基于深度增强学习的卫星姿态控制方法

针对卫星在执行丢弃载荷或捕获目标等复杂任务时遭遇的姿态突然发生变化的问题,采用深度增强学习方法对卫星姿态进行控制,使卫星恢复稳定状态。具体来说,首先搭建飞行器的姿态动力学环境,并将连续的控制力矩输出离散化,然后采用Deep Q Network算法进行卫星自主姿态控制训练,以姿态角速度趋于稳定作为奖励获得离散行为的最优智能输出。仿真试验表明,面向空间卫星姿态控制的深度增强学习算法能够在卫星受到突发随机扰动后稳定卫星姿态,并能有效解决传统PD控制器依赖被控对象质量参数的难题。所提出的方法采用自主学习的方式对卫星姿态进行控制,具有很强的智能性和一定的普适性,在未来卫星执行复杂空间任务中的智能控制方面有着很好的应用潜力。     

GPS/INS超紧组合载波跟踪环路最优带宽在线设计

在全球定位系统和惯性导航系统组成的超紧耦合系统中,卫星信号的跟踪性能直接取决于载波跟踪环路的带宽。为提高最优带宽的计算精度,在对惯导辅助下载波跟踪环路跟踪特性进行分析的基础上,详细推导了载波多普勒频率估计误差、多普勒频率变化率估计误差的计算方式,建立了惯导辅助下的环路跟踪误差模型;在实时估计跟踪载噪比的基础上,应用离散牛顿二阶梯度法迭代解算最优带宽,并进行实时调整。仿真结果表明,所设计最优带宽迭代解算方法的计算精度能够在11次迭代内达到99.6%,以此作为环路的带宽,能够在弱信号、辅助信息精度较低的情况下有效提高环路的跟踪精度。     

基于星体边缘和轨道投影的光学自主导航算法

针对光学导航中存在的通过星体（球体）图像部分边缘点拟合椭圆参数计算轨道参数产生中间误差的问题,提出利用边缘点映射轨道参数的直接投影模型,避免拟合椭圆参数的方法。在小孔成像模型基础上,建立了边缘点与轨道参数的直接投影数学模型,对其映射过程进行了理论推导,利用列文伯格-马夸尔特迭代算法进行求解轨道参数。用实际探测器以及镜头参数进行数值仿真验证,结果表明：该方法在相同边缘点的条件下,轨道精度可以达到5‰。与传统方法相比,这种方法避免了椭圆的拟合过程,减少了引入中间误差过程。     

TopPixelLoss:类别不均衡的遥感影像语义分割损失函数

针对遥感影像中类别不均衡的小目标分割效果不理想的问题，提出了一种类别不均衡小目标二分类分割的损失函数——TopPixelLoss损失函数。首先计算出每个像素的交叉熵，然后将所有像素的交叉熵按从大到小进行排序，随后确定一个K值作为阈值，筛选出前K个交叉熵最大的像素，最后对于筛选出的K个像素交叉熵取平均，做为损失值。在ISPRS 提供的 Vaihingen 数据集上，使用PSPNet网络与普通交叉熵、FocalLoss、TopPixelLoss三种损失函数分别对车辆进行二分类分割试验。结果表明，不同的K值，使用TopPixelLoss损失函数的平均交并比（MIoU）、F1-score、准确度（ACC）都最高;当K值为5×104时效果最佳，MIoU、F1-score、ACC分别比FocalLoss提高了3.0%、5.0%、0.1%。TopPixelLoss损失函数是一种针对类别不均衡分割非常有效的损失函数     

基于时标分解的弹性高超声速飞行器智能控制

考虑弹性高超声速飞行器纵向动力学模型，提出了一种基于时标分解的智能控制方法。考虑刚体状态和弹性模态具有不同的时标特性，采用奇异摄动理论进行快慢时标分解，将模型转换为刚体慢变子系统和弹性快变子系统。针对刚体子系统考虑动力学不确定，基于平行估计模型构造表征不确定逼近效果的预测误差，结合跟踪误差给出复合学习控制策略。针对弹性子系统设计自适应滑模控制稳定弹性模态。通过李雅普诺夫稳定性分析可证系统状态一致终值有界。仿真表明所提出的控制方法能够实现刚弹模态的稳定收敛，且具有更高的跟踪精度、更好的学习性能和更快的收敛速度。     

Propulsionless planar phasing of multiple satellites using deep reinforcement learning

This work creates a framework for solving highly non-linear satellite formation control problems by using model-free policy optimisation deep reinforcement learning (DRL) methods. This work considers, believed to be for the first time, DRL methods, such as advantage actor-critic method (A2C) and proximal policy optimisation (PPO), to solve the example satellite formation problem of propellantless planar phasing of multiple satellites. Three degree-of-freedom simulations, including a novel surrogate propagation model, are used to train the deep reinforcement learning agents. During training, the agents actuated their motion through cross-sectional area changes which altered the environmental accelerations acting on them. The DRL framework designed in this work successfully coordinated three spacecraft to achieve a propellantless planar phasing manoeuvre. This work has created a DRL framework that can be used to solve complex satellite formation flying problems, such as planar phasing of multiple satellites and in doing so provides key insights into achieving optimal and robust formation control using reinforcement learning.     

