The increase in battery electric vehicles (BEVs) leads to a growing number of traction batteries reaching the end of their first life in the upcoming yeras. According to the European Battery Regulation (EU) 2023/1542, manufacturers are required to take back these batteries after vehicle use, raising the question of appropriate end-of-life strategies. Although batteries are typically removed when they no longer meet user requirements, they often retain a remaining capacity of 70–80%, indicating a strong potential for second-life applications instead of immediate recycling. A key challenge lies in enabling a fast, reliable, and non-invasive assessment of battery state to support decision-making.
This doctoral research project investigates a methodology for rapid battery state classification based on electrochemical impedance spectroscopy (EIS) data combined with machine learning (ML) techniques. EIS enables non-destructive characterization of battery behaviour, while ML-based clustering facilitates automated condition assessment without extensive testing procedures. To evaluate the feasibility of this approach, a comprehensive battery aging dataset from the Karlsruhe Institute of Technology (KIT) was used, documenting the aging behaviour of 228 lithium-ion cells over two years, including periodic EIS measurements. Each cell was labeled as “Survivor” or “Failed” at the end of the aging period.
Starting from the KIT dataset, two derived datasets were created: a “full” dataset, including measurements across different temperatures and states of charge, and a “filtered” dataset, restricted to 25 °C and 50% state of charge. Four different feature representations were then generated (real/imaginary and magnitude/phase, each in grouped and interleaved form), resulting in a total of eight data–feature combinations. These combinations served as input for three supervised models (Random Forest, Feedforward Neural Network, and Long Short-Term Memory) as well as one unsupervised approach (LSTM autoencoder with K-means clustering).
Initial results indicate that EIS data alone contain sufficient information for reliable class discrimination, with both supervised and unsupervised models achieving classification accuracies above 90%. Among the supervised approaches, Random Forest slightly outperformed the other two models, reaching an accuracy of 97%, compared to approximately 95% for both the Feedforward Neural Network and the Long Short-Term Memory model. The filtered dataset showed only marginally lower performance across all models.
Ongoing work focuses on validating the approach across additional datasets to assess its robustness and generalizability for practical second-life battery screening.