Battery health diagnosis and prognosis using physics-informed data-driven methods

Thesis

Estimation and prediction of battery state of health (SOH) are critical to ensuring operational safety and reliability. Due to complex degradation mechanisms and the limitations of measurement techniques in real-world environments, accurate SOH estimation and prediction remain challenging. This thesis begins with a review of health diagnosis and prognosis approaches in the battery literature, focusing on two problem settings: battery lifetime prediction and SOH estimation.

For battery lifetime prediction, a hierarchical Bayesian regression model is proposed and thoroughly evaluated. The approach is motivated by the observed variation in the relationship between extracted health features and lifetime labels under different aging conditions. Health features are categorized into two groups: individual cell-level features (reflecting intrinsic variability across cells) and population-level features (capturing the influence of cycling conditions on the average behaviour of the population). It is shown how this relationship can be explicitly modeled through a hierarchical dependency between these two types of features. The proposed method is validated using a public dataset consisting of cells subjected to fast-aging experiments, as well as a self-tested dataset that covers more realistic and diverse cycling conditions.

For state of health estimation, an aging-aware equivalent circuit model is proposed that combines the flexibility of data-driven methods with a model-based framework. Gaussian process regression is used to include parameter dependencies on operating conditions and lifetime for the equivalent circuit model. Both capacity and resistance are estimated from operational data without requiring ground-truth labels. The close relationship between the estimated resistance function and aging-induced changes of the open circuit voltage curve allows further estimation of degradation modes. Results from 114 battery packs deployed in off-grid solar systems indicate resting at high voltage will accelerate battery aging. Robust state of heath estimation for large fleets of field-deployed batteries enables a deeper understanding about how customer usage profiles influence battery degradation, which in turn has the potential to inform optimized operational strategies to prolong battery lifetime.

Authors: Zhou, Z

• DOI: 10.5287/ora-rdn0qmqxm
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English