Research on MAE Prediction Method for Early Cycle Data-Based Quality Sorting of Lithium-Ion Batteries

锂离子电池的早期质量分类对提升电池整体质量至关重要。然而，仅依赖于预测电池单体的寿命终止(EOL)并不足够。由于锂离子电池老化机理多样、化学变化复杂，以及环境和使用条件的差异，导致相同使用条件下电池单体的老化速率仍然存在显著差异。这一问题直接影响了电池早期质量分类的准确性，并降低了后续电池成组的安全性和可靠性。为应对这一挑战，研究生成了一个数据集，包括154个在相同加速老化条件下循环的18650三元锂离子电池。并提出了一种基于半监督学习的寿命终止(EOL)预测方法。首先利用大量无标签的前3个循环数据对生成式无监督结构掩码自编码器(MAE)进行预训练，以实现自动特征提取，避免了充放电数据人工特征提取的失真和高成本。随后，通过少量标记数据微调预训练模型，最终实现高精度EOL预测。评估结果表明，该方法具有高度准确性：仅使用20%的标记数据即可将验证集上的RMSE降低至2.3%。该研究为利用未标记数据进行电池早期质量分类提供了一种新方法，展示了其应用前景。Early quality classification of lithium-ion batteries is crucial for improving their overall performance. However, relying solely on end-of-life (EOL) prediction of individual cells is insufficient. Due to the diverse aging mechanisms, complex chemical changes, and variations in environmental and usage conditions, even cells operated under identical accelerated aging protocols can exhibit significantly different degradation rates. This variation directly affects the accuracy of early quality classification and compromises the safety and reliability of battery pack assembly. To address this challenge, a dataset comprising 154 18650-type ternary lithium-ion batteries cycled under identical accelerated aging conditions was developed. A semi-supervised learning approach is proposed for EOL prediction. Specifically, a generative Masked Autoencoder (MAE) is pre-trained using abundant unlabeled data from the first three cycles to enable automatic feature extraction, thereby avoiding the distortion and high costs associated with manual feature engineering from charge/discharge data. The pre-trained model is then fine-tuned using a small amount of labeled data to achieve high-accuracy EOL prediction. Evaluation results show that this method is highly effective: using only 20% of the labeled data reduces the RMSE on the validation set to 2.3%. This study presents a novel approach for early quality classification of lithium-ion batteries using unlabeled data and demonstrates promising application potential.