基于生成对抗网络的航班起飞风险预测  

Flight Departure Risk Prediction Based on Generative Adversarial Networks

作　　者：秦雷 谢华[1] 陈海燕[1] 于颖 QIN Lei;XIE Hua;CHEN Hai-yan;YU Ying(Nanjing University of Aeronautics and Astronautics,Nanjing 211000,China)

机构地区：[1]南京航空航天大学,江苏南京211000

出　　处：《航空计算技术》2023年第2期80-83,87,共5页Aeronautical Computing Technique

基　　金：国家自然科学基金项目资助(61903187)。

摘　　要：航班的安全运行是民航的第一要素。根据历史数据对航班运行过程中可能存在的风险进行提前评估和预测是保障航班安全运行的有效手段。针对航班起飞阶段的风险问题,提出了一种基于生成对抗网络的航班起飞风险预测模型。对收集到的机组、飞机和环境3方面数据进行清洗、变换和融合,得到初始的起飞风险预测数据集;针对数据集存在的严重类不平衡问题,使用生成对抗网络进行数据增强以提高数据质量;在不同增强程度的数据集上,采用3种不同结构的神经网络训练航班起飞风险预测模型。实验结果表明,生成对抗网络能有效提高实测数据集的质量从而提高了航班起飞风险预测的精度。Safety is the first principle of civil aviation operation system.Prediction or evaluation of possible risks during flight operation based on historical data is an important method to ensure the safe operation.For the risk prediction of flight departure phase,this paper proposes a flight departure risk prediction model based on generative adversarial network(GAN).Firstly,the collected data from crew,aircraft and environment are cleaned,transformed and merged to obtain the initial dataset for risk prediction.Then,in order to solve the class imbalance problem in the dataset,data augmentation is performed by using GAN.Finally,the flight departure risk prediction model is trained on the datasets with different augmentation levels using three neural networks with different structures.The results of the comparison experiments show that GAN can effectively improve the quality of the real world data and thus improve the accuracy of flight departure risk prediction.

关 键 词：航班起飞风险预测 数据增强 生成对抗网络 神经网络 

分 类 号：V328[航空宇航科学与技术—人机与环境工程] O242[理学—计算数学]