基于天/地同步进化神经网络的智能SAR卫星技术研究

There are some problems in the existing SAR satellites, including the data-image-information conversion link is long, the timeliness efficiency is low, it is difficult to store, manage, and transmit the massive original data, the on-board task response link is not closed loop, the autonomous work capability is poor, the on-orbit performance improvement capability is weak, and it is difficult to continue to evolve. This project carries out the research of intelligent SAR satellite technology based on the space-to-earth synchronous evolutionary neural networks, reconstructs the SAR satellite's architecture, information flow links, and task flow links, breaking through the existing SAR satellite technology and application bottlenecks. Using two space-to-earth synchronous evolutionary neural networks, the spaceborne application neural network is applied on the satellite to achieve high-performance information extraction and task planning. The ground tutor neural network is trained and studied under the supervision of the users, and is continuously evolved. The two neural networks evolve synchronously, which not only achieves the continuous evolvement of SAR satellites, but also avoids the massive resource requirements for the training of the spaceborne neural network. Data acquisition, imaging processing and information extraction are all completed on board. It compresses the information conversion link and reducing the pressure of data transmission on space-to-earth. In addition, combined with the on-board information extraction results, the task is generated and planned autonomously. A closed-loop task execution link with self-feedback is formed to improve the autonomous running capability of SAR satellites. This project will enhance the intelligence level of SAR satellites and meet increasingly diverse and time-sensitive user needs.

针对现有SAR卫星数据-图像-信息转换链路长、时效低，海量原始数据存储、管理和下传困难，星上任务响应链路不闭环、自主工作能力差，在轨性能提升能力弱、难于持续进化等问题。开展基于天/地同步进化神经网络的智能SAR卫星技术研究，重建SAR卫星的体系架构、信息流链路和任务流链路，突破现有SAR卫星技术和应用瓶颈。采用天/地同步进化的两组神经网络，星上应用神经网络实现高性能信息提取和任务规划，地面导师神经网络在用户监督下训练进化，所得参数上注卫星，实现天/地神经网络进化同步，既实现了SAR卫星持续升级，又避免了星上神经网络训练的大量资源需求。数据获取、成像处理和信息提取全部星上完成，压缩了信息转换链路，降低星地间数传压力。结合星上信息提取结果，自主完成任务生成和规划，形成具有自反馈的闭环任务执行链路，提升SAR卫星自主工作能力。本课题将提升SAR卫星的智能化水平，满足日益多样化、高时效的用户需求。

本项目经过3年的研究，达到了预期研究目标，研究成果得到初步验证和应用。针对现有SAR卫星数据-图像-信息转换链路长、时效低，海量原始数据存储、管理和下传困难，星上任务响应链路不闭环、自主工作能力差，在轨性能提升能力弱、难于持续进化等问题，完成了基于天/地同步进化神经网络的智能SAR卫星技术研究，重建SAR卫星的体系架构、信息流链路和任务流链路，突破了现有SAR卫星技术和应用瓶颈。采用天/地同步进化的两组神经网络，星上应用神经网络实现高性能信息提取和任务规划，地面导师神经网络在用户监督下训练进化，所得参数上注卫星，实现天/地神经网络进化同步，既实现了SAR卫星持续升级，又避免了星上神经网络训练的大量资源需求。数据获取、成像处理和信息提取全部星上完成，压缩了信息转换链路，降低星地间数传压力。结合星上信息提取结果，自主完成任务生成和规划，形成具有自反馈的闭环任务执行链路，提升SAR卫星自主工作能力。预期采用本课题研究的体制和架构，经过持续的技术攻关和工程实践，具备将我国SAR卫星任务响应速度和信息时效性由天级提升至分钟级的能力。本课题成果将有助于提升我国SAR卫星的智能化水平，具有潜在广阔的应用前景，能够满足广大用户日益多样化、高时效的需求。同时，研究的高效星上数据处理、智能信息提取和自主任务规划等技术，也可在其它相关领域推广应用，具有较大的价值。