基于小波散射网络-贝叶斯优化门控循环单元的电力变压器声纹识别方法

The voiceprint recognition method for power transformers based on wavelet scattering network-Bayesian optimized gated recurrent unit

针对小规模样本下电力变压器的声纹识别问题,本文提出一种基于小波散射网络-贝叶斯优化门控循环单元(GRU)的声纹识别方法。首先,为滤除干扰分量,提高声纹识别的正确率,通过经验小波变换(EWT)与快速独立成分分析算法(FastICA)对原始信号进行盲源分离,得到变压器本体声纹信号。然后,为降低模型输入数据的复杂度,采用小波散射网络提取声纹信号的特征向量作为声纹识别模型的输入,并采用GRU作为模型分类器。最后,通过贝叶斯算法完成对GRU网络层数与初始学习率的超参数优化。实验结果表明,在样本规模偏小的情况下,相较于当前普遍使用的声纹时频谱——深度卷积神经网络模型,本文所构建的模型收敛用时缩短,识别正确率提高,性能得到了明显改善。

This paper proposes a voiceprint recognition method for power transformers with small-scale samples based on wavelet scattering network-Bayesian optimized gated recurrent unit(GRU).Firstly,in order to filter out interference components and improve the accuracy of voiceprint recognition,the original signal extracted from the transformer is subjected to blind source separation through empirical wavelet transform(EWT)and fast independent component analysis algorithm(FastICA),resulting in the voiceprint signal of the transformer itself.Then,the feature vector of the voiceprint signal is extracted using the wavelet scattering network as the input of the voiceprint recognition model,and a GRU is applied as the classifier.To improve the recognition accuracy,Bayesian algorithm is utilized to optimize the hyperparameters of GRU layers and initial learning rate.The experimental results show that in the case of small sample size,compared with the commonly used voiceprint time-frequency spectrum and deep convolutional neural network,the model constructed in this paper converges faster and the recognition accuracy increases,significantly improving its performance.