Abstract
With the growing adoption of unmanned aerial vehicle (UAV)-assisted Internet of Things (IoT), its resilience against cascading failures has garnered significant attention. Cascading failures can severely compromise the topological integrity of such networks, making efficient recovery a significant challenge. To address this challenge, a Network Recovery scheme with Heterogeneous Graph neural network (NRHG) is proposed. The proposed scheme employs a Heterogeneous Graph Neural Network (HGNN), which includes graph perception layers processing local observations from individual UAVs, and graph communication layers enabling information exchange among UAVs. A multi-agent reinforcement learning (MARL) framework is further employed to enable collaborative action decisions for UAVs. Experimental results demonstrate that the proposed NRHG scheme can efficiently schedule surviving UAVs to cover the network blind spots caused by cascading failures. Compared to other schemes, the proposed scheme shows superior performance in both network coverage recovery and system throughput restoration.