Hybrid Machine Learning for Imbalanced Landing-Phase Safety Event Prediction from Flight Data

Taha Asghar; Bilal Mufti; Syed Qasim Zaheer; Luqman Ahmed

cs.LGNov 5, 2026

Hybrid Machine Learning for Imbalanced Landing-Phase Safety Event Prediction from Flight Data

Authors: Taha Asghar, Bilal Mufti, Syed Qasim Zaheer, Luqman Ahmed

Organizations: College of Aeronautical Engineering, National University of Sciences and Technology (NUST), Risalpur, Pakistan

Abstract

Landing-phase safety-event prediction requires the detection of rare abnormal approaches within strongly imbalanced multivariate flight data. Capturing coupled temporal behavior while distinguishing event-specific signatures across airport operating contexts remains a central modeling challenge. This study proposes an interpretable hybrid feature-fusion framework for multiclass landing-phase safety-event prediction. A long short-term memory (LSTM) autoencoder trained on nominal flight windows characterizes deviations from routine approach behavior. Its continuous reconstruction-error descriptor is fused with parameter-specific statistical flight-history features and supplied to a class-weighted XGBoost classifier to distinguish Normal, Speed High, Path High, and Flaps Late windows. The framework is evaluated using 53,750 NASA DASHlink approach windows from three arrival airports. Pooled and airport-specific learning strategies are compared alongside component-level benchmarks, repeated pipeline retraining, and feature-attribution analysis. The results demonstrate strong class-balanced detection, with mean macro recall above 0.90 across all airports under repeated partitioning and retraining. Component comparisons establish the importance of class weighting and show that continuous feature fusion outperforms reconstruction-error-only and hard-gate alternatives. Airport-level comparisons reveal benefits from local specialization at the two airports with larger datasets and from pooled training at the airport with the smallest dataset. Feature attributions associate predictions with operationally relevant vertical-path, configuration, speed, and approach-context variables. The framework provides an interpretable approach to rare-event detection and practical guidance for adapting flight-data monitoring models to airport-level data availability.

Figures & tables

Figure 1: Proposed hybrid feature-fusion framework for landing-phase safety-event prediction. Statistical flight-history descriptors and an LSTM-autoencoder reconstruction-error descriptor are extracted from the same multivariate flight window, concatenated, and supplied to a class-weighted XGBoost classifier for multiclass prediction.

Descriptor	Purpose
Mean	Average parameter level
Standard deviation	Overall variability
Maximum	Upper extreme
Minimum	Lower extreme
Range	Peak-to-peak variation
Linear slope	Direction and rate of temporal change

Table 1: Statistical descriptors extracted from each flight-parameter history.

Airport	Normal	Speed High	Path High	Flaps Late	Total
KMSP	23,758	1,267	272	99	25,396
KDTW	17,037	1,454	157	149	18,797
KMEM	8,374	968	61	154	9,557
Pooled	49,169	3,689	490	402	53,750

Table 2: Arrival-airport-only dataset composition used in the pooled and airport-specific experiments.

Figure 2: Arrival-airport-only dataset composition for KMSP, KDTW, and KMEM: (a) total class composition at each airport and (b) minority-event support shown using a logarithmic vertical scale.

No.	Recorded parameter	Unit/type	No.	Recorded parameter	Unit/type
1	Left aileron position	deg	11	Selected heading	deg
2	Right aileron position	deg	12	Localizer deviation	%
3	Corrected angle of attack	deg	13	Average core speed	%
4	Baro-corrected altitude	ft	14	Total pressure	mbar
5	Computed airspeed	kt	15	Pitch angle	deg
6	Selected course	deg	16	Roll angle	deg

Table 3: Flight parameters retained in each landing-phase window.

Figure 3: Class-specific landing-phase signatures at KMSP: (a) computed airspeed for Normal and Speed High windows, (b) glideslope deviation for Normal and Path High windows, and (c) trailing-edge flap position for Normal and Flaps Late windows. Curves show the class-wise median and shaded regions indicate the interquartile range. Time is measured relative to touchdown.

Figure 4: Illustrative multivariate landing-phase histories for sampled KDTW windows: (a) Normal flight and (b) Flaps Late flight. The four rows show altitude, computed airspeed, trailing-edge flap position, and vertical speed over the 160-s interval preceding touchdown. Vertical speed is obtained by numerically differentiating the altitude history with respect to time and is shown for illustration only; it is not a model input.

