THz Wave Detection of Gap Defects Based on a Convolutional Neural Network Improved by a Residual Shrinkage Network

Internal air gap is a serious type of defect in the insulation equipment, which threatens the safe operation of the power grid. In order to diagnose the position and thickness of the internal air gap, this paper proposes a terahertz wave detection method based on wavelet analysis and a CNN (convolution neural network) model. According to the time-frequency characteristics of the wavelet cluster, the calculation method of air gap depth is proposed. To determine the thickness of the internal air gap, the performances of several classification methods, such as waveform feature analysis, Bayes, MLP (Multi-layer Perceptron), SVM (Support Vector Machine) and CNN are compared. The results show that the CNN modified by a residual shrinkage network and SVM (CNN-RSN-SVM) has the best performance. By adjusting the parameters, the classification accuracy of the CNN-RSN-SVM model can be elevated to 98.91%. Furthermore, the 3D imaging method of air gap defect based on wavelet analysis and CNN-RSN-SVM classification model is formed.