用于吊弦故障检测的CycleGAN样本生成方法研究  

Research on CycleGAN sample generation method for dropper fault detection

作　　者：肖昊宇 顾桂梅[1] 曹文翔 XIAO Haoyu;GU Guimei;CAO Wenxiang(School of Electrical Engineering and Automation,Lanzhou Jiaotong University,Lanzhou 730070,China)

机构地区：[1]兰州交通大学自动化与电气工程学院,甘肃兰州730070

出　　处：《传感器与微系统》2024年第8期49-52,共4页Transducer and Microsystem Technologies

基　　金：甘肃省科技计划项目(20JR10RA216)。

摘　　要：针对深度学习算法在接触网吊弦缺陷识别时,存在数据样本不平衡、缺陷样本少、难以准确体现识别算法有效性的问题,提出了一种基于改进循环生成对抗网络(CycleGAN)模型的吊弦缺陷样本生成方法。首先,在生成器中用密集卷积块替换残差块,使得生成模型表现更稳定,收敛速度更快;然后,在生成器卷积层和密集卷积块后添加坐标注意力机制,使得生成的吊弦缺陷样本更清晰;最后,将常见的缺陷吊弦数据迁移到正常吊弦数据上,生成吊弦缺陷样本。仿真实验结果表明:所提出的方法比深度卷积生成对抗网络(DCGAN)算法和CycleGAN算法生成的图像更清晰,最终所生成的样本可以替代真实样本。Aiming at the problems of unbalanced data samples,few defect samples and difficulty in accurately reflecting the effectiveness of the identification algorithm when the depth learning algorithm is used to identify the dropper defects of the catenary,a method for generating dropper defect samples based on the improved cycle generative adversarial network(CycleGAN)model is proposed.Firstly,the residual blocks are replaced with dense convolutional blocks in the generator,so that the generatived model behaves more stably and converges faster.Then,a coordinate attention mechanism is added after the generator convolutional layer and the dense convolutional block to make the generated dropper defect sample more clearer.Finally,the common defect dropper data is migrated to the normal dropper data to generate a sample of the dropper defect.The simulation results show that the proposed method generates clearer images than deep convolutional GAN(DCGAN)algorithm and CycleGAN algorithm,and the finally generated samples can replace the real samples.

关 键 词：吊弦 循环生成对抗网络 密集卷积网络 坐标注意力机制 样本扩充 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] U225.4[自动化与计算机技术—计算机科学与技术]