Aiming at the shortcomings that the important feature information of the original vibration signal of the rolling bearing is submerged by strong background noise, and the extracted time domain features have high redundancy and strong relevance, this paper proposes a new rolling bearing fault diagnosis research method based on maximum relevance-minimum redundancy ( mRMR) feature selection and random forest. First, the original signal is subjected to complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) to obtain a series of intrinsic modal functions ( IMFs), analyze IMF and remove high frequency noise and part of false component, then reconstruct the signal and extract its time domain characteristics, mRMR is used to remove redundant and highly correlated feature vectors, so that the selected feature subset has the greatest dependence on the label, and finally the feature subset is input to the random forest classifier for classification. Experiments show that mRMR has an excellent feature search strategy, the important features are selected first. Only three features are needed to achieve a higher classification accuracy, and the efficiency is higher than other feature selection algorithms.