Research On Remaining Useful Life Prediction For Satellite Lithium-ion Battery

The lithium-ion battery becomes the third generation power system for satellite dueto its high voltage, small size, light weight, high energy density, long lifetime, and lowself-discharge rate. The state monitoring in-orbit, remaining useful life (RUL)prediction, and system management have become the research focuses and challenges inaerospace because the power system is the key component of the satellite. In this work,therefore, the degradation modeling and remaining useful life prediction for satellitelithium-ion battery based on data-driven method is implemented.Firstly, the capacity is difficult to monitor when the lithium-ion battery in in-orbitfor space application, as a result, it is not easy for researchers to predict the performanceof lithium-ion battery. Thus, we propose an indirect RUL prediction method based onthe external parameters that can be measured directly. The indirect health indicator isextracted with the time interval of equal discharging voltage difference. Secondly, theframework for indirect RUL prediction of lithium-ion battery based on Echo StateNetwork (ESN) is proposed. This framework can establish an accurate degradationmodel for lithium-ion battery based on the highly nonlinear approximation ability ofESN. Thirdly, due to the first prediction of time interval of equal discharging voltagedifference cannot track the decreasing tread of the true data in the predicting process,which leads to the error accumulation step by step. We propose Monotonic Echo Statenetwork (MONESN) which combines monotonic prior knowledge to the trainingprocess of ESN to solve this drawback. The output and input unit of MONESN areforced to meet the monotonic relationship, which improves the accuracy of RULprediction. Fourthly, we find that the RUL prediction of single MONESN model isunstable, which can lead to larger RUL prediction error some times. So, we adoptensemble method for RUL prediction of lithium-ion battery. The Ensemble MonotonicEcho State Network (EnMONESN) is proposed to improve the stability of predictionmodel. At the same time, the EnMONESN can reduce the generalization error by thediversity of predictions. Lastly, we apply the above methods in this thesis, Gaussianprocess regression, and NDAR in the software for indirect RUL prediction of lithium-ion battery. The software is based on LabVIEW and MATLAB.A large number of experiments, which use NASA and satellite lithium-ion batterymonitoring data are carried out, the results indicate that the indirect RUL predictionalgorithm based on EnMONESN can achieve more accurate and stable RUL prediction.The stable operation of the software for indirect RUL prediction of lithium-ion batteryproved that the indirect RUL prediction algorithm is an effective way to achieve theperformance prediction of lithium-ion battery in-orbit.