Low-data Machine Learning Framework for Generalizable Battery Degradation and Fault Analysis

The widespread adoption of lithium-ion batteries (LIBs) as the core of modern energy storage has introduced complex challenges related to their reliability, longevity, and safety, especially under varying operating conditions and abuse scenarios. These challenges are worsened by the scarcity of high-quality labeled data, differences in cell chemistries and usage patterns, and the urgent need for predictive tools that are adaptable, scalable, and efficient with data. Solving these issues requires combining advanced machine learning methods with domain-specific knowledge of battery behavior.This dissertation introduces a unified AI framework that addresses three main challenges in battery diagnostics: precise state prediction, early degradation forecasting, and fault detection under domain shift. First, CD-Net enables efficient, chemistry-aware lifetime prediction using only a small number of past cycles, making it well-suited for real-time, onboard applications. Second, DegradAI conducts early-life prognosis with just a few hours of data, while also functioning as a generative model to expand training datasets with synthetic yet realistic degradation data. Third, AccidentAI focuses on detecting rare, safety-critical anomalies through contrastive pretraining on unlabeled sensor data and few-shot learning to adapt to new domains with minimal labeled examples.Together, these methods create a flexible, modular, and data-efficient machine learning framework that advances the limits of predictive diagnostics for batteries. Beyond batteries, the techniques developed here have wider relevance in other safety-critical, data-scarce fields such as aerospace, medical devices, and industrial systems.Future work could investigate multi-modal fusion of sensor data, develop continual learning for online adaptation, and enhance generalization across different hardware platforms. As battery systems become increasingly vital to clean energy and electrified transportation, the ability to diagnose and predict failures early, reliably, and at scale will be crucial, and this work provides a foundation for creating such intelligent systems.