A Bio-Inspired Hierarchical Federated Learning Framework With Joint Computing Communication Control Optimization for Green UAV Inspection

Abstract

With the development of UAV communications and aerial edge intelligence, Uncrewed Aerial Vehicles (UAVs) are increasingly used for inspection in high-risk environments. However, limited onboard energy and data-privacy constraints make efficient collaborative learning challenging. Hierarchical Federated Learning (HFL) provides a privacy-preserving paradigm, yet existing energy-optimization studies often treat computing, communication, and client selection separately, lacking a unified system perspective. Inspired by biological neural systems, where neurons compute in an event-driven manner, synapses transmit information sparsely, and organisms make reward-modulated decisions, we propose a Bio-Inspired Hierarchical Federated Learning (BIO-HFL) framework that takes UAV energy consumption as the central driving objective. In the computation module, BIO-HFL employs Spiking Neural Networks (SNNs) to realize neuron-like event-driven computing and reduce onboard energy; in the communication module, it integrates a Critical Tensor mechanism into Deep Gradient Compression (DGC-CT) to mimic synapse-like sparse but selective transmission and maintain stability under high compression ratios; and in the control module, it uses a Distributed Multi-Armed Bandit (DMAB) strategy as an organism-level decision module to select clients with minimal expected energy consumption. These three components are mutually coupled, SNN and DGC-CT energy statistics serve as inputs to DMAB, DMAB determines the optimization targets of SNN and DGC-CT through energy-aware scheduling, and CT selection in DGC-CT further depends on SNN-driven spike distributions. Experiments demonstrate that DGC-CT ensures stable training even at a 10% compression ratio, and DMAB reduces inefficient client participation. Across both CIFAR-10 and DAGM2007, BIO-HFL consistently improves energy efficiency while maintaining competitive macro-average F1, achieving up to 81.1% lower total energy consumption compared with a conventional SNN-HFL baseline.