Remaining Useful Life Prediction Of Lithium-ion Battery Based On Multiple Model Combination Method

Complex engineering systems have the properties,including miscellaneous structure,high integration and strong coupling.To realize the safe and reliable operation of the system,the research on system fault prediction and health management has made considerable progress.As the core subject of fault prediction,remaining useful life prediction is to analyze the system degradation trend and the time when the fault may occur according to the operation state and early fault tip of the system,so as to provide a theoretical basis for condition-based maintenance and to avoid the continued deterioration of the failure’s symptoms which will lead to serious disasters.This paper sorts out the existing remaining useful life prediction methods in detail.Considering the huge internal structure of complex systems and the correlation between the components.There are difficulties to accurately obtain the failure mechanism.Moreover,the degradation process of the system will be interfered by many factors,which will have a negative impact on the historical observation data and the current system state.Both model-based methods and data-based methods use incomplete system information to build a single degradation model,which has obvious limitations for the life prediction of complex systems,and their prediction effects need to be improved in accuracy,stability and universality.Therefore,this paper regards the Lithium-ion battery as a representative of complex nonlinear system and predict its remaining useful life.First of all,the battery health status from the perspective of battery capacity was defined,and a double exponential empirical degradation model based on the data of historical battery capacity was established.The parameters of the model are continuously modified by unscented Kalman filter until the end of the observation phase.Secondly,the residual sequence of the unscented Kalman filter in the observation stage is used as the training data to construct a Relevance Vector Machine model,and correct the prior estimates by the predicted residual data in the prediction stage.The fusion method of the unscented Kalman filter and the Relevance Vector Machine is used to recover the update function of the filter.Then,in order to improve the generalization performance of the prediction model.We build an ensemble prediction model based on fusion way and establish several sub-fusion models through the sample disturbance.The average prediction results of all sub-fusion models are applied as the final output of the ensemble model.Finally,Considering the diversity of sub-models will cause more uncertainty,and some of the prediction results of sub-fusion models deviate abnormally.Isolation Forest algorithm is used to identify and mark anomalies.We compare the integration results under the three types of anomaly detection standards,analyze the impact of the abnormal prediction value on the overall degradation trend of the battery capacity,and use the multiple model combination method to predict the remaining battery life.