As the advance and foundation of prognostics， the bearing performance degradation assessment （PDA） is of great significance for online condition monitoring. Aiming at the problems that similarity-based methods are complex and time consuming， an online PDA method for rolling bearings is proposed based on autoregressive model （AR） and multivariate state estimation technique （MSET）. The coefficients of the AR model serve as feature vectors to depict bearing performance states and the MSET model is used to reconstruct AR coefficients. For that purpose， firstly， the historical memory matrix of the MSET model is constructed with the AR coefficients of vibration signals under normal operation， and then the AR coefficients of the signals under consideration are input into the MSET model as observation vectors to obtain the reconstructed estimation vectors. By inputting the signal into the two autoregressive models， which are composed of original AR coefficients and reconstructed AR coefficients， respectively， the corresponding residual sequences are obtained. Finally， the performance degradation index is constructed by exploiting the difference between the root mean square values of the two residual sequences. Hence， artificially induced defects and run-to-failure data set from rolling bearings are processed to demonstrate the advantages of the method in terms of the trendability， consistency and sensitivity of early failure.