基于改进MFCC和3D-CNN的变压器铁心松动故障声纹识别模型  

Voiceprint recognition model of transformer core looseness fault based on improved MFCC and 3D-CNN

作　　者：崔佳嘉 马宏忠[1] CUI Jia-jia;MA Hong-zhong(School of Energy and Electrical Engineering,Hohai University,Nanjing 211100,China)

机构地区：[1]河海大学能源与电气学院,江苏南京211100

出　　处：《电机与控制学报》2022年第12期150-160,共11页Electric Machines and Control

基　　金：国家自然科学基金(51577050);江苏省电力有限公司重点科技项目(J2021053)。

摘　　要：为了准确提取变压器铁心松动故障时的声纹特征,提出一种基于改进的梅尔频率倒谱系数(MFCC)和三维卷积神经网络(3D-CNN)的变压器声纹识别模型。首先对变压器噪声信号进行分帧加窗处理,提取梅尔频率倒谱系数(MFCC);然后运用局部线性嵌入算法(LLE)对现有的MFCC特征向量降维改进;最后使用三维卷积神经网络对变压器铁心松动故障进行识别。以某10 kV变压器为对象进行空载试验,采集铁心在不同松动程度下的声纹信号。计算结果表明,使用改进后的MFCC特征向量提取算法及3D-CNN模型具有良好的识别效果,准确率可达到98.33%,且平均迭代的时间可降至8.511 26 s。最终研究结果可为变压器的噪声治理提供依据。In order to accurately extract the voiceprint features of transformer core looseness fault, a transformer voiceprint recognition model based on improved Mel frequency cepstrum coefficient(MFCC) and three-dimensional convolutional neural network(3 D-CNN) is proposed. Firstly, the noise signal of transformer was processed by frame and windowing, and the Mel frequency cepstrum coefficient(MFCC) was extracted. Then the local linear embedding algorithm(LLE) was used to reduce the dimension of the existing MFCC feature vector;Finally, three-dimensional convolution neural network was used to identify the loose fault of transformer core. Taking a 10 kV transformer as the object, the no-load test was carried out, and the voiceprint signals of the iron core under different degrees of looseness were collected. The calculation results show that the improved MFCC feature vector extraction algorithm and 3 D-CNN model have good recognition effect. The accuracy can reach 98.33%, and the average iteration time can be reduced to 8.511 26 s. The final research results can provide a basis for transformer noise control.

关 键 词：变压器 声纹 铁心松动 梅尔倒谱系数 局部线性嵌入算法 降维 三维卷积神经网络 

分 类 号：TH212[机械工程—机械制造及自动化] TH213.3