Deep Neural Network Routing with Dynamic Space Division for 3D½ UAV FANETs

With unmanned aerial vehicles (UAVs) being widely used, the rapidly changing network topology and vertical height changes of UAVs have been bottlenecks for many wild applications, such as battlefield communication. These problems lead to the frequent communication interruptions and poor stability of 3D UAV networks. Facing these challenges, we propose deep neural network routing (DNNR) that is characterized by a dynamic 3D two-subspace division (i.e., vertical-axis cylinder and horizontal-plane divisions) and deep neural network (DNN) forwarding. With the trajectories of base station and nodes changing, vertical-axis cylinder and horizontal-plane divisions also change dynamically according to the broadcast information. Different from multi subspace division, this kind of two subspace divisions could reduce the complexity of routing discovery and make full use of the dynamic adaptability of 3D space division against the rapidly changing network topology. Due to the DNN flexibility, DNN forwarding is a promising scheme to improve the probability of recognizing the available links and select the rational next-hop node. We implement four compared protocols and DNNR in NS3 network simulator and test them for various application scenarios, when changing base station speed, node speed, horizontal plane size, and vertical height. Comparing with four protocols, DNNR achieves better performance in terms of packet delivery rate and energy-saving performance. These indicate that 3D space division is a concise and feasible scheme in flight ad hoc networks which may be extended to other fields. Besides, owing to the flexibility and prevalent availability, machine learning routing protocols are becoming a popular technology.