UC San Diego

Lithium ion (Li-ion) batteries degrade with cyclic usage and storage duration. Batteries close to their end-of-life can no longer fulfill their performance requirements, and have an increased likelihood of catastrophic failures. Different usage conditions, complex manufacturing, and lack of essential data contribute to the complex degradation of second life batteries and hinder accurate analysis of battery capacity degradation. Therefore, a quick and precise diagnosis of used batteries has become an important research area for battery management, specifically in large-scale power storage systems. This thesis illustrates potential approaches for diagnosing battery degradation, considering both a physical-based and a data-driven model. The main objective is to boost the degradation prediction with the data-driven model by leveraging artificially generated data from the physical model. The approach is divided into two steps with two different battery cell models. A gradient-free optimization approach is introduced with the most widely used battery model (21700), which comes with published data of its battery cell structure, to optimize inside-of-the-cell structure. We estimate the physical battery cell structure to acquire artificial data and compare the performance of the estimated battery against the original battery. Then the degradation prediction is investigated with the A123 battery, which has extensive and high quality of battery degradation data but lacks exact cell physical structural information. We acquire artificial degradation data by estimating the physical battery cell structure with the optimization approach, and then utilize these data to boost the original battery life cycle prediction. Linear regression and backpropagation methods (resilient backpropagation, conjugate backpropagation, and bayesian regularization backpropagation) are applied along with various test matrices and cross-validation to compare the life cycle predictions. The life cycle prediction was first conducted with only the actual data. Then the artificial data was added to the training sets to conduct the life cycle prediction. The life cycle prediction error with actual data was improved with resilient backpropagation and a sigmoid activation function by 4 - 6\% compared to linear regression. Backpropagation performs better on test sets containing a large amount of extreme values than the linear regression. Lastly, the prediction results with artificial data and actual data were compared by using the regression and resilient backpropagation. Adding artificial data reduced the life cycle prediction error by 0.3 - 1.3\% in the regression method; there was a greater improvement when the life cycle prediction error with actual data was over 15\%. However, with the backpropagation method, adding artificial data resulted in larger prediction errors.