Abstract: A phase segmentation method is proposed to solve the problem that a small amount of data is difficult to form a map when extracting the characteristic parameters of partial discharges of cable accessories. At the same time, multidimensional scaling (MDS) is used to reduce the dimension of the eigenvalues and to extract the best features with higher classification ability. Through the division of partial discharge (PD) signals at equal angles in the power frequency cycle, the eigenvalue extraction is carried out in each of the divided regions to obtain more detailed and specific characteristic parameters, and then the feature values are optimized by MDS to improve the recognition speed and identify accuracy. The experimental results of the typical defects of various power cable accessories show that this method can extract better eigenvalues under fewer data and get better recognition results.