基于稀疏表示理论的特高压交流变压器声信号盲分离研究

Blind Separation of UHV Power Transformer Acoustic Signal Preprocessing Based on Sparse Representation Theory

运行中的变压器声信号由其机械结构振动经空气传播形成,易受周围环境影响,且所含干扰信号具有较大的不确定性,给基于声信号的变压器状态监测带来不利影响。为有效获取反映变压器运行状态的声信号,从特高压交流变压器声信号的声学特征出发,依据声信号的集合经验模态分解(ensembled empirical mode decomposition,EEMD)结果,提出了基于稀疏表示理论的特高压交流变压器声信号盲分离方法,即通过构建K-奇异值分解(K-singularvalue decomposition,K-SVD)字典和应用正交匹配追踪算法,得到了成分相对单一的特高压交流变压器声信号,有效抑制了特高压交流变压器声信号中的干扰成分。对某1000kV变电站主变声信号的计算结果表明:基于EEMD算法和稀疏表示理论能有效地对特高压交流变压器声信号中包含的环境噪声和人说话声等干扰成分进行盲分离,具有自适应性强、计算效率高和重构误差低的优点,可为基于声信号的特高压交流变压器状态监测提供重要数据支持。

Acoustic signals of an operating transformer are caused by its mechanical structure vibration and transmitted in the air, which is inevitably affected by their surrounding environments. Besides, the interference components in the acoustic signals appear in relatively large uncertainty, which would eventually lead to some negative effects of the condition monitoring of power transformer. According to the acoustic features of the transformer and the Ensembled Empirical Mode Decomposition(EEMD) results of acoustic signals of transformer, this paper presents a blind separation method for acoustic signals of UHV power transformer based on the sparse representation theory in order to effectively obtain the real acoustic signals of power transformer. With the construction of K-SVD dictionary and the orthogonal matching pursuit algorithm, the relatively pure acoustic signals of UHV power transformer are obtained and the inference components are greatly suppressed. The calculation results of the acoustic signals of a 1000 kV transformer show that the combination of EEMD algorithm and sparse representation theory can effectively separate the inference components in the acoustic signals of UHV power transformer such as the environment noise or voice of human talking. This method has the advantage of high adaptation, high calculation efficiency and low reconstruction error. And it provides the important data support for the condition monitoring of UHV power transformer.