基于Hankel矩阵的复小波–奇异值分解法提取局部放电特征信息

Feature Information Extraction of Partial Discharge Signal With Complex Wavelet Transform and Singular Value Decomposition Based on Hankel Matrix

气体绝缘组合电器(gas insulated switchgear,GIS)在产生局部放电(partial discharge,PD)时,会向外辐射特高频(ultra-high frequency,UHF)电磁信号,有效提取UHF PD信号的特征信息可实现GIS的在线监测与故障诊断。针对UHF PD信号经过复小波变换后,层间奇异信息分布和层内奇异信息复杂度的差异性,采用二元树复小波变换(combined dual-tree complex wavelet transform,DT-CWT)和奇异值分解(singular value decomposition,SVD)相结合的信号处理方法,提取了UHF PD信号的特征信息。采用Birge-Massart阈值策略对DT-CWT分解后的复小波系数模值序列进行压缩,并构造复合矩阵,分析复合矩阵的奇异熵和复小波分解层数的关系,提出一种求解复小波最优分解层数的算法;利用最优分解层数下的压缩后的各高频系数模值序列构造Hankel矩阵,提取各Hankel矩阵的最大奇异值和奇异熵作为PD辨识的特征参量。结果表明:该特征可以有效识别4种典型绝缘缺陷,且识别率都到达了92%及以上。

Gas insulated switchgear（GIS） would radiate the ultra-high frequency（UHF） electromagnetic waves to the surroundings when partial discharge（PD） occurs. Effective extraction of the feature information of UHF PD signals could conduct and realize online monitoring and fault diagnosis of GIS. Considering the differences of the distribution and complexity of singular information between the wavelet coefficients of each layer, a novel signal processing method was proposed, which combined dual-tree complex wavelet transform（DT-CWT） with singular value decomposition（SVD）. The wavelet magnitude sequences were compressed by applying Biarge-Massart threshold strategy, and a composite matrix by using the magnitude sequences was constructed and transformed by SVD. An optimal complex wavelet decomposition algorithm was proposed based on the relationship between the singular entropy of the matrix and the wavelet decomposition level. Then, the UHF PD signals were processed in the optimal decomposed level, and the compressed magnitude sequences of each high frequency layer were used to construct a Hankel matrix, and the largest singular values and singular entropies of each Hankel matrix were extracted. Finally, the recognition results show that these feature parameters extracted by the new method can identify four kinds of typical insulation defects in GIS effectively, and all the recognition rates have reached 92% or above.