A Collaborative Inference Algorithm in Low-Earth-Orbit Satellite Network for Unmanned Aerial Vehicle

• Published in: Drones (Basel) Vol. 7; no. 9; p. 575
• Main Authors: Xu, Zhengqian, Zhang, Peiying, Li, Chengcheng, Zhu, Hailong, Xu, Guanjun, Sun, Chenhua
• Format: Journal Article
• Language: English
• Published: Basel MDPI AG 01.09.2023
• Subjects: Algorithms; Artificial neural networks; Artificial satellites; Bandwidths; Collaboration; collaborative inference; Communication; Communications traffic; Comparative analysis; Computation offloading; Data processing; Drone aircraft; Edge computing; Inference; Information dissemination; Low earth orbit satellites; Low earth orbits; Machine learning; Neural networks; Power; Remote sensing; Rocket launches; satellite network; Satellite networks; Servers; Unmanned aerial vehicles
• ISSN: 2504-446X; 2504-446X
• DOI: 10.3390/drones7090575

In recent years, the low-Earth-orbit (LEO) satellite network has achieved considerable development. Moreover, it is necessary to introduce edge computing into LEO networks, which can provide high-quality services, such as worldwide seamless low-delay computation offloading for unmanned aerial vehicles (UAVs) or user terminals and nearby remote-sensing data processing for UAVs or satellites. However, because the computation resource of the satellite is relatively scarce compared to the ground server, it is hard for a single satellite to complete massive deep neural network (DNN) inference tasks in a short time. Consequently, in this paper, we focus on the multi-satellite collaborative inference problem and propose a novel COllaborative INference algorithm for LEO edge computing called COIN-LEO. COIN-LEO manages to split the complete DNN model into several submodels consisting of some consecutive layers and deploy these submodels to several satellites for inference. We innovatively leverage deep reinforcement learning (DRL) to efficiently split the model and use a neural network (NN) to predict the time required for inference tasks of a specific submodel on a specific satellite. By implementing COIN-LEO and evaluating its performance in a highly realistic satellite-network-emulation platform, we find that our COIN-LEO outperforms baseline algorithms in terms of inference throughput, time consumed and network traffic overhead.