Ground maneuver for front-wheel drive aircraft via deep reinforcement learning

The maneuvering time on the ground accounts for 10%–30% of their flight time, and it always exceeds 50% for short-haul aircraft when the ground traffic is congested. Aircraft also contribute significantly to emissions, fuel burn, and noise when taxiing on the ground at airports. There is an urgent need to reduce aircraft taxiing time on the ground. However, it is too expensive for airports and aircraft carriers to build and maintain more runways, and it is space-limited to tow the aircraft fast using tractors. Autonomous drive capability is currently the best solution for aircraft, which can save the maneuver time for aircraft. An idea is proposed that the wheels are driven by APU-powered (auxiliary power unit) motors, APU is working on its efficient point; consequently, the emissions, fuel burn, and noise will be reduced significantly. For Front-wheel drive aircraft, the front wheel must provide longitudinal force to tow the plane forward and lateral force to help the aircraft make a turn. Forward traction effects the aircraft’s maximum turning ability, which is difficult to be modeled to guide the controller design. Deep reinforcement learning provides a powerful tool to help us design controllers for black-box models; however, the models of related works are always simplified, fixed, or not easily modified, but that is what we care about most. Only with complex models can the trained controller be intelligent. High-fidelity models that can easily modified are necessary for aircraft ground maneuver controller design. This paper focuses on the maneuvering problem of front-wheel drive aircraft, a high-fidelity aircraft taxiing dynamic model is established, including the 6-DOF airframe, landing gears, and nonlinear tire force model. A deep reinforcement learning based controller was designed to improve the maneuver performance of front-wheel drive aircraft. It is proved that in some conditions, the DRL based controller outperformed conventional look-ahead controllers.     

基于强化学习的航空兵认知行为模型

航空兵的认知行为模型为仿真航空兵的空战决策提供支持,通过强化学习积累战术决策经验.在虚拟战场环境中,作战态势通过多个属性进行描述,这使得强化学习过程将面临一个高维度的问题空间.传统的空间离散化方法处理高维空间时将对计算资源和存储资源产生极大需求,因此不可用.通过构造一个基于高斯径向基函数的拟合网络解决了这个问题,大大减少了对资源的需求以及强化学习周期,并最终产生了合理的机动策略.模型的有效性和自适应性通过一对一的空战仿真进行了验证,产生的交战轨迹与人类飞行员产生的交战轨迹类似.     

基于深度学习的无人机航拍目标检测研究综述

目标检测是提高无人机（UAV）感知能力的关键技术之一,其研究对于无人机的应用有着重要意义。与基于手工特征的传统方法相比,基于卷积神经网络的深度学习方法具有强大的特征学习和表达能力,成为目前目标检测任务的主流算法。近年来,目标检测技术已经在自然场景图像上取得了一系列突破性进展,在无人机领域的研究也逐渐成为热点。首先系统阐述了基于深度学习的目标检测算法的研究进展,并总结了相关算法的优缺点。对常见的航空影像数据集进行了梳理并介绍了迁移学习的方法;从无人机影像背景复杂、目标较小、视场大、目标具有旋转性的特点出发,对无人机目标检测在近期的研究进行了归纳和分析。最后讨论了存在的问题和未来可能的发展方向。     

基于改进SSD算法的航拍目标检测算法研究

研究基于深度学习技术的无人机航拍图像目标检测算法,首先介绍目标检测算法SSD(Single Shot MultiBox Detector),并对其特征提取网络进行改进,采用稠密特征提取网络替换原网络的主干特征提取网络,提高算法的特征提取能力,从而提升了算法的检测精度。针对网络实时性问题,在算法中引入分组卷积,极大地减少了网络参数量,提升了网络推理速度。为解决训练中出现的正负样本不均衡问题,利用焦点损失（Focal Loss）改进了原算法的损失函数,进一步提升了网络的收敛速度和精度。最后,通过仿真验证了改进算法在目标检测精度上的优越性。     

基于YOLOv5的红外目标检测算法

针对红外图像的特点,提出了一种YOLOv5-IF算法,采用基于残差机制的特征提取网络,实现了不同特征层之间信息的高效交互,能够得到更丰富的目标语义信息。通过改进YOLOv5的检测方案,增加更大尺度的检测头,有效提升了红外图像中小目标的检测概率。针对计算平台资源有限、算法实时性要求高等问题,设计了Detection Block模块,并由此构建了特征整合网络,该模块不仅能提升算法检测精度,还可有效缩减模型参数量。在FLIR红外自动驾驶数据集上,该算法的平均准确率(mAP)为74%,参数量仅19.5MB,优于现有算法。     

