Optimization strategy for ensemble learning models based on fusing resampling, adaptive dimensionality reduction, and Optuna in intelligent flight technology evaluation

The flight techniques of pilots directly impact flight safety and constitute the primary factors influencing aviation accidents. However, existing research is limited by several challenges, including overly subjective evaluation criteria, the inability to implement large-scale data-driven approaches, and suboptimal outcomes. In this paper, a novel dynamic assessment framework for flight technique is proposed based on QAR data. The framework employs multiple ensemble learning models as baselines, fusing a three-stage optimization strategy. First, resampling techniques are introduced to address data imbalance. Second, a Stepwise Feature Selection-Adaptive Dimensionality Reduction (SFS-ADR) method is developed to efficiently reduce flight parameter dimensionality. Finally, Optuna hyperparameter optimization is adopted to further enhance model precision. Experiments on four public datasets demonstrate that the optimization strategy achieves accuracy improvements of 8.45%, 7.51%, 7.37%, and 7.22% respectively, with consistent superiority over baseline models. Model interpretability is further validated through SHapley Additive exPlanation (SHAP) method. The framework provides a high-precision dynamic solution for flight technique evaluation, offering valuable insights for Human Factors (HF) research in aviation safety.