基于卷积神经网络的激光雷达湍流预警

为实现湍流的自动化预警，提出了一种基于卷积神经网络的激光雷达湍流预警算法。首先，该方法将激光雷达获取的风速数据进行速度结构函数的构建；然后，拟合出涡流耗散率，进而将涡流耗散率构建为像素数据集。将数据集输入一种由两个卷积层、两个全连接层、一个softmax层、若干激活函数组成的卷积神经网络分类模型进行湍流识别；最后，采用学习率递减的方法来调整模型的参数对网络进行训练，网络收敛后，其损失度低至3%，通过对比实验表明网络的准确度可达到85%。运用中川机场2016年机组报告进行对比分析，结果表明:文中方法对大气湍流的预警命中率可达80%、误报率为13.3%、虚警率为6.7%，该方法与Hog-SVM分类方法相比，命中率显著提高，从而证明了该卷积网络模型在湍流预警中泛化能力强，提高了预警效率，能够为管制员和气象预报人员提供一种判断依据。

Turbulence warning based on convolutional neural network by lidar

In order to realize automatic turbulence warning, a novel turbulence warning algorithm based on convolution neural network(CNN) by lidar was proposed. Firstly, the velocity structure function was constructed from the wind speed data obtained by lidar; Then, the eddy dissipation rate was fitted, and then the eddy dissipation rate was constructed as a pixel data set. The data set was input into the CNN model composed of two convolution layers, two fully connected layers, one softmax layer and several activation functions for turbulence identification. The learning rate decreasing method is used to adjust the parameters of the model to train the network. After the network converges, the loss is as low as 3%. The comparative experiment shows that the accuracy of the network reaches 85%. Based on the flight crew report of Zhongchuan airport in 2016, the results show that the hit rate of this method for atmospheric turbulence warning is 80%, the false alarm rate is 13.3%, and the distort alarm is 6.7%. Compared with the Hog-SVM classification method, the hit rate of this method is significantly improved, which proves that the convolution network model has strong generalization ability in turbulence warning, and improves the warning efficiency significantly. It can provide a judgment basis for relevant weather forecasters.