Machine learning-based aging models for estimating battery state of health and predicting future degradation

Thumbnail Image
Thelen, Adam
Major Professor
Hu, Chao
Laflamme, Simon
Wang, Zhaoyu
Hu, Shan
Pint, Cary
Committee Member
Journal Title
Journal ISSN
Volume Title
Research Projects
Organizational Units
Journal Issue
Is Version Of
Mechanical Engineering
Lithium-ion (Li-ion) batteries are everywhere, from portable electronics to the latest electric vehicles, because of their unmatched energy density and rechargeability. Unfortunately, the performance of Li-ion batteries degrades over time as their continued use and operating environment cause irreversible chemical reactions that decrease cell capacity and power. Estimating and predicting the health of Li-ion batteries in consumer devices is imperative to ensure safe and reliable operation over the product’s expected lifetime. What's more, the sudden deployment of millions of new electric vehicles recently has created a plethora of aging batteries that will soon be retired, with few plans to capture and deploy the packs to a second life. My graduate thesis focuses on machine learning-based modeling methods for diagnosing Li-ion cell health and predicting future cell degradation. Diagnosing cell degradation modes is essential for understanding the probability of different failure modes an aging Li-ion cell might experience. Understanding the degradation modes present in a cell provides engineers with valuable information to improve cell design and performance. Further, estimating the health of aged Li-ion cells provides more insight into how they might be optimally used in their second life after being retired from an electric vehicle. In the first chapter of my thesis, I discuss a cell health estimation model capable of rapidly assessing the state of health of used Li-ion batteries from electric vehicles so they can be optimally assigned to recycling, rejuvenation, or redeployment. To accurately estimate the health of aged cells, I developed a deep neural network to estimate the remaining capacity and both charge/discharge resistances at three different states of charge. The algorithm uses only 5 minutes of raw voltage time-series data taken from the constant current portion of a battery's charge curve. The rapid capability of this algorithm makes it suitable for deployment in an industrial battery recycling and rejuvenation facility. The rapid health estimation model can maintain an average prediction error of no more than $4.0\%$ under all tested conditions. The methods discussed in chapter one of my thesis show promise for improving the longevity of Li-ion battery packs and cells by accurately assessing their usability at the end of their first life. In the second chapter of my thesis, I focus on the challenge of predicting the life of cells operating under different conditions. Predicting a cell’s future degradation trajectory can inform engineers of the cell’s lifetime, which can be used to inform future cell designs, screen for materials, set warranties, and schedule product maintenance. Accurate cell lifetime models can rapidly accelerate high-impact material optimization and control projects by substantially reducing time spent aging cells. In this chapter, I discuss a hybrid modeling approach to predicting the remaining useful life of Li-ion batteries. The approach leverages the best aspects of traditional model-based trajectory prediction and data-driven learning. My approach decomposes the task of RUL prediction into two steps: 1) Offline training of data-driven models for RUL error correction and 2) Online data-driven correction of model-based RUL prediction. The approach is evaluated on five datasets consisting of 237 cells: 1) three open-source datasets, 2) one proprietary dataset, and 3) a simulated out-of-distribution dataset. Results show that data-driven error correction reduces root-mean-square error by 40\% and mean uncertainty calibration error by 34\% compared to a model-based approach alone. The proposed approach is also shown to be more conservative in its uncertainty estimates than a purely data-driven RUL prediction approach. The machine learning-based battery aging models presented in this thesis are just a few examples of models capable of enhancing our understanding of battery degradation. To ensure a fast transition to renewable energy technology focused on battery energy storage, more battery aging modeling will have to be done. What's more, as batteries evolve in chemistry and design, new techniques will need to be developed and deployed to effectively monitor and predict cell degradation. In the conclusion of this thesis, I revisit some of the major modeling challenges that the industry and academia face and propose a handful of novel ideas worth researching further.