Design And Research Of A Class Of Filter-based Prediction For Remaining Useful Life Of Lithium-ion Battery

Predicting its remaining useful life(RUL)is an effective means for fault diagnosis and health management of lithium-ion batteries.At present,the RUL prediction approaches of lithium batteries are divided into three main categories: model-based approaches,datadriven approaches and data-model linkage hybrid approaches.The model-based approaches use the stability of the model to make the prediction results robust.The data-driven approaches rely on the objectivity of the data to enable prediction results real-time.The data-model linkage hybrid approaches have a better prospect of application,which have the stability of models and use historical data to reflect the real environment.For a class of filter-based data-model linkage hybrid approaches,the main research work in this thesis is as follows:(1)For the problem of short prediction periods,a multi-step prediction method for lithium-ion battery RUL based on the prediction of innovation by multi-dimensional Taylor network and prediction update by Kalman filter is proposed.Firstly,the RUL multi-step prediction iterative model is built;Secondly,the multi-step innovation is predicted based on the multi-dimensional Taylor network;Finally,the fixed-step extended Kalman filter is built to achieve the multi-step prediction of RUL.The proposed method presented in the NASA data was validated.(2)For the phenomenon of "capacity augmentation ",a strong tracking filter based on the adaptive compensation model is proposed to predict the RUL;Firstly,the RUL model is modeled with higher-order states and the measurements are modeled with "capacity augmentation" to form an adaptive compensation model;Secondly,a long short-term memory neural network is used to update the long-term innovation;Finally,a strong tracking filter is designed based on the "capacity augmentation" of the adaptive compensation model.The proposed method presented in the CALCE data was validated.(3)For the problem that the mutation information of "capacity augmentation" is lagged,the smoothing back method based on characteristic function filter is proposed.Firstly,the RUL multi-step smoothing back model is built;Secondly,the multi-step smoothing back model based on the inverse backtracking provides lagged state information to the characteristic function filter that is being predicted;Finally,the fusion filter design of forward prediction and smoothing backward is completed.The proposed method was validated separately in two types of data sets.