面向DSP平台的CiSSA-CSP特征提取算法的移植与优化  

Transplant and Optimization of CiSSA-CSP Feature Extraction Algorithm on DSP Platform

作　　者：刘哲贤 赵金库[2] 赵玉峰 王鹏[1] LIU Zhexian;ZHAO Jinku;ZHAO Yufeng;WANG Peng(Department of Precision Instrument,Tsinghua University,Beijing 100084,China;Heilongjiang North Tool Co.,Ltd.,Mudanjiang 157000,China)

机构地区：[1]清华大学精密仪器系,北京100084 [2]黑龙江北方工具有限公司,黑龙江牡丹江157000

出　　处：《计算机测量与控制》2024年第1期260-267,共8页Computer Measurement &Control

摘　　要：为实现便携式信号二分类解析系统的在线实时处理,采用DSP平台完成CiSSA-DSP特征提取算法的嵌入式移植;CiSSA-CSP特征提取算法具有出色的时-频-空域特征提取性能,适合于提取实时二分类系统中非平稳信号的特征;相比于PC机,嵌入式系统具有小型化、便携性、低功耗和低延时的特点,而嵌入式平台处理器的计算资源和内存受到限制,必须优化移植特征提取算法,才能保证二分类解析系统的分类精度和低延时;通过优化CiSSA-CSP算法流程,使用编译器优化、关键字和库函数等手段提高编译效率,将CiSSA-CSP特征提取算法移植到TMS320C6678DSP嵌入式平台,并利用公共数据库数据验证了其用于实时分类系统的有效性;相比于PC机的Matlab实现,DSP平台实现的二分类系统分类准确度下降小于0.5%,且单次实验信号解析耗时少于0.15 s。In order to realize the on-line and real-time processing of portable signal binary classification analysis systems,DSP platform is used to implement the embedded transplantation of CiSSA-DSP feature extraction algorithm.A CiSSA-CSP feature extraction algorithm has the excellent performance of time-frequency-spatial feature extraction,and it is suitable for extracting non-stationary signals in real-time binary classification system.Compared with PC,an embedded implementation has the characteristics of miniaturization,portability,low power consumption and low delay,while the computing resources and memory of embedded platform processor is limited,and the transplant feature extraction algorithm must be optimized to ensure the binary classification accuracy and low delay.The steps of the CiSSA-CSP algorithm,the compiler optimization,keywords and library functions are optimized to improve the compilation efficiency,the CiSSA-CSP feature extraction algorithm is transplanted to the TMS320C6678 embedded DSP platform,and its effectiveness for real-time classification system is verified by the public dataset.Compared with Matlab on PC,the classification accuracy of the binary classification system is reduced by less than 0.5%on the DSP platform,and the computing time for singel experiment is less than 0.15s.

关 键 词：二分类 特征提取 CiSSA-CSP DSP 优化移植 

分 类 号：TP274[自动化与计算机技术—检测技术与自动化装置]