Data-driven State Of Health Estimation And Remaining Useful Life Prediction For Lithium-ion Batteries

Electrified transportation is currently an effective way to alleviate global warming and environmental pollution.The energy storage system is the key to determine the driving range,insure safe and reliable operation.Lithium-ion batteries have become important components of energy storage systems due to their excellent performance.However,lithium-ion batteries will inevitably experience aging and deterioration during usage and storage,which will affect their performance and safety.In order to effectively perform predictive maintenance on the battery system,the estimation of the state of health and the prediction of the remaining useful life are significant.This thesis studies the two key scientific and technical problems of the data-driven health prediction problem,which are health indicators extraction and data-driven algorithms design.The main research contents are as follows:(1)Aging experiment and aging analysis of battery system.An experimental platform has been established.Then,the battery cell and battery pack aging experiments have been designed and carried out to build battery system aging experiment data sets.Finally,the influence of different aging conditions on the battery aging rate is analyzed.(2)Research on battery state of health estimation based on feature selection and machine learning.A novel feature subsets selection method based on correlation analysis and sequential search fusion is proposed,and compared with conventional methods based on the accuracy of health state estimation and feature dimensions.The accuracy and computational requirements of different machine learning algorithms are evaluated.The results show that the method based on the combination of the proposed subset selection method and Gaussian process regression has the best accuracy and the lowest computational requirements.(3)Research on General health indicator extraction method for the battery and a modified Gaussian process regression algorithm.The partial charge or discharge data are used to extract health indicators that have strong correlations with battery capacity.For the series battery packs,a further health indicators generation strategy is proposed,which fully considers the pack capacity degradation and battery capacity inconsistency of the connected cells.A dual time-scale filtering method is proposed to expand this strategy to dynamic conditions.Finally,an improved kernel function is proposed to improve the estimation accuracy.(4)State of health estimation of battery packs based on model migration and fusion.The method of integrating multiple sub-neural networks is used to improve the accuracy and reliability of the self-training model.A method based on the fusion of migration model and self-training model is proposed to improve the accuracy and reliability of different model applications.The applicability of the model under different structures,different aging stages,and different discharge modes is verified.(5)Battery life prediction based on transfer learning.An online battery cell lifetime prediction based on the combination of online health indicators extraction and transfer learning is proposed.An algorithm combining Gaussian process regression and long-short-term memory neural network is used to improve the prediction accuracy.A lifetime prediction method based on the migration of the cell information to the battery pack is proposed,and the prediction of the future capacity change of the whole battery pack and each cell in the pack is achieved.