基于极坐标化与改进深度学习的GIS局部放电类型识别OA

As an important part of online monitoring and fault diagnosis for high-voltage GIS equipment,the accuracy of partial discharge type recognition is crucial to ensure safe operation of the equipment.In order to eliminate the distortion of phase resolved partial discharge(PRPD)map features caused by phase deviation and the problems of low recognition accuracy and high missed detection rate of traditional neural networks,this paper proposes a GIS partial discharge type recognition method based on polar coordinate transformation and improved deep learning.Firstly,a sample dataset is constructed,and based on this,the PRPD map of the sample dataset is transformed into polar coordinates to correct feature distortion,and data augmentation is performed through multi angle rotation.Secondly,using ResNet50 as the backbone network,a rotation prediction branch is introduced to enhance the expression ability of convolutional features for directional changes,achieving rotation invariance of the network model.Furthermore,the Fisher discriminant regularization term is introduced to improve the prediction accuracy of the network model for various types of graphs and reduce the missed detection rate by utilizing its characteristics of intra class aggregation and inter class dispersion.Finally,the model is trained and validated using experimental data and its effectiveness is further verified by combining it with on-site measured data.The experimental results show that the model has a high accuracy of 0.98 in type recognition on the experimental dataset and a recognition accuracy of 0.91 on the measured dataset.The missed detection rate of partial discharge signals mistakenly identified as external interference signals is only 0.012,demonstrating good robustness.