基于DDPG算法的无人机集群追击任务

无人机的集群化应用技术是近年来的研究热点，随着无人机自主智能的不断提高，无人机集群技术必将成为未来无人机发展的主要趋势之一。针对无人机集群协同执行对敌方来袭目标的追击任务，构建了典型的任务场景，基于深度确定性策略梯度网络（DDPG）算法，设计了一种引导型回报函数有效解决了深度强化学习在长周期任务下的稀疏回报问题，通过引入基于滑动平均值的软更新策略减少了DDPG算法中Eval网络和Target网络在训练过程中的参数震荡，提高了算法的训练效率。仿真结果表明，训练完成后的无人机集群能够较好地执行对敌方来袭目标的追击任务，任务成功率达到95%。可以说无人机集群技术作为一种全新概念的作战模式在军事领域具有潜在的应用价值，人工智能算法在无人机集群的自主决策智能化发展方向上具有一定的应用前景。     

Modeling irregular small bodies gravity field via extreme learning machines and Bayesian optimization

Close proximity operations around small bodies are extremely challenging due to their uncertain dynamical environment. Autonomous guidance and navigation around small bodies require fast and accurate modeling of the gravitational field for potential on-board computation. In this paper, we investigate a model-based, data-driven approach to compute and predict the gravitational acceleration around irregular small bodies. More specifically, we employ Extreme Learning Machine (ELM) theories to design, train and validate Single-Layer Feedforward Networks (SLFN) capable of learning the relationship between the spacecraft position and the gravitational acceleration. ELM-base neural networks are trained without iterative tuning therefore dramatically reducing the training time. Analysis of performance in constant density models for asteroid 25143 Itokawa and comet 67/P Churyumov-Gerasimenko show that ELM-based SLFN are able learn the desired functional relationship both globally and in selected localized areas near the surface. The latter results in a robust neural algorithm for on-board, real-time calculation of the gravity field needed for guidance and control in close-proximity operations near the asteroid surface.     

Bayesian network-based extraction of lunar impact craters from optical images and DEM data

Impact craters are among the most noticeable geomorphological features on the planetary surface and yield significant information about terrain evolution and the history of the solar system. Thus, the recognition of impact craters is an important branch of modern planetary studies. Aiming at addressing problems associated with the insufficient and inaccurate detection of lunar impact craters, a decision fusion method within the Bayesian network (BN) framework is developed in this paper to handle multi-source information from both optical images and associated digital elevation model (DEM) data. First, we implement the edge-based method for efficiently searching crater candidates which are the image patches that can potentially contain impact craters. Secondly, the multi-source representations of an impact crater derived from both optical images and DEM data are proposed and constructed to quantitatively describe the two-dimensional (2D) and three-dimensional (3D) morphology, consisting of Histogram of Oriented Gradient (HOG), Histogram of Multi-scale Slope (HMS) and Histogram of Multi-scale Aspect (HMA). Finally, a BN-based framework integrates the multi-source representations of impact craters, which can provide reductant and complementary information, for distinguishing craters from non-craters. To evaluate the effectiveness and robustness of the proposed method, experiments were conducted on three lunar scenes using both orthoimages from the Lunar Reconnaissance Orbiter (LRO) and DEM data acquired by the Lunar Orbiter Laser Altimeter (LOLA). Experimental results demonstrate that integrating optical images with DEM data significantly decreases the number of false positives compared with using optical images alone, with F₁-score of 84.8% on average. Moreover, compared with other existing fusion methods, our proposed method was quite advantageous especially for the detection of small-scale craters with diameters less than 1000 m.     

月球表层采样样品智能确认方法

表层采样是月球采样探测的重要方式，样品智能确认有助于提升工作效率与复杂问题处理能力。结合月球表层采样铲挖工作过程，分析了铲挖过程中臂载相机图像的特点，模仿有人参与识别过程，提出了层次解耦的月球样品智能识别流程，利用深度学习方法构建了一类深度卷积识别网络，完整地描述了图像、特征、标记在网络中的正反传递关系，并在月球表层采样地面试验中进行了验证，结果表明该方法对不同光照、不同背景、不同过程、不同形态的样品，具有较好的泛化识别能力，误识别率优于8.1%，平均单幅识别时间约0.7 s。