Research On Optimizing Cellular-Connected UAV Communication System Based On Machine Learning

In recent years,Cellular-connected Unmanned Aerial Vehicle(UAV)has received more and more attention from industry and academia.Cellular-connected UAV has the advantages of robust navigation,wide-range access,flexible configuration,cost-effeciency,convenient supervision and reusable base stations(BSs).Although cellular-connected UAV has the above advantages,compared with the traditional cellular network serving ground users,it also has many problems that need to be solved urgently.On the one hand,due to the unique air-to-ground channel model,the influence of ground communication interference,and the nonuniform of the three-dimensional antenna gain pattern of the ground BSs in practical applications,the three dimensional(3D)coverage analysis of the base stations(BSs)becomes challenging.In addition,how to ensure stable channel conditions between cellular-connected UAV and the cellular network is a challenging problem.Especially,the antennas of the BSs usually have certain downtilt angle as most of the existing cellular networks are dispatched to serve ground users.The UAV flight altitude is generally higher than that of ground BSs,which results in weak antenna gains from the BSs antenna side-lobes.Therefore,the UAV may not be covered by the cellular-connected network when it is performing tasks in the sky.Therefore,in view of the above analysis,i)To obtain channel state information(CSI)in real situations,we construct a radio map utilizing federated learning(FL).User selection and resource allocation are optimized in the process of building a radio map.ii)We used deep reinforcement learning(DRL)to design the trajectory of the UAV,which addresses the limitations of traditional methods.The innovation and main research contents of this paper are summarized as follows:Firstly,aiming at the problem of quickly and safely constructing a large-area radio map,a method to protect user data privacy and allow a large number of users to participate in a high-precision radio map is proposed.In this paper,in the UAV communication network,radio map model,energy consumption model,wireless transmission model and FL model are established to minimize the sum of the local computing energy and wireless transmission energy of the system.Considering the limited resource blocks of the UAV communication system,user selection,bandwidth allocation and UAV deployment are jointly optimized to save the total energy consumption of the system.When optimizing user selection,we propose a probabilistic user selection method,so that users with a large influence on the global model parameters have a greater probability of being selected in each iteration.This speeds up the convergence of the algorithm.Besides,the communication bandwidth is allocated reasonably according to the location of the selected user and the UAV.Then,the position of the UAV is deployed utilizing Deep Reinforcement Learning(DRL)method,which reduces the consumption of communication energy.Secondly,aiming at the scenario of cellular-connected UAVs,in order to maintain reliable connection between UAVs and ground base stations(BSs)when performing tasks,a method of UAV trajectory optimization is proposed,which reduces the interruption time of connected UAVs when performing tasks.The channel between the UAV and the BSs and the energy consumption of the UAV are modeled,and then the problem of minimizing the interruption time of the UAV when completing the task is established.To solve the established problem,it is first transformed into a Markov decision problem(MDP),and then the multi-step Dueling DDQN algorithm is adopted to optimize the trajectory of the UAV.Numerical results demonstrate the proposed design with DRL contributing to significant performance enhancement compared with the benchmark.