一种基于改进的WGAN模型的电缆终端局部放电识别准确率提升方法

Accuracy Improvement of Cable Termination Partial Discharging Recognition Based on Improved WGAN Algorithm

目前,基于机器学习的电缆终端局部放电模式识别常因标注数据匮乏或不均衡而导致泛化能力不足,识别准确率较低。为解决该问题,该文提出了一种基于改进的Wasserstein生成对抗网络(Wasserstein generative adversarial network,WGAN)模型的电缆终端局部放电识别准确率提升方法。该方法以单个脉冲对应的小波时频谱图为对象,首先训练具有条件生成能力且训练过程稳定的改进WGAN模型并生成新的样本;然后利用新样本对原始样本集进行扩充,以提高样本多样性;最后,利用扩充后数据集训练得到新的局放分类器。实验结果表明,所提方法相较于其他条件生成模型能够更稳定生成新的高质量样本;利用该方法分别对典型电缆终端缺陷数据进行扩充,训练出的新分类器具有更优的泛化能力,且对不同分类器具有适用性。该方法有效抑制了工程中局放类型识别时由于数据匮乏或不均衡所导致的过拟合风险,有效提升了在小数据量下的识别准确率。

Nowadays,the recognition of the partial discharging(PD)pattern at the cable terminations based on machine learning often leads to the insufficient generalization ability and low recognition accuracy due to the lack or imbalance of the marked data.In order to solve the problem,a novel method for improving the recognition accuracy of the cable termination PD based on the improved Wasserstein generative adversarial network(WGAN)is presented in this paper.Taking the wavelet T-F spectrums corresponding to the pulses as the object,this method first trains the improved WGAN model with the conditional generation capabilities and stable training process to generate the new samples.Then it uses the samples to expand the original data set and improve the sample diversity.Finally,the expanded data set is used to train the new PD classifier.The experiment results show that,compared with the other generative models,the method in this paper generates new high-quality samples more stably.When this method is used to expand the typical cable termination defect data,the new classifier trained has better generalization ability and is applicable to different classifiers.This method effectively suppresses the over-fitting risks caused by the lack or imbalance of data in the identification of partial discharging types in engineering,improving the recognition accuracy.