Recognition of Plasma Discharge Patterns Based on CNN and Visible Images

Abstract

Electrical discharge is a sign of insulation damage of power equipment and an essential part of power inspection. For the recognition of plasma discharge patterns, the accuracy rate of traditional manual inspection relies heavily on the experience of the staff, and misjudgment is inevitable. The emerging method based on manual feature extraction of discharge signals also has its limitations. Thus, a high-accurate recognition method based on convolutional neural network (CNN) and visible images is proposed in this paper. Under the deep learning framework-TensorFlow, CNN model is established and pre-trained weights are loaded through transfer learning. The visible image dataset used for training and evaluation includes four discharge patterns: arc discharge, corona discharge, creeping discharge and plasma jet. In data processing, data amplification and data enhancement are used to expand the sample. Besides, typical network models with fully connected layers are adopted and compared. Three optimization methods of ReLU, BN and Dropout are applied to solve the overfitting problem, and the corresponding effects are discussed in the model of VGG16 and ResNet152. The test accuracy of the optimized VGG16 reaches 99.6% without overfitting, which indicates the promising potential in the recognition of plasma discharge patterns in power inspection.