关键词: 强化学习, 神经网络, 多无人机, 分布式控制, 林火侦察, 林火蔓延仿真

Abstract: In order to improve the practicality and autonomy in multiple UAV collaborative control for persistent wildfire surveillance, based on the Spread Vector Induced Celluar Automata (SVICA) fire spread algorithm, UAV kinematic, and sensor modeling, a more realistic three-dimension simulation environment for multiple UAV fire surveillance and the surveillance effectiveness indicators are constructed. A two-layer distributed control architecture for multiple UAV is proposed. Based on the reinforcement learning trained Artificial Neural Networks (ANN), the operational layer control realized the autonomous wildfire surrounding and terrain following under windy conditions. In the tactical layer, through the temporal even-distribution algorithm, the discrete airspeed adjustment of each UAV is performed to achieve the uniformity and immediacy of temporal distribution of UAVs in persistent wildfire surveillance. Then through a series of numerical experiments, the adaptability of the proposed distribution control method is verified under sudden UAV loss and supplement during the surveillance. In addition, based on the research on the relationship between the number of UAVs and the surveillance effectiveness, the UAV dispatch threshold is defined and enduring UAV dispatch and recycling strategy is examined. The final simulation results show that the proposed multiple UAV distributed control algorithm is effective and practical for persistent surveillance of wildfire.

Key words: reinforcement learning, artificial neural network, multiple UAVs, distributed control, wildfire surveillance, fire spread simulation