摘要
为准确识别变压器的工作状态,该文从运行中的变压器声音信号出发,提出一种基于压缩观测与判别字典学习的变压器声纹识别方法。首先采用稀疏随机矩阵对变压器声音信号进行压缩观测来获取观测信号,然后将变压器不同状态下的观测信号组成样本集,经迭代算法训练获得子字典、公共字典和判别字典。同时,为提升判别字典的分类性能,引入了Fisher判别约束项和公共字典低秩约束项来优化判别字典学习的目标函数,并对目标函数中的惩罚系数进行了优选。最后,求解待识别样本在判别字典上的稀疏表示系数,根据其在子字典上的重构误差最小原则进行识别。对某10k V干式变压器正常与典型故障下声音信号的分析结果表明,压缩观测可将声音信号数据量减小至原来的10%,大幅提升了运算效率。所构建的字典学习目标函数能有效增强字典的判别性能,对变压器不同工作状态的总体识别准确率可达96%,从而为变压器的声纹识别提供了一种新的思路。
To accurately recognize the working condition of power transformers, a transformer voiceprint recognition method based on compressed observation and discriminant dictionary learning was proposed in this paper considering the acoustic signals of power transformers. At first, the sparse random matrix was applied to compress the acoustic signals of power transformers to obtain the observed signals. Then the sample set was built and trained by the iteration algorithm to obtain the sub-dictionaries, a public dictionary and a discriminant dictionary according to the observed signals resulted from the different working conditions of transformers. Meanwhile, to enhance the classification performance of the discriminant dictionary, the objective function of dictionary learning was optimized by introducing the Fisher discriminant constraint term and the public dictionary low rank constraint term, and the penalty coefficients were optimally selected. Finally, the sparse representation coefficients on the discriminant dictionary was calculated for the test sample. And the corresponding condition was recognized according to the minimum reconstruction error of sparse representation coefficients on the sub-dictionaries. The calculation results of the acoustic signals of a 10 k V dry type transformer under normal conditions and typical faults show that the compressed observation can greatly reduce the data size of the transformer acoustic signals to 10% of the original signals with high calculation efficiency. The objective function of dictionary learning can effectively enhance the discriminant performance of the dictionary, and the recognition accuracy of the transformer working condition can reach 96%. The calculated results could provide a new idea for the voiceprint recognition of power transformers.
作者
周东旭
王丰华
党晓婧
张欣
刘顺桂
ZHOU Dongxu;WANG Fenghua;DANG Xiaojing;ZHANG Xin;LIU Shungui(Department of Electrical Engineering,Shanghai Jiaotong University,Minhang District,Shanghai 200240,China;Electric Power Research Institute,Shenzhen Power Supply Co.,Ltd.,Shenzhen 518000,Guangdong Province,China)
出处
《中国电机工程学报》
EI
CSCD
北大核心
2020年第19期6380-6389,共10页
Proceedings of the CSEE
关键词
干式变压器
声纹识别
压缩感知
字典学习
稀疏表示
dry type transformer
voiceprint recognition
compressed sensing
dictionary learning
sparse representation
