Drogue detection for autonomous aerial refueling based on convolutional neural networks

Drogue detection is a fundamental issue during the close docking phase of autonomous aerial refueling(AAR). To cope with this issue, a novel and effective method based on deep learning with convolutional neural networks(CNNs) is proposed. In order to ensure its robustness and wide application, a deep learning dataset of images was prepared by utilizing real data of ‘‘Probe and Drogue" aerial refueling, which contains diverse drogues in various environmental conditions without artificial features placed on the drogues. By employing deep learning ideas and graphics processing units(GPUs), a model for drogue detection using a Caffe deep learning framework with CNNs was designed to ensure the method's accuracy and real-time performance. Experiments were conducted to demonstrate the effectiveness of the proposed method, and results based on real AAR data compare its performance to other methods, validating the accuracy, speed, and robustness of its drogue detection ability.     

改进的 YOLOv4在飞机蒙皮损伤检测中的应用

针对飞机蒙皮检测中存在的小目标检测欠佳、漏检等问题,提出了 1种基于增强特征融合和 ATSS的 YO. LOv4飞机蒙皮图像目标检测算法。首先,增加用于目标预测的大尺度浅层特征层,以提高模型对小目标的检测效果;其次,增加特征融合网络层数,通过浅层与深层特征层的深度融合,丰富多尺度特征图中的特征信息;然后,通过 K-means++聚类算法对数据集的真实框聚类,获得更具代表性的先验框尺寸,以提高预测框对目标的定位准确度;最后,引入 ATSS对 YOLOv4的样本选择策略进行优化,通过自适应获取最优的 IoU阈值,实现正负样本自动划分,提升模型的检测性能。实验表明,在增加少量计算成本的情况下,算法的检测性能得到有效提升,mAP提升 7.7%,检 测的准确率达到 80%以上。     

Deep learning enabled localization for UAV autolanding

This article concentrates on ground vision guided autonomous landing of a fixed-wing Unmanned Aerial Vehicle (UAV) within Global Navigation Satellite System (GNSS) denied environments. Cascaded deep learning models are developed and employed into image detection and its accuracy promoting for UAV autolanding, respectively. Firstly, we design a target bounding box detection network BboxLocate-Net to extract its image coordinate of the flying object. Secondly, the detected coordinate is fused into spatial localization with an extended Kalman filter estimator. Thirdly, a point regression network PointRefine-Net is developed for promoting detection accuracy once the flying vehicle’s motion continuity is checked unacceptable. The proposed approach definitely accomplishes the closed-loop mutual inspection of spatial positioning and image detection, and automatically improves the inaccurate coordinates within a certain range. Experimental results demonstrate and verify that our method outperforms the previous works in terms of accuracy, robustness and real-time criterions. Specifically, the newly developed BboxLocate-Net attaches over 500 fps, almost five times the published state-of-the-art in this field, with comparable localization accuracy.     

基于注意力YOLO-V3的合成孔径雷达舰船检测识别一体化方法

传统的合成孔径雷达舰船检测识别需要分两步实现，检测识别精度和效率难以满足实际应用需求。本文结合注意力机制和YOLO-V3网络提出了注意力YOLO-V3网络实现合成孔径雷达舰船检测识别一体化。同时，利用公开的AIR-SARShip-1.0数据集和OpenSARShip数据集构建了大场景舰船检测识别数据集，用于验证目标检测识别性能。实验结果表明，本文提出的注意力YOLO-V3网络可以获得较高的检测识别精度，证明了本文方法的有效性。     

基于深度学习的PIV流场图像修复技术

粒子图像测速技术（PIV）是空天动力装置研究中常用的流场测试方法。但对具有复杂流动特征的燃烧室,通过传统互相关算法处理得到的流场结果往往具有一定缺陷。本文将深度学习应用于PIV后处理中,以实现流场数据的异常检测和修复。在甲烷预混对冲火焰数据集上,将异常划分为两种类型,并搭建U-Net卷积神经网络架构。经过训练和优化,模型以较高置信水平识别两类异常并使用不同策略自适应修复,过滤噪声并保留原始正常数据。同时模型具有较好的可迁移性,可以为其它种类的流场数据修复提供参考。与POD迭代法和中值滤波相比,神经网络强大的非线性特征具有明显的优势,这种方法不仅修复率高,而且在不同工况下鲁棒性好。     