LSTM Autoencoder		XGBoost Classifier
Hyperparameter	Value	Hyperparameter	Value
Encoder LSTM units	64	Objective	multi:softprob
Latent dimension	32	Boosting iterations	1,000
Dropout rate	0.2	Maximum depth	6
$L_{2}$ regularization	0.001	Learning rate	0.1
Optimizer	Adam	Row subsampling	0.8

Table 4: Hyperparameter configurations of the LSTM autoencoder and class-weighted XGBoost classifier used in the proposed framework.

Dataset	Split	Total	Normal	Speed High	Path High	Flaps Late
KMSP	Total	25396	23758	1267	272	99
	Train	17776	16630	886	191	69
	Validation	2540	2376	127	27	10
	Test	5080	4752	254	54	20
KDTW	Total	18797	17037	1454	157	149
	Train	13157	11925	1018	110	104

Table 5: Training, validation, and test partitions for the three arrival airports and the combined pooled dataset. Class-wise counts document the support available for each safety-event class.

Model	Feature input / decision logic	Purpose
M1	Statistical features + unweighted XGBoost	Baseline classifier
M2	Statistical features + weighted XGBoost	Effect of class weighting
M3	AE reconstruction error + weighted XGBoost	AE-error-only baseline
M4	Statistical features + AE reconstruction error + weighted XGBoost	Proposed hybrid fusion
M5	Hard AE gate + statistical-feature event classifier	Hard-threshold baseline

Table 6: Benchmark and ablation configurations used to assess the contributions of class weighting, autoencoder-based temporal deviation, and hybrid feature fusion.

Airport	Test	$\mathbf{Acc.}_{\mathrm{spec}}$	$\mathbf{Acc.}_{\mathrm{global}}$	$\mathbf{MR}_{\mathrm{spec}}$	$\mathbf{MR}_{\mathrm{global}}$	$\boldsymbol{1-\mathrm{MR}}_{\mathrm{spec/global}}$
KMSP	5080	0.9856	0.9852	0.9572	0.9474 (-1.02%)	0.0428 / 0.0526
KDTW	3760	0.9779	0.9771	0.8950	0.8886 (-0.72%)	0.1050 / 0.1114
KMEM	1912	0.9702	0.9754	0.8599	0.8818 (+2.55%)	0.1401 / 0.1182

Table 7: Comparison of pooled global and airport-specific specialist models under arrival-airport-only evaluation. Parenthetical percentages indicate the relative change in global macro recall compared with the specialist model.

Figure 5: Row-normalized confusion matrices for the airport-specific specialist models under arrival-airport-only evaluation. Each cell reports the fraction of samples in the true class assigned to the predicted class, with the corresponding raw count shown in parentheses.

Figure 6: Row-normalized confusion matrices for the pooled global model evaluated on the same airport-specific arrival-only test sets. Each cell reports the fraction of samples in the true class assigned to the predicted class, with the corresponding raw count shown in parentheses.

Airport	Macro Recall	Macro F1
KMSP	$0.9313\pm 0.0142$	$0.9347\pm 0.0062$
KDTW	$0.9303\pm 0.0190$	$0.9229\pm 0.0129$
KMEM	$0.9022\pm 0.0212$	$0.9010\pm 0.0242$

Table 8: Performance of the proposed airport-specific hybrid-fusion models across five random data splits and retraining runs. Values are reported as mean

\pm

standard deviation.

Figure 7: Macro-recall variability of the proposed airport-specific hybrid-fusion models across five random data splits and retraining runs. Individual points represent the five runs for each arrival airport, and diamonds denote their means.

Airport	Model	Accuracy	Macro Recall	Macro F1	$\mathbf{1-\mathrm{MR}}$	Event-Class Recall Avg.
KMSP	M1	0.9852	0.8939	0.9249	0.1061	0.8601
	M2	0.9837	0.9529	0.9389	0.0471	0.9404
	M3	0.6364	0.4815	0.2883	0.5185	0.4264
	M4	0.9848	0.9542	0.9400	0.0458	0.9417
	M5	0.9000	0.6186	0.6517	0.3814	0.5103
KDTW	M1	0.9766	0.8808	0.9121	0.1192	0.8437

Table 9: Benchmarking comparison of model variants under the arrival-airport-only specialist setting. The event-class recall average assigns equal weight to Speed High, Path High, and Flaps Late, excluding Normal.

