Identification and analysis of fatal accidents of training aircraft under uneven distribution of accident types

doi:10.16265/j.cnki.issn1003-3033.2026.08.0848

Abstract

Abstract:

To clarify the lethality mechanisms of trainer aircraft accidents with significantly imbalanced accident data, the influencing factors of their lethality were analyzed. The accident data for general aviation trainer aircraft from the National Transportation Safety Board (NTSB) were utilized. Six dimensions of variables, namely time, flight, pilot, aircraft operation, meteorology, and airport, were extracted. Given that fatal accidents accounted for 9.66% of the data, a sample imbalance existed. Resampling techniques, including the Synthetic Minority Over-sampling Technique (SMOTE) and Adaptive Synthetic (ADASYN) sampling, were employed to address this issue. Multiple interpretable machine learning models were subsequently trained to analyze accident causes and fatal accident influencing factors. The results indicate that the LightGBM model combined with Borderline-SMOTE (synthetic sample ratio 0.12) achieves the optimal performance in predicting accident fatality. Its accuracy is 96.04%, and its macro-F₁ score is 88.19%. Among the characteristic influencing factors, accidents occurring in the flight-maneuvering phase, the flight-takeoff phase, and in-flight loss-of-control events are positively correlated with fatal accidents. The flight-landing phase is negatively correlated with fatal accidents. Pilots aged over 65 produces a mixed effect. This study, through interpretable machine learning models, reveals the key influencing factors for the fatality of trainer aircraft accidents. It provides an empirical basis for understanding the accident mechanism.

Key words: data imbalance, trainer aircraft, fatal accident, machine learning, aviation accidents, light gradient boosting machine (LightGBM) model, Shapley additive explanations (SHAP)

CLC Number:

X928.9其他

Zhang Qingfeng, Cao Yufei, Zhou Yue, Fu Chuanyun. Identification and analysis of fatal accidents of training aircraft under uneven distribution of accident types[J]. China Safety Science Journal, 2026, 36(8): 124-132.

Figures/Tables 7

Table 1

Descriptions of key variables

特征变量名称	特征的多变量分类变量及描述	二元变量数
事故等级	非死亡事故、死亡事故	2
时间-季节	春、夏、秋、冬	4
时间-时间段	早晨、上午、下午、晚上、夜晚	5
飞机-飞机类型	直升机、滑翔机、未标明	3
飞机-起火爆炸	是否发生起火爆炸	1
飞机运营-发动机额定功率/kW	[0,112)、[112,149)、[149,187)、≥187	4
飞机运营-发动机类型	往复式发动机、双发发动机	2
飞机运营-发动机制造商	莱康明、罗塔克、大陆、其他	4
飞机运营-起落架类型	后三点尾轮式、前三点、其他	3
飞机运营-适航证书	适航证书是否为正常	1
飞机运营-最大起飞质量/kg	[0,680)、[680,908)、[908,1 134)、[1 134,1 361)≥1 361	5
飞机运营-座位数	[0,2]、[3,4]、≥5	3
飞行-本场	是否为本场	1
飞行-飞行阶段	滑行道、起飞、机动、进近、着陆	5
飞行-关键事件	可控飞行撞地、发动机失效、硬着陆、起落架问题、空中失去控制、地面失控、非动力系统失效、其他	8
飞行员-安全带	两点式、三点式、四点式、五点式、未标记	5
飞行员-飞行时长/h	[0,100)、[100,200)、[200,500)、[500,1 000)、 [1 000,2 000)、[2 000,5 000)≥5 000	7
飞行员-执照类型	单发、多发、未标明	3
飞行员-飞行员资质	航空运输、商业、教员、学员、私人、其他	6
飞行员-教员评级	单发教员、多发教员、其他教员、仪表教员、未注明	5
飞行员-年龄/岁	[0,25)、[25,35)、[35,45)、[45,55)、[55,65)、≥65	6
飞行员-性别	飞行员是否为女性	1
飞行员-仪表能力	是否具有仪表飞行能力	1
飞行员-职业飞行员	飞行员是否为职业飞行员	1
飞行员-座位位置	左侧、右侧、其他	3
飞行员-副驾驶类	副驾驶飞行员变量与飞行员一致,此类变量仅在双人机组存在	31
机场-机场标高/ft	[0,500)、[500,1 000)、[1 000,1 500)≥1 500、未知	5
机场-跑道表面类型	沥青、混凝土、干跑道、其他	4
机场-跑道长度/ft	[0,2 625)、[2 625,3 938)、[3 938,5 250) ≥5 250、未知	5
气象-风速/kn	[0,4]、[5,8]、[9,12]、[13,16]、[17,20]≥20、未知	7
气象-光照条件	光照条件是否为夜晚	1
气象-温度/℃	<0、[0,8)、[8,16)、[16,24)、[24,32)≥32	6