Pre-locate net for object detection in high-resolution images

Small-object detection has long been a challenge. High-megapixel cameras are used to solve this problem in industries. However, current detectors are inefficient for high-resolution images. In this work, we propose a new module called Pre-Locate Net, which is a plug-and-play structure that can be combined with most popular detectors. We inspire the use of classification ideas to obtain candidate regions in images, greatly reducing the amount of calculation, and thus achieving rapid detection in high-resolution images. Pre-Locate Net mainly includes two parts, candidate region classification and behavior classification. Candidate region classification is used to obtain a candidate region, and behavior classification is used to estimate the scale of an object. Different follow-up processing is adopted according to different scales to balance the variance of the network input. Different from the popular candidate region generation method, we abandon the idea of regression of a bounding box and adopt the concept of classification, so as to realize the prediction of a candidate region in the shallow network. We build a high-resolution dataset of aircraft and landing gears covering complex scenes to verify the effectiveness of our method. Compared to state-of-the-art detectors (e.g., Guided Anchoring, Libra-RCNN, and FASF), our method achieves the best mAP of 94.5 on 1920 × 1080 images at 16.7 FPS.     

物理启发的深度学习模型及其在热端部件散热中的应用

热端部件散热是众多空天设备的关键技术。表面温度分布是散热设计中用到的重要信息,常规的解析建模手段和机器学习方法均无法有效地表达此类高维信息。近年来兴起的图像深度学习算法是解决表面温度信息预测的有效手段。然而,现有的基于大数据的深度学习方法往往对于物理数据和小样本数据不适用,体现为泛化精度差、数据兼容性差、可解释性差。因此,有必要结合传热的先验知识发展物理启发的新型深度学习算法,以增强高自由度、高复杂度散热对象上的设计能力。本文基于卷积算子和有限差分求解方式的类比关系,提出了一种物理启发式的循环卷积神经网络。以横向出流的冲击冷却为例,开展了变计算域大小、变工况、变尺寸的批量数值模拟,获取了冲击冷却关键特征的小样本图像数据。进一步通过神经网络的训练,构建了多参数、大范围内有较好拟合能力的温度、传热系数、压力代理模型。研究结果表明,本文提出的物理启发神经网络模型,对于计算域大小没有限制,可以统一表达不同空间范围内获取的物理数据的共性规律。模型的各类超参设定均具有明确的物理意义,且与经典的微分方程求解理论有一定的类比关系,增强了神经网络调参的方向性。通过传热物理规律与黑箱模型的融合,本文实现了小样本多参数物理数据的共性建模。该方法可以迅速重构热端部件的高维分布信息,可服务于热端部件的快速分析设计以及优化。     

Prior-guided GAN-based interactive airplane engine damage image augmentation method

Deep learning-based methods have achieved remarkable success in object detection, but this success requires the availability of a large number of training images. Collecting sufficient training images is difficult in detecting damages of airplane engines. Directly augmenting images by rotation, flipping, and random cropping cannot further improve the generalization ability of existing deep models. We propose an interactive augmentation method for airplane engine damage images using a prior-guided GAN to augment training images. Our method can generate many types of damages on arbitrary image regions according to the strokes of users. The proposed model consists of a prior network and a GAN. The Prior network generates a shape prior vector, which is used to encode the information of user strokes. The GAN takes the shape prior vector and random noise vectors to generate candidate damages. Final damages are pasted on the given positions of background images with an improved Poisson fusion. We compare the proposed method with traditional data augmentation methods by training airplane engine damage detectors with state-of-the-art object detectors, namely, Mask R-CNN, SSD, and YOLO v5. Experimental results show that training with images generated by our proposed data augmentation method achieves a better detection performance than that by traditional data augmentation methods.     

