Développement d’une plateforme de diagnostic non-intrusive des grands alternateurs hydroélectriques par mesures vibroacoustiques

Abstract

This thesis presents a contribution to the monitoring and diagnosis of large hydroelectric generators using an approach based on the variational autoencoder, a deep learning technique. To implement this technique, it is necessary to have a database containing signals both with and without faults. Since it is impossible to interrupt the operation of the machine and create faults in the power plant, a digital model of a large hydroelectric generator on ANSYS WORKBENCH is used to generate the frequency signatures of the faults. These signatures are then injected into real signals collected on-site. The database thus created will serve for the training and validation of the artificial intelligence model, which is the Variational Autoencoder. Real-time monitoring of electrical machines is essential for early detection of faults. To this end, two metrics have been established based on this technique to signal the presence of faults while minimizing false alarms. The use of the model’s latent space, which reduces dimensionality, has allowed for effective diagnosis. In this approach, each state of the machine is indicated by a colored cluster, thus simplifying the visualization and interpretation of the data. Additionally, the addition of a desirability term to the model’s objective function has standardized the diagnosis of faults in different machines. A major advancement of this research has been the exploration of the model’s latent space with various levels of fault severity, leading to a first monitoring metric more effective than traditional methods based on the temporal analysis of vibratory signals, even with different levels of Gaussian noise. Another significant contribution lies in the creation of a second monitoring metric, based on a probability calculation, which allows for an early alert for a predefined risk level. The exploration of the latent space included multiple levels of severity for various types of faults, thereby enhancing the model’s ability to detect various anomalies. Another notable contribution was the integration of a term based on the desirability function into the model’s objective function, making it suitable for different designs of hydroelectric machines and compatible with an online test bench. This improves the model’s ability to identify a wide range of faults. These works, conducted in collaboration with the Hydro-Quebec Research Centre (CRHQ), are based on data provided by this entity. This close cooperation facilitates the transfer of technology to the industry, thus optimizing the monitoring and diagnosis of electrical machines. It also paves the way for the adoption of alternative measurement systems within the company, benefiting from the advances of this research.

Date	17 Jun 2024
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Antoine Tahan (Supervisor), Kamal Al-Haddad (Co-supervisor) & Arezki Merkhouf (Co-supervisor)