Traditional and neural network-based methods being in existence of rolling bearing fault prediction, a dual adaptive sliding time window fault prediction model is proposed. Firstly, the rolling bearing vibration signal is mapped into fault features characterized as its degradation state by setting up a state estimation non-linear operator capable of removing correlations. Secondly, taking a loss function as a criterion, an adaptive update mechanism for the model parameters is set up, and a sliding time window capable of adaptively selecting the data length is constructed. Finally, the validity of the proposed failure prediction model is verified by simulating the occurrence of failures under the combined sudden and gradual failures in practice using the whole life cycle data of rolling bearings released by Xi’an Jiaotong University. The experimental results show that the prediction model proposed can accurately identify at the beginning moment and at the failure moment for the rolling bearing at the degradation stage, and truly reflect the trend of equipment performance degradation. The prediction error is only 0.068% and the prediction time is only 1.385% of the interval between failures, meeting the needs of rolling bearing failure prediction under condition of complex operation.