Due to the use of the spectral kurtosis based on the fourth sample moments, the Wavelet packet transform filter based method might suffer from the non-robustness problem in practical application. Therefore, robust spectral kurtosis coefficients are defined firstly. Subsequently, the robust Kurtogram based on wavelet packet transform is proposed. The practical data have been applied to demonstrate that the improved method not only can measure the real spectral kurtosis, but also can precisely detect the fault transient components used for the bearing fault diagnosis.