Aiming at the problems such as high loss of characteristic information, weak generalization ability, and strong data dependence commonly existed in existing data-driven based methods for bearings health status evaluation, a new evaluation model based on variational auto-encoder (VAE) which allowing high-entropy characteristic input is proposed. By learning the high-dimensional potential probability distribution of the bearing vibration signal spectrum point in characteristic space, our model can quantitatively evaluate the bearing operating health state. First, the health status evaluation model based on VAE is theoretically elaborated. Afterward, a state assessment index based on the lower bound of variational evidence is established. As a consequence, it is proved that the variational auto-encoder has good accuracy in dealing with the evaluation of bearing running state and is more sensitive to the abnormal state through comparative experiments. Additionally, there is no need to extract features and set complex parameters artificially, also without setting and adjusting specific parameters for a specific system. Furthermore, it retains good robustness even with small training data sets as well as showcases certain promotion value of engineering application.