| 全振宇* **,尹龙祥*,陈晓明*,韩银和*.OODAFlow：面向智能无人系统的流式数据处理框架[J].高技术通讯(中文),2024,34(9):905~920 |
| OODAFlow：面向智能无人系统的流式数据处理框架 |
| OODAFlow： streaming data processing framework for intelligent unmanned systems |
| |
| DOI：10. 3772 / j. issn. 1002-0470. 2024. 09. 001 |
| 中文关键词: 智能无人系统； 深度学习加速卡； 观察-判断-决策-行动(OODA)； 流式数据处理框架； 异构计算资源 |
| 英文关键词: intelligent unmanned system, deep learning accelerator card, observe-orient-decide-act (OODA), streaming data processing framework, heterogeneous computing resource |
| 基金项目: |
| 作者 | 单位 | | 全振宇* ** | （*中国科学院计算技术研究所智能计算机研究中心北京 100190）
（**中国科学院大学北京 100190） | | 尹龙祥* | | | 陈晓明* | | | 韩银和* | |
|
| 摘要点击次数: 555 |
| 全文下载次数: 783 |
| 中文摘要: |
| 智能无人系统是一种能够在复杂环境中自主进行实时推理、决策和制定行动方案的计算系统。智能无人系统实现实时决策的关键在于对流式数据的实时处理，然而随着人工智能技术和传感器技术的快速发展，智能无人系统需要处理的数据规模不断增长，数据类型变得更加复杂。面对不断增长的数据处理性能需求，智能无人系统需要一个充分优化的专用流式数据处理框架来提升其数据处理性能。针对该问题，本文提出了一种面向智能无人系统的流式数据处理框架OODAFlow，该框架将智能无人系统的硬件特征和智能计算任务的数据特征与观察-判断-决策-行动(OODA)模型思想相融合，实现了OODA任务创建、任务调度、资源调度等功能，能够实现对智能无人系统异构资源的调度和智能计算任务的处理。本文在智能无人系统上搭建了一套OODA任务处理系统，验证了所提OODAFlow框架的可行性。通过提出的图像预处理过程优化、流水线优化以及判断节点并行加速优化等方法，提高了系统的数据吞吐性能和资源利用率。无人机智能控制任务的实验表明，采用本文提出的OODAFlow框架后，智能无人系统的数据处理性能提升了73倍。 |
| 英文摘要: |
| Intelligent unmanned system is a type of computing system that can autonomously perform real-time inferences and decisions, and formulate action plans in complex environments. The key to realizing real-time decision-making in intelligent unmanned systems lies in the real-time processing of streaming data. However, with the rapid development of artificial intelligence technology and sensor technology, the scale of data that intelligent unmanned systems need to process is constantly increasing, and the types of data are becoming more complex. Faced with the growing demand for data processing performance, intelligent unmanned systems need a well-optimized dedicated streaming data processing framework to enhance their data processing performance. To address this problem, this article proposes a stream data processing framework for intelligent unmanned systems called OODAFlow, which integrates the hardware features of intelligent unmanned systems and the data features of intelligent computing tasks with the observe-orient-decide-act (OODA) model to enable functions such as OODA task creation, task scheduling, and resource scheduling, it can schedule heterogeneous resources of intelligent unmanned systems and process intelligent computing tasks. This paper builds an OODA task processing system on intelligent unmanned systems to verify the feasibility of OODAFlow framework. Then, by proposing methods such as image preprocessing optimization, pipeline optimization, and orient node multi-process parallel optimization, the data throughput performance and resource utilization of the system have been improved. The experimental results in unmanned aerial vehicle intelligent control tasks show that when the proposed OODAFlow framework is used, the data processing performance of the intelligent unmanned system is improved by 73 times. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
