Low density and low convergence implosion occurs in the exploding-pusher target experiment, and generates neutrons isotropically to develop a high yield platform.In order to validate the performance of ShenGuang(SG) l...Low density and low convergence implosion occurs in the exploding-pusher target experiment, and generates neutrons isotropically to develop a high yield platform.In order to validate the performance of ShenGuang(SG) laser facility and test nuclear diagnostics, all 48-beam lasers with an on-target energy of 48 kJ were firstly used to drive room-temperature, DT gas-filled glass targets.The optimization has been carried out and optimal drive uniformity was obtained by the combination of beam repointing and target.The final irradiation uniformity of less than 5% on polar direct-drive capsules of 540 μm in diameter was achieved, and the highest thermonuclear yield of the polar direct-drive DT fuel implosion at the SG was 1.04 × 10^(13).The experiment results show neutron yields severely depend on the irradiation uniformity and laser timing,and decrease with the increase of the diameter and fuel pressure of the target.The thin CH ablator does not impact the implosion performance, but the laser drive uniformity is important.The simulated results validate that the cos γ distribution laser design is reasonable and can achieve a symmetric pressure distribution.Further optimization will focus on measuring the symmetry of the hot spot by self-emission imaging, increasing the diameter, and decreasing the fuel pressure.展开更多
在自动驾驶感知系统中视觉传感器与激光雷达是关键的信息来源,但在目前的3D目标检测任务中大部分纯点云的网络检测能力都优于图像和激光点云融合的网络,现有的研究将其原因总结为图像与雷达信息的视角错位以及异构特征难以匹配,单阶段...在自动驾驶感知系统中视觉传感器与激光雷达是关键的信息来源,但在目前的3D目标检测任务中大部分纯点云的网络检测能力都优于图像和激光点云融合的网络,现有的研究将其原因总结为图像与雷达信息的视角错位以及异构特征难以匹配,单阶段融合算法难以充分融合二者的特征.为此,本文提出一种新的多层多模态融合的3D目标检测方法:首先,前融合阶段通过在2D检测框形成的锥视区内对点云进行局部顺序的色彩信息(Red Green Blue,RGB)涂抹编码;然后将编码后点云输入融合了自注意力机制上下文感知的通道扩充PointPillars检测网络;后融合阶段将2D候选框与3D候选框在非极大抑制之前编码为两组稀疏张量,利用相机激光雷达对象候选融合网络得出最终的3D目标检测结果.在KITTI数据集上进行的实验表明,本融合检测方法相较于纯点云网络的基线上有了显著的性能提升,平均mAP提高了6.24%.展开更多
自动驾驶车辆在行驶过程中,需要对行人和车辆同时完成目标检测、实例分割和目标跟踪三个任务。提出一种基于深度学习的环境感知模型同时对三个任务进行多任务学习。首先,通过卷积神经网络对连续帧图像提取时空特征;然后,通过注意力机制...自动驾驶车辆在行驶过程中,需要对行人和车辆同时完成目标检测、实例分割和目标跟踪三个任务。提出一种基于深度学习的环境感知模型同时对三个任务进行多任务学习。首先,通过卷积神经网络对连续帧图像提取时空特征;然后,通过注意力机制对时空特征进行去耦再融合,充分利用任务间的相关性,实现不同任务对时空特征的差异化选择;最后,为平衡不同任务间的学习速率,使用动态加权平均的方式对模型进行训练。在KITTI数据集上的实验结果表明,所提模型在目标检测方面,比CenterTrack模型F1得分提高了0.6个百分点;在目标跟踪方面,比TraDeS(Track to Detect and Segment)模型多目标跟踪精度(MOTA)提高了0.7个百分点;在实例分割方面,比SOLOv2(Segmenting Objects by LOcations version 2)模型AP_(50)和AP_(75)分别提高了7.4和3.9个百分点。展开更多
Setting engine emission targets to meet diesel car requirements is particularly important in engine performance development phase. Many researches are focused on associating vehicle performance with engine targets, bu...Setting engine emission targets to meet diesel car requirements is particularly important in engine performance development phase. Many researches are focused on associating vehicle performance with engine targets, but most work is done by testing, which is time and cost consuming, furthermore, the relationship of vehicle and engine will change when either engine or vehicle changes. A GT-Drive model to simulate New European Driving Cycle (NEDC) for passenger car is developed and calibrated by testing data, model precision is controlled within 5%. Time distribution of engine operating conditions when car running NEDC cycle has been analyzed, 10 critical major engine operating points are summarized according to running time proportion. Emission of NOx and smoke control regions containing these 10 points for target engine are set. Vehicle emissions are simulated and evaluated during engine development after engine performance test data are got, and engine combustion system layout and calibration are adjusted until vehicle targets are met. Vehicle is tested in chassis dynamometer finally, the testing results show a good agreement with the simulated results with an error of less than 5%, which proves that the emission value exchange of vehicle and engine is reliable. Performance target decomposition method for passenger car diesel presented can greatly shorten the development cycle and save costs.展开更多
