基于生成对抗网络的自爆绝缘子检测模型设计

随着输电线路的持续建设，无人机逐步代替人工成为巡线工作的主要工作方式。绝缘子在输电线路中具有重要作用，然而，因自爆绝缘子导致的事故尤为频繁，从大量的航拍图像中识别自爆绝缘子，是一个亟待解决的任务。在航拍图像中，大部分绝缘子数据均是无损绝缘子，自爆绝缘子数量较少，因而无法满足识别算法的训练要求。针对现有输电线路无人机巡检中自爆绝缘子数据量稀缺的问题，该文提出了一种基于生成对抗网络的自爆绝缘子检测模型。通过生成器和鉴别器的对抗训练，该模型仅使用无损绝缘子数据训练即能完成对自爆绝缘子的检测。在此基础上，该文优化了生成对抗网络的训练过程。通过引入指导网络，解决了生成对抗网络的模式崩塌问题，提高了对自爆绝缘子检测的召回率；通过对鉴别器的输入添加扰动，解决了生成对抗网络中的样本不均衡问题，提高了对自爆绝缘子检测的精确度。通过与其他异常检测算法的对比实验，证明了该文方法的可靠性。并通过对模型各部分的消融实验，证明了该文方法各部分的可靠性。实验结果证明，该生成对抗网络模型有效避免了传统生成对抗网络中的缺陷，完成了对自爆绝缘子的高效自动检测。

New Design of Self-Explosive Insulator Detection Model Based on Generative Adversarial Network

With the continued construction of transmission lines in China, the manual work on line patrol is being replaced gradually by the unmanned aerial vehicle, Insulators play important role in transmission lines, however, in view of that the accidents caused by self-bursting insulators particularly frequently occur, so identifying self-bursting insulators from aerial images is an urgent task to be solved. In the aerial images, most of the insulator data belong to lossless insulators and less data belong to the self-bursting insulators, which number is small, thus not meeting the training requirements of the recognition algorithm. In allusion to the scarcity of self-exploding insulator data in existing transmission line unmanned aerial vehicle (abbr. UAV) inspections, based on generative adversarial networks a self-exploding insulator detection model was proposed. Through adversarial training between the generator and the discriminator, the proposed model could complete the detection of self-exploding insulators by only using lossless insulator data during training. The training process of the generative adversarial network was optimized. By means of introducing a guidance network, the mode collapse problem of the generative adversarial network could be solved, and the recall rate of the self-exploding insulator detection was improved; through adding a perturbation to the input of the discriminator, the sample imbalance problem in the generative adversarial network was solved, and the accuracy of the self-exploding insulator detection was improved. The reliability of the put forward method was demonstrated through comparison experiments with other anomaly detection algorithms. The reliability of each part of the put forward method was also demonstrated by ablation experiments on each part of the model. The experimental results demonstrate that the defects in traditional generative adversarial networks can be effectively avoided by the proposed generative adversarial network model, and the efficient automatic detection of self-exploding insulators is accomplished.