Constellation, le dépôt institutionnel de l'Université du Québec à Chicoutimi

Résumé

Failures in power transformers generates high financial loss inherent to equipment cost and power interruption. Thus, the industry has been working into developing efficient and reliable monitoring techniques with the objective of detecting incipient faults. Frequency Response Analysis (FRA) is among well-known methods to detect faults in transformers’ active parts. FRA is very sensitive to changes in the winding geometry and a comparison between healthy and faulty measurements can present deviations that leads to the fault identification. However, until now, FRA interpretation is performed with the aid of experts. The development of FRA and the spreading of its application in transformers’ diagnostics, have led studies to focus on the advancement of objective interpretation techniques. To contribute to this matter, intelligent classifiers are evaluated over its capabilities to classify transformers faults. For this purpose, a database of measurements performed on a laboratory winding model including different faults is used. Numerical indices are used as input for machine learning classifiers. This analysis indicated that the use of one individual index is appropriate for a good classification performance using Radial Basis Function neural network. Meanwhile Support-vector machine and Backpropagation neural network performed better with a combination of indices.