Cooperative Path Planning For Region Surveillance Of Multi-UAV Based On Genetic Algorithm And Deep Reinforcement Learning

With the rapid development and wide application of the multi unmanned aerial vehicle(UAV)technology,it has played an irreplaceable role in the military fields and civilian fields.In various tasks,regional surveillance is a very important task which can be executed by the UAV system in the future battlefield.When the area that needs to be monitored is relatively wide,it is difficult for a single UAV to complete the task of monitoring the entire target area,so multi-UAV are often required to be controlled through reasonable path planning.However,the path planning of multi-UAV cooperative for regional surveillance has not been well solved.Based on genetic algorithm and deep reinforcement learning,two path planning methods for multi-UAV cooperative area surveillance are proposed in this thesis,which can dynamically plan and adjust the multi-UAV route of the cooperative work in real time.and a better coverage effect was obtained in a period of time.The main works of this thesis are as follows:1.A path planning model for multi-UAV cooperative regional surveillance is established.This model takes into account the characteristics of the route which does not have a fixed destination when the UAV group performs surveillance flight mission in the target area.And under the restrict constraints of the UAV flight,the flight position of the UAV at the next moment was determined.2.A multi-UAV path planning method is presented based on genetic algorithm.Through the genetic encoding to the turning angle of the UAVs,the coverage percent of surveillance area of multiple UAVs is used as fitness function,and the corresponding genetic operation is carried out.According to the principle of maximizing the real-time monitoring area coverage,the route planning design of multi-UAV cooperative regional monitoring is carried out.This encoding method can guarantee each individual in the new generation of population is still an achievable coordinated flight way regardless of whether crossover or mutation operation is performed,which has a greater advantage in both the calculation process and the calculation quantity than the way of directly encoding the position as the gene.Through experimental simulation,the results show that the genetic algorithm can solve the problem well.In addition,two optimization methods which named hyperopia and multi-step are presented to further improve the coverage percent.The multi-step method is better than the hyperopia method,but it requires a longer time to get the results.3.A multi-UAV path planning method based on deep reinforcement learning is presented.The corresponding deep neural network is designed for each UAV,and through the trained neural network the action of the UAV can be deduced based on the current state of the UAV group.Then by repeating this learning process,the UAV will can execute the flight action which can make the next surveillance area hold the largest coverage percent until the end of the task.The simulation results show that this method can solve the problem to a certain extent,but the monitoring coverage percent of the target area is not ideal.Then,by using the Deep Q-Network(DQN)method,the deficiencies mentioned above are improved and the monitoring area coverage percent is effectively improved.In this thesis,the route planning method of multi-UAV cooperative area monitoring which is based on genetic algorithm and deep reinforcement learning can effectively maximize the coverage percent of the surveillance area under the restricted condition of UAV flight.By comparing the two presented methods,we find that the coverage percent obtained through the genetic algorithm design is higher than that of the depth reinforcement learning method,but the time consume of planning the route is longer.However,for deep reinforcement learning based method,the time is mainly consumed in training the deep neural network,and once it is trained well the route planning can be calculated directly through the forward propagation of the neural network.And by increasing data samples to train deep neural networks,the coverage percent of multi-UAV surveillance area can be effectively improved through this method.