航班运行控制风险评估精度提升方法  

A method of flight operations control risks assessment accuracy improvement

作　　者：谢春生[1] 杨志远 刘锟 王岩韬[1] XIE Chun-sheng;YANG Zhi-yuan;LIU Kun;WANG Yan-tao(Key Laboratory of Airline Artificial Intelligence Civil Aviation Administration,Civil Aviation University of China,Tianjin 300300,China)

机构地区：[1]中国民航大学航空公司人工智能民航局重点实验室,天津300300

出　　处：《安全与环境学报》2022年第3期1227-1234,共8页Journal of Safety and Environment

基　　金：国家自然科学基金项目(U1933103)。

摘　　要：为了解决航班风险评估精度不足的问题,对某航空公司225个航班运行数据进行统计和分析,运用Lasso和随机森林算法、粗糙集分析和支持向量机、主成分分析与RBF神经网络结合3类算法,使用相同训练集和测试集构建风险评价模型。结果表明:随机森林算法分类精度为88%;主成分分析与支持向量机算法合用分类精度由64%提升至86%;非线性主成分分析与RBF神经网络算法合用精度由52%提升至80%。综合3类算法的精度适用范围,构建混合模型,其最终分类结果精度可高达94%;并且,经过K折稳定性检验验证了方案的可用性和可靠性。This paper aims to solve the problem of insufficient flight risk assessment accuracy.Based on 225 sets of standard data of an airline,the risk level of the flight operation can be found,including 15 indexes by quantifying the values from 1 to 10.First of all,Lasso was adopted to sort and filter risk indexes according to their importance and Random Forest algorithm was used to select a certain amount of risk indexes and training sets to construct a risks assessment model that the accuracy was tested using the testing set.Then,according to Principal Component Analysis(including Linear Principal Component Analysis and Nonlinear Principal Component Analysis)and Rough Set Theory to reduce the dimension of risk indexes,Support Vector Machine was adopted to build a risk assessment model and calculate its accuracy with the same training set and testing set.Besides,the RBF Neural Network algorithm was also adopted to construct the risk assessment model and calculate accuracy with the data after the dimensionality reduction process.The accuracies and the scopes of application of each algorithm were compared,and the mixed algorithm model was built within their respective scopes of application.Then its availability and reliability were validated via the K-fold cross-validation.The calculation results show that the classification accuracy of the Random Forest algorithm is 88%;the accuracy of Support Vector Machine algorithm can be increased to 86%from 64%using the data pre-processing of the Linear Principal Component Analysis;the accuracy of RBF Neural Network can be increased to 80%from 52%using the data pre-processing of Nonlinear Principal Component Analysis;the final classification accuracy of the mixed model can be as high as 94%.Thus,it is proved that the mixed algorithm model can effectively improve the accuracy of flight operation control risk assessment,compared with a single algorithm.

关 键 词：安全工程 航班运控风险 风险评估 Lasso与RF 支持向量机 RBF神经网络 

分 类 号：X949[环境科学与工程—安全科学] X951