基于卷积注意力网络的卫星观测任务序贯决策方法

Satellite Observation Task Sequential Decision-making Method Based on Convolutional Attention Neural Network

星上自主任务规划能够提高对地观测卫星应对星上任务、资源变化等动态不确定因素的快速响应能力,是卫星任务规划领域的一个重要研究热点。考虑到星上计算资源有限这一特点,现有研究主要采用启发式搜索算法对卫星星上自主任务规划问题进行求解,优化性还有待提升。基于观测任务序贯决策框架,利用卷积神经网络并行计算和注意力机制更易获取到高价值信息的优势,提出了一种基于卷积注意力神经网络的观测任务序贯算法,并设计了与之匹配的输入特征表示方法,实现对观测任务的实时决策。最后将提出算法和两种深度学习算法、两种启发式搜索算法进行了实验比较。实验结果表明,提出方法的平均响应时间不到已有深度学习算法的1/2,收益误差远低于启发式搜索算法,证实了所提方法的可行性和有效性。

Onboard task planning could improve the response of Earth observation satellite to dynamic uncertainties such as onboard task and resource changes.It has become an important research hotspot in the field of satellite task planning.Considering the limited computing resources onboard,most studies mainly used the rule-based heuristic search algorithms to solve the satellite onboard task planning problem.The optimization of the solution was difficult to meet needs,and should be further improved.An observation task sequential decision-making algorithm based on the convolutional attention neural network was proposed,which integrated the advantage of parallel computing of convolutional neural network and the advantage of attention mechanism to obtain key information.Then,a matching input feature representation method was designed.Based on the proposed method,the satellite could decide the observation task to execute in real-time.Finally,the proposed algorithm was compared with two deep learning algorithms and two heuristic search algorithms based on five scenarios with different number of tasks.The experimental results showed that the average response time of the proposed method was less than 1/2 of the existing deep learning algorithms,and the profit gap was much lower than that of the heuristic search algorithm,which confirmed the feasibility and effectiveness of the proposed method.