Research On The Remaining Useful Cycle Life Prediction Of Lithium-ion Battery Based On Unscented Particle Filtering

With the highlighting advantages of high energy density, high-potential, low self-discharge, long cycle life and no memory effect, etc., Lithium-ion battery has become a core component of the storage of many systems, such as electric vehicles, electrical equipment, spacecraft electronic equipment. However, there are also security and reliability issues, especially a dead battery will not only cause System failure and huge losses and even catastrophic consequences. In order to prevent the weakening of adverse consequences, huge breakthrough must be taken in lithium-ion battery health management. Therefore, lithium-ion battery health management technology like characterization of health state and the remaining cycle life prediction is becoming a study of cutting-edge technology. Effective remaining cycle life (Remaining Useful Cycle Life, RUCL) prediction method can optimize lithium-ion battery maintenance strategy and greatly improve system reliability.Lithium-ion battery is a rather complex electrochemical system of strongly nonlinear non-Gaussian, but its specific chemical reaction inside is very difficult to directly monitor, meanwhile its application environments and uncertainties in load conditions will lead to performance degradation state instability. Firstly, from a purely data-driven perspective, this thesis usd the global search optimization capabilities of particle swarm to the optimum position relevant parameters fitness goal to optimize penalty coefficient and kernel function parameters under LSSVM framework, then set up PSO-LSSVM model and used in lithium-ion battery cycle life remaining prediction. Secondly, from the view of combining model-driven and data-driven, this thesis select a RUCL prediction method under particle filter framework based on degradation empirical model, then used evidence theory to determine the initial parameter degradation model and combined with degradation sample data to realize RUCL prediction. However, in particle filter algorithm implementation process, there are inevitably particle degradation and re-sampling particle deprivation, in view of this, this thesis proposed regularization particle filter and unscented particle filter to improve the particle filter shortcomings. Through algorithm simulation and degradation status tracking of sample data to improve the prediction accuracy and adaptability issues, here providea new technical ideas for remaining lithium-ion battery cycle life prediction.The main researches and the acquired innovative achievements in the tarticle are as follows(1) Study of failure mechanisms and degradation data analysis.Firstly, based on the composition structure of lithium-ion battery, this thesis introduces the its working principle and analysis of failure mechanisms to familiar with its failure modes and failure of internal and external factors; Secondly, take the battery capacity attenuation as the characteristics, introduce the degradation empirical models based on electrolyte resistance, charge transfer resistance and the double layer capacitance as total parameters to reflect the actual capacity attenuation. Finally, the excellent uncertainty reasoning ability of evidence theory is used for empirical model parameters initialization, combining with the degradation samples.(2) Support vector machine use structural risk minimization principle instead of empirical risk minimization principle, but it still have a better identification accuracy in small samples, while LSSVM replaced slack variable squared training error, and with equality constraints instead of SVM inequality constraints, which can convergence speed and improve the efficiency of training, but its accuracy is closely related to its penalty coefficient and kernel function parameters. In order to avoid the adverse effects of basic parameters of space exhaustive search optimization method, this method combined particle swarm algorithm with LSSVM, which use the PSO global search ability to optimize LSSVM parameters to establish the battery degradation model and apply to the prediction of RUCL finally.(3) As to the particle deprivation in resampling process of particle filter, this article proposed a new RUCL prediction method based on regularized particle filter. RPF increased Nth generation process from nuclear density function on the basis of the PF, which can effectively solve the particles deprivation, ensuring less complexity under the premise and improving filtering performance simultaneously.(4) As to the particle degradation of particle filter, this article proposed a new RUL prediction method based on regularized unscented particle filter. Taking into account the approximate distribution function which is produced by unscented Kalman algorithm, has great overlap with real posterior distribution, and it is also capable statistic heavy-tailed estimation error high-end components of heavy-tailed posterior distribution. So under the framework of particle filtering, this UPF method proposed to use unscented Kalman to generate PF proposal distribution function, which can effectively solve the particle degeneracy problem and more substantial increase filtering accuracy at the expense of part of the computational complexity of the premise.This thesis introduces the capacity degradation empirical model, then combines the evidence theory with sample data to construct a priori model. The unscented particle filter is used to update model parameters and track data degradation trend. Finally the last updated model is used to achieve the RUCL prediction and probability density distribution based on fault threshold.