According to the view that the rolling bearing fault signal is a periodic pulsing signal, a method based on maximum correlated kurtosis deconvolution (MCKD) and spectral kurtosis (SK) called the MCKD-SK method is proposed in order to diagnose early faults of rolling bearing. This method can effectively suppress noise in the signal in order to extract the bearing early fault signal from the signal submerged by noise and improve the kurtosis of the original signal. The SK method is then used to select a reasonable frequency band and demodulate the low frequency information from the high frequency band. Through simulation and actual monitoring, as well as data analysis and validation, the MCKD-SK method has being shown to accurately and effectively diagnose early faults of rolling bearing. This method is also suitable for on line monitoring early motor bearing faults.