随机森林算法在擦机尾事件预测中的应用研究  

Research on the application of random forest algorithm in predicting tail strike events

作　　者：卢飞[1] 宋佳佳 LU Fei;SONG Jiajia(College of Air Traffic Management,Civil Aviation University of China,Tianjin 300300,China)

机构地区：[1]中国民航大学空中交通管理学院,天津300300

出　　处：《安全与环境学报》2024年第8期3028-3035,共8页Journal of Safety and Environment

基　　金：国家自然科学基金项目(52272356);中央高校基本业务费项目(3122022101);民航安全能力建设项目(ASSA2023/29)。

摘　　要：为实现对擦机尾事件的预测,研究提出一种以非超限飞行数据为基础,运用随机森林算法建立擦机尾事件的预测模型。首先,以B737-800机型的数据为基础,根据对着陆运行过程及擦机尾形成机理的分析得出影响着陆擦机尾事件的特征参数。其次,利用袋外数据计算不同特征的重要性分数,并以此为基础提取适用于随机森林预测的最优特征组合。最后,运用随机森林算法建立着陆擦机尾事件预测模型并与支持向量回归和长短期记忆网络模型进行对比。运用随机森林所建立的航空器擦机尾预测方法可以有效利用大量存在潜在价值的数据,实现对擦机尾风险告警的关口前移的目的。结果显示:在着陆前7 s预测的情况下,所建模型测试集的决定系数在0.85以上,且在改变数据规模的情况下仍能保持预测精确率,与支持向量回归和长短期记忆网络相比,随机森林模型具有较好的拟合度和更高的预测精度,可为飞行员提供足够的反应时间来采取必要的措施,进而避免擦机尾事件的发生。Ensuring flight safety is paramount in civil aviation operations.A tail strike incident involving an aircraft can result in structural damage to the airframe,posing significant risks that could potentially lead to accidents.To facilitate the prediction of tail strike events,a non-exceedance flight database was introduced.By integrating flight state analysis,the random forest algorithm was employed to develop a predictive model for tail strike events.The effectiveness of the model was validated across various dataset sizes.Initially,pertinent preprocessing was carried out on the obtained QAR data for the B737800 model to safeguard the integrity of the dataset.Furthermore,thirteen distinctive parameters influencing landing tail strike events were derived through an analysis of the landing operation process and the mechanism of tail strike formation.Subsequently,the importance scores of various features were calculated using out-of-bag data.The optimal feature combination parameters were extracted through calculations of importance scores.Finally,the authors utilized the random forest algorithm to establish a predictive model for strike events during the landing phase,which was then compared with the support vector regression and long-short memory network models.Aircraft tail strike prediction methods employing random forests can effectively harness a substantial amount of potentially valuable data to provide proactive alerts regarding tail-strike risk.The results demonstrate that the coefficient of determination of the test set for the proposed model exceeds 0.85 when predicting 7 seconds before landing.Furthermore,the model maintains its accuracy even when the scale of the data is adjusted.For instance,the coefficient of determination for the test set of the proposed model exceeds 0.9.When compared to support vector regression and long short-term memory models,this approach showcases superior fitting and higher predictive accuracy.It affords pilots ample time to react and implement necessary measures to prevent ta

关 键 词：安全工程 事故征候 机器学习算法 预测模型 快速存储记录器(QAR)数据 

分 类 号：X951[环境科学与工程—安全科学]