基于改进生成对抗网络的玻璃绝缘子自爆缺陷检测方法

Self-explosion Defect Detection Method of Glass Insulator Based on Improved Generative Adversarial Network

为实现玻璃绝缘子自爆缺陷的无监督检测,该文提出一种基于改进生成对抗网络的无监督玻璃绝缘子自爆缺陷检测模型,该模型仅用正常玻璃绝缘子数据进行训练,以学习正常样本的特征分布,并基于目标与重建之间的偏差来检测缺陷。该方法利用U-Net网络搭建新生成器,使用跳跃连接结构捕捉正常样本的多尺度特征,并在图像空间和潜在空间重建目标,通过改进的差异评分来评估目标与正常分布的偏离程度,较高的差异分数代表缺陷的存在。研究结果表明:在该文所提出模型下,测试集的受试者工作特性曲线下的面积值达到0.846,相较已有的无监督学习检测方法平均提升35%左右,相较有监督学习检测方法平均提升4.4%左右。该文研究结果可为电力设备缺陷的无监督检测提供参考。

In order to realize the unsupervised detection of self-explosion defects of glass insulators,this paper proposes an unsupervised self-explosion defect detection model of glass insulators based on an improved Generative Adversarial Network.The model is only trained with normal glass insulator data to learn the feature distribution of normal samples and detect defects based on the deviation between the target and the reconstruction.In this method,the U-Net network is adopted to build a new generator,the skip connection structure is adopted to capture the multi-scale features of the normal sample,and the target is reconstructed in the image space and the latent space.The improved difference score is used to evaluate the deviation of the target from the normal distribution.A high difference score indicates the presence of defects.The research results show that the area under curve value of the test set under the proposed model reaches 0.846,which achieves an average increase of about 35%compared with the existing unsupervised learning detection method,and an average increase of about 4.4%compared with the supervised learning detection method.The research results can provide reference for the unsupervised detection of power equipment defects.