基于DCGAN的终端区激光雷达晴空湍流识别

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

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。

A study was conducted on the problem of insufficient turbulence samples and low recognition rate when us⁃ing 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 con⁃structed using the radial wind speed data of nine months from the Doppler wind lidar experimental platform of the Lan⁃zhou 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 convolu⁃tional 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.