Securing aviation networks: A generalized and robust machine learning solution for monitoring abnormalities in commercial aircraft

• Comprehensive dataset utilization: Analyzed 9,949,927 flight records, significantly larger than datasets in prior studies, ensuring real-world applicability and broad representation of flight patterns. • High predictive accuracy without data reduction: Achieved 93.63% accuracy while maintaining dataset integrity, avoiding the pitfalls of aggressive filtering or subsampling. • Scalable ML framework for large datasets: Optimized machine learning training strategies for massive aviation datasets, demonstrating effective handling of computational challenges. • Robust feature selection: Identified key predictive features (timestep, baroaltitude, velocity, longitude, and latitude) using Gradient Boosting Regression and Bonferroni correction, enhancing model interpretability. • Handling imbalanced data effectively: Employed SMOTEENN to balance the dataset, ensuring fair representation of rare events and improving predictive reliability. • Interpretability and trustworthiness: Validated model decisions using SHAP and LIME, confirming logical and explainable predictions based on selected features. • Analysis of dataset size vs. performance trade-offs: Investigated why smaller datasets report higher accuracy (up to 99%) and discussed the implications for generalizability and real-world deployment. • Advancing aviation safety: Provides a robust, scalable, and generalizable framework for flight anomaly detection, supporting safer operational practices and enhanced predictive maintenance. • Practical guidance for industry adoption: Offers insights into large-scale flight data handling, helping aviation authorities and airlines implement data-driven monitoring solutions effectively. Accurate abnormality prediction in commercial aircraft is critical for aviation safety, yet existing approaches often compromise reliability due to limited datasets or suboptimal feature selection. While several machine learning models have been proposed, most suffer from small sample sizes or inadequate handling of imbalanced data, limiting real-world applicability. This study addresses these gaps by developing an XGBoost-based predictive model trained on 9,949,927 flight samples, significantly larger than prior research. Our methodology incorporates SMOTEENN for data balancing and employs rigorous feature selection using Gradient Boosting Regressor with Bonferroni correction, identifying five key predictors: timestep, baroaltitude, velocity, longitude, and latitude. The model achieved robust performance with an accuracy of 0.9363 (95% CI: 0.9360–0.9366), precision of 0.9222, recall of 0.9719, PR-AUC of 0.9755, Balanced Accuracy of 0.9295, Log Loss of 0.1902, and Brier Score of 0.0516, indicating scalability, generalizability, and robustness, thereby outperforming existing approaches. Nested cross-validation confirmed robustness with accuracies of 0.927 ± 0.010 and 0.934 ± 0.008 for 5-fold and 10-fold schemes. SHAP and LIME analyses validated the decision logic, showing predictions driven by meaningful and logically consistent feature patterns. These results demonstrate that comprehensive data utilization, proper feature selection, and interpretability yield more reliable abnormality detection compared to models trained on filtered datasets. The study provides aviation safety systems with a scalable, transparent framework for large-scale anomaly monitoring and establishes a practical benchmark for future research and deployment of data-driven aircraft monitoring systems.