Figure 8: Macro-recall comparison of the benchmark model variants across the three arrival airports. Model definitions are given in Table 6 .

Figure 9: Class-wise recall comparison between the weighted statistical-feature model (M2) and the proposed hybrid-fusion model (M4).

Figure 10: Top five parameter-level SHAP drivers for each class and arrival airport, with reconstruction error retained as a separate feature. Values are normalized within each airport–class panel by the largest aggregated mean absolute SHAP value in that panel, emphasizing relative importance within each panel.

Aviation Safety

See all 23 papers

Can Large Language Models Explain Flight Safety Events? A Prior-Guided Semantic LLM-based Approach
Aug 18, 2026Lu Xu, Xu Li, Linjiang Zheng +3Aviation Safety LLM Interpretability
College of Computer Science, Chongqing University, Chongqing, China · School of Statistics and Data Science, Shanghai University of International Business and Economics, Shanghai, China
Data-Driven Runway and Taxiway Exits Prediction of Landing Aircraft: A Case Study at Hartsfield-Jackson Atlanta International Airport
Jun 9, 2026Alex Porcayo, Yutian Pang, Maria Thomas +1Air Traffic Management Trajectory Prediction
Department of Aerospace Engineering and Engineering Mechanics, The University of Texas at Austin, Austin, 78712, TX, USA
Towards Automated Air Traffic Safety Assessment Around Non-Towered Airports Using Large Language Models
May 12, 2026Torsten Darrell, Mahyar Ghazanfari, Jordan Kam +3Air Traffic Management Aviation Safety
Department of Mechanical and Aerospace Engineering, George Washington University · Aerospace Program, University of California, Berkeley · Department of Electrical Engineering and Computer Science, University of California, Berkeley +1

Aug 18, 2026J/K move · Enter open · S save

Can Large Language Models Explain Flight Safety Events? A Prior-Guided Semantic LLM-based Approach

Lu Xu, Xu Li, Linjiang Zheng, Fan Li+2

College of Computer Science, Chongqing University, Chongqing, China · School of Statistics and Data Science, Shanghai University of International Business and Economics, Shanghai, China

Aviation Safety LLM Interpretability Anomaly Detection

Open paper AskPDF arXiv

Improving flight safety with flight data requires not only accurate detection of risk events, but more importantly, clear interpretation of their underlying causes at the level of pilot control behavior. Existing explainable AI techniques, such as feature importance maps, often require considerable domain knowledge to translate them into operationally meaningful explanations. Large Language Models (LLMs), which excel at language reasoning, bring a promising solution to this issue. However, applying LLMs in this domain presents key challenges such as modal inconsistency, limited classification ability, scarcity of task-specific data for fine-tuning, and lack of domain knowledge. To overcome these challenges, we propose FlightLLM, a prior-guided semantic LLM-based approach for interpretable flight safety analysis. Specifically, we first perform feature engineering to address modal inconsistency, combining statistical descriptors with physically meaningful flight indicators. This representation is further processed by a Semantic Discretization module, which converts abstract numerical patterns into qualitative descriptions that are more compatible with language reasoning. In addition, since LLMs are not inherently strong classifiers, CatBoost is incorporated as a statistical expert, and its prediction results are injected into the prompt as prior guidance. A contrastive few-shot learning strategy is further adopted to compensate for limited data. Finally, we design structured prompts to embed aviation-specific knowledge into the inference process. Using hard landing, a representative risk event with complex causal mechanisms, as an anchor point, we evaluate FlightLLM on a dataset of 704 real-world A320 flight samples. Experimental results show that the proposed approach achieves competitive classification performance while generating direct and reasonable explanations for event causes.

Abstract

Aviation Safety

Explore similar work

Can Large Language Models Explain Flight Safety Events? A Prior-Guided Semantic LLM-based Approach

Data-Driven Runway and Taxiway Exits Prediction of Landing Aircraft: A Case Study at Hartsfield-Jackson Atlanta International Airport

Towards Automated Air Traffic Safety Assessment Around Non-Towered Airports Using Large Language Models