By studies conducted by CIGRE in 2009, it was found that the main source of hydro generator failures is correlated to the machine electrical insulation. Due to this fact, monitoring the stator winding conditions became an important supervising procedure. A very used practice to accomplish this supervision is through the measurement and analysis of partial discharges (PDs), being this practice one of the most effective and secure methods for analysis of generator stator insulation. However, although PDs have well-defined standards, it is not trivial to classify the obtained PDs signals in these patterns, mainly due to the large number and variety of PDs occurrences. Today, the significant increase in the amount of PDs data available was due to improvements in equipment and software for PDs monitoring, as for example the system IMA-DP, which has contributed to better planned and more frequent measurement campaigns. So, this work proposes the use of an intelligent tool to facilitate the process of identification and diagnosis of partial discharges, based on data mining techniques using decision trees (DT), which is a solution for analyzing large amount of data. In the specific case presented in this dissertation it was used 2,435 measurements obtained for phase A of a hydro generator of the Tucurui Hydro Plant, which was essential to validate the proposed method, because they represent real data obtained from the Hydro Plant operation. A hybrid approach (supervised/unsupervised) was used to identify and rank PDs patterns among the well-known forms of DPs. A fast and very satisfactory PDs classification procedure was achieved, especially when converting data from statistical maps into amplitude histograms, thus, obtaining well-defined clusters and a created decision tree that achieved global indices of accuracy above 98%.
|
