Conception d’un modèle de reconstruction de données de vol fondé sur des approches avancées d’intelligence artificielle

Abstract

Detecting anomalies in aircraft sensor data is critical for ensuring aircraft safety and reliability during flight test campaigns. However, the wide range of flight maneuvers and variability across different aircraft types presents significant challenges to the design of robust predictive models. The dynamic nature of flight conditions leads to non-stationary data distributions, thus complicating traditional modeling approaches. In this work, we propose a novel framework that combines meta-learning with parameter-efficient fine tuning to address these challenges. Our approach leverages TSMixer, a state-of-the-art Deep Learning architecture for time-series modeling, to reconstruct the normal behavior of aircraft sensors and identifies deviations indicative of anomalies. Additionally, the performance of a State Space Model, specifically Mamba is evaluated to assess its effectiveness in reconstructing physical quantities such as forces. We demonstrate the applicability of our framework by use of two cases : anomaly detection in aircraft dynamics using TSMixer, and anomaly detection in force measurements using Mamba. To enhance model generalization across diverse flight conditions, we employ Model- Agnostic Meta-Learning (MAML), which enables the model to adapt rapidly using limited data. Furthermore, the Low-Rank Adaptation (LoRA) is integrated in the model, that is a lightweight and efficient fine-tuning technique, to transfer knowledge across different maneuver types or aircraft platforms. Experimental results show that this dual adaptation strategy not only improves the model reconstruction accuracy but it also significantly reduces the training time required when adapting to new flight conditions. Overall, our framework provides a scalable, data-efficient, and high-performing solution for anomaly detection in complex and heterogeneous aeronautical environments, pushing forward the state of the art in aircraft sensor data analysis.

Date	18 Nov 2025
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Georges Ghazi (Supervisor) & Ruxandra Botez (Co-supervisor)