Supervision : Bertrand GRANADO,Christophe MARSALA

Co-supervision : DENOULET Julien, NIKOLAJEVIC Konstanca

Flight loads recognition using machine learning in order to increase flight safety and improve maintenance predictive service

In the aviation industry, aircraft maintenance is crucial, it must be done at the right time. Today, in the case of helicopter maintenance and component life monitoring, a severe usage assumption is associated with typical flight case measurements, defining a flight spectrum. This spectrum and the corresponding maintenance planning are defined using several parameters, as the flight parameters describing the flight context, and the flight loads, acting on the mechanical components, that are measured over flight tests performed during the development of the helicopter. To ensure the safety of the aircraft, the flight spectrum must cover the most severe usages of the helicopter, which leads to both unnecessary conservatism, that changes components even when they are still operational, and cumbersome use.
This thesis aims at introducing a predictive maintenance approach, based on the real usage of each helicopter, where flight loads virtual sensors producing a reliable estimation of the helicopter flight loads could help to derive the damage of the helicopter components and therefore to determine when to replace them. In other words, direct recognition of loads in flight would allow adjustment of maintenance phases according to the helicopter use.
The ambition is to achieve a maintenance process more closely based on the reality of the aircraft usage and therefore to reduce the conservatism of the method currently used.
This thesis takes place in a multidisciplinary context, combining signal processing, Machine Learning (ML) and helicopter expertise. As part of this thesis work, a Direct Load Recognition (DLR) methodology is built for helicopters load estimation using signal decomposition and ML algorithms to build prediction models from prototype flight test data points. The methodology is tested and evaluated on prototype flight tests to estimate major flight loads of the main rotor, enabling them to compute the damage of several components of the helicopter main rotor. This first approach shows interesting results, especially for the estimation of the Mast bending moment. The application of a load having very dynamic evolution, reveals the limits of the built methodology, showing interest to improve the methodology. Several methods of improvement are investigated by focusing on some parts of the methodology: the building of the dataset, the method of decomposition, the building of the ML model. This investigation shows two main conclusions. First, some changes in the DLR methodology improved the load estimation. Second, it shows the necessity to adapt the methodology to the studied load.
Finally, load estimators are built to estimate three main rotor loads of two helicopters, and are applied to operational (customer) flights. The derived damage estimations are compared to the current spectrum, the estimated damages being well below the current spectrum. That shows the real interest of considering the helicopter usage by implying potential gains in lifetime and maintenance tasks. This method may be employed for predictive maintenance and would reduce the conservatism. Additionally, for customers, these advancements will improve flight safety, while potentially optimizing component inspection intervals.

2021-2024 Publications

• 2024
  • C. Del cistia Gallimard, K. Nikolajevic, F. Beroul, J. Denoulet, B. Granado, Ch. Marsala : “Direct Load Recognition Application to Main Rotor Pitch-Link Load on the H175 Fleet: A New Wavelet Approach”, Vertical Flight Society, Montréal (Québec), Canada (2024)
• 2023
  • C. Del cistia Gallimard, K. Nikolajevic, F. Beroul, J. Denoulet, B. Granado, Ch. Marsala : “Direct Load Recognition to Estimate the Damper Load on the H175 Fleet”, European Rotorcraft Forum, Bückeburg, Germany (2023)
  • J. Jouve, C. Del cistia Gallimard, K. Nikolajevic, H. Morel : “Estimation of confidence margins for Direct Load Recognition (DLR) using supervised and unsupervised machine learning”, Vertical Flight Society, West Palm Beach, Florida, United States (2023)
• 2022
  • C. Del cistia Gallimard, F. Beroul, J. Denoulet, K. Nikolajevic, A. Pinna, B. Granado, Ch. Marsala : “Harmonic Decomposition to Estimate Periodic Signals using Machine Learning Algorithms: Application to Helicopter Loads”, IEEE World Congress on Computational Intelligence, Padua, Italy (2022)
• 2021
  • C. Del cistia Gallimard, K. Nikolajevic, J. Jouve, F. Beroul : “Direct Load Recognition and Damage Estimation using Supervised Learning”, Vertical Flight Society, Online, United States (2021)