Abstract: In order to solve the problem of complex insulation fault type recognition and redundancy of feature selection of cable accessories insulation defects, a method to recognize the fault of insulation defects by directly using the partial discharge signal containing the noise is proposed in this paper. The signals in three power frequency cycles are selected as a sample. The discharge pulse is firstly extracted by using mathematical morphological filtering technology to obtain two statistical characteristics of average discharge and discharge times. At the same time, the Hurst exponent is used to judge the fractality of partial discharge signals. If the condition is satisfied, the box dimension is directly obtained as a fractal feature. Finally, three features are imported into the extension neural network for pattern recognition to verify the feasibility and effectiveness of the method. The results show that the three features of box dimension, average discharge, and discharge times have strong separability, which can solve the problem of redundancy of feature selection, and the extension neural network can identify different types of insulation defects, and the recognition rate is better than that of similar methods based on support vector machine and BP neural network.