基于GFCC-SVM-RFE的电力设备声音特征提取方法  

Sound Feature Extraction Method of Power Equipment Based on GFCC-SVM-RFE

作　　者：王赵国 韦存海 彭雅妮 武明路 李军彬 翟永杰[3] WANG Zhaoguo;WEI Cunhai;PENG Yani;WU Minglu;LI Junbin;ZHAI Yongjie(State Power Investment Group Power Station Operation Technology(Beijing)Co.,Ltd.,Beijing 100000,China;Shijiazhuang Liangcun Thermal Power Co.,Ltd.,Shijiazhuang 050000,China;North China Electric Power University,Baoding 071003,China)

机构地区：[1]国家电投集团电站运营技术(北京)有限公司,北京100000 [2]石家庄良村热电有限公司,河北石家庄050000 [3]华北电力大学自动化系,河北保定071003

出　　处：《电力信息与通信技术》2022年第9期34-42,共9页Electric Power Information and Communication Technology

基　　金：国家自然科学基金联合基金项目重点支持项目(U21A20486)。

摘　　要：火电厂电力设备声音包含了丰富的有效信息,但受限于复杂环境噪声,使得有效特征提取极其困难。为解决这一问题,文章提出一种基于GFCC-SVM-REF的特征提取方法,并在现场采集的6种设备运行声音、加入ESC-50公共数据集中的环境干扰音以及对原始设备声音加入不同分贝的高斯白噪声这3类数据上对2种经典语音识别领域的特征提取方法进行抗噪性和准确性的对比。仿真结果表明,针对所研究的数据集,GFCC-SVM-RFE方法在10 dB和20 dB的高斯白噪声下分别达到了81.04%和96.88%的准确率。The sound of electrical equipment in thermal power plants contains rich and effective information,but limited by complex environmental noise,it is extremely difficult to extract effective features.In order to solve this problem,this paper proposes a feature extraction method based on GFCC-SVM-REF,and two feature extraction methods in classical speech recognition fields are compared from noise immunity and accuracy using three types of data where six kinds of equipment operating sounds collected on site are mixed with environmental interference sounds in the ESC-50 public data set and Gaussian white noise of different decibels.The simulation results show that for the studied dataset,the GFCC-SVM-RFE method achieves 81.04% and 96.88% accuracy under 10 dB and 20 dB Gaussian white noise.

关 键 词：电力设备 声音特征提取 环境噪音 梅尔频率倒谱系数 伽马通频率倒谱系数 支持向量机递归特征消除 

分 类 号：TB535[理学—物理]