This paper presents a novel insulator defect detection scheme based on Deep Convolutional Auto-Encoder (DCAE) for small negative samples. The proposed DCAE scheme combines the advantages of supervised learning and unsupervised learning. In order to reduce the high cost of training Deep Neural Networks, this paper pre-trained the Convolutional Neural Networks (CNN) through open labelled datasets. Through transferring learning, the encoder part of the traditional Convolutional Auto-Encoder was replaced by the first three layers of the CNN, and a small number of defect samples were used to fine-tune the parameters. A threshold discrimination scheme was designed to evaluate the model detection, realising the self-explosion detection of insulator by judging the residual result and abnormal score. The experimental results show that compared with the existing insulator self-explosion detection schemes, the proposed scheme can reduce the model training time by up to 40%, and the recognition accuracy can reach 97%. Moreover, this model does not need a large number of insulator labelled data and is especially suitable for small negative sample application.

中文翻译：

---

一种基于深度卷积自动编码器的小负样本绝缘子缺陷检测方案

本文针对小负样本提出了一种基于深度卷积自动编码器（DCAE）的新型绝缘子缺陷检测方案。提出的 DCAE 方案结合了监督学习和无监督学习的优点。为了降低训练深度神经网络的高成本，本文通过开放标记数据集对卷积神经网络（CNN）进行了预训练。通过迁移学习，将传统卷积自编码器的编码器部分替换为CNN的前三层，并使用少量缺陷样本对参数进行微调。设计了阈值判别方案对模型检测进行评价，通过判断残差结果和异常分值实现绝缘子自爆检测。实验结果表明，与现有的绝缘子自爆检测方案相比，该方案可以将模型训练时间缩短40%，识别准确率可达97%。而且，该模型不需要大量的绝缘体标记数据，特别适合小负样本应用。