Most reasons causing the onload tap changer(OLTC) unfunctional are mechanical faults. In order to ensure the reliable operation of the OLTC, a kind of vibration feature extraction method of OLTC based on the empirical mode decompositionpower spectral density(EMDPSD) algorithm is proposed. Firstly, several mechanical faults of the OLTC is simulated and the vibration signal is collected. The noise of vibration signal is eliminated by the wavelet packet firstly. And then a reasonable order for noise reduction is selected according to the singular entropy which is to eliminate the noise of vibration signal by singular value decomposition(SVD). Secondly, the noisereduced signals are decomposed by EMD and the power spectral density of the IMF is obtained. Finally, the PSD energy is used to construct the eigenvectors. It is shown that the signaltonoise ratio of the vibration signal can be improved by selecting the SVD order of noise reduction based on the singular entropy. Based on the combination of EMD and PSD methods, the fault features of the OLTC can be effectively extracted.