Identification des segments de trim d’un aéronef à l’aide d’un modèle XGBoost entraîné sur des données de vol

Abstract

This research aims to develop a supervised learning method capable of automatically detecting equilibrium phases of an aircraft from large flight data sets. An equilibrium phase, also known as a trim segment, refers to a portion of the flight during which the aircraft is in a stable state and can maintain its trajectory without pilot input. These phases are particularly relevant for stability analysis, aircraft behavior modeling, and supporting flight test campaigns. To achieve this, we implemented a complete pipeline based on real data from 2,147 flight tests. This process includes, among other steps, the selection and aggregation of relevant parameters, the creation of an annotated dataset, and the optimization of a XGBoost model’s hyperparameters, evaluated via cross-validation. Special attention was given to preprocessing quality, including data cleaning, temporal feature engineering, and handling class imbalance, in order to maximize model robustness. The results confirm the effectiveness of our approach : on an independent test set, the classifier achieves an accuracy of 90%, a recall of 91%, an F1-score of 0.91, and an area under the ROC curve of 0.96. These performance metrics reflect a strong ability to detect trim phases while minimizing both false positives and false negatives. Beyond these results, our study provides a reproducible labeling protocol, an annotated database, and a thorough analysis of limitations related to sensor noise and class imbalance. A complementary study, based on the IsolationForest algorithm, also enables the detection of potential anomalies within the detected segments, adding an additional layer of reliability to the method. These contributions pave the way for practical applications such as the automation of flight test analysis, the improvement of control systems, or support for aircraft certification. They also provide a solid methodological foundation for future work incorporating temporal dynamics, deep learning, or adaptation to other types of aircraft with varied architectures and behaviors.

Date	4 Sept 2025
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Georges Ghazi (Supervisor) & Ruxandra Botez (Co-supervisor)