基于改进生成对抗网络的变压器声纹故障诊断

Transformer Voiceprint Fault Diagnosis Based on Improved Generative Adversarial Network

为了提高小样本条件下变压器声纹故障诊断的准确率,提出了一种基于梅尔声谱图和改进的Wasserstein生成对抗网络(IW-GAN)的变压器声纹诊断模型。提取变压器声信号的梅尔声谱图,将声谱图输入到IW-GAN中进行样本扩充。其中,IW-GAN使用更具表达能力的Transformer网络,判别器采用满足Lipschitz连续性约束的SN-CNN,从而使IW-GAN能够稳定生成多样性和高质量的样本;将扩充后的数据输入不同的分类器中进行故障分类。实验证明,所提方法在有效扩充变压器故障声纹数据的同时,显著提升了小样本情况下变压器声纹故障诊断的整体性能。该方法对不同分类器的识别准确率均有显著提升,特别是对卷积神经网络分类准确率的提升达到了6.9%。

In order to improve the accuracy of transformer voiceprint fault diagnosis under small sample conditions,a transformer voiceprint diagnosis model based on Mel spectrogram and IW-GAN is proposed.Firstly,the Mel spectrogram of transformer sound signal is extracted,and then the spectrogram is input into IW-GAN for sample expansion.Among them,IW-GAN uses a more expressive transformer network,and the discriminator uses SN-CNN that satisfies Lipschitz continuity constraints.This improvement enables IW-GAN to stably generate diverse and high-quality samples;Finally,the expanded data is input into different classifiers for fault classification.Experimental results show that the proposed method effectively expand the transformer fault voiceprint data and significantly improve the overall performance of transformer voiceprint fault diagnosis under small sample conditions.This method significantly improves the recognition accuracy of different classifiers,especially the recognition accuracy of CNN classifier has been improved by 6.9%.