西南交通大学学报2025，Vol.60Issue(3)：781-792,12.DOI:10.3969/j.issn.0258-2724.20230177
基于SBSS与CNN的750 kV变压器和尖板的放电信号声纹识别
Voiceprint Recognition of Discharge Aliasing Signals from 750 kV Transformer and Pin-Plate Based on Sparse Representation Theory and Convolutional Neural Network
摘要
Abstract
Transformer insulation level and health state are crucial to the safety and stability of the power grid.In order to study the practical engineering problem that the audible acoustic signals collected outside the box may be mixed with other interference signals,such as corona sound and bird song when there is a discharge fault inside the 750 kV transformer,a voiceprint recognition of 750 kV transformer and pin-plate discharge aliasing signals based on sparse representation theory(SBSS)and convolutional neural network(CNN)was proposed.Firstly,the normal operation sound signal of Wusheng 750 kV Substation was collected as the background sound,and the discharge sound signal and the common interference sound in the field were used as the foreground sound by constructing the pin-plate discharge model.The aliasing sound signal was constructed by adding the foreground sound with different signal-to-noise ratios to the background sound.Secondly,the blind separation algorithm based on SBSS was used to realize the separation of target foreground and redundant background voiceprint spectra.Finally,the hyperparameters of the CNN model were optimized to improve the classification performance of the model on the separated various types of foreground voiceprint spectra.The results show that the blind source separation algorithm can eliminate the redundant background sound interference so that the neural network can focus on the classification and recognition of foreground sound.The proposed method can separate foreground voiceprint in the aliasing sound signals,and the recognition accuracies of the CNN,the support vector machine(SVM),and the back-propagation neural network(BPNN)after separation are improved by 7.6%,17.2%,and 14.3%,respectively.关键词
局部放电/时频谱图/稀疏表示/盲分离/卷积神经网络/深度学习Key words
partial discharge/time-frequency spectrum/sparse representation/blind separation/convolutional neural network/deep learning分类
动力与电气工程引用本文复制引用
包艳艳,杨广泽,陈伟,冯婷娜..基于SBSS与CNN的750 kV变压器和尖板的放电信号声纹识别[J].西南交通大学学报,2025,60(3):781-792,12.基金项目
甘肃省电力公司电力科学研究院科技项目(52272219000Q)的支持. （52272219000Q）
