A Pattern Recognition Method for Partial Discharge of Generator Stator Bars Under the Condition of High Similarity Samples Based on DenseNet

Existing deep learning-based methodologies for pattern recognition of partial discharge of generators suffer from underutilization of network parameters, high training resource consumption, and potential issues of gradient vanishing or exploding. To address these challenges, a pattern recognition method based on DenseNet for partial discharge of stator bar is proposed. Six types of generator stator bars with different defects were prepared, and a high-voltage testing platform was set up. Partial discharge signals were detected using high-voltage coupling capacitor, and phase-resolved partial discharge (PRPD) patterns under different defects were plotted. DenseNet was employed for pattern recognition of PRPD from stator bars, with adjustments made to network parameters to optimize recognition performance. Finally, the effectiveness of this method was validated by comparing recognition results with those from other models. The results indicate that under the experimental data, DenseNet's accuracy in pattern recognition of partial discharge of the stator bars is 9.92%, 4.25%, and 3.80% higher than AlexNet, GoogLeNet, and VGG-19, respectively, demonstrating its effectiveness in partial discharge pattern recognition for generators.