge of dynamically managing these resources—bits, transmission power, and sampling sets—to maintain a stable network that provides the best possible estimate of a signal over time.

Methods

The authors frame the problem through the lens of Decentralized Estimation. They assume a network of $N$ sensors observing a signal that lives in a low-dimensional subspace (a common property of physical fields like temperature or pollution levels).

Key technical components include:

  • Probabilistic Quantization: Sensors don't just send raw data; they convert it into bits using a probabilistic uniform quantizer. This helps model the "distortion" introduced by limited bandwidth.
  • Bayesian LMMSE Estimation: The Fusion Center uses a Linear Minimum Mean-Square Error (LMMSE) estimator to reconstruct the global signal from the noisy, quantized bits sent by individual nodes.
  • Stochastic Lyapunov Optimization: This is the core innovation. Instead of solving a complex, long-term optimization problem (which usually requires knowing the future), the authors use "virtual queues" to represent battery levels. By minimizing a "drift-plus-penalty" function, they turn a long-term stability problem into a series of simple, per-time-slot decisions.
  • Linearization for Low Complexity: To make the algorithms runnable on low-power IoT hardware, the authors use a first-order Taylor expansion (linearization) of the Mean-Square Error function. This transforms a non-convex, "hard" problem into a linear, "easy" one that can be solved in closed form.

Datasets

The paper relies on numerical simulations rather than a fixed public dataset. The authors simulate a Wireless Sensor Network (WSN) with the following parameters:

  • Network: 50 nodes distributed over a 100m disk.
  • Signal: A graph signal based on the first six eigenvectors of a Laplacian matrix (representing spatial correlation).
  • Radio Model: Free-space propagation at 10 MHz with Rayleigh fading, consistent with realistic Bluetooth WSN specifications.
  • Energy: Harvested energy is modeled as a uniform distribution between 0 and a maximum value ($R_{max}$), and an "ON/OFF" profile is used to test non-stationary environments.

Key Insights and Ideas

  • The Battery-Accuracy Tradeoff: There is a quantifiable linear relationship between the required battery size and the desired estimation accuracy. To get closer to the theoretical "optimal" performance, you must be willing to tolerate larger battery fluctuations (and thus need larger physical batteries).
  • Model-Free Adaptability: Because the approach is based on stochastic optimization, the system doesn't need to know the "statistics" of the wind or the radio channel. It learns the best behavior simply by observing instantaneous realizations.
  • Sparsity as a Feature: The algorithms naturally lead to "sparse sensing." At any given time, the system automatically selects only the most informative sensors to transmit, saving the energy of others for future use.
  • Virtual Queues as Stabilizers: By treating the battery level as a queue that shouldn't "overflow" (get too full to harvest more) or "empty," the math naturally enforces energy causality—you can't spend what you haven't harvested.

Key Results

  • Convergence to Optimality: Algorithm 1 was shown to approach the optimal Bayesian Mean-Square Error (BMSE) as the control parameter $V$ increases.
  • Energy Efficiency: Algorithm 3 (the low-complexity version) achieved nearly the same performance as the more complex Algorithm 2 but with significantly lower computational overhead, making it ideal for real-time IoT applications.
  • Robustness: In non-stationary scenarios (where energy harvesting suddenly stops and starts), the proposed strategies successfully adapted the BMSE in real-time, "throttling" the estimation quality to keep the network alive during energy droughts (see Fig. 4 in the paper).

Limitations and Future Work

  • I.I.D. Assumption: The theoretical proofs assume that radio channels and energy arrivals are independent and identically distributed (i.i.d.) over time. While simulations show the algorithms work in non-stationary settings, a more rigorous theoretical analysis for those cases is still needed.
  • Fixed Overhead: The model assumes a fixed energy cost ($e_{o,i}$) for processing and signaling. In very complex networks, this overhead might itself be dynamic.
  • Future Directions: The authors suggest extending this framework beyond simple estimation to more complex tasks like decentralized machine learning, anomaly detection, and data classification at the "edge" of the network.

Summary updated

Abstract

We study decentralized estimation of time-varying signals at a fusion center, when energy harvesting sensors transmit sampled data over rate-constrained links. We propose dynamic strategies to select radio parameters, sampling set, and harvested energy at each node, with the aim of estimating a time-varying signal while ensuring: 1) the accuracy of the recovery procedure and 2) the stability of the batteries around a prescribed operating level. The approach is based on stochastic optimization tools, which enable adaptive optimization without the need of apriori knowledge of the statistics of radio channels and energy arrivals processes. Numerical results validate the proposed approach for decentralized signal estimation under communication and energy constraints typical of Internet-of-Things scenarios.
BibTeX
@article{battiloro2021dynamic,
  author = {Claudio Battiloro and Paolo Di Lorenzo and Paolo Banelli and Sergio Barbarossa},
  title = {{Dynamic Resource Optimization for Decentralized Estimation in Energy Harvesting IoT Networks}},
  journal = {IEEE Internet of Things Journal},
  year = {2021},
  url = {https://openreview.net/forum?id=Hqp9UbeZBn},
}

Loading related content...

Figure