Research On Multi-Mode Reconnaissance Task Planning Method For Unmanned Aerial Vehicle Swarms In Complicated Scenes

As a new operational style,unmanned operations can change the warfare.The unmanned aerial vehicle(UAV)swarm reconnaissance is one of the most typical unmanned operation application problem,that has been continuously concerned by world military powersand.This paper focuses on the wide-area persistent reconnaissance problem of the UAV swarm in complex scenes.It proposes a formal description of the operational system-of-systems(So S)architecture and the exploration algorithm of the architecture scheme space,and studies the swarm online task planning of three typical command and control modalities.The main work and innovations of this thesis are as follows:(1)For the uncertainty of the potential capability of the operational So S architecture,an optimal polynomial time space exploration algorithm that searches architecture scheme is proposed.The architecture defines the swarm task scope and specifies the swarm decision modality in different scenarios.First,the formal definition of the operational So S architecture is given based on the So S capability generation elements,and the super-network model of the operational So S architecture is constructed.Second,a exploration problem model of the operational architecture scheme is established,where the problem is then transformed into a dynamic programming problem.Given this,a dynamic exploration algorithm based on greedy search is proposed.Through theoretical analysis and simulation experiments,it is proved and verified that the proposed algorithm is polynomial time optimal.(2)For the centralized-interconnected weakly-coupled swarm reconnaissance problem,an optimal online action link planning algorithm is proposed.In the centralized-interconnected swarm,the joint action link space grows double exponentially as the number of UAVs and the planning horizon increase.In order to solve this challenge,the swarm reconnaissance problem is first cast as a multi-agent partially observable Markov decision process.Furthermore,an online planning algorithm is proposed by extending Monte Carlo tree search,which is suitable for the weakly-coupled swarm.The innovation of this algorithm is to construct a local look-ahead tree in parallel,and use the variable elimination method to calculate the optimal joint action at specific positions of local look-ahead trees.(3)For the decentralized fixed-connected large-scale swarm reconnaissance problem,an approximate optimal action link planning algorithm is proposed.In the decentralized fixed-connected swarm,it is a challenging problem to design a simple but efficient cooperation mechanism.In order to solve this challenge,the swarm reconnaissance problem is first cast as a transition-decoupled partially observable Markov decision process.Furthermore,an action link online planning algorithm suitable for large-scale UAV swarm reconnaissance is proposed.The algorithm constructs a local look-ahead tree for each UAV independently,and sequentially determines the action strategy of each UAV through a distributed sequential allocation mechanism.(4)For the decentralized unfixed-connected self-learning swarm reconnaissance problem,an online learning and task planning algorithm with approximate optimal action links under specific constraints is proposed.It is challenging to dispatch a UAV swarm to perform well in unknown environments.In order to solve this challenge,the swarm reconnaissance problem is first cast as a Bayesian adaptive transition-decoupled partially observable Markov decision process.Second,a swarm reconnaissance method adapted to the unknown environment is proposed based on Bayesian learning,Monte Carlo tree search and sequential allocation method.The outstanding feature of this algorithm is that it performs the online learning algorithm and the action link planning algorithm iteratively under a flexible cooperation mechanism.