Meta-learning collaborative optimization for lifetime prediction of lithium-ion batteries considering label noise

Accurate lifetime prediction for lithium-ion batteries requires training the prediction model with adequate high- quality lifetime data. However, generating high-quality datasets with accurate lifetimes labels is often costly and time-consuming in practice, and noisy datasets with inaccurate pseudo-lifetime labels are more easily obtained. Training on such noisy datasets may lead to degraded prediction accuracy due to model overfitting to label noise. To address this issue, a lithium-ion battery lifetime prediction framework based on meta-learning collaborative optimization (MLCO) is proposed, which can achieve the adaptive co-optimization of model parameters and inaccurate labels. In the framework, a mixup error loss function is proposed, where mixup augmentation can serve as regularization to avoid the model overfitting to noisy-labeled samples. Then, a soft label generation method based on improved K-nearest neighbor (KNN) is developed for adaptively updating the loss function. The method employs shape distance and stratified sampling, enhancing the accuracy and diversity of soft label generation compared to conventional KNN methods. Additionally, an information entropy-based hard label correction criterion improves the identification and correction of high-noise pseudo-lifetime labels by simplifying the high-noise sample identification in the regression task to a binary source entropy problem. Through applying to two types of lithium-ion battery datasets with label noise, the proposed framework is proven highly effective; the mean absolute percentage errors are 14.9 % and 13.4 %, respectively, showing significant superiority over other four conventional methods.