Machine Learning techniques applied to joint state of charge, state of health, and remaining useful life estimation of lithium battery, by monitoring a reduced number of parameters

This Ph.D. thesis is based on the development of machine learning techniques for the accurate estimation of key internal states of lithium-ion batteries, including State of Charge, State of Health, and Remaining Useful Life, using a reduced number of measurable parameters. Deep learning architectures, such as Convolutional Neural Networks, Long Short-Term Memory, and Gated Recurrent Units, are investigated using voltage, current, temperature and electrochemical impedance spectroscopy data, leveraging the electrochemical information contained in the impedance spectra, to improve the accuracy and robustness of battery states estimation. In particular, both equivalent circuit model parameterization and direct impedance-based feature extraction are explored to enable reliable State of Health prediction across different degradation stages. Furthermore, a Tiny Machine Learning framework is developed for embedded State of Charge and State of Health estimation, where deep learning models are optimized for deployment on resource-constrained microcontrollers through projection based neural network compression technique, representing a novel approach for reducing model complexity and memory footprint while preserving estimation accuracy. This enables real-time, low-power, and distributed battery diagnostics directly at the edge, supporting practical implementation in real Battery Management Systems. Additionally, numerical thermal modeling and optimization based on the finite element method are conducted to analyze the spatial temperature distribution and investigate passive cooling and housing design configurations, generating high-fidelity temperature data and providing insight into thermal behavior for improved battery safety, lifetime, and future data-driven model development. The proposed methodologies demonstrate accurate, computationally efficient, and scalable solutions for battery state estimation and thermal aware design, contributing to the development of intelligent, reliable, and deployable battery management systems

Questa tesi di Ph.D. si basa sullo sviluppo di tecniche di machine learning per la stima accurata dei principali stati interni delle batterie lithium-ion, inclusi State of Charge, State of Health e Remaining Useful Life, utilizzando un numero ridotto di parametri misurabili. Vengono investigate architetture di deep learning, come Convolutional Neural Networks, Long Short-Term Memory e Gated Recurrent Units, utilizzando dati di voltage, current, temperature ed electrochemical impedance spectroscopy, sfruttando le informazioni elettrochimiche contenute negli impedance spectra, per migliorare l’accuratezza e la robustezza della stima degli stati della batteria. In particolare, vengono esplorati sia la parameterization degli equivalent circuit models sia l’estrazione diretta di feature basate su impedance, per consentire una previsione affidabile dello State of Health attraverso diversi stadi di degradazione. Inoltre, viene sviluppato un framework di Tiny Machine Learning per la stima embedded di State of Charge e State of Health, in cui i modelli di deep learning sono ottimizzati per il deployment su microcontrollers con risorse limitate attraverso una projection based neural network compression technique, rappresentando un approccio innovativo per ridurre la model complexity e il memory footprint preservando al contempo l’accuratezza della stima. Ciò consente battery diagnostics real-time, low-power e distribuite direttamente at the edge, supportando l’implementazione pratica in real Battery Management Systems. Inoltre, vengono condotti modeling termico numerico e ottimizzazione basati sul finite element method per analizzare la distribuzione spaziale della temperatura e investigare configurazioni di passive cooling e housing design, generando dati di temperatura ad alta fedeltà e fornendo una comprensione più approfondita del thermal behavior per migliorare battery safety, lifetime e il futuro sviluppo di data-driven models. Le metodologie proposte dimostrano soluzioni accurate, computationally efficient e scalabili per battery state estimation e thermal-aware design, contribuendo allo sviluppo di battery management systems intelligenti, affidabili e concretamente deployable.