Research On Lithium-ion Battery Health Management Based On Data-Driven Modeling

The performance of lithium-ion battery continues to degrade with the operating time,while reducing the power conversion efficiency of the lithium-ion battery,which causes the phenomenon of overcharging,over-discharging and the temperature exceeding their own range,and increase the risk of spontaneous combustion due to the thermal runaway.Therefore,in order to ensure the safe and reliable operation of lithium-ion battery,it is important to monitor the battery operating state.Since the ageing process of lithium-ion battery is a dynamic coupling processes,its internal electrochemical mechanism is very complicated,and the battery parameters are not easy to obtain,which greatly increase the difficulty of establishing a mechanism model.At present,the establishment of lithium-ion battery model based on data-driven method provides another solution for monitoring the battery state,but the premise of the feasibility of this method is sufficient data information and consistent stability.However,the generalization performance of the model will be greatly reduced when the labeled data samples are rare and the operating conditions are complex.Hence,in order to improve the accuracy prediction of battery state of health in complex environments,while aiming at the problem of the insufficient sample data information,difficult evaluation of effective ageing features and variable operating conditions of lithium-ion battery,the following researches have been carried out in this dissertation:1.Research on lithium-ion battery state of health prediction methods under small sample condition.Under laboratory conditions,the high cost of collecting lithium-ion battery ageing data and insufficient data information affect the accuracy of the battery state of health prediction.This dissertation proposes a lithium-ion battery state of health prediction model based on semi-supervised transfer component analysis.By reconstructing the feature space to eliminate redundant information between aging features and minimize the difference between different data distributions,so as to realize the maximum alignment of the source domain data and target domain data distribution.Furthermore,the correlation between ageing feature and lithium-ion battery state of health are quantified by mutual information,which provides a basis for training the lithium battery state of health prediction model.Under the same verification conditions,compared with other models,the proposed method in this dissertation can guarantee the prediction accuracy when the training set only account for the first 35% of the entire set.2.Research on lithium-ion battery state of health prediction methods under variable temperature condition.The accuracy of the prediction model is related to the effectiveness of the model input.It is important to evaluate the correlation of ageing feature to improve the accuracy of the battery state of health prediction.This dissertation proposes a lithium-ion battery state of health prediction model based on Gaussian process regression with Matern correlation automatic relevance determination(Automatic Relevance Determination,ARD)kernel function.By using ARD kernel function,the correlation degree of each ageing feature is directly to quantify,while selecting the best aging feature automatically,and provide a basis for choosing the extraction interval of optimal ageing feature.Furthermore,taking into account the diversity of operating temperature modes of lithiumion battery,it is verified that the proposed method has the ability for predicting the battery state of health under different operating temperature modes while extracting the optimal aging feature automatically.3.Research on lithium-ion battery state of health prediction methods under different aging mode changes.The diversity of ageing modes of lithium-ion battery affects the generalization ability of the battery health state prediction model.This dissertation proposes a battery state of health prediction model based on random forest–relevance vector machine,and by using random forest method to quantify the contribution of aging features,while extracting the optimal aging features automatically,which provide a basis for selecting the appropriate model input.In addition,the output posterior probability value(standard deviation)of the relevance vector machine is selected as the weight factor of the prediction result under different verification set.Combined with the concept of ensemble learning,a lithium-ion battery state of health prediction model that can automatically extract aging features is constructed,and the battery state of health prediction under different aging mode changes is realized.4.Research on lithium-ion battery state of health prediction methods under current pulse test condition.The variability of dynamic operating condition makes the source domain data more diverse,which affects the effectiveness of the battery state of health state prediction model.Consider how to select the effective data which related to the target domain from a variety of source domain data,and realize the sharing of data information among different working conditions is the problem that needs to be solved.This dissertation proposes a lithium-ion battery state of health prediction model based on the Tradaboost.R2 algorithm.The distribution of source domain data is adjusted by modifying the weight samples in the source domain data.At the same time,in order to ensure the mapping relationship of the prediction model under different state of charge is not affect,by using different source domain data to build the ageing feature,when constructing the auxiliary data set and the source domain data set in the Tradaboost.R2 algorithm,six division methods of the data set is considered.It is verified that the proposed method can guarantee the accuracy the battery state of health prediction under dynamic multi-working conditions,while the feature dimension and the demand for the target domain label data is reduced.