高噪声环境下电厂设备声音融合特征生成方法研究与实现  

Research and implementation of sound fusion feature generation method for power plant equipment in high noise environment

作　　者：张强 武明路 韩文学 包伟伟 张楠 李明轩 翟永杰 ZHANG Qiang;WU Minglu;HAN Wenxue;BAO Weiwei;ZHANG Nan;LI Mingxuan;ZHAI Yongjie(Shijiazhuang Liangcun Electric Power Company Limitd,Shijiazhuang 050000,China;SPIC Central Research Institute Co.,Ltd.,Beijing 102209,China;Department of Automation,North China Electric Power University(Baoding),Baoding 071003,China)

机构地区：[1]石家庄良村热电有限公司,河北石家庄050000 [2]国家电投集团科学技术研究院有限公司,北京102209 [3]华北电力大学(保定)自动化系,河北保定071003

出　　处：《热力发电》2022年第12期39-47,共9页Thermal Power Generation

基　　金：国家自然科学基金项目(U21A20486)。

摘　　要：基于声音信号的电厂设备状态监测是一种新型的设备状态故障诊断技术,电厂的高噪声环境是该技术应用过程中需要解决的技术难点。针对高噪声环境下电厂设备声音识别问题,提出一种基于Mel频率倒谱系数(MFCC)和Gammatone频率倒谱系数(GFCC)的电厂设备声音融合特征生成方法,以实现对不同设备运行声音的准确识别。首先,使用96通道声像仪在电厂高噪声环境中采集处于运行状态的6种设备声音信号,并对声音信号进行预处理,得到多帧样本信号;然后,利用MFCC和GFCC对预处理后的声音信号进行特征提取,得到原始高维特征;为消除噪声影响并降低特征维度,分别使用主成分分析(PCA)、降噪自动编码器(DAE)和t分布-随机近邻嵌入(t-SNE)算法对原始高维特征进行降维处理,计算对应类别可分性准则函数值,将同种降维方法下的降维特征融合得到融合特征;最后,向数据集中加入高斯白噪声,基于融合特征采用支持向量机算法进行设备声音识别,验证融合特征的准确性和鲁棒性。实验结果表明,融合特征既具有MFCC特征的准确性也具有GFCC特征的鲁棒性,相较于降维前的MFCC和GFCC特征提取方法,识别成功率明显提高,可为电厂设备状态监测与故障预警方法的进一步研究提供理论基础。Power plant equipment condition monitoring based on sound signal is a new type of equipment condition fault diagnosis technology. The high noise environment of power plant is the technical difficulty to be solved in the application of this technology. Aiming at solving the problem of sound recognition of power plant equipment in high noise environment, a fusion feature generation method of power plant equipment sound based on Mel frequency cepstrum coefficient(MFCC) and gamma tone frequency cepstrum coefficient(GFCC) is proposed to realize accurate recognition of different equipment operation sounds. Firstly, the 96-channel sound imager is used to collect the sound signals of 6 kinds of equipment in operation in the high noise environment of the power plant,and preprocess the sound signals to obtain multi-frame sample signals. Then, MFCC and GFCC are applied to extract the features of the preprocessed audio signal to obtain the original high-dimensional features. In order to eliminate the influence of noise and reduce the feature dimension, principal component analysis(PCA), de-noise automatic encoder(DAE) and t-SNE algorithm are respectively used to reduce the dimension of the original highdimensional features, calculate the separability criterion function value of the corresponding category, and fuse the reduced dimension features in the same dimension reduction method to obtain the fused features. Finally, Gaussian white noise is added to the data set, and support vector machine algorithm is used for equipment voice recognition based on the fusion features to verify the accuracy and robustness of the fusion features. The experimental results show that, the fusion features have both the accuracy of MFCC features and the robustness of GFCC features.Compared with the MFCC and GFCC feature extraction methods before and after dimension reduction, the recognition success rate has been significantly improved, which can provide a theoretical basis for the further research of power plant equipment condition monitoring

关 键 词：状态监测 故障诊断 声音识别 融合特征 去噪 降维 

分 类 号：TP277[自动化与计算机技术—检测技术与自动化装置]