A Study on Aircraft Safety Target Levels Based on Lasso-Random Forest Model
-
摘要: 随着航空安全水平的不断提升，运输事故呈现出小样本、低概率特征，传统基于历史数据的预测方法难以充分刻画当前航空运行风险演化规律，难以满足安全管理的精细化和个性化需求。针对低概率事故样本不足、直接预测不稳定的问题，研究了基于最小绝对收缩与选择算子回归（least absolute shrinkage and selection operator，Lasso）-随机森林模型预测运输事故征候的安全目标水平计算方法。在综合考虑航空运输规模、运行效率、资源投入，以及运行强度等多维因素的基础上，初步构建运输事故征候影响因素集，引入Lasso回归通过时间序列交叉验证方法进行特征筛选，有效缓解小样本条件下多变量共线性问题，提高特征选择的稳定性与合理性。采用随机森林模型对运输事故征候进行预测，通过特征重要性分析与误差驱动的模型简化策略，提高模型的预测精度并且在保证精度的同时降低模型复杂度，提高实用性。以中国2003—2022年民航运行数据为样本进行验证，结果表明：Lasso-随机森林模型具有最低标准化均方根误差（standardized root mean square error，SRMSE）值（45.2）和最高决定系数R2值（0.834），在预测精度上显著优于线性回归和支持向量机（supportvector regression，SVR）预测模型。模型简化后SRMSE比原模型进一步降低6.14%。基于简化后的模型对2023年飞行时间和事故征候次数进行预测，得航路上航空器碰撞的安全目标水平为符合标准。Abstract: As aviation safety continuously improves, transportation accidents exhibit small-sample and low-probability characteristics. Traditional prediction methods based on historical data struggle to characterize the evolution of aviation operational risks and refined safety management demands. To address prediction instability caused by insufficient accident samples, a method for calculating the target level of safety using a Lasso-random forest model is proposed. The method integrates Lasso regression and a random forest model to improve robustness under low-probability conditions. An influencing factor set for transportation incident precursors is constructed by considering transport scale, operational efficiency, resource input, and operational intensity. Lasso regression combined with time-series cross-validation is applied for feature selection to alleviate multicollinearity under small-sample conditions. This procedure improves the stability and rationality of selected features. A random forest model is employed to predict transportation incident precursors. Feature importance analysis is applied to improve prediction accuracy. An error-driven model simplification strategy is used to reduce model complexity and enhance practical applicability. Civil aviation operational data of China from 2003 to 2022 are used for validation. Results indicate that the Lasso-random forest model achieves the lowest SRMSE value of 45.2 and the highest R2 value of 0.834. The model significantly outperforms linear regression and support vector regression models. After simplification, the SRMSE is further reduced by 6.14%. Based on the simplified model, flight hours and incident precursor occurrences for 2023 are predicted. The resulting en-route aircraft collision safety target level is, which satisfies applicable safety standards. The proposed method provides a robust and operational framework for low-probability aviation risk assessment and safety target level formulation.
-
Key words:
- aviation safety /
- transportation incident /
- Lasso regression /
- random forest /
- target level of safety
-
图 1 Lasso方法影响因素筛选的运输事故征候预测问题流程图
