基于分数阶RCMDE和参数优化LSSVM的开关柜故障声纹识别方法

A Fault Acoustic Signature Recognition Method for Switchgear Based on Fractional-Order RCMDE and Parameter-Optimized LSSVM

中文关键词:电力开关柜  故障检测  声纹识别  精细复合多尺度散布熵  瞪羚优化算法  最小二乘支持向量机

keywords:power switchgear  fault detection  acoustic signature recognition  refined composite multiscale dispersion entropy （RCMDE）  gazelle optimization algorithm （GOA）  least square support vector machine（LSSVM）

中文摘要:开关柜发生故障时会产生不同的异常声音，声纹识别技术可以实现对开关柜的不停电检测。提出了基于分数阶精细复合多尺度散布熵（refined composite multiscale dispersion entropy，RCMDE）和参数优化最小二乘支持向量机（least square support vector machines，LSSVM）的开关柜故障声纹识别方法。首先，提出分数阶RCMDE熵特征提取方法计算开关柜声纹信号的熵特征；其次，对瞪羚优化算法的位置更新模块进行了优化，以确定LSSVM的最优分类参数；最后，利用参数优化的LSSVM分类器对开关柜声纹数据的分数阶RCMDE熵特征进行分类，识别开关柜故障。为了验证方法的有效性，采集了开关柜正常状态、分合闸不到位导致的间歇性放电、间断放电和悬浮放电在内的4种声纹数据，并进行了分类识别。实验结果表明，所提方法对这4种样本识别的准确率和召回率最高可达100%，最低不低于97%。与其他熵特征相比，分数阶RCMDE对声纹数据特征区分度最大，参数优化后的LSSVM分类器对声纹故障分类的准确性更高。在跨域开关柜故障识别中，故障识别的准确率和召回率不低于90%，且对噪声有较好的鲁棒性。

Abstract:Abnormal sounds of different types will be generated when switchgear malfunctions，and acoustic signature recognition technology can realize live detection of switchgear. A fault acoustic signature recognition method for switchgear based on fractional-order refined composite multiscale dispersion entropy （RCMDE） and parameter-optimized least square support vector machines （LSSVM） is proposed in this paper. First，a fractional-order RCMDE entropy feature extraction method is proposed to calculate the entropy features of switchgear acoustic signature signals. Second，the position update module of the gazelle optimization algorithm is optimized to determine the optimal classification parameters of LSSVM. Finally，the fractional-order RCMDE entropy features of switchgear acoustic signature data are classified by the parameter-optimized LSSVM classifier to identify switchgear faults. To verify the effectiveness of the proposed method，four types of acoustic signature data are collected，including those of switchgear under normal state，intermittent discharge caused by improper switching-on and switching-off，discontinuous discharge and floating potential discharge，and classification and recognition are conducted on the data. Experimental results show that the accuracy and recall of the proposed method for identifying these four types of samples can reach up to 100% and not be lower than 97% at the minimum. Compared with other entropy features，fractional-order RCMDE exhibits the maximum discriminability for the features of acoustic signature data，and the parameter-optimized LSSVM classifier achieves higher accuracy in the classification of acoustic signature faults. In cross-domain switchgear fault recognition，the accuracy and recall of fault recognition are not lower than 90%，and the method has good robustness to noise.