Exploiting unlabeled data for battery state-of-health estimation using transformer-LSTM neural network with semi-supervised learning

Accurate estimation of the state of health (SOH) is critical to ensure the safe and reliable operation of lithium-ion batteries. Existing SOH estimation methods are typically based on supervised learning and require large-scale labeled battery aging datasets. However, the acquisition of accurate data labels is costly and labor-intensive, thereby limiting the effective utilization of the abundant unlabeled aging data. To address this challenge, a hybrid framework integrating semi-supervised learning (SSL) with a Transformer-LSTM neural network is proposed in this study. First, recursive feature elimination is employed to optimize feature selection related to charge capacity and energy, thereby reducing data redundancy. Next, a Transformer-LSTM architecture is constructed to capture both local and global temporal dependencies. Transfer learning is used to adapt pre-trained knowledge from the source domain to the target battery, thereby providing a robust initial representation for the subsequent SSL stage. Finally, a self-training-based SSL strategy is introduced to leverage the synergy between the limited labeled data and abundant unlabeled data. High-confidence pseudo-labeled samples are iteratively added to the training process, enabling the unlabeled data to enhance the supervisory signal and progressively improve the model performance. Experimental results show that the proposed method reduces the root mean square error by 26.27 % relative to the baseline Transformer-LSTM model. These findings confirm the effectiveness of the proposed SSL-based framework in decreasing dependence on labeled data, improving generalization, and enhancing SOH estimation accuracy.