To address the challenge of remaining useful life (RUL) prediction of lithium‐ion battery under the condition of insufficient small‐sample training data, a new two‐stage based method is proposed to improve the learning of degradation features and the accuracy of RUL prediction. First, a two‐stage adaptive Wiener process degradation model is developed, which can integrate the expectation–maximization algorithm and Kalman filtering to achieve online dynamic updating of model parameters. It can deeply enable the accurate characterization of nonlinear and multiphase degradation behaviors. On this basis, a Monte Carlo‐based simulation reconstruction strategy is employed to generate a set of statistically representative degradation trajectories, thereby effectively augmenting the training dataset. To mitigate the inconsistency between simulated trajectories and actual degradation behavior, a dynamically updated data reconstruction mechanism is introduced to accurately capture the time‐varying degradation characteristics of individual batteries under small‐sample conditions. Finally, a long short‐term memory network enhanced with a convolutional block attention module is constructed to fully exploit the long‐range dependencies and spatiotemporal degradation patterns embedded in the reconstructed data. Experimental results demonstrate that the proposed method achieves an average root mean square error of 0.83% and mean absolute error of 0.63% under small‐sample conditions, significantly enhancing the accuracy and robustness of RUL prediction.
Wang et al. (Thu,) studied this question.