实时低功耗飞行器神经网络  

Real-time Low Power Consumption Aircraft Neural Network

作　　者：张英 陶磊岩[4] 曹健[1] 王世会 赵茜[2,3] 张兴[1] ZHANG Ying;TAO Lei-yan;CAO Jian;WANG Shi-hui;ZHAO Qian;ZHANG Xing(School of Software and Microelectronics,Peking University,Beijing 100871,China;Beijing Aerospace Automatic Control Institution,Beijing 100854,China;National Key Laboratory of Science and Technology on Aerospace Intelligent Control,Beijing 100854,China;Beijing Institute of Remote Sensing Equipment,Beijing 100854,China)

机构地区：[1]北京大学软件与微电子学院,北京100871 [2]北京航天自动控制研究所,北京100854 [3]宇航智能控制技术国家级重点实验室,北京100854 [4]北京遥感设备研究所,北京100854

出　　处：《计算机科学》2021年第3期196-200,共5页Computer Science

基　　金：国家自然科学基金(51877008)。

摘　　要：为了满足飞行器实时飞行过程中对大量异构输入数据的信息处理需求,文中提出了一种神经网络,其包括卷积定点滑动核、池化压缩量化核以及全连接压缩融合核,将飞行器异构传感器多路并行数据作为系统的输入,将辨识结果作为系统的输出。卷积滑动窗口核通过排除冗余数据的滑动窗快速实现数据特征的提取;池化压缩量化核使用压缩量化技术来提高系统的执行效率;全连接压缩融合核经删减量化后压缩融合并输出。该设计满足了飞行器对高可靠性、低功耗的在线智能集成需求。使用所提压缩量化方法,准确率最高可达98.54%,压缩率为77.8%,运行速度提升了40倍。In order to meet the information processing requirements of a large amount of heterogeneous input data in the real-time flight of aircraft,this paper proposes a neural network,including convolution core with fixed-point sliding,pooling core with compression quantization and fully connected core with compression fusion.The input of the system is heterogeneous sensor data,and the output of the system is the identification results.Convolution core can extract data features quickly by eliminating redundant data sliding window.Pooling core improves system execution efficiency by using compression quantization technology.The design meets the on-line intelligent integration requirements of high reliability and low power consumption.With the proposed compression quantization method,the peak accuracy is 98.54%,the compression rate is 77.8%,and the running speed increases by 40 times.

关 键 词：低功耗 神经网络 实时在线 飞行器 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]