According to the modulation characteristics of roller bearing fault vibration signals，the limitation of the traditional envelope analysis，a roller bearing fault diagnosis method using improved envelope analysis based on empirical mode decomposition(EMD) and spectrum kurtosis (SK) was proposed in this paper. Firstly, roller bearing fault vibration signals were decomposed into a finite number of intrinsic mode functions(IMFs); secondly, using fft transformation to make each IMF into complex signal and calculatting these complex signal absolute value, then, spectrum kurtosis value was calculated of each IMF spectrum square envelope; finally, using the filter function of spectrum kurtosis, the best frequency band for demodulation was automatically choseen by the criterion of IMFs spectrum kurtosis. The proposed method has been applied into simulated signal and actual signal, the analysis results demonstrated the effectiveness of the proposed method.