By analyzing the root-mean-square （RMS） and sample entropy evolution on the run-to-failure data of rolling bearing， it is found that there is a certain cointegration relationship between the two features， and a method for extracting degradation feature of rolling bearing is proposed based on cointegration theory. The proposed feature has a good two-stage property， and it can unify the run-to-failure data of different rolling bearings and obtain a consistent evolution process. By using several run-to-failure datasets of rolling bearings， the advantages of the proposed feature are verified. At the same time， the reasons for the two-stage and consistency of the proposed features are studied in depth. Compared with RMS and sample entropy， the proposed feature has better monotonicity and better prediction ability in non-stationary stage. Finally， the differences between cointegration fusion and common fusion algorithms are compared. By comparison and analysis， the proposed feature has good two-stage property and consistency， and has better prognostic ability than RMS and sample entropy.