航空学报 > 2024, Vol. 45 Issue (16): 329748-329748   doi: 10.7527/S1000-6893.2024.29748

基于DCGAN的终端区激光雷达晴空湍流识别

庄子波1, 张春辉2, 陈星3, 邵靖媛1, 陈柏纬4()   

  1. 1.中国民航大学 航空气象研究所,天津 300300
    2.中国民航大学 安全科学与工程学院,天津 300300
    3.中国民航大学 空中交通管理学院,天津 300300
    4.香港天文台,香港 999077
  • 收稿日期:2023-10-20 修回日期:2023-12-28 接受日期:2024-01-06 出版日期:2024-01-15 发布日期:2024-01-11
  • 通讯作者: 陈柏纬 E-mail:pwchan@hko.gov.hk
  • 基金资助:
    天津市自然科学基金(21JCYBJC00740);气象软科学项目(2023ZZXM29);江苏省重点研发项目(社会发展)(BE2021685)

Clear⁃air turbulence recognition by Doppler⁃wind⁃lidar in terminal area based on DCGAN

Zibo ZHUANG1, Chunhui ZHANG2, Xing CHEN3, Jingyuan SHAO1, Pakwai CHAN4()   

  1. 1.Aviation Meteorological Research Institute,Civil Aviation University of China,Tianjin 300300,China
    2.College of Safety Science and Engineering,Civil Aviation University of China,Tianjin 300300,China
    3.Air Traffic Management College,Civil Aviation University of China,Tianjin 300300,China
    4.Hong Kong Observatory,Hong Kong 999077,China
  • Received:2023-10-20 Revised:2023-12-28 Accepted:2024-01-06 Online:2024-01-15 Published:2024-01-11
  • Contact: Pakwai CHAN E-mail:pwchan@hko.gov.hk
  • Supported by:
    Natural Science Fund Project of Tianjin(21JCYBJC00740);Meteorological Soft Science Project(2023ZZXM29);Key Research and Development-Social Development Program of Jiangsu Province(BE2021685)

摘要:

针对在飞行终端区采用激光雷达进行晴空湍流识别时湍流样本不充足、识别率低的问题,提出了一种改进的深度卷积生成对抗网络(DCGAN)算法。使用中川机场雷达实验平台9个月的径向风速数据构建涡流耗散率(EDR)图像,筛选存在晴空湍流的样本构建湍流样本集;通过扩充卷积层和转置卷积层改进了DCGAN结构,实现样本的扩充,并使用对抗训练后的判别器进行识别。结果表明:用原始样本集和增广样本集训练的对抗后判别器识别准确率均优于卷积神经网络(CNN)和对抗前判别器,分别提高了6.55%、8.25%和0.31%、1.9%;最后,用实测样本进行检验,识别准确率分别提高了3.33%、6.67%,验证了所提方法的可行性。

关键词: 激光雷达, 晴空湍流, 涡流耗散率(EDR), 深度卷积生成对抗网络(DCGAN), 判别器

Abstract:

A study was conducted on the problem of insufficient turbulence samples and low recognition rate when using Doppler wind lidar for clear-air turbulence recognition in the flight terminal area. An improved Deep Convolutional Generative Adversarial Network (DCGAN) algorithm was proposed. Eddy Dissipation Rate (EDR) images were constructed using the radial wind speed data of nine months from the Doppler wind lidar experimental platform of the Lanzhou Zhongchuan International Airport. Samples with clear-air turbulence were selected to construct a turbulence sample set, the DCGAN structure was improved by expanding the convolutional layer and transposing the convolutional layer, so as to achieve sample expansion. The post-confrontation was then used for recognition. The results show that the recognition accuracy of the post-confrontation discriminator trained with the original sample set and that trained with the augmented sample set are both better than that trained with the Convolutional Neural Network (CNN) and the pre-confrontation discriminator, with improvements of 6.55%, 8.25%, and 0.31%, 1.9%, respectively. A comparison with the measured samples shows that the recognition accuracy was improved by 3.33% and 6.67%, verifying the feasibility of the proposed method.

Key words: Doppler wind lidar, clear-air turbulence, Eddy Dissipation Rate (EDR), Deep Convolutional Generative Adversarial Network (DCGAN), discriminator

中图分类号: