Advancing Battery State of Health Estimation and Remaining Useful Life Prediction Through Deep Learning

<p dir="ltr">Lithium-ion batteries are crucial for modern energy storage applications, particularly in electric vehicles and renewable energy systems. However, these batteries inevitably degrade with repeated charging and discharging, resulting in capacity loss, reduced power output, and safety concerns. Accurate estimation of battery state of health (SOH) and prediction of remaining useful life (RUL) are essential for reliable battery management. Existing methods, including model-based and data-driven approaches, often face challenges such as limited data availability, high computational costs, and inadequate generalisation across different battery chemistries and operating conditions. This thesis presents a series methods to systematically tackling these challenges by developing advanced deep learning frameworks for improved SOH estimation and RUL prediction.</p><p dir="ltr">First, a novel battery ageing tag-free training (BATT) method is introduced for SOH estimation using only incomplete charging curves. By eliminating the need for complete charge-discharge cycles, BATT reduces the data collection burden while maintaining high prediction accuracy. The method is validated across multiple datasets, demonstrating its effectiveness in estimating battery health indicators with minimal data requirements.</p><p dir="ltr">Second, a semi-supervised learning approach is developed for RUL prediction, leveraging both labelled and unlabelled data to enhance model generalisation. The proposed architecture incorporates an encoder-decoder structure with multiple decoder heads, enabling robust feature extraction from partially labelled datasets. Comparative analyses with supervised learning models show that this approach improves prediction performance while addressing data scarcity issues.</p><p dir="ltr">Further extending the semi-supervised learning paradigm, this thesis proposes a federated learning method to enable multiple battery users to collaboratively train RUL predictive models without sharing raw data. This method preserves data privacy while enhancing model generalisation across diverse battery chemistries and operating conditions by incorporating massive unlabelled data. Experimental results confirm that federated semi-supervised learning outperforms traditional centralised learning approaches in terms of RUL predictive performance.</p><p dir="ltr">Finally, a physics-informed machine learning approach is proposed to incorporate domain knowledge into the RUL prediction process. By integrating physically meaningful ageing parameters into deep learning models, this approach improves interpretability and prediction accuracy. The incorporation of domain-specific features reduces reliance on purely data-driven techniques, offering a more generalisable and efficient solution for battery health monitoring.</p><p dir="ltr">Overall, this thesis addresses the challenges in short- and long-term battery degradation prediction, including the scarcity of run-to-failure test data and the need for privacy protection. Domain knowledge is further unlocked to enhance the performance of traditional deep learning methods in degradation prediction. The proposed methodologies contribute to the development of scalable, data-efficient battery health management systems, paving the way for more reliable and sustainable energy storage solutions.</p>