GEO卫星机动识别的自适应深度学习方法  

An Adaptive Deep Learning Approach for Maneuver Identification of GEO Satellite

作　　者：龙洗 杨乐平[1] 乔琛远 黄涣[1] LONG Xi;YANG Leping;QIAO Chenyuan;HUANG Huan(College of Aerospace Science and Engineering,National University of Defense Technology,Changsha 410073,China)

机构地区：[1]国防科技大学空天科学学院,长沙410073

出　　处：《宇航学报》2025年第1期51-67,共17页Journal of Astronautics

摘　　要：针对地球同步轨道(GEO)卫星未来机动状态识别问题,提出了一种“经度预测+机动识别”的自适应深度学习方法。首先,构建自动优化时空因果注意力深度学习框架(AOSTCADL),引入长短记忆神经网络(LSTM)、卷积模块、传递熵和注意力机制,并采用智能算法优化网络参数,以提高经度预测的精度。其次,基于经度预测结果,提出非参数化阈值动态更新方法,引入滚动时域策略提高机动识别F1分数。针对3个典型GEO卫星的仿真结果表明:与当前流行的两类CNN-LSTM模型相比,所提方法经度预测最小平均误差、均方根误差平均降低一个数量级,机动识别平均准确率达97.52%,F1分数达77.09%。An adaptive deep learning approach,encompassing longitude prediction and maneuver identification,is proposed for the identification of geosynchronous Earth orbit(GEO)satellite incoming maneuvers.Initially,an automatic optimization spatio-temporal causal attention deep learning framework(AOSTCADL)is developed.This framework incorporates long short-term memory(LSTM)neural networks,convolutional neural network(CNN)modules,transfer entropy,and attention mechanisms.Intelligent algorithms are employed to optimize network parameters,thereby enhancing longitude prediction accuracy.Subsequently,based on longitude prediction results,a non-parametric threshold dynamic updating method is introduced.The rolling horizon strategy is utilized to boost the F1 score for maneuver identification.Simulations are conducted using three typical GEO satellites,and the results demonstrate that the proposed framework significantly reduces the mean absolute error(MAE)and mean square error(MSE)by an order of magnitude compared to two commonly used CNN-LSTM models in longitude prediction.The average accuracy of maneuver identification is 97.52%,with an F1 score of 77.09%.

关 键 词：太空态势感知(SSA) GEO卫星 机动识别 长短记忆神经网络(LSTM) 自适应学习 

分 类 号：V11[航空宇航科学与技术—人机与环境工程]