• Citer

Design of model-based fault diagnosis systems from concatenated data ─ application to wind turbines

par Vasquez Rodriguez, Sandra Paola
Président du jury Hendrick, Patrick
Promoteur Kinnaert, Michel ;Pintelon, Rik
Publication Non publié, 2021-09-08

Thèse de doctorat

• Accès en ligne
• Détails
• Contenu
• Statistiques

Résumé :

The blade pitch system is a critical subsystem of Variable-Speed Variable-Pitch wind turbines that is characterized by a high failure rate. It has an essential function on the power regulation, mitigation of operational loads, and the aerodynamic braking of wind turbines. To contribute to a higher reliability and establishment of predictive maintenance, this thesis presents a model-based Fault Diagnosis (FD) system for blade pitch systems with hydraulic actuators. The proposed model-based FD system addresses incipient multiplicative faults. To account for the normal variability of the hydraulic pitch actuators dynamics and the non-ideal quality (for FD purposes) of their excitation signal, we propose a parameter estimation approach based on Linear Time-Invariant (LTI) and Linear Parameter-Varying models identified from concatenated input/output data, and complementary passive and active FD strategies.The FD method relies on frequency-domain estimators for the identification of continuous-time models in a user defined frequency band. The frequency selection can notably simplify the complexity of the modeling problem, and continuous-time models facilitate the design of the FD algorithms. Moreover, the estimation of confidence bounds on the parameters used for FD is included in order to ensure robustness with respect to noise in measurements and stochastic nonlinear distortions. The proposed FD system is thoroughly validated on a wind turbine simulator.The research on the data concatenation technique derived on an important outcome of this work. Being data concatenation a well-known technique for identifying LTI models from multiple records, the study of the asymptotic properties of the estimator is limited. Therefore, we investigated consistency and asymptotic normality as the number of records tend to infinity, with focus on the identification of parametric models.