Strain state estimation of lithium-ion batteries based on strain decoupling and a collaborative deep learning network

Mechanical strain generated during the charge/discharge process is a critical physical quantity reflecting internal state and safety margin for lithium-ion batteries. Accurate strain monitoring is valuable for battery health management and safety early warning. However, battery strain signals consist of reversible and irreversible components that are deeply coupled, with fundamental differences in driving mechanisms and time scales, posing severe challenges to high-accuracy estimation over the full battery lifecycle. This paper proposes a strain decoupling strategy and a convolutional neural network–sequence-to-sequence (CNN-Seq2Seq) collaborative network architecture based on the voltage-strain physical coupling mechanism, enabling full-lifecycle strain state estimation from voltage signals. The CNN branch extracts irreversible strain feature vectors from voltage curves and embeds them into the Seq2Seq gated recurrent unit (Seq2Seq-GRU) network as degradation priors, guiding the network to capture strain responses under varying aging conditions. Compared with existing methods that treat strain as a single macroscopic response, the proposed decoupling strategy achieves higher estimation accuracy. The strain decoupling and the degradation-prior conditioning of the network constitute the two main contributions of this work. For full-lifecycle strain estimation, the proposed method achieves a root mean square error (RMSE) of 1.40% and a mean absolute error (MAE) of 1.21%, reducing RMSE by 0.68%, 1.08%, 1.99%, and 3.06% relative to Transformer, GRU, CNN, and Autoencoder long short-term memory (Autoencoder-LSTM), respectively. In addition, cross-condition transfer experiments yield average RMSEs ranging from 1.87% to 2.76% across all transfer directions, validating the generalization of the proposed method.