Low-cost diagnostic framework for second-life battery grading: from short-duration testing to machine learning classification

<p dir="ltr">With the increasing adoption of electric vehicles (EVs), a growing volume of lithium-ion batteries is expected to enter the recycling and repurposing chain. Repurposing offers a practical route to extend battery life, reduce waste, and ease the burden on recycling systems. Although many of these batteries retain 70-80% of their original capacity, determining which cells are suitable for second-life applications remains a key challenge. Current grading approaches are slow, inconsistent, or rely on laboratory-grade equipment, making them difficult to scale. This thesis investigates whether short-duration electrical tests and diagnostic curve analysis can support reliable, scalable grading of second-life batteries, focusing on their Potential of Reusability Level (PORL).</p><p dir="ltr">The research was structured around five core knowledge contributions, each addressing a specific gap in the development of scalable second-life battery grading methods. The first contribution explored short charge-discharge tests within a real production energy storage system, revealing that although moderate correlations could be observed between voltage response and capacity, results were highly sensitive to starting voltage. To address hardware and accessibility constraints, the second contribution involved the design and evaluation of a low-cost automated battery testing rig. While repeatable, the rig could not replicate results from short charge-discharge tests, suggesting that the complexity of the charge-voltage relationship limits its effectiveness for robust grading.</p><p dir="ltr">The third contribution focused on the development of a standardised incremental capacity analysis (ICA) methodology, establishing a consistent approach for generating IC curves and extracting health features from battery cycling data. Existing literature uses non-standardised methods for generating these curves resulting in inconsistency. This methodology was validated across three independent datasets. </p><p dir="ltr">The fourth contribution centred around the creation and processing of a new long-term sweat testing second life battery dataset, representing six realistic second-life use cases. This dataset, alongside two public datasets was used to evaluate whether meaningful ICA features could be reliably extracted and whether these features were generalisable across different battery types, chemistries, and degradation histories. Finally, the fifth contribution applied machine learning models to these extracted features to classify batteries into five PORL bands. A stacking ensemble model achieved the best performance, with a weighted F1 score of 0.91. However, interpretability and class imbalance remained challenges, and mid-range PORL bands were harder to predict.</p><p dir="ltr">The thesis concludes that while short-duration voltage tests alone are insufficient for grading without prior voltage equalisation, ICA-derived features can support accurate and generalisable PORL classification. Although this thesis did not develop a robust low-cost method of grading SLBs, the framework developed in this thesis provides a foundation for future work looking into automated second-life battery handling systems.</p>