Multiscale entropy has begun to play an increasingly important role in the analysis of non-stationary and nonlinear vibration signals. Changes in the sample entropy of different scales can reflect changes in the transformer windings of different runnings. In this paper, a novel feature extraction is proposed, and a new and effective feature parameter is provided to efficiently and quantitatively describe faulty signals of the transformer winding. The results of analyzing the experimental data of the winding vibration show that compared to sample entropy, multiscale entropy can efficiently realize the feature extraction of faulty signals. Therefore, it is feasible to introduce the effective feature parameter into the use of transformer winding vibration signal analysis.