Quantitative identification of bearing fault severity is the basis of condition-based maintenance. The vibration signals will exhibit nonstationarity and nonlinearity in the presence of bearing faults. On the base of the multi-scale entropy, a new index quantitatively describing bearing fault severity is constructed which takes into account the mean value and the variations of the entropies over multiple scales. The new index is termed Partial Mean of Multi-scale Entropy (PMME). Simulation and experimental results show that the new index is able to detect incipient bearing fault and trend the fault development well.