Table 1

Table 2

Table 3

Table 4

Fig.1

Fig.2

Fig.3

References 21

[1]	National Transportation Safety Board. Aviation accident database query[EB/OL]. [2025-05-28]. https://www.ntsb.gov/Pages/AviationQueryV2.aspx.
[2]	中国民用航空局. 2013—2021年民航行业发展统计公报[R], 2022.
[3]	中国民用航空局航空安全办公室. 中国民用航空安全信息系统[EB/OL]. [2026-03-26]. https://safety.caac.gov.cn/indexnewsmore/init.act?args=353BC05F44C7451FA7C1F8BC050D8867.
[4]	靳慧斌, 朱孟昌, 马明霞. 基于ACE-BN的通勤飞行事故/事件诱因分析[J]. 中国安全科学学报, 2023, 33(2): 96-102. doi: 10.16265/j.cnki.issn1003-3033.2023.02.1109
	Jin Huibin, Zhu Mengchang, Ma Mingxia. Cause analysis of commuter flight accidents/incidents based on ACE-BN[J]. China Safety Science Journal, 2023, 33(2): 96-102. doi: 10.16265/j.cnki.issn1003-3033.2023.02.1109
[5]	Subagia R, Saleh J H, Churchwell J S, et al. Statistical learning for turboshaft helicopter accidents using logistic regression[J]. PLoS One, 2020, 15(1): DOI:10.1371/journal.pone.0227334.
[6]	Yang Ruiliang, Yang Libin, Zhao Lina. Distribution and risk factors of water-related accidents for general aviation aircraft[J]. International Journal of Aerospace Engineering, 2021, 2021(1): DOI:10.1155/2021/8812881.
[7]	Burnett R A, Dong Si. Prediction of injuries and fatalities in aviation accidents through machine learning[C]. Proceedings of the International Conference on Compute and Data Analysis, 2017: DOI:10.1145/3093241.3093288.
[8]	Zhang Xiaoge, Mahadevan S. Ensemble machine learning models for aviation incident risk prediction[J]. Decision Support Systems, 2019, 116: DOI:10.1016/j.dss.2018.10.009.
[9]	殷豪, 林淼, 王鹏, 等. 基于XGBoost的二轮车碾压事故致因研究[J]. 安全与环境工程, 2023, 30(5): 19-27, 45.
	Yin Hao, Lin Miao, Wang Peng, et al. Study on the cause of two-wheeler run-over accidents based on XGBoost[J]. Safety and Environmental Engineering, 2023, 30(5): 19-27, 45.
[10]	李嘉璐. 基于CTGAN及模型可解释性的不平衡交通事故严重程度影响因素研究[D]. 西安: 长安大学, 2023.
	Li Jialu. Analysis of factors contributing to imbalanced crash severity based on CTGAN and model interpretability[D]. Xi'an: Chang'an University, 2023.
[11]	Kabashkin I. Framework for addressing imbalanced data in aviation with federated learning[J]. Information, 2025, 16(2): DOI:10.3390/info16020147.
[12]	江玉杰, 万征, 陈继红. 基于有序选择模型的船舶事故严重度致因分析[J]. 中国安全科学学报, 2023, 33(12): 131-139. doi: 10.16265/j.cnki.issn1003-3033.2023.12.0245
	Jiang Yujie, Wan Zheng, Chen Jihong. Causation analysis of ship accident severity based on ordered choice model[J]. China Safety Science Journal, 2023, 33(12): 131-139. doi: 10.16265/j.cnki.issn1003-3033.2023.12.0245
