In the state prediction combination model of rolling bearings, the proportion weights are mostly fixed weights, and there are few combined state prediction methods that adaptively dynamically adjust the weights. To solve this problem, an adaptive combined bearing state prediction method based on ARIMA and Elman was proposed; the bearing acceleration performance degradation data set provided by IMS was used for verification. The results show that by using a single ARIMA model and Elman model for prediction, the average relative errors are 3.95% and 5.62% respectively, and the average relative error of variable weight Elman-ARIMA combined prediction model is 3.22%, which is lower than the two single prediction models, and the prediction results have higher reliability, which proves the feasibility of the combined prediction method.