基金Project supported by the National Natural Science Foundation of China(Grant No.11605178)the Science Challenging Project,China(Grant Nos.JCKY2016212A505 and TZ2016001)
文摘Low density and low convergence implosion occurs in the exploding-pusher target experiment, and generates neutrons isotropically to develop a high yield platform.In order to validate the performance of ShenGuang(SG) laser facility and test nuclear diagnostics, all 48-beam lasers with an on-target energy of 48 kJ were firstly used to drive room-temperature, DT gas-filled glass targets.The optimization has been carried out and optimal drive uniformity was obtained by the combination of beam repointing and target.The final irradiation uniformity of less than 5% on polar direct-drive capsules of 540 μm in diameter was achieved, and the highest thermonuclear yield of the polar direct-drive DT fuel implosion at the SG was 1.04 × 10^(13).The experiment results show neutron yields severely depend on the irradiation uniformity and laser timing,and decrease with the increase of the diameter and fuel pressure of the target.The thin CH ablator does not impact the implosion performance, but the laser drive uniformity is important.The simulated results validate that the cos γ distribution laser design is reasonable and can achieve a symmetric pressure distribution.Further optimization will focus on measuring the symmetry of the hot spot by self-emission imaging, increasing the diameter, and decreasing the fuel pressure.
文摘在自动驾驶感知系统中视觉传感器与激光雷达是关键的信息来源,但在目前的3D目标检测任务中大部分纯点云的网络检测能力都优于图像和激光点云融合的网络,现有的研究将其原因总结为图像与雷达信息的视角错位以及异构特征难以匹配,单阶段融合算法难以充分融合二者的特征.为此,本文提出一种新的多层多模态融合的3D目标检测方法:首先,前融合阶段通过在2D检测框形成的锥视区内对点云进行局部顺序的色彩信息(Red Green Blue,RGB)涂抹编码;然后将编码后点云输入融合了自注意力机制上下文感知的通道扩充PointPillars检测网络;后融合阶段将2D候选框与3D候选框在非极大抑制之前编码为两组稀疏张量,利用相机激光雷达对象候选融合网络得出最终的3D目标检测结果.在KITTI数据集上进行的实验表明,本融合检测方法相较于纯点云网络的基线上有了显著的性能提升,平均mAP提高了6.24%.
文摘自动驾驶车辆在行驶过程中,需要对行人和车辆同时完成目标检测、实例分割和目标跟踪三个任务。提出一种基于深度学习的环境感知模型同时对三个任务进行多任务学习。首先,通过卷积神经网络对连续帧图像提取时空特征;然后,通过注意力机制对时空特征进行去耦再融合,充分利用任务间的相关性,实现不同任务对时空特征的差异化选择;最后,为平衡不同任务间的学习速率,使用动态加权平均的方式对模型进行训练。在KITTI数据集上的实验结果表明,所提模型在目标检测方面,比CenterTrack模型F1得分提高了0.6个百分点;在目标跟踪方面,比TraDeS(Track to Detect and Segment)模型多目标跟踪精度(MOTA)提高了0.7个百分点;在实例分割方面,比SOLOv2(Segmenting Objects by LOcations version 2)模型AP_(50)和AP_(75)分别提高了7.4和3.9个百分点。
基金supported by National Hi-tech Research and Development Program of China (863 Program, Grant No. 2008AA11A115)
文摘Setting engine emission targets to meet diesel car requirements is particularly important in engine performance development phase. Many researches are focused on associating vehicle performance with engine targets, but most work is done by testing, which is time and cost consuming, furthermore, the relationship of vehicle and engine will change when either engine or vehicle changes. A GT-Drive model to simulate New European Driving Cycle (NEDC) for passenger car is developed and calibrated by testing data, model precision is controlled within 5%. Time distribution of engine operating conditions when car running NEDC cycle has been analyzed, 10 critical major engine operating points are summarized according to running time proportion. Emission of NOx and smoke control regions containing these 10 points for target engine are set. Vehicle emissions are simulated and evaluated during engine development after engine performance test data are got, and engine combustion system layout and calibration are adjusted until vehicle targets are met. Vehicle is tested in chassis dynamometer finally, the testing results show a good agreement with the simulated results with an error of less than 5%, which proves that the emission value exchange of vehicle and engine is reliable. Performance target decomposition method for passenger car diesel presented can greatly shorten the development cycle and save costs.
