Intelligent frameworks for battery health prognostics: from field variability to scalable solutions

Wong, Kei Long (2026) Intelligent frameworks for battery health prognostics: from field variability to scalable solutions, [Dissertation thesis], Alma Mater Studiorum Università di Bologna. Dottorato di ricerca in Automotive engineering for intelligent mobility, 39 Ciclo.
Documenti full-text disponibili:
[thumbnail of unibo-keilongwong-phd-thesis-20260709.pdf] Documento PDF (English) - Richiede un lettore di PDF come Xpdf o Adobe Acrobat Reader
Disponibile con Licenza: Creative Commons: Attribuzione 4.0 (CC BY 4.0) .
Download (24MB)

Abstract

The widespread adoption of lithium-ion batteries in electric vehicles has increased the need for reliable, scalable, and application-oriented battery health prognostics under real-world operating conditions. Despite advances in data-driven methods, practical deployment remains limited by field operation variability, scarce labeled aging data, privacy and communication constraints in centralized learning, and weak generalization across heterogeneous battery systems. This thesis proposes intelligent frameworks for battery health prognostics that address these challenges from field data analysis to model generalization. First, extensive electric vehicle field-testing data are collected and analyzed to characterize realistic battery usage and operational variability. Based on these data, a comprehensive set of battery health indicators is identified, considering the dynamic, incomplete, and irregular nature of real-world charging and discharging processes. Correlation analysis shows that these indicators maintain strong relationships with degradation trends, providing an interpretable and robust feature basis for prognostic modeling. To mitigate data scarcity, a synthetic battery operation data generation framework is developed by integrating generative adversarial networks with a deep learning-based state-of-charge estimator. The framework generates realistic battery data that preserve temporal dynamics and statistical characteristics of real measurements, improving downstream task performance. To address privacy and scalability concerns, a decentralized battery health forecasting framework based on federated learning is proposed. It enables collaborative training across distributed electric vehicles without sharing raw operational data. Results show competitive forecasting accuracy and robustness under heterogeneous and incomplete data conditions. Finally, this thesis explores generalized battery health prognostics using large language models. By encoding battery operational features into structured natural language prompts, the framework enables few-shot and zero-shot forecasting without task-specific training or fine-tuning, demonstrating generalization across unseen battery types and operating conditions. Overall, this thesis contributes practical and deployable solutions by addressing field variability, data scarcity, decentralized learning, and cross-domain generalization.

Abstract
Tipologia del documento
Tesi di dottorato
Autore
Wong, Kei Long
Supervisore
Dottorato di ricerca
Ciclo
39
Coordinatore
Settore disciplinare
Settore concorsuale
Parole chiave
lithium-ion batteries, battery health prognostics, state-of-health forecasting, data-driven modeling, electric vehicle !eld data, synthetic data generation, generative adversarial networks, federated learning, model generalization
Data di discussione
24 Aprile 2026
URI

Altri metadati

Statistica sui download

Gestione del documento: Visualizza la tesi

^