par Wauthion, Benjamin 
Président du jury Hendrick, Patrick
Promoteur Kinnaert, Michel
Co-Promoteur Garone, Emanuele
Publication Non publié, 2023-12-21
Président du jury Hendrick, Patrick
Promoteur Kinnaert, Michel
Co-Promoteur Garone, Emanuele
Publication Non publié, 2023-12-21
Thèse de doctorat
| Résumé : | Hydraulic actuators are replaced by electromechanical actuators (EMAs) for the control of airplane flight surfaces in order to limit weight and avoid resorting to highly corrosive and polluting fluids notably. However, when used for the actuation of primary flight surfaces, the EMAs must be supervised to reach the required level of reliability. The supervision or health monitoring system aims at detecting and localizing possible degradations so that a maintenance operation can be performed before any failure arises. Similarly, health monitoring systems are needed for the EMAs used in future reusable launchers to ensure rapid decisions regarding potential maintenance before a new launch. The aim of the present work is to develop systematic methodologies to design health monitoring systems (also called fault diagnosis systems) for such applications. The focus will be on an aeronautic case study and mechanical faults, like an increase in friction and backlash, as such faults may induce jamming and have catastrophic consequences.To design and validate a health monitoring system, there is a need for data corresponding to both the healthy and the faulty operation of the device. Experimental data sets recorded on a fleet of healthy EMAs are available, but no data is available for faulty operation. Therefore, an EMA simulator is built to be able to generate synthetic data corresponding to operation in healthy and faulty modes. The simulator relies on a grey box model merging a physics-based description of the electromechanical system with a data-driven model for the friction and backlash phenomena. A grey box identification method is used to estimate the friction parameters including their confidence level for the different EMAs. Production variability is then characterized by showing that each parameter estimate can be represented by a Gaussian distribution at a given temperature. Next thermal effects are accounted for through a Gaussian process characterizing parameter variability with respect to temperature. This modeling framework is the cornerstone of the thesis: on the one hand, it is the basis for the above-mentioned simulator. On the other hand, the model is partially or fully exploited in the different model-based health monitoring methods.The developed health monitoring methods are based on three specific scenarios corresponding to three moments in the life of the actuator. The first one corresponds to the in-flight phase of the vehicle. As an EMA only moves during specific phases of flight, the data subsets containing relevant information for the identification of at least some parameters are determined by using tools from sensitivity analysis. The resulting data analysis is used to govern the parameter estimation of a Dual Extended Kalman Filter (DEKF) in such a way that all informative data are exploited at best for parameter estimation. The resulting passive monitoring process then consists of comparing the estimated parameters to the healthy ones while accounting for their confidence bounds. The second health monitoring method exploits the opportunity to perform test bench tests on the EMAs. Exploiting the analytical redundancy within the dynamical model of the EMA, a residual generator is designed to be sensitive to a potential friction change or backlash increase, triggered by specifically designed measurement sequences. A classifier based on a multiclass Support Vector Machine approach is used to detect and identify the nature of the faults. The classifier validity is corroborated by a K-fold cross-validation. The third monitoring process makes use of the taxi phase of the vehicle. During this phase, the EMA is asked to follow a dedicated excitation profile that maximizes the sensitivity of the outputs regarding the parameters of interest while respecting the mechanical constraints of the EMA. Then, using a combination of a Dual Unscented Kalman Filter and an EKF, the EMA parameters are estimated and compared to the healthy one. Given the optimized input profile this last approach is an active fault diagnosis method. |
