PD Pattern Recognition for Generator Stator Bar: A Data-driven Learning Approach.

Generators are one of the most critical electrical equipment in power systems, which are vulnerable to faults due to irreversible insulation degradation of their stator windings. Hence, accurately evaluating the stator insulation condition is of great importance for ensuring the safe operation of generators. Partial discharge (PD) testing has long been used as a quality control test for stator winding insulation, which can indicate defects in the insulation of generator stator windings based on the PD pattern recognition results. In this paper, a novel data-driven learning approach is developed for accurately recognizing PD patterns. Firstly, a Multi-dimensional feature vector is derived from the collected discharge spectrograms to distinguish the heterogeneity of different discharge signals. Then, the Genetic algorithm (GA) is adopted to construct the balanced training sample space to eliminate the influence of unbalanced sample space on the recognition accuracy. Finally, an advanced Radial Basis Function (RBF) network is trained for linking the mapping relationship between the different feature vectors and PD Patterns. To demonstrate the accuracy and usefulness of the proposed approach, some generator model bars were made and typical PD phenomena are simulated to collect the testing data. The results show that the proposed data-driven learning approach has a promising capability for PD Pattern Recognition and bears great practicality in the real world.