Pan Weijun, Wang Yuhao, Deng Leilei, Jiang Yanqiang, Leng Yuanfei
School of Air Traffic Management, Civil Aviation Flight University of China, Guanghan 618307, China.
School of Transportation and Logistics, Southwest Jiaotong University, Chengdu 611756, China.
Sensors (Basel). 2025 May 4;25(9):2909. doi: 10.3390/s25092909.
This paper proposes a hybrid deep learning network architecture (Inception-VGG16) to address the challenge of accurate aircraft wake vortex identification. The model first employs a Feature0 module for preliminary feature extraction of two-dimensional Doppler radar radial velocity data. This module comprises convolution, batch normalization, ReLU activation, and max pooling operations. Subsequently, improved InceptionB and InceptionC modules are utilized for parallel extraction of multi-scale features. The InceptionB former module adopts two parallel branches, combining 1 × 1 and 3 × 3 convolutions, and outputting 64-channel feature maps, while the InceptionC latter module expands the number of channels number to 128, enhancing the model's feature representation capability. The backend employs the VGG16's hierarchical structure, performing deep feature extraction through multiple convolution and pooling operations, and ultimately achieving wake vortex classification through fully connected layers. Experimental validation based on 3530 wind field samples collected at the Chengdu Shuangliu Airport demonstrates that compared to traditional methods (SVM, KNN, RF) and single deep networks (VGG16), the proposed hybrid model achieves a classification accuracy of 98.8%, significantly outperforming comparative traditional methods (SVM, KNN, RF) and single deep networks (VGG16). The model not only overcomes the limitations of single networks in processing multi-scale wake features but also enhances the model's ability to identify wake vortices in complex backgrounds through deep feature hierarchies, providing a new technical solution for aviation safety monitoring systems based on deep learning.
本文提出了一种混合深度学习网络架构（Inception-VGG16），以应对准确识别飞机尾流涡旋的挑战。该模型首先采用Feature0模块对二维多普勒雷达径向速度数据进行初步特征提取。该模块包括卷积、批量归一化、ReLU激活和最大池化操作。随后，利用改进的InceptionB和InceptionC模块并行提取多尺度特征。前一个InceptionB模块采用两个并行分支，结合1×1和3×3卷积，并输出64通道特征图，而后一个InceptionC模块将通道数扩展到128，增强了模型的特征表示能力。后端采用VGG16的层次结构，通过多次卷积和池化操作进行深度特征提取，并最终通过全连接层实现尾流涡旋分类。基于在成都双流机场收集的3530个风场样本的实验验证表明，与传统方法（支持向量机、K近邻、随机森林）和单一深度网络（VGG16）相比，所提出的混合模型实现了98.8%的分类准确率，显著优于对比的传统方法（支持向量机、K近邻、随机森林）和单一深度网络（VGG16）。该模型不仅克服了单一网络在处理多尺度尾流特征方面的局限性，还通过深度特征层次增强了模型在复杂背景下识别尾流涡旋的能力，为基于深度学习航空安全监测系统提供了一种新的技术解决方案。
