Prognostics and health management (PHM) play a key role in increasing the reliability and safety of systems especially in key sectors (military, aeronautical, aerospace, nuclear, etc.). This paper presents a new methodology which combines data-driven and experience-based approaches for the PHM of roller bearings. The proposed methodology uses time domain features extracted from vibration signals as health indicators. The degradation states in bearings are detected by an unsupervised classification technique called artificial ant clustering. The imminence of the next degradation state in bearings is given by hidden Markov models, and the estimation of the remaining time before the next degradation state is given by the multistep time series prediction and the adaptive neuro-fuzzy inference system. A set of experimental data collected from bearing failures is used to validate the proposed methodology. Experimental results show that the use of data-driven and experience-based approaches is a suitable strategy to improve the PHM of roller bearings.