Figure 1. Flowchart of the transportation incident prediction process based on influencing factor selection using the lasso method
图 4 原始和化简后Lasso-随机森林模型运输事故征候拟合情况
Figure 4. Original and simplified Lasso-random forest model commercial aviation incident fit
图 5 2003-2022年运输事故征候和事故次数变化图
Figure 5. Changes in transportation incidents and number of accidents from 2003 to 2022
表 1 部分原始数据
Table 1. Part of the raw data
影响因素 2003年 2004年 2005年 影响因素 2003年 2004年 2005年 运输总周转量/($\mathrm{kg} \cdot \mathrm{m}$) $1.71 \times 10^{16}$ $2.31 \times 10^{16}$ $2.61 \times 10^{16}$ 民航基本建设和技术改造投资/元 $1.1 \times 10^{10}$ $2.11 \times 10^{9}$ $2.12 \times 10^{9}$ 旅客运输量/人 $8.759 \times 10^{7}$ $1.212\;3 \times 10^{8}$ $1.382\;7 \times 10^{8}$ 民用航空运输机场/个 143 137 142 货邮运输量/kg $2.19 \times 10^{9}$ $2.77 \times 10^{9}$ $3.07 \times 10^{9}$ 完成旅客吞吐量/万人 $1.743\;3 \times 10^{8}$ $2.419\;4 \times 10^{8}$ $2.844\;1 \times 10^{8}$ 旅客周转量/($\mathrm{kg} \cdot \mathrm{m}$) $1.263 \times 10^{16}$ $1.782\;2 \times 10^{16}$ $2.044\;9 \times 10^{16}$ 货邮吞吐量/kg $4.52 \times 10^{9}$ $5.53 \times 10^{9}$ $6.33 \times 10^{9}$ 货邮周转量/($\mathrm{kg} \cdot \mathrm{m}$) $4.47 \times 10^{15}$ $5.278 \times 10^{15}$ $5.651 \times 10^{15}$ 飞行时间/h 1 837 129 2 3933 46 2 733 312 航班正常率/% 79.8 79.9 81.99 飞行班次/次 833 428 1 075 879 1 367 568 收入水平（/元/kg·m） 4 990 5 200 5 130 航线条数/条 1 155 1 279 1 257 燃油消耗（/kg/kg·m） 0.354 0.341 0.336 航线里程/m $1.749 545 \times 10^{9}$ $2.049 394 \times 10^{9}$ $1.998 501 \times 10^{9}$ 年末全行业在册运输飞机数/架 661 754 863 起降架次/次 2 118 790 2 666 309 3 056 521 运输飞机平均日利用率/h 7.8 9.4 93 从业人数/人 $1.7\;882 \times 10^{5}$ $2.0\;502 \times 10^{5}$ $2.0\;987 \times 10^{5}$ 正班客座率/% 64.4 69.6 71.5 国内生产总值/元 $1.374\;220\;349 \times 10^{14} 1$ $1.374\;220\;349 \times 10^{14} 1$ $1.873\;189\;031 \times 10^{14}$ 正班运载率/% 61.9 64 65 运输事故征候/起 100 106 116 
下载: 导出CSV
表 2 Lasso最优调和参数下的回归系数
Table 2. Regression coefficients under Lasso's s optimal reconciliation parameters
影响因素 相关性 Lasso回归系数 $A_{1}$ 正相关 164.508 7 $A_{2}$ 正相关 41.202 6 $A_{3}$ 正相关 35.518 0 $A_{4}$ 正相关 3.979 5 $A_{5}$ 正相关 1.062 1 $A_{6}$ 正相关 0.610 9 $A_{7}$ 负相关 -6.097 7 $A_{8}$ 负相关 -3.829 7
下载: 导出CSV
表 3 模型参数优化结果
Table 3. parameter optimization results
模型 最优模型结构 随机森林 决策树的数量=200
分裂内部节点所需的最小样本数=2
叶节点所需的最小样本数=1
最大深度=4SVR-高斯径向基核函数 Ggamma（核系数）=0.101
C（惩罚参数）=1 000SVR-线性核函数 $C=262.7$ SVR-多项式核函数 Ggamma=0.202
C=10.2SVR-sigmoid核函数 Ggamma=0.101
C=303.1
下载: 导出CSV
表 4 不同预测模型效果
Table 4. Effectiveness of different prediction models
预测模型 $S_{\text {RMSE }}$ $R^{2}$ 线性回归 104.7 0.111 SVR-线性核函数 98.0 0.221 SVR-多项式核函数 110.7 0.005 SVR-高斯径向基核函数 47.6 0.816 SVR-Sigmoid核函数 148.9 -0.800 随机森林 45.2 0.834
下载: 导出CSV
表 5 Lasso-随机森林模型特征重要性评分
Table 5. Lasso-random forest model feature importance scores
影响因素 重要度 $A_{5}$ 0.189 143 $A_{4}$ 0.173 267 $A_{7}$ 0.161 224 $A_{8}$ 0.151 983 $A_{1}$ 0.144 976 $A_{3}$ 0.132 971 $A_{2}$ 0.042 606 $A_{6}$ 0.003 830
下载: 导出CSV
表 6 化简Lasso-随机森林模型效果
Table 6. Simplified Lasso-random forest model effects
