Agile Imaging Satellite Task Planning Method Based On Deep Reinforcement Learning

The agile imaging satellite has strong attitude maneuvering ability.Compared with the common imaging satellite,it has the ability to swing along the pitch axis,and can produce a longer observable time window for the ground observation target.With the enlargement of the application of agile imaging satellites and user task demand increasing,how to plan the input sequence of user tasks,formulate rational and effective task execution sequence,give full play to the efficiency of agile imaging satellite,to implement agile imaging satellites orbiting intelligent manufacture,so it has the vital significance.Focusing on the task planning problem of agile imaging satellite,this paper established a task planning model of agile imaging satellite with time window constraints.Based on the deep reinforcement learning method,research is carried out on how to improve the training speed,generalization ability,convergence speed and reward rate of the algorithm.The specific work includes the following aspects:1)Considering the time window constraint,attitude adjustment time during task transfer,memory constraint and power constraint,the ordinary imaging satellite task planning problem is modeled.Given under the intensive observation scene,agile imaging satellites task corresponding observable time window between the overlapping degree is higher,for agile imaging satellites placed along the side of the pitch axis ability,can be within the time window to select any period of time to perform the characteristics of observation,the agile imaging satellite task planning design to meet the time window constraints,to achieve the agile imaging satellites is longer than the time the efficient use of the window.2)In view of the complex constraints of imaging satellite task planning,large solution space and variable input task sequence length,the Pointer Networks mechanism is improved by referring to the multi-head attention mechanism,and the MHA-PN sequence decision algorithm model is proposed to solve the established imaging satellite task planning problem model.The Mask vector is used to consider various constraints in the imaging satellite task planning problem,and the algorithm model is trained through the Actor Critic reinforcement learning algorithm to obtain the maximum reward rate.According to the experimental results,it can be concluded that the MHA-PN algorithm model has a faster training speed and stronger generalization ability.3)On the basis of completing the MHA-PN algorithm to solve the task planning problem of ordinary imaging satellites,in view of the large solution space and long input task sequence of the agile imaging satellite task planning problem in dense observation scenarios,and take advantage of Ind RNN’s advantages in processing longer sequences.A fusion of Ind RNN and Pointer Networks Ind-PN algorithm model is proposed to solve the established agile imaging satellite task planning problem model,using Ind RNN as the decoder of the MHA-PN algorithm model.Experimental results show that for agile imaging satellite task planning in dense observation scenarios,the Ind-PN algorithm converges faster and can obtain higher observation reward rate.4)Establish an agile imaging satellite task planning simulation verification system based on the Django framework to meet the task planning requirements of agile imaging satellites for ground observation targets in the ground observation scenario,and simulate the algorithm functions.The system implements management functions such as adding,deleting,parameter setting and viewing for agile imaging satellites,observation targets,and observation target collections,provides a selection interface for task planning reasoning algorithms,supports embedding multiple algorithms,and provides multiple perspectives for planning results the visualization function.