机器学习辅助湍流建模在分离流预测中的应用

数据驱动湍流建模是近年来发展的提高雷诺平均N-S方程预测精度的有效手段,通过机器学习算法能够从高置信度数据中自动提取特征,建立准确的从平均流动特征到雷诺应力的预测模型。针对高雷诺数下积冰翼型绕流这一类典型的复杂流动分离问题,基于此前研究者提出的机器学习预测框架,从输入输出特征选择和翼型绕流中数据分布特性两个方面出发,对机器学习预测结果的光滑性和准确性进行改善。提出了基于雷诺应力张量分析和流动特征辨识的输入特征选择准则;提出了局部区域建模方法以及基准模型和机器学习预测模型混合的代入计算方法。将改进方法应用于积冰翼型绕流问题之中,结果表明改进的方法能够准确给出训练集和预测集上的雷诺应力结果,并且代入平均流计算可以得到和真实分布更加接近的流动和机翼表面压力分布。     

行星软着陆GPS有模型强化学习制导方法

由于距离地球较远、测控延时误差较大、飞行环境十分复杂且难以提前预测，行星软着陆的自主制导技术目前存在水平位置估计困难、导航参考信息匮乏、复杂地形着陆困难等挑战。针对行星软着陆存在的困难和挑战，提出了基于引导策略搜索算法的有模型强化学习制导方法，实现了着陆器在初始状态受到扰动时，无需重新规划，仍能在满足约束条件的情况下降落在指定位置。该方法将迭代线性二次调节器作为控制器，产生初始轨迹；其次，使用多层神经网络拟合制导策略；最后，利用控制器监督策略学习，进而收敛产生可行策略。针对行星表面软着陆的仿真验证结果显示该算法仅通过几次循环，即可以实现初始状态变化的快速软着陆。一方面表明了基于有模型强化学习的数据高效利用率，另一方面也证明了强化学习方法在深空探测领域中具有广阔的应用前景。     

基于PCA-MPSO-ELM的空战目标威胁评估

目标威胁评估是空战对抗过程中的关键环节。由于影响空战目标威胁评估的因素复杂多样，且指标之间存在相关性，导致传统的评估算法无法得到准确客观的评估结果。由此，提出了一种基于主成分分析法和改进粒子群算法优化的极限学习机（PCA-MPSO-ELM）的目标威胁评估算法。首先，综合分析了影响目标威胁程度的指标，利用主成分分析法对原始评估指标进行线性变化处理得到综合变量，消除了评估指标之间的相关性，实现了对评估数据的降维；在此基础上，构建ELM神经网络并利用改进的粒子群算法优化极限学习机的输入权值和阈值，提高了目标威胁评估模型的精度。最后，在空战训练测量仪中选取空战对抗数据，利用威胁指数法构造了目标威胁评估样本数据，通过仿真实验分析了PCA-MPSO-ELM算法的精度和实时性，结果表明所提算法可以快速准确地进行空战目标威胁评估。     

自主空战技术中的机动决策:进展与展望

自主空战（AAC）是指飞机依靠机载等相关设备，自主进行战场感知、决策及控制，以执飞空战的技术，其核心是机动决策模块。针对该模块的研究，调研了国内外的现有方法，并按照有关方法的核心内涵，将其分为基于数学求解、机器搜索、以及数据驱动3类分别简要概述；针对每一类方法列出了具有代表性的示例技术，并讨论了其优缺点。指出了自主空战决策的研究应立足现有相关方法及技术基础，同时积极吸纳机器学习、人工智能等新兴技术，合理利用各类方法的优点，取长补短，实现自主空战决策的长足发展；有关于该研究设想的关键科学问题及其潜在解决方案也在文中进行了初步讨论。希冀基于针对自主空战决策研究框架及实现方法的概念性讨论，促进其技术发展，推广相关思路在有关国家安全、国民经济建设等领域的应用，以满足国家重大需求和工程实践。     

Heading south with a north-up map: A choice processing analysis of map alignment effects

This study employed an information accumulation model of choice reaction times to investigate alignment effects in mental representations of maps. University students studied a map from a single orientation (with north at the top). In a subsequent two-choice reaction time task, the students’ spatial knowledge of the map was assessed employing spatial left/right judgments, which were made from imagined perspectives that were either north-aligned or south-aligned. Data showed a standard alignment effect, favoring north- over south-aligned trials. To examine the locus of this effect, data were fit using the Linear Ballistic Accumulator (LBA) model of speeded decisions (Brown & Heathcote, 2008). Of interest were three model parameters: drift rate, the speed at which evidence accumulates toward a response; response threshold, the amount of evidence demanded from the decision maker before selecting a response; and non-decision time, the time consumed by pre- and postdecisional processes. The best-fitting model suggested that non-decision time accounted for the alignment effect. The difference in non-decision time between north and south-aligned judgments suggests a mental alignment stage on south-aligned trials, accounting for the longer reaction times for judgements misaligned with the presented north orientation of the map.