It is difficult to accurately extract the health index of non-stationary signals of rolling bearings under variable rotational speed, which also leads to greater prediction error for bearing degradation models with fixed parameters. For this reason, an angular domain unscented particle filter model with time-varying degradation parameters is proposed to deal with the remaining useful life (RUL) prediction of rolling bearings. Order analysis is first performed to transform the variable-speed signal from time domain to angular domain for extracting the health index in the angular domain representation. To track the bearing degradation state, a real-time finite element model is established to guide the parameters updating the procedure of the performance degradation model. Finally, the bearing degradation state is estimated by the unscented particle filter (UPF), and then the remaining useful life of the bearing is predicted. In this way, the time-varying degradation model is developed by considering both non-stationary feature extraction and dynamic state tracking for rolling bearings. The proposed method is verified by both benchmarks: bearing experimental data, and a bearing accelerated life experiment. Compared with state-of-the-art prognostic methods, the present model can predict the bearing remaining useful life (RUL) more accurately under variable rotational speed.