Due to the high costs of long-term testing and discrepancies between real-world operating conditions and laboratory environments

labeled data for lithium-ion battery state-of-health (SOH) estimation remains scarce. To address this challenge

this study proposes a semi-supervised SOH estimation method based on self-convolutional neural network (CNN)-transformer with loss regularization (SCT-LR). First

health indicators are extracted through incremental capacity analysis

and a CNN-Transformer hybrid architecture is employed for establishing a mapping model between these indicators and SOH. Second

this paper adopts a self-training semi-supervised learning framework to generate pseudo-labels for unlabeled samples

enabling iterative annotation and utilization of unlabeled data. Finally

to mitigate the impact of label noise in pseudo-labels

this paper introduces a semi-supervised prediction method incorporating a label noise learning module with loss function regularization. The results demonstrate that the proposed method significantly improves SOH estimation accuracy by leveraging unlabeled data. Specifically

it achieves a 12.11% or greater improvement in estimation accuracy compared to supervised learning benchmarks

while the integration of the label noise learning module reduces the average error by 14.41%. This work provides a robust framework for accurate battery SOH prediction in real-world scenarios with limited labeled samples.