In the light of extracting fault features of a rolling bearing in early failure period, an incipient fault feature extraction method based on k-optimized variational mode decomposition (VMD) and kurtogram is proposed, where k value is selected based on energy criteria. First, the vibration signals are decomposed by VMD, the energy of intrinsic mode function components is summed under different k values. Based on the energy criterion, the best k value is determined, which is used as the parameter of VMD in decomposing the signals. The reconstructed signals are demodulated at last. The simulated and measured signals of fault bearing are analyzed by this method, which can effectively extract the weak inner ring fault feature. The result shows that the method can determine the incipient fault feature with the best k value, which is significantly valuable in applications.