Abstract: To realize bearing life prediction based on vibration data, a health index set containing time, frequency statistical indices and inverse trigonometric ones was constructed in this paper. By comparing the cosine similarity and correlation coefficient with the root mean square index, the index set was screened and reduced. For solving the problem of low accuracy of basic extreme learning machine (ELM) model caused by the random configuration of parameters, a meta-heuristic algorithm represented by bat algorithm (BA) was introduced. As the traditional BA was prone to fall into local optimum, an improved strategy of BA was established, which used levy flight strategy to improve the search ability of BA and enhanced the diversity of the population through competitive learning, thereby improving the search efficiency while ensuring search accuracy. The contribution of the BA improved strategy to the life prediction model was verified using the full life cycle test data of a rolling bearing. Results show that no matter what working conditions the test data set is in, the above prediction framework can obtain a relatively high prediction accuracy, not only keeping a high consistency with the true life curve but also showing stronger generalization performance and stability than the RBF and CS-ELM predictive models.