Insulation defect is an important factor to induce partial discharge of enclosed switching equipment. As a non-electric partial discharge detection method

the gas component analysis method has no problems such as introduction of noise and environmental interference. Consequently

four kinds of common insulation defect models were simulated to study the characteristics of gas decomposition components. A P-CNN fusion fault diagnosis algorithm was proposed

and a data preprocessing method was proposed to rapidly extract the features of characteristic data components and equalize the gray scale map. Three gas component features were introduced

respectively

and were trained by probabilistic neural network and convolutional neural network. Meanwhile

the results were fused to the conclusion level by DS evidence theory. Simulation results show that

compared with traditional machine learning

this algorithm can effectively improve the recognition accuracy while ensuring the recognition speed.