基于CycleGAN和CNN的GIS振动信号去噪与机械缺陷识别

针对现场气体绝缘开关设备(gas insulated switchgear,GIS)振动检测结果易受外界背景噪声干扰的不足,文中提出基于生成对抗网络和卷积神经网络的现场GIS接触缺陷抗干扰检测框架。首先,开展GIS通流试验,获取在触指缺失、螺栓松动、存在分解物和导体对接深度不足4种典型缺陷下的振动波形,并收集包含背景噪声干扰的现场GIS振动波形作为参考,通过对振动数据进行图谱转化,构建用于背景噪声干扰去除和缺陷分类的数据集;其次,将现场振动图谱作为输入,采用周期一致生成对抗网络(cycle-consistent generative adversarial network,CycleGAN)对GIS进行现场背景噪声干扰去除;然后,采用AlexNet和ResNet18卷积网络结构对振动图谱特征进行提取;最后,采用全连接层对图谱特征进行分类,并对比不同振动信号图谱算法对分类结果的影响。结果表明,对于现场数据,所提模型的最大均值差异(maximum mean discrepancy,MMD)可达0.956 0,弗雷谢特起始距离(Fréchet inception distance,FID)可达62.09;Mel-ResNet18模型对GIS接触缺陷分类的准确率达99.43%。文中所提方法对于提高现场GIS振动检测和接触缺陷诊断结果的有效性具有重要应用价值。

GIS vibration signal denoising and mechanical defect identification based on CycleGAN and CNN

To overcome the influence of background noise interference on vibration detection efficiency,an anti-interference framework based on generation adversarial networks and convolutional neural networks (CNNs) is proposed to realize the contact defect detection for field gas insulated switchgear (GIS). Firstly,by current-carrying tests on prototype GIS platform,vibration waveforms of GIS with four artificial designed contact defects (missing finger,loosening bolt,with decomposed products and insufficient conductor insert depth) are acquired. Vibration waveforms on field GIS which contain background noise interference are also collected as a reference. Dataset for background noise interference removal and contact fault classification is built through spectrogram transform. Secondly,a cycle-consistent generative adversarial network (CycleGAN) with field vibration spectrogram as input is adopted to remove background noise interference on GIS. Then,two classical CNN architectures (AlexNet,ResNet18) are empirically designed to extract defeat features hidden in vibration spectrograms. Finally,the contact faults are identified via fully connected classifier. Influence of different time-frequency transformation algorithms on fault classification results are also compared. The results show that the proposed model can obtain maximum mean discrepancy (MMD) with 0.956 0 and Fréchet inception distance (FID) with 62.09 on field dataset,and the Mel-ResNet18 model could obtain 99.43% contact defect classification accuracy on test dataset. The proposed method in this paper can bring sound application value on improving the effectiveness of vibration detection and contact defect diagnosis results of field GIS.