Research On Lithium-ion Battery Life Prognostics And Health Management Method Based On Machine Learning

With the booming development of the electric vehicle(EVs)under the environment and energy crisis,the power batteries,as the core of EVs have been more and more attractive.Hence,it is essential to conduct research on the battery life prognostics and health management(PHM)technology in battery management system to ensure the driving safety of EVs.Firstly,considering the direct measure of capacity,which is the index to envluate the battery state of health(SOH)or life,the full-life health indicators(HIs)and early HIs,such as voltage increment at the same charging time interval and the difference of constant current charging time,are extracted to describe the battery degradation through the analysis of measurable parameters in charging-discharging.The correlation of HIs was quantitatively analyzed by Pearson to ensure that they could be applied in the following work.Secondly,a novel SOH estimation based on imporved long-short term memory(LSTM)combined with multi-health indicators fusion is proposed.Considering the influence of multi-dimensional HIs input on the complexity of network model calculation,the feature redundancy is eliminated by neighborhood component analysis(NCA).Aiming at the hyperparameter optimization in LSTM and potential local optimum,a differential evolution gray wolf optimizer(DEGWO)is proposed for global hyperparameter search,which improved SOH estimation accuracy.Its feasibility is verified by multi-group simulation.Then,in view of the problem that the model trained by the historical data is not suitable for new scene data due to the inconsistency of data caused by the variety of battery aging data and complex working conditions,this paper proposes a novel battery health management framework combined with transfer learning and hybrid deep learning.Based on the common characteristics of battery data in different domains,the new source domain data is obtained by combining joint distributed adaptation(JDA)learning to train the hybrid model.And then SOH estimation is achieved online based on the target domain.Deep belief network is utilized to mine the deep aging information of battery degradation,and then the degradation model based on LSTM is established.And the method considering knowledge transfer in this section has been analyzed through many aspects for the verification.Finally,an ensemble learning framework oriented to battery life early prediction is proposed for the difficulties of obtaining complete battery data.Clustering by Fast Search(CFS)is adopted to filter the central candidate features in original early HIs obtained from early degradation data and eliminate the redundant information,in order to make up the negative influence on prediction precision caused by redundant and low correlation information in multi-dimensional HIs.Aiming at the problems of low prediction accuracy and poor robustness of a single learner in long-term prediction,the integrated learning prediction method is adopted to improve the long-term prediction accuracy of power battery life.The feasibility of the framework is testified by the simulation with the precison significantly improved.