Partial discharge （PD） on-line monitoring is a common technology in condition of high-voltage motor monitoring. However, it is difficult to avoid noise interference on site. The most common noises are white noise and periodic narrowband noise. A new denoising method combining singular value decomposition and wavelet transform (SVD-WT) is proposed. The original signal is decomposed by SVD. Based on calculating the kurtosis value of the singular value sequence, the periodic narrowband noise is removed by adaptively selecting the singular value to be reconstructed. Then, the starting position of PD signal is determined by calculating the variance of the signal in the sliding window. Finally, the PD signal after denoising is obtained by zeroing the no PD location. The simulated and measured PD signals are denoised and compared with empirical mode decomposition and wavelet transform (EMD-WT) and adaptive singular value decomposition (ASVD). The results of simulated and measured PD signals show that the SVD-WT method has excellent performance.