[13]	周悦, 蒋维安, 付川云, 等. 考虑因素异质影响的直升机事故严重性分析[J]. 交通运输系统工程与信息, 2024, 24(2): 323-332.
	Zhou Yue, Jiang Weian, Fu Chuanyun, et al. Analyzing heterogeneous effects of contributing factors affecting injury severity of helicopter accidents[J]. Journal of Transportation Systems Engineering and Information Technology, 2024, 24(2): 323-332.
[14]	Lu Xinying, Zhang Doudou, Xiao Jianli. A hybrid model for traffic incident detection based on generative adversarial networks and transformer model[J]. arXiv, 2024: DOI:10.48550/arXiv.2403.01147.
[15]	马辉, 贺鹰霞, 陈杨杨. 基于XGBoost的城市污水管道缺陷发生概率预测[J]. 中国安全科学学报, 2024, 34(11): 163-171. doi: 10.16265/j.cnki.issn1003-3033.2024.11.0368
	Ma Hui, He Yingxia, Chen Yangyang. Prediction of urban sewage pipeline defect probability based on XGBoost[J]. China Safety Science Journal, 2024, 34(11): 163-171. doi: 10.16265/j.cnki.issn1003-3033.2024.11.0368
[16]	周克, 闫苗苗. 基于集成LightGBM的交通事故持续时间预测[J]. 现代计算机, 2024, 30(19): 52-55, 60.
	Zhou Ke, Yan Miaomiao. Prediction of traffic accident duration based on integrated LightGBM[J]. Modern Computer, 2024, 30(19): 52-55, 60.
[17]	王劲夫, 郭松杰, 厉林聪, 等. 基于AdaBoost学习策略的污水近红外光谱快速检测[J]. 控制工程, 2024, 31(7): 1314-1323.
	Wang Jinfu, Guo Songjie, Li Lincong, et al. Rapid detection of sewage by near-infrared spectroscopy based on AdaBoost learning strategy[J]. Control Engineering of China, 2024, 31(7): 1314-1323.
[18]	吴贤国, 刘俊, 王静怡, 等. 基于CatBoost-MOEAD的大直径泥水盾构姿态多目标预测与优化[J]. 中国安全科学学报, 2024, 34(10): 50-57. doi: 10.16265/j.cnki.issn1003-3033.2024.10.1718
	Wu Xianguo, Liu Jun, Wang Jingyi, et al. Multi-objective prediction and optimization of large-diameter slurry shield posture based on CatBoost-MOEAD[J]. China Safety Science Journal, 2024, 34(10): 50-57. doi: 10.16265/j.cnki.issn1003-3033.2024.10.1718
[19]	Vitianingsih A V, Othman Z, Kama S S, et al. Application of the synthetic over-sampling method to increase the sensitivity of algorithm classification for class imbalance in small spatial datasets[J]. International Journal of Intelligent Engineering & Systems, 2022, 15(5): DOI:10.22266/ijies2022.1031.58.
[20]	Zhou Yue, Fu Chuanyun, Wei Lin, et al. An integrated approach for addressing data imbalance in predicting fatality of helicopter accident[J]. Reliability Engineering & System Safety, 2026, 267: DOI:10.1016/j.ress.2025.111921.
[21]	郑彤. 飞行员决策能力特征模型构建研究[D]. 天津: 中国民航大学, 2023.
	Zheng Tong. Research on the construction of pilot decision-making ability characteristic model[D]. Tianjin: Civil Aviation University of China, 2023.

