DIAGNOSIS AND PROGNOSIS OF USED LITHIUM-ION BATTERIES

Abstract

The growing deployment of electric vehicles is expected to produce a large stream of retired lithium-ion batteries. Reusing these batteries can reduce material demand and support sustainable electrification, but practical reuse requires more than a pass-or-fail capacity test. Retired batteries must be diagnosed rapidly enough for sorting, and their future degradation must be estimated well enough to assign them to appropriate second-life applications and operate them safely. This dissertation develops diagnosis and prognosis methods for this purpose, with emphasis on rapid nondestructive measurements and physics-informed learning from limited data. The first part of the dissertation develops rapid state-of-health (SoH) diagnosis for used lithium-ion batteries. Mechanical expansion of pouch cells is measured using digital image correlation (DIC), and the results show that surface expansion contains strong SoH information despite cell-to-cell variation and spatial heterogeneity. Representative surface points are then used to approximate full-field expansion, enabling a simpler thickness-based diagnosis method. Fusing mechanical thickness features with electrochemical impedance spectroscopy (EIS) further improves SoH estimation, demonstrating that mechanical and electrochemical measurements provide complementary diagnostic information. The diagnosis framework is then extended from direct thickness measurement to ultrasound testing (UT), where multi-point acoustic measurements provide internal mechanical information for large-format or constrained cells. Combining UT-derived features with EIS preserves the same central diagnostic principle: representative mechanical information and electrochemical information can be fused to improve rapid SoH estimation. The second part of the dissertation develops physics-informed prognosis methods for estimating hidden states and parameters. A physics-informed neural network (PINN) framework with an integration-based loss is first applied to an equivalent-circuit battery model. Instead of penalizing each differential equation separately, predicted states are integrated through the governing model and compared with measured outputs, reducing the number of competing loss terms. The method identifies model parameters during training and provides a neural-network state estimator after training. The framework is then extended to a single-particle model with electrolyte dynamics (SPMe) with degradation mechanism due to solid electrolyte interphase (SEI) formation. In this electrochemical setting, separate long short-term memory (LSTM) networks estimate lithium-ion concentration in anode, electrolyte and cathode from current and voltage histories, while the SPMe dynamics constrains those estimates and identifies degradation-related parameters such as SEI thickness. To demonstrate that the physics-informed prognosis concept is not limited to batteries, the dissertation also applies the method to motor-bearing health monitoring. In this application, physics-informed machine learning is used to infer hidden lubrication-related states, including lubricant film thickness, surface roughness, and a lubrication regime parameter, from vibration measurements. The learned physically meaningful trajectory supports remaining-useful-life estimation using interpretable degradation thresholds. Finally, the dissertation describes how these technical contributions were translated into a broader module-level second-life battery reuse program. Funding proposals, intellectual-property activity, and related collaborative projects connected rapid diagnosis, physics-informed model construction, heterogeneous pack operation, and automated disassembly. Overall, the dissertation argues that sustainable battery reuse requires both rapid diagnosis of present condition and physics-informed prognosis of future behavior, and it provides methods that move these capabilities toward practical deployment.