In order to solve the problem that the signal in the online detection of high-voltage motor partial discharge (PD) is submerged by Gaussian white noise and narrow-band periodic interference, an algorithm based on improved variational mode decomposition (VMD) is used to filter out the noise. The common VMD algorithm may cause signal under-decomposition or over-decomposition. In order to overcome this shortcoming, the energy loss coefficient is used as the criterion to optimize the decomposition mode number. After the optimal decomposition, the kurtosis value is used to eliminate invalid modes and reconstruct the PD signal. The improved algorithm is compared with the existing methods in simulation. The results show that the improved VMD algorithm can filter out PD signal noise more effectively.