子集 $S_{\text {RMSE }}$ $\left\{A_{5}\right\}$ 85.9 $\left\{A_{5}, A_{4}\right\}$ 47.2 $\left\{A_{5}, A_{4}, A_{7}\right\}$ 64.3 $\left\{A_{5}, A_{4}, A_{7}, A_{8}\right\}$ 48.2 $\left\{A_{5}, A_{4}, A_{7}, A_{8}, A_{1}\right\}$ 39.8 $\left\{A_{5}, A_{4}, A_{7}, A_{8}, A_{1}, A_{3}\right\}$ 50.3 $\left\{A_{5}, A_{4}, A_{7}, A_{8}, A_{1}, A_{3}, A_{2}\right\}$ 51.9 $\left\{A_{5}, A_{4}, A_{7}, A_{8}, A_{1}, A_{3}, A_{2}, A_{5}\right\}$ 45.2
下载: 导出CSV
表 7 飞行时间的Lasso-随机森林模型特征重要性评分
Table 7. Lasso-random forest model feature importance scores for flight time
影响因素 重要度 起降架次 0.239 278 年份 0.227 152 飞行班次 0.216 282 航线里程 0.198 201 正班客座率 0.119 086
下载: 导出CSV
-
[1] DAVIES L, SHARPE A. Review of target level of safety for north Atlantic MNPS airspace[J]. Air Traffic Control Quarterly, 1994, 2(4): 249-272. doi: 10.2514/atcq.2.4.249 [2] 张健, 赵嶷飞, 卢飞, 等. 城市物流无人机平行航线横向间隔研究[J]. 交通信息与安全, 2024, 42(4): 118-124. doi: 10.3963/j.jssn.1674-4861.2024.04.013ZHANG J, ZHAO Y F, LU F, et al. Lateral separation of parallel routes for urban logistics unmanned aerial vehicles[J]. Journal of Transport Information and Safety, 2024, 42(4): 118-124. (in Chinese) doi: 10.3963/j.jssn.1674-4861.2024.04.013 [3] CAMPOS L, MARQUES J. On a dimensionless alternative to the ICAO target level of safety[J]. Proceedings of the Institution of Mechanical Engineers, Part G: Journal of Aerospace Engineering, 2016, 230(9): 1548-1557. doi: 10.1177/0954410015621336 [4] ICAO. Safety management manual(SMM): Doc 9859[S]. Montreal: ICAO, 2013. [5] 王剑辉, 李翰之. 基于改进灰色模型预测民航事故征候数[J]. 微型电脑应用, 2021, 37(12): 134-136, 140.WANG J H, LI H Z. Prediction of civil aviation accident incident numbers based on an improved grey model[J]. Microcomputer Applications, 2021, 37(12): 134-136, 140. (in Chinese) [6] 熊明兰, 王华伟, 徐怡, 等. 基于鸟击事故征候预测的通用航空安全研究[J]. 系统工程与电子技术, 2020, 42(9): 2033-2040.XIONG M L, WANG H W, XU Y, et al. General aviation safety analysis based on bird strike incident prediction[J]. Journal of Systems Engineering and Electronics, 2020, 42 (9): 2033-2040. (in Chinese) [7] 吴雨桐, 方学东. 基于组合模型的民航事故征候预测研究[J]. 综合运输, 2022, 44(12): 112-116.WU Y T, FANG X D. Prediction of civil aviation accident incidents based on a combination model[J]. China Transportation Review, 2022, 44(12): 112-116. (in Chinese) [8] BLEU LAINE M H, PURANIK T G, MAVRIS D N, et al. Multiclass multiple-instance learning for predicting precursors to aviation safety events[J]. Journal of Aerospace Information Systems, 2022, 19(1): 22-36. doi: 10.2514/1.I010971 [9] TAKALE D G, GUNJAL S D, KHAN V N, et al. Road accident prediction model using data mining techniques[J]. NeuroQuantology, 2022, 20(16): 2094-2101. [10] 文军, 陈建勤, 罗心雨, 等. 基于IOWGA的运输航空飞行事故征候组合预测[J]. 安全与环境学报, 2022, 22(1): 256-262.WEN J, CHEN J Q, LUO X Y, et al. Combined prediction of transport aviation accident incidents based on IOWGA[J]. Journal of Safety and Environment, 2022, 22(1): 256-262. (in Chinese) [11] 梁文娟, 李雪艳. 