The present invention discloses an EEMD (ensemble empirical mode decomposition)-DFA-based cable oscillation wave partial discharge signal feature extraction method. The method includes the following steps that: 1) ensemble empirical mode decomposition (EEMD) is adopted to perform partial discharge signal decomposition, so that a decomposition coefficient IMF is obtained; 2) a DFA method is adopted to optimize and reconstruct the IMF; and 3) the sample entropies and approximate entropies of each sub-band coefficient are extracted and calculated, the sample entropies and approximate entropies are adopted as feature vectors and are inputted into a support vector machine classifier model to verify the effect of the extraction method.