样本标签污染条件下的雷达辐射源个体识别技术

针对辐射源个体识别(Specific Emitter Identification,SEI)中由于数据集存在错误标签导致识别率下降的问题,提出了 1种有监督和无监督融合的错误标签识别和纠正方法。首先采用无监督密度峰值聚类方法将数据集中出现的标签错误样本找出,再使用 K折交叉实验对这些标签异常的样本进行预测投票,将得票数多的标签作为错误标签纠正的结果。经过清洗的数据集再通过卷积神经网络进行训练,得到 1个较为理想的辐射源个体识别的网络模型,保证了在样本污染条件下,辐射源个体识别网络仍能具有较好的识别率。文章所提方法的识别率相比未经处理的数据集的识别率在标签错误率小于 30%时平均提高 3.3%;在标签错误率大于 30%时,也能使个体识别率达到 90%左右,验证了文章所提方法在对错误标签的识别和纠正上可以取得较好的效果。     

基于CenterNet 的航空遥感图像目标检测

为实现高精度的航空图像目标检测，将Anchor free 的目标检测算法CenterNet 应用到检测中，同时 使用Resnet50 主干网络，并引入CIoU 损失替代原有损失函数对网络模型做出了改进。改进后的算法在RSOD 与DIOR 数据集上进行测试，结果显示在保证网络轻量化的前提下检测精度有明显的提高，证明了算法在航空 目标检测方面的可行性与准确性。     

Three-Axes Attitude Determination of Spacecraft Using a Laser

A new method for attitude determination of spacecraft is proposed. The distinctive feature of this method is the ability to determine with high accuracy three elementary angles of the attitude by detection of the electromagnetic wave transmission from a single point. The system consists of a transmitter of a linearly polarized laser beam on the earth (or spacecraft) and receiving equipment on a relevant spacecraft. When the system is used for geosynchronous satellites, the possible accuracies of determination are 10-4 rad or higher for the angles which correspond to roll and pitch, and 10-2 rad or higher for the angle which corresponds to yaw, with the period of 1 s. The system margin for atmospheric attenuation is estimated to be about 50 dB (midnight) to about 20 dB (midday) on the basis of commercially available components. Consequently, it becomes possible to orient antennas or detectors toward arbitrary points around the laser transmitting point on the earth with a high pointing accuracy.     

基于多特征图像增强深度卷积神经网络的航天用电子元器件分类算法

为了实现航天用电子元器件的全自动及非接触识别,并减少由照明系统造成的图像亮度不均、偏色等问题对检测结果的影响,通过结合局部、区域和总体三个层次特征提升物体检测精度,提出了一种基于多特征图像增强深度卷积神经网络(MFIE-DCNN)的航天用电子元器件分类算法。MFIE-DCNN算法包含多特征学习和深度学习,其学习过程类似于人类视觉系统,能够对形状、方向和颜色特征进行深度挖掘,突出元器件边界信息,抑制背景杂波干扰。实验结果表明,该算法能够区分电路板板载元器件的种类,检测准确度优于传统算法。对比基于稀疏自动编码器的深度神经网络,检测结果提高了近20%。     

GroupNet: Learning to group corner for object detection in remote sensing imagery

Due to the attractive potential in avoiding the elaborate definition of anchor attributes,anchor-free-based deep learning approaches are promising for object detection in remote sensing imagery. Corner Net is one of the most representative methods in anchor-free-based deep learning approaches. However, it can be observed distinctly from the visual inspection that the Corner Net is limited in grouping keypoints, which significantly impacts the detection performance. To address the above problem, ...     

Visual inspection of aircraft skin: Automated pixel-level defect detection by instance segmentation

Skin defect inspection is one of the most significant tasks in the conventional process of aircraft inspection. This paper proposes a vision-based method of pixel-level defect detection, which is based on the Mask Scoring R-CNN. First, an attention mechanism and a feature fusion module are introduced, to improve feature representation. Second, a new classifier head—consisting of four convolutional layers and a fully connected layer—is proposed, to reduce the influence of information around the area of the defect. Third, to evaluate the proposed method, a dataset of aircraft skin defects was constructed, containing 276 images with a resolution of 960 × 720 pixels. Experimental results show that the proposed classifier head improves the detection and segmentation accuracy, for aircraft skin defect inspection, more effectively than the attention mechanism and feature fusion module. Compared with the Mask R-CNN and Mask Scoring R-CNN, the proposed method increased the segmentation precision by approximately 21% and 19.59%, respectively. These results demonstrate that the proposed method performs favorably against the other two methods of pixel-level aircraft skin defect detection.     