基于ARIMA、LS-SVM和BP神经网络组合模型的航空运输飞行事故征候预测[J]. 安全与环境工程, 2018, 25(1): 130-136.LIANG W J, LI X Y. Flight incidents prediction of air transportation based on the combined model of ARIMA, LS-SVM and BP neural network[J]. Safety and Environmental Engineering, 2018, 25(1): 130-136. (in Chinese) [12] 陈芳, 张迪, 卫微, 等. 基于ARIMA和BP神经网络的民航事故征候率预测[J]. 安全与环境学报, 2019, 19(6): 2040-2047.CHEN F, ZHANG D, WEI W, et al. Integrated forecasting model for civil aviation incident rate based on the ARIMA and BP neural network[J]. Journal of Safety and Environment, 2019, 19(6): 2040-2047. (in Chinese) [13] 叶雪飞. 基于数据挖掘的民航旅客运输事故征候研究[D]. 广汉: 中国民用航空飞行学院, 2023.YE X F. Study on civil aviation passenger transport incident based on data mining[D]. Guanghan: Civil Aviation Flight University of China, 2023. (in Chinese) [14] 王永刚, 吕学梅. 民航事故征候的灰色马尔可夫预测[J]. 安全与环境学报, 2008, 8(1): 163-165.WANG Y G, LYU X M. Grey Markov model for forecasting civil aviation incidents[J]. Journal of Safety and Environment, 2008, 8(1): 163-165. (in Chinese) [15] 刘智玉, 陈南梁, 汪军. 基于随机森林算法的小样本纱线质量预测[J]. 东华大学学报(自然科学版), 2023, 49(6): 80-86.LIU Z Y, CHEN N L, WANG J. Yarn quality prediction for small samples based on random forest algorithm[J]. Journal of Donghua University(Natural Science), 2023, 49(6): 80-86. (in Chinese) [16] 曹冲, 程林松, 张向阳, 等. 基于多变量小样本的渗流代理模型及产量预测方法[J]. 力学学报, 2021, 53(8): 2345-2354.CAO C, CHENG L S, ZHANG X Y, et al. Seepage proxy model and production forecast method based on multivariate small samples[J]. Chinese Journal of Theoretical and Applied Mechanics, 2021, 53(8): 2345-2354. (in Chinese) [17] MIOLANE L, MONTANARI A. The distribution of the Lasso[J]. The Annals of Statistics, 2021, 49(4): 2313-2335. [18] 随泽远, 王军, 彭宏, 等. 基于门控脉冲神经P系统模型的概率负荷预测[J/OL]. 中国电力: 1-10(2025-08-28)[2025-12-14]. https://link.cnki.net/urlid/11.3265.tm.20250828.0927.002 .SUI Z Y, WANG J, PENG H, et al. Probabilistic load forecasting based on gated spiking neural P system model[J/OL]. Electric Power: 1-10(2025-08-28)[2025-12-14].https://link.cnki.net/urlid/11.3265.tm.20250828.0927.002 . (in Chinese)[19] BREIMAN L. Random forests[J]. Machine Learning, 2001, 45 (1): 5-32. doi: 10.1023/A:1010933404324 [20] 卢飞, 宋佳佳. 随机森林算法在擦机尾事件预测中的应用研究[J]. 安全与环境学报, 2024, 24(8): 3028-3035.LU F, SONG J J. Application of the random forest algorithm in predicting tail strike events[J]. Journal of Safety and Environment, 2024, 24(8): 3028-3035. (in Chinese) [21] 吴潇雨, 和敬涵, 张沛, 等. 基于灰色投影改进随机森林算法的电力系统短期负荷预测[J]. 电力系统自动化, 2015, 39 (12): 50-55.WU X Y, HE J H, ZHANG P, et al. Short-term load forecasting of power systems based on an improved random forest algorithm using gray projection[J]. Automation of Electric Power Systems, 2015, 39(12): 50-55. (in Chinese) [22] PETER S, PHILIP S G. Safety minima study: review of existing standards and practices[R]. Brussels: EUROCONTROL, 2000. -
点击查看大图
图(5) / 表(7)
计量
- 文章访问数: 13
- HTML全文浏览量: 9
- PDF下载量: 2
- 被引次数: 0