Optimal Resource Management for Hierarchical Federated Learning Over HetNets With Wireless Energy Transfer

Rami Hamdi, Ahmed Ben Said, Emna Baccour, Aiman Erbad, Amr Mohamed, Mounir Hamdi, Mohsen Guizani

TL;DR

This paper introduces a novel framework for Hierarchical Federated Learning (HFL) designed for industrial IoT environments where devices are powered via Wireless Energy Transfer (WET). By optimizing device-to-edge association and energy management through a new heuristic algorithm (H2RMA), the researchers achieved significant reductions in grid energy costs and training loss, even when accounting for worker mobility.

What Problem Does This Paper Address?

As industrial IoT grows, training AI models on thousands of resource-constrained devices (like wearables or smart cameras) becomes a logistical nightmare. Standard Federated Learning (FL)—where devices train locally and send updates to a central server—suffers from communication bottlenecks and high energy drain.

Hierarchical Federated Learning (HFL) adds an intermediate layer: Multi-Access Edge Computing (MEC) units. Devices sync with nearby MECs, which then sync with a Central Unit (CU). However, two major hurdles remain:

  1. Energy Scarcity: IoT devices have tiny batteries.
  2. Data Skewness (Non-I.I.D.): If a device only sees one type of data, the global model becomes biased.

The authors tackle these by integrating Wireless Energy Transfer (WET), where MECs wirelessly "beam" power to devices, and by creating a resource management strategy that balances energy needs with data diversity.

Methods

The authors formulate the resource management as a Mixed-Integer Non-Linear Programming (MINLP) problem. Since MINLP is NP-hard and too slow for real-time use, they developed two key practical solutions:

  • H2RMA (Heuristic Resource Management Algorithm): This algorithm decides which device connects to which MEC. Instead of just picking the closest MEC, it selects one that minimizes "divergence" (a measure of how much the local data distribution differs from the global one) while staying within the MEC's wireless charging range.
  • Device Scheduling: To save power, the system can "deactivate" devices that require excessive energy for a specific round, provided their absence doesn't hurt the model's accuracy too much.
  • DHDA (Dynamic HFL Device Association): To handle mobility, the authors used Hidden Markov Models (HMMs) to categorize worker movement (Static, Normal, or Risky). The system dynamically re-associates devices with MECs as workers move across a construction site or factory floor.

The architecture assumes a Massive MIMO wireless backhaul between the MECs and the CU, using Zero-Forcing (ZF) and Block Diagonalization (BD) to manage interference.

Datasets

The framework was evaluated using two standard computer vision benchmarks:

  1. MNIST: 60,000 images of handwritten digits (0-9).
  2. CIFAR-10: 60,000 color images across 10 classes (airplane, bird, cat, etc.).

These datasets were distributed across devices in a non-i.i.d. (non-independent and identically distributed) fashion to simulate real-world data skewness.

Key Insights and Ideas

  • The Power-Data Trade-off: The most innovative part of the H2RMA algorithm is that it doesn't just look at signal strength. It prioritizes MECs that help reduce the "weight divergence" of the model. This means the system might choose a slightly further MEC if it helps the AI learn faster.
  • WET as an Enabler: By using WET, the MECs act as both data aggregators and power stations. This creates a closed-loop system where the cost of "grid energy" at the MEC is the primary constraint, rather than the device's battery life.
  • Mobility Matters: The study shows that "fixed" association schemes fail quickly in industrial settings. By predicting movement states (like "Risky" workers moving faster), the DHDA algorithm prevents energy spikes caused by devices trying to communicate with an MEC that is now too far away.

Key Results

  • Energy Efficiency: The H2RMA algorithm significantly outperformed random device association. In simulations with 60 devices, the grid energy cost was orders of magnitude lower than the random baseline.
  • Accuracy & Convergence: The proposed HFL approach reached higher test accuracy (approx. 85-90% on MNIST) faster than random association.
  • Scheduling Gains: By deactivating high-energy devices, the system saved a "significant amount of energy" without deteriorating the learning accuracy, thanks to the divergence-aware selection.
  • Optimality: While the optimal MINLP solution (found via Brute-Force) is the gold standard, the H2RMA heuristic achieved near-optimal performance with much lower computational complexity.

Limitations and Future Work

  • Complexity of WET: The paper assumes perfect Channel State Information (CSI) for the WET beamforming, which is difficult to maintain in highly dynamic, cluttered industrial environments.
  • Scalability: While the heuristic is polynomial-time, the impact of interference in extremely dense deployments (thousands of devices per MEC) remains an open question.
  • Future Direction: The authors suggest exploring Reinforcement Learning (RL) to further automate the clustering and association strategies in HFL over HetNets.

Summary updated

Abstract

Remote monitoring systems analyze the environment dynamics in different smart industrial applications, such as occupational health and safety, and environmental monitoring. Specifically, in Industrial Internet of Things (IIoT) systems, the huge number of devices and the expected performance put pressure on resources, such as computational, network, and device energy. Distributed training of machine and deep learning (ML/DL) models for intelligent industrial IoT applications is very challenging for resource limited devices over heterogeneous wireless networks (HetNets). Hierarchical federated learning (HFL) performs training at multiple layers offloading the tasks to nearby multiaccess edge computing (MEC) units. In this article, we propose a novel energy-efficient HFL framework enabled by wireless energy transfer (WET) and designed for heterogeneous networks with massive multiple-input–multiple-output (MIMO) wireless backhaul. Our energy-efficiency approach is formulated as a mixed-integer nonlinear programming (MINLP) problem, where we optimize the HFL device association and manage the wireless transmitted energy. However due to its high complexity, we design a heuristic resource management algorithm, namely, H2RMA, that respects energy, channel quality, and accuracy constraints, while presenting a low-computational complexity. We also improve the energy consumption of the network using an efficient device scheduling scheme. Finally, we investigate device mobility and its impact on the HFL performance. Our extensive experiments confirm the high performance of the proposed resource management approach in HFL over HetNets, in terms of training loss and grid energy costs.
BibTeX
@article{hamdi2023optimal,
  author = {Rami Hamdi and Ahmed Ben Said and Emna Baccour and Aiman Erbad and Amr Mohamed and Mounir Hamdi and Mohsen Guizani},
  title = {{Optimal Resource Management for Hierarchical Federated Learning Over HetNets With Wireless Energy Transfer}},
  journal = {IEEE Internet of Things Journal},
  year = {2023},
  url = {https://openreview.net/forum?id=PHsyOpiPSk},
}

Loading related content...

Figure