Abstract：Over the years，machine learning has made some breakthroughs in the insulation defects of gas

insulated switchgear（GIS），but the traditional methods have the disadvantages of incomplete information，

excessive reliance on artificial feature extraction and low diagnosis rate. In order to solve these problems，a

diagnosis method based on deep graph convolutional neural network（DGCN）was proposed. Firstly，a partial

discharge（PD）experimental platform was built on a 220 kV real GIS and the partial discharge signals collected by

ultra high frequency sensor were converted into frequency domain spectrogram samples by Fourier transform.

Then，the spectrogram samples were input into the DGCN，which undergoes graph convolution，coarsening and

pooling operations to make the spectrogram structure was clearer and enrich the input information. Finally，the test

samples were used to test the DGCN with set parameters. The experimental results show that the proposed method

can achieve a recognition rate of 98.77% for GIS fault defects，which is significantly higher than other methods and

has good robustness.

      Key words：gas insulated switchgear（GIS）；partial discharge（PD）；fault diagnosis；insulation defect；deep graph convolutional neural network（DGCN）；simple linear clustering method