基于MFCC和MDE-SVDD的滚动轴承音频信号异常检测方法  

Anomaly Detection Method for Acoustic Signal of Rolling Bearing Based on MFCC and MDE-SVDD

作　　者：高原[1] 邓艾东[1,2] 范永胜 梁志宏 傅行军[2] GAO Yuan;DENG Aidong;FAN Yongsheng;LIANG Zhihong;FU Xingjun(National Engineering Research Center of Power Generation Control and Safety,Southeast University,Nanjing 210096,China;School of Energy and Environment,Southeast University,Nanjing 210096,China;CHN Energy Jiangsu Power Co.,Ltd.,Nanjing 215433,China)

机构地区：[1]东南大学大型发电装备安全运行与智能测控国家工程研究中心,南京210096 [2]东南大学能源与环境学院,南京210096 [3]国家能源集团江苏电力有限公司,南京215433

出　　处：《动力工程学报》2024年第2期277-283,共7页Journal of Chinese Society of Power Engineering

基　　金：江苏省碳达峰碳中和科技创新专项资金资助项目(BA2022214);江苏省重点研发计划资助项目(BE2020034)。

摘　　要：针对传统振动传感器安装不易,而声信号分析易受环境噪声干扰的问题,提出一种基于梅尔倒谱系数(MFCC)和马氏距离加权改进支持向量数据描述(MDE-SVDD)的音频信号异常检测方法,用于滚动轴承运行状态监测。该方法从轴承运行声信号中提取MFCC作为特征向量,进而使用马氏距离加权改进SVDD,以增强对噪声样本的抗干扰性,从而提高算法的检测精度,然后在实验音频信号中添加多种强度的高斯白噪声以模拟现场噪声环境,并将所提方法的测试结果与传统SVDD等异常检测方法进行比较。结果表明:在低信噪比(-5 dB)场景下,MDE-SVDD的异常检测平均准确率达到91.99%,相较于传统SVDD提升了7.73百分比。Aiming at the problems that traditional vibration sensor is not easy to install and the acoustic signal analysis is easily interfered by environmental noise,an anomaly detection method for audio signal was used to test the running state of the rolling bearings based on Mel-Frequency cepstral coefficients(MFCC)and Mahalanobis distance weighting support vector data description(MDE-SVDD).In this method,MFCC was extracted from the running sound signal of bearings as the feature vector,and then Mahalanobis distance weighting was used to improve SVDD,so as to enhance the anti-interference of noise samples and improve the detection accuracy of the algorithm.Gaussian white noise of various intensities was added to the experimental sound signal to simulate the field noise environment,and the test results of the proposed method were compared with traditional anomaly detection methods such as SVDD.Results show that the anomaly detection accuracy of MDE-SVDD reaches 91.99%in the scene of low signal-to-noise ratio by-5 dB,which is 7.73%higher than that of the traditional SVDD model.

关 键 词：滚动轴承 声纹识别 梅尔倒谱系数 支持向量数据描述 异常检测 

分 类 号：TH133[机械工程—机械制造及自动化]