基于改进残差网络和InfoGAN的变压器局部放电故障诊断方法研究

Research on Transformer Partial Discharge Fault Diagnosis MethodBased on Improved Residual Network and InfoGAN

油浸式电力变压器作为电力系统的重要设备,对其绝缘状态进行科学诊断具有重要意义。目前,变压器局部放电故障诊断存在因样本匮乏或不均衡造成分类器泛化能力不足,最终导致识别准确率较低的问题。为此,提出了一种基于改进残差网络(ResNet)和信息生成对抗网络(information maximizing generative adversarial nets,InfoGAN)的变压器局部放电故障诊断方法。首先通过试验平台采集4类典型局放故障的PRPD谱图,然后利用InfoGAN对谱图进行样本增强并评估增强效果;最后,通过在残差块中嵌入注意力机制挤压激励(squeeze and excitation,SE)模块,得到改进残差网络模型,并利用该模型对故障类型进行识别。结果表明,InfoGAN可以有效丰富样本及其多样性;在样本数充足的条件下,改进的残差网络SE-ResNet18识别率可达99%,相比于其他经典CNN网络识别效果更优,且显著优于传统机器学习算法BPNN和SVM。

As an important equipment in power system,it is of great significance to scientifically diagnose the insulation state of oil-immersed power transformers.At present,the pattern recognition of partial discharge(PD)in transformer has the problem of insufficient generalization ability of classifier due to the lack of sample or unbalanced sample,resulting in low recognition accuracy.Therefore,we proposed a transformer PD fault diagnosis method based on improved residual network and InfoGAN.Firstly,the PRPD of four typical PD faults were collected by the experimental platform,and then the InfoGAN was used to enhance the sample and evaluate the enhancement effect.Finally,the improved residual network model was obtained by embedding the attention mechanism SE module in the residual block,and the model was used to identify the fault type.The results show that InfoGAN can effectively enrich the samples and their diversity;under the condition of sufficient samples,the recognition rate of the improved residual network SE-ResNet18 can reach 99%,which is better than that of other classical CNN networks,and is significantly better than the traditional machine learning algorithms BPNN and SVM.