Robust Fault Detection For Lithium-ion Batteries With Random Parameter Uncertainties

Battery system is the energy equipment of the modern power grid and electric vehicles(EV).Due to numerous advantages in economy and safety,lithium-ion batteries have been widely used in smart grids and transportation.The status of the lithium-ion battery system is related to the electricity supply quality in power grid and the safety issues of EVs.Therefore,a quickly and accurate fault detection method for lithium-ion battery is necessary.However,current lithium-ion battery fault detection methods consider inadequate of the uncertainties in model,resulting in a high false alarm rate(FAR)or a low fault detection rate(FDR).In summary,quantifying the uncertainties of the lithium-ion battery model,and then systematically designing a robust fault detection method is an urgent problem and has rich research value and practical significance.Aiming at the modeling problem of lithium-ion battery,this thesis establishes the 1-order RC equivalent circuit model and the 2-order thermal model for the lithium-ion battery electric and thermal dynamics respectively,and considers the changes of the model parameters with the State-of-Charge(SOC)and temperature.For the parameter identification of the models,the impulse response method and the prediction-errorminimization method are used to identify the parameters respectively.To quantify the uncertainties of the model,this thesis uses the chi-square test method to verify the probability distribution of the parameter uncertainties.For the problem of lithium-ion battery fault detection,this thesis proposes a robust fault detection method of lithium-ion battery considering the uncertainties of model parameters.Firstly,derive the residual signal and use the Savitzky-Golay filter to reduce the discretization error.Then the linear dependence of the residual signal on the parameter uncertainties is derived,and the Gaussian mixture model(GMM)is used to quantify the parameter uncertainties distribution,then the online GMM parameters of the residual signal is calculated.Finally,a weighted distance sum(WDS)is proposed as a residual evaluation function and the threshold of WDS is set through the fault-free simulation.In order to compare with the above methods,this thesis also gives two nominal model fault diagnosis methods based on discretization and SG filter respectively.In this thesis,the LIONSIMBA electrochemical simulation platform is used to verify the above theoretical methods,and the abrupt fault and incipient fault of current,voltage and temperature sensors are used for simulation testing.The simulation results indicate that the robust fault detection method considering the parameter uncertainties improve the FDR with the FAR almost unchanged,especially for the temperature sensor fault detection.