The invention discloses a rolling bearing compound fault extraction method. According to the rolling bearing compound fault extraction method, the maximum correlation kurtosis deconvolution is regarded as a filter for filtering an original vibration signal, the original vibration signal is subjected to dimension raising to form four groups of signals after deconvolution processing through reasonably designing deconvolution parameters, the four groups of signals correspond to four fault types separately, and the four groups of signals obtained through dimension raising each correspond to a fault resonance frequency band, a single each correspond to signal is processed by means of a fast spectral kurtosis graph, thus a fault sensitive frequency band can be accurately identified and filtered, further envelope spectra of the four groups of signals can be obtained, and a fault type of an original compound fault signal can be determined through comparing corresponding relations of the envelope spectra of the four groups of signals and a fault feature frequency obtained through theoretical calculation. Meanwhile, the rolling bearing compound fault extraction method adopts a particle swarm optimization method in an empirical interval when selecting the maximum correlation kurtosis deconvolution parameters, and can improve the accuracy and self-adaptability of parameter selection.