采样方法	方法介绍	上采样	下采样
随机过采样	随机复制少数类样本以增加数量	√	×
SMOTE	特征空间中对少数类样本插值生成新样本	√	×
自适应合成采样	根据样本分布稀疏程度生成不同数量的合成样本	√	×
边界线- SMOTE	生成位于决策边界附近的合成样本	√	×
SVM SMOTE	使用SVM确定边界样本,在周围生成合成样本	√	×
SMOTE+邻域链接	采用SMOTE生成合成少数类样本,并利用邻域链接识别并移除噪声样本与不必要多数类样本	√	√
SMOTE+ 编辑最近邻 (Edited Nearest Neighbors,ENN)	应用SMOTE后,利用ENN检查生成样本与原样本的最近邻,移除被多数类近邻包围的少数类样本	√	√

模型	特征集	宏F₁分数	准确率	召回率
SVM	13	0.790 3	0.927 4	0.790 3
	7	0.790 3	0.927 4	0.790 3
	10	0.790 0	0.930 7	0.776 7
	18	0.790 0	0.930 7	0.776 7
	20	0.790 0	0.930 7	0.776 7
DT	9	0.703 8	0.861 4	0.784 6
	34	0.697 6	0.861 4	0.769 2
	35	0.697 6	0.861 4	0.769 2
	36	0.697 6	0.861 4	0.769 2
	38	0.697 6	0.861 4	0.769 2
KNN	5	0.830 0	0.943 9	0.814 8
	7	0.810 7	0.940 6	0.782 2
	15	0.803 6	0.937 3	0.780 3
	6	0.803 6	0.937 3	0.780 3
	8	0.803 6	0.937 3	0.780 3
RF	22	0.723 1	0.891 1	0.754 8
	49	0.712 5	0.894 4	0.725 8
	51	0.710 0	0.904 3	0.700 4
	52	0.704 0	0.894 4	0.710 4
	42	0.703 0	0.887 8	0.722 1
GBM	1	0.783 6	0.927 4	0.774 9
	4	0.782 9	0.930 7	0.761 3
	7	0.782 9	0.930 7	0.761 3
	3	0.766 9	0.930 7	0.730 4
	2	0.763 9	0.920 8	0.755 8
XGBoost	8	0.751 1	0.927 4	0.713 2
	7	0.744 7	0.924 1	0.711 4
	9	0.744 7	0.924 1	0.711 4
	14	0.741 2	0.927 4	0.697 8
	12	0.734 7	0.924 1	0.695 9
LightGBM	4	0.796 4	0.940 6	0.751 3
	6	0.782 0	0.934 0	0.747 7
	2	0.772 0	0.937 3	0.718 7
	15	0.751 1	0.927 4	0.713 2
	7	0.747 6	0.920 8	0.725 0
AdaBoost	30	0.723 9	0.924 1	0.680 5
	33	0.712 1	0.924 1	0.665 1
	34	0.712 1	0.924 1	0.665 1
	52	0.711 7	0.917 5	0.676 9
	12	0.705 9	0.920 8	0.663 3
CatBoost	7	0.789 6	0.934 0	0.763 1
	6	0.766 9	0.930 7	0.730 4
	20	0.744 7	0.924 1	0.711 4
	22	0.734 7	0.924 1	0.695 9
	3	0.728 5	0.920 8	0.694 1

分类器	重采样方式	特征集	采样比例	宏F₁分数	少数类F₁分数
SVM	SMOTE	13	0.3	0.809 3	0.655 2
DT	SMOTE ENN	9	0.12	0.778 0	0.600 0
KNN	随机过采样	6	0.14	0.789 1	0.612 2
RF	SMOTE邻域链接	22	0.11	0.808 6	0.655 7
GBM	SVM SMOTE	1	0.12	0.822 9	0.678 6
XGBoost	SMOTE邻域链接	7	0.19	0.840 4	0.711 9
LightGBM	边界线SMOTE	6	0.12	0.881 9	0.785 7
AdaBoost	边界线SMOTE	12	0.12	0.840 4	0.711 9
CatBoost	SMOTE	6	0.13	0.810 7	0.653 8