Détection et diagnostic des défauts pour les machines électriques multi-phasés

Compared to conventional three-phase drives, multi-phase machines have several advantages. However, these advantages come with increased system cost, due to the need for additional current sensors, gate drive circuits, and an appropriate power supply. Among the various multi-phase machine configurations, seven-phase machines have attracted increasing attention. Thus, this thesis focuses on developing robust Fault Detection and Diagnosis (FDD) methodologies for seven- phase permanent magnet synchronous machines. Regardless of the fault type and its evolution, faults directly or indirectly affect the current flowing in the windings, causing changes in three parameters: gain, phase shift, and mean value change. To capture these effects, this thesis proposes three analytical models of the currents flowing in the windings in three reference frames (natural, stationary, and synchronous) and four fictitious machines (Principal Machine, PM ; Secondary Machine, SM ; Tertiary Machine, TM ; and Homopolar Machine, HM), with a particular focus on the analytical models in the stationary reference frame. After the models are validated, fault detection techniques are first explored in the time and frequency domains, focusing on the trajectories of the current vectors in the stationary reference frame in healthy and faulty modes. Then, for fault isolation, the simulation results show that in the presence of low noise levels, the multi-step classification methodology allows for low noise levels to effectively separate severe faults from incipient ones and to discriminate them. For multi-parameter changes in one or more faulty phases, the advanced t-SNE clustering method is used, showing superior clustering performance when combined with Mahalanobis distance to correct for feature correlations. The last task, fault estimation, focuses on quantifying fault severity using entropy- and sampling-based approaches, adopting univariate and multivariate perspectives. Jensen-Shannon divergence (JSD) and Wasserstein distance (WD) are used to evaluate fault estimation performance under different noise and fault severity conditions. Finally, the slope of the decision function from the CUSUM algorithm is introduced as a fast fault assessment metric, showing particular effectiveness in scenarios with Gaussian-distributed data.