Imbalance-robust retired battery sorting via semi-supervised contrastive representation learning

Effective sorting of retired lithium-ion batteries is a prerequisite for safe and reliable second-life deployment, yet it remains challenging in practice due to heterogeneous degradation behaviours, limited labels, and pronounced class imbalance. To achieve imbalance-robust sorting, this paper proposes a semi-supervised contrastive representation learning framework built upon a unified Visual Geometry Group 16-layer network (VGG16). Specifically, state of charge–discharge voltage curves are transformed into pseudo-colour images and processed by a pre-trained VGG16 backbone for efficient end-to-end feature extraction. To reduce redundancy and highlight discriminative information, convolutional feature responses are further exploited to identify discharge segments exhibiting the most prominent inter-class differences, which are then used as the key inputs for subsequent learning. On this basis, a semi-supervised contrastive objective is introduced to improve intra-class compactness and enlarge inter-class separability, with particular emphasis on strengthening minority-class representations under imbalanced data. Experiments on the public dataset and our private dataset validate the effectiveness of the proposed method, achieving 94.34% accuracy and 0.942 F1-score on the public dataset, and 95.35% accuracy and 0.924 F1-score on the private dataset, respectively. These results suggest that the proposed framework can support fast and reliable retired-battery sorting through a rapid test procedure, thereby reducing sorting uncertainty and safety risks, improving pack consistency for second-life assembly, and facilitating scalable and cost-effective battery reuse.