南京航空航天大学(Nanjing University of Aeronautics and Astronautics,NUAA)雷达探测与成像团队利用自主研发的无人机载微小型合成孔径雷达(Synthetic aperture radar,SAR)系统针对不同型号的坦克、装甲车和战机等十余类典型军事目...南京航空航天大学(Nanjing University of Aeronautics and Astronautics,NUAA)雷达探测与成像团队利用自主研发的无人机载微小型合成孔径雷达(Synthetic aperture radar,SAR)系统针对不同型号的坦克、装甲车和战机等十余类典型军事目标构建了圆周SAR数据集。通过对多次外场试验数据的高精度成像处理,在多俯仰角单基圆周SAR图像数据集的基础上,扩展了不同双基角组合的双基圆周SAR图像数据集。基于该数据集,本文结合团队在SAR图像目标检测和识别方法及应用方面的研究成果,对基于深度学习的SAR目标检测识别技术进行了回顾和综述,对比了不同神经网络模型在南航无人机载圆周SAR数据集上的检测和识别性能。具体地,在目标检测方面,利用SAR图像固有属性获得目标位置信息并结合单阶段轻量级检测算法,提出利用信息分布规律并结合全局注意力机制捕捉小目标位置信息的检测算法,以提高复杂背景下的小目标检测准确率和效率。在目标识别方面,在通过SAR图像先验信息抑制干扰噪声的基础上,提出利用SAR目标多视角信息联合Transformer的目标识别算法,通过设计视角正则化项以约束多视角之间的关联性从而实现不同视角间的特征融合,提高SAR小目标识别的准确率。从无人机载微型SAR系统对地面目标进行实时检测和识别的实际需求出发,本文还探讨了轻量化检测和识别网络在数字信号处理(Digital signal processing,DSP)平台上的部署方案,同时展示了初步试验结果。最后,本文展望了SAR目标智能检测和识别领域面临的挑战和发展趋势。展开更多
The high-quality assembly of Large Aircraft Components(LACs)is essential in modern aviation manufacturing.Numerical control locators are employed for the posture adjustment of LAC,yet the system's multi-input mult...The high-quality assembly of Large Aircraft Components(LACs)is essential in modern aviation manufacturing.Numerical control locators are employed for the posture adjustment of LAC,yet the system's multi-input multi-output,nonlinearity,and strong coupling presents significant challenges.The substantial internal force generated during the adjustment process can potentially damage the LAC and degrade the assembly quality.Hence,a workspace-based hybrid force position control scheme was developed to achieve high quality assembly with high-precision and lower internal force.Firstly,an offline workspace analysis with inherent geometric characteristics to form time-varying posture error constraint.Then,the posture error is integrated into the online position axis control to ensure tracking the ideal posture,while the force control axis compensates for posture deviation by minimizing internal force,thereby achieving high precision and low internal force.Finally,the effectiveness was demonstrated through experiments.The root mean square errors of orientation and position are 104 rad and 0.1 mm,respectively.A reduction in internal force can range from 10.96%to 57.4%compared to the traditional method.Key points'max position error is decreased from 0.32 mm to 0.18 mm,satisfying the 0.5 mm tolerance.Therefore,the proposed method will help promote the development of high-performance manufacturing.展开更多
文摘南京航空航天大学(Nanjing University of Aeronautics and Astronautics,NUAA)雷达探测与成像团队利用自主研发的无人机载微小型合成孔径雷达(Synthetic aperture radar,SAR)系统针对不同型号的坦克、装甲车和战机等十余类典型军事目标构建了圆周SAR数据集。通过对多次外场试验数据的高精度成像处理,在多俯仰角单基圆周SAR图像数据集的基础上,扩展了不同双基角组合的双基圆周SAR图像数据集。基于该数据集,本文结合团队在SAR图像目标检测和识别方法及应用方面的研究成果,对基于深度学习的SAR目标检测识别技术进行了回顾和综述,对比了不同神经网络模型在南航无人机载圆周SAR数据集上的检测和识别性能。具体地,在目标检测方面,利用SAR图像固有属性获得目标位置信息并结合单阶段轻量级检测算法,提出利用信息分布规律并结合全局注意力机制捕捉小目标位置信息的检测算法,以提高复杂背景下的小目标检测准确率和效率。在目标识别方面,在通过SAR图像先验信息抑制干扰噪声的基础上,提出利用SAR目标多视角信息联合Transformer的目标识别算法,通过设计视角正则化项以约束多视角之间的关联性从而实现不同视角间的特征融合,提高SAR小目标识别的准确率。从无人机载微型SAR系统对地面目标进行实时检测和识别的实际需求出发,本文还探讨了轻量化检测和识别网络在数字信号处理(Digital signal processing,DSP)平台上的部署方案,同时展示了初步试验结果。最后,本文展望了SAR目标智能检测和识别领域面临的挑战和发展趋势。
基金co-supported by the National Natural Science Foundation of China(No.52125504)the Liaoning Revitalization Talents Program(No.XLYC2202017)Dalian Support Policy Project for Innovation of Technological Talents(No.2023RG001)。
文摘The high-quality assembly of Large Aircraft Components(LACs)is essential in modern aviation manufacturing.Numerical control locators are employed for the posture adjustment of LAC,yet the system's multi-input multi-output,nonlinearity,and strong coupling presents significant challenges.The substantial internal force generated during the adjustment process can potentially damage the LAC and degrade the assembly quality.Hence,a workspace-based hybrid force position control scheme was developed to achieve high quality assembly with high-precision and lower internal force.Firstly,an offline workspace analysis with inherent geometric characteristics to form time-varying posture error constraint.Then,the posture error is integrated into the online position axis control to ensure tracking the ideal posture,while the force control axis compensates for posture deviation by minimizing internal force,thereby achieving high precision and low internal force.Finally,the effectiveness was demonstrated through experiments.The root mean square errors of orientation and position are 104 rad and 0.1 mm,respectively.A reduction in internal force can range from 10.96%to 57.4%compared to the traditional method.Key points'max position error is decreased from 0.32 mm to 0.18 mm,satisfying the 0.5 mm tolerance.Therefore,the proposed method will help promote the development of high-performance manufacturing.