基于改进Yolov3 算法的舰船目标检测识别系统

针对复杂战场环境下对海目标检测识别的需求,设计了一种基于改进Yolov3 算法的海面舰船目标实 时检测识别系统。使用微调分类网络、增加训练尺度、聚类目标边框维度、二级特征分类等方法对Yolov3 检 测识别网络模型进行了优化,在提高识别精度的同时有效降低了漏检率和虚警率。实验结果表明,优化后的网 络模型在自建的舰船图像数据库中将检测识别平均准确率提高到了79.3%,对真实海上航拍视频中舰船目标识 别的平均准确率达到了81% 以上。     

面向飞行员疲劳状态监测的脑认知深度模型研究

通过分析飞行员脑电信号，构建了疲劳状态的彩色脑功率图，设计基于高斯牛顿在线变分方法的卷 积神经网络参数优化方法，形成一种新型脑功率图深度网络模型，有效实现脑功率图深度网络的模型分类识别 能力。相比于其它基于脑电信号的疲劳检测深度模型，疲劳状态认知的准确度提升了3%~5%。     

基于改进YOLOv4的航空发动机小目标损伤检测研究

智能化的航空发动机损伤检测是飞机故障诊断重要的研究方向,针对现有目标检测模型对航空发动机的小目标损伤检测效果差的问题,提出了一种改进的基于You Only Look Once version 4(YOLOv4)的多尺度目标检测方法。在路径聚合网络（PANet）中构建低层次的特征融合层,将更浅层的特征与深层特征融合,提高网络对小目标损伤的检测性能。为减少网络中的冗余参数,在颈部结构中引入了深度可分离卷积,将标准卷积重构为深度可分离卷积的形式。实验表明：改进后的YOLOv4对小目标损伤的检测精度提升了3.43%,模型大小降低了54.06 MB,同时检测速度提高了31.03%。研究结果表明改进的YOLOv4模型对小目标损伤具有更好的检测性能。     

An inverse design method for supercritical airfoil based on conditional generative models

Inverse design has long been an efficient and powerful design tool in the aircraft industry.In this paper, a novel inverse design method for supercritical airfoils is proposed based on generative models in deep learning. A Conditional Variational Auto Encoder(CVAE) and an integrated generative network CVAE-GAN that combines the CVAE with the Wasserstein Generative Adversarial Networks(WGAN), are conducted as generative models. They are used to generate target wall Mach distributions for the inve...     

基于深度学习的非合作航天器姿态估计

针对空间非合作航天器姿态测量时受光照和地球背景影响大的问题,提出了一种基于卷积神经网络的端到端姿态估计方法.在该方法中,主干网络采用AlexNet与ResNet.首先,移除主干网络末端的全连接层,并列连接3个全连接层,采用三分支网络分别对姿态角进行估计.然后,设计了将分类问题与回归问题相结合的损失函数,通过分类方法将姿态估计限定在某一范围内,再使用回归方法进一步微调姿态.姿态分类损失函数确定姿态角度基准点,姿态回归损失函数对估计角度进行微调.相较于仅采用回归方法进行姿态估计,此方法能够有效减小姿态估计平均绝对误差、标准差与最大误差.实验对比了不同主干网络的测量精度,平均绝对误差在0.376°～0.746°之间,最优标准差为0.474°.     

基于综合应力工作态试验和神经网络的CMG失效边界域预测

本文设计了一种可同时模拟真空热环境和CMG与航天器角动量交换工况的试验设备,提出了模拟在轨真空环境下温度、CMG框架转速、航天器转速3种应力的工作态试验方法,给出了适用于神经网络的CMG运行状态定量表达方法,利用少量试验数据和神经网络方法对工作极限转速矩阵、失效边界、失效边界域进行预测,分析了经验样本对预测结果的影响,以及各应力对其他应力工作域的耦合影响,并给出了预测结果的可信度分析方法。研究结果表明,所提出的方法可以更真实模拟CMG在轨工作状态的同时显著节省试验经费和时间,并具有较高的预测准确性和多应力工作场景适应性,对I和Ⅱ两类训练数据集分别获得100%和98.8%的预测正确率,给出了仅凭试验数据无法得到的55℃下的转速失效边界,并且可以内化试验数据背后的工程经验。