According to the limitation of traditional envelope spectrum and kurtogram, a novel approach to fault diagnosis of bearing based on dual-tree complex wavelet packet transform and kurtogram is presented. The dual-tree complex wavelet packet transform is substituted for the filter in spectral kurtosis. The shortcomings of traditional kurtogram based on the FIR and short time Fourier transform filters is overcome and its accuracy in detecting transients in a signal from strong background noise is improved. Firstly, the bearing fault vibration signals were decomposed into various frequency band signals. Then the spectral kurtosis was computed and the maximum kurtosis was found. In the end, the filtered signal that maximizes kurtosis and its envelope spectrum were obtained. Therefore, the characteristics of the bearing faults can be recognized according to the envelope spectrum. The experimental results show that not only the frequency band selection accuracy and signal noise ratio are improved, but also the faults of the bearing can be effectively detected.