基于改进MFCC-OCSVM和贝叶斯优化BiGRU的GIS异常工况声纹识别算法  

Voiceprint Recognition Algorithm for GIS Anomalous Conditions Based on Improved MFCC-OCSVM and Bayesian Optimized BiGRU

作　　者：庄小亮 李乾坤 刘紫罡 张禄亮[2] 季天瑶 张长虹 ZHUANG Xiaoliang;LI Qiankun;LIU Zigang;ZHANG Luliang;JI Tianyao;ZHANG Changhong(EHV Power Transmission Company,CSG,Guangzhou 510663,China;School of Electric Power Engineering,South China University of Technology,Guangzhou 510640,China;Electric Power Research Institute of EHV Power Transmission Company,CSG,Guangzhou 510663,China)

机构地区：[1]南方电网超高压输电公司,广州510663 [2]华南理工大学电力学院,广州510640 [3]南方电网超高压输电公司电力科研院,广州510663

出　　处：《南方电网技术》2025年第1期30-40,共11页Southern Power System Technology

基　　金：国家自然科学基金资助项目(52077081);中国南方电网有限责任公司科技项目(CGYKJXM20220069)。

摘　　要：为了准确识别气体绝缘开关柜(gas insulated switchgear,GIS)设备的异常工况,提出了一种基于加权梅尔频率谱系数单类支持向量机(Mel frequency cestrum coefficient-one class support vector machine,MFCC-OCSVM)和贝叶斯优化的门控循环单元(bidirectional gate recurrent unit,BiGRU)声纹识别算法。首先,利用基于F统计量的MFCC对声纹数据进行加权特征提取,突出重要特征并减弱噪声的影响,然后利用OCSVM对加权后的特征进行异常检测并去除异常值,提高数据质量。为解决样本不平衡问题,采用合成少数类过采样技术(synthetic minority over-sampling technique,SMOTE)进行声纹样本的均衡。最后,应用基于贝叶斯优化的BiGRU模型进行声纹识别。以某气体绝缘全封闭组合电器(gas insulated switchgear,GIS)为例,采集了20类不同工况下操纵机构的声音样本,与多种经典分类模型进行对比。结果显示,所提算法取得的最高平均识别准确率达到了92.8%,相比于自适应增强、朴素贝叶斯和线性判别分析算法分别提升了30.1%、14.7%和11.5%。通过消融实验进一步评估和验证了所提算法各个流程对声纹识别的实际效果和性能影响,研究成果可为GIS设备异常工况的声纹识别提供高效技术路线。To accurately identify abnormal conditions in gas insulated switchgear(GIS)equipment,a voiceprint recognition algorithm is proposed based on weighted Mel frequency cestrum coefficient-one class support vector machine(MFCC-OCSVM)and Bayesian optimized bidirectional gate recurrent unit(BiGRU).Firstly,weighted extractions of voiceprint data are performed using MFCC based on the F-statistic,highlighting important features and reducing the influence of noise.Subsequently,OCSVM is utilized to detect anomalies and remove anomalous values from the weighted features to improve data quality.To address the issue of sample imbalance,synthetic minority over-sampling technique(SMOTE)is employed to balance voiceprint samples.Finally,voiceprint recognition is carried out using a BiGRU model based on Bayesian optimization.Taking a certain GIS equipment as an example,sound samples from 20 different operating conditions are collected and compared with various classical classification models.The results demonstrate that the proposed algorithm achieves the highest average recognition accuracy of 92.8%,resulting in improvements of 30.1%,14.7%and 11.5%compared to adaptive boosting,Naïve Bayes,and linear discriminate analysis,respectively.Ablation study further assesses and validates the practical effects and performance impacts of each process in the proposed algorithm.Research results provide an efficient technical approach for voiceprint recognition of anomalous conditions in GIS.

关 键 词：GIS设备 梅尔频谱倒谱系数 单类支持向量机 双向门控循环单元 声纹识别 贝叶斯优化 

分 类 号：TM41[电气工程—电器]