融合MFCC特征的声发射故障检测

Acoustic Emission Fault Detection Fused with MFCC Features

为了探索机电设备潜在声发射(AE)故障的有效预测方法,提出了融合梅尔频率倒谱系数(MFCC)与信号典型时频特征用于支持向量机(SVM)模式识别来进行故障诊断的方法。基于气体绝缘金属封闭开关(GIS)绝缘结构表面金属颗粒堆积缺陷模型,采集了多种局部放电(PD)故障诱发的AE信号;对获取的AE信号添加多种强度的高斯白噪声以模拟受现场干扰的低信噪比AE信号;进而通过信号降噪与特征提取,将提取的MFCC特征与时频域特征值作为SVM的输入向量。实验结果显示,对-15 dB信噪比AE信号的故障识别准确率可达到98.75%,对-10 dB信噪比以上AE信号的故障识别准确率均达到100%。该研究分析了MFCC特征结合时频域特征作为模式识别算法输入时,对各种信噪比信号的故障识别精度,适合用于强噪声环境下对机电故障的种类和进展进行准确预测。

In order to explore an effective prediction method for potential acoustic emission(AE)fault of mechanical and electrical equipment,this paper proposed a method for fault diagnosis by fusing Mel frequency cepstrum coefficient(MFCC)and signal typical time-frequency features for support vector machine(SVM)pattern recognition.Based on the defect model of metal particle accumulation on the insulation structure surface of gas insulated metal enclosed switchgear(GIS),AE signals induced by various partial discharge(PD)faults were collected.Gaussian white noise of various intensities was added to the acquired AE signals to simulate AE signals with low signal-to-noise ratio interfered by the scene.And then the extracted MFCC features and time-frequency domain eigenvalues were used as the input vectors of SVM after signal denoising and feature extraction.The experimental results show that the fault recognition accuracy of-15 dB signal-to-noise ratio AE signal can reach 98.75%,and the fault recognition accuracy of above-10 dB signal-to-noise ratio AE signal can reach 100%.This study analyzes the fault recognition accuracy of MFCC features combined with time-frequency features as the input of pattern recognition algorithm for various signal-to-noise ratio signals,which is suitable for accurately predicting the types and progress of electromechanical faults in strong noise environment.