Vibration signals resulting from rolling element bearings present a rich content of physical information, the proper analysis of which can lead among others to the identification of possible faults. Traditionally, this analysis is performed by the use of signal processing methods, which assume statistically stationary signal features. The paper proposes an alternative framework for analyzing bearing vibration signals, based on cyclostationary analysis. This framework, being able to model additionally signals with periodically varying statistics, is better able to exhibit the underlying physical concepts of the modulation mechanism, present in the vibration response of bearings. The basic concepts of the approach are demonstrated both in illustrative simulation results, as well as in experimental results and industrial measurements for two different types of bearing faults.