To achieve autonomous decision-making in the process of drone power inspection and solve the issues of slow convergence and susceptibility to local optima in traditional reinforcement learning traj-ectory planning,this paper propose an improved deep reinforcement learning approach based on the Trans-former model,which designs a drone charging inspection decision-making algorithm under the constraint of battery capacity.Firstly,an energy consumption model and a Markov decision model are established for the power inspection task scenario.Then,static and dynamic encoders based on graph neural networks(GNN)and gated recurrent units(GRU)are designed to extract different types of environmental data.The multi-head pointer network is employed to plan a global charging inspection strategy and predict future rewards.Finally,the converged inference model is validated in a power inspection simulation environment.Simulation results demonstrate that compared to traditional reinforcement learning,the proposed algorithm can extract deep-level map features,path energy consumption reduced by 26.61％,while achieving better convergence.
维普
万方数据