Multi-Modal Feature Fusion Network for State-of-Health Estimation of Lithium-Ion Batteries

For the safe and reliable deployment of lithium-ion batteries, accurate state of health (SOH) estimation is paramount. However, most existing data-driven methodologies depend exclusively on single-modal data, such as voltage-capacity or incremental capacity (IC) curves. Such limited data frequently fails to offer a holistic understanding of the complex battery degradation process. To address this limitation, this paper proposes a novel multi-modal feature fusion network. This network can effectively combine three different but complementary data modalities: historical point features, voltage-capacity and IC sequence features, as well as degraded image features. To this end, the framework incorporates a one-dimensional convolutional neural network (1D-CNN) for analyzing point features, leverages a Transformer encoder to process sequence features, and employs ResNet for identifying spatio-temporal patterns in degraded images. These heterogeneous features are then collaboratively integrated through a fusion network. This model was validated on the CSIE dataset, and the results showed that its performance was superior to that of the single-modal method. At 0.5C discharge rate, the average RMSE of the fusion model was 0.34%, the MAPE was 0.31%, and the R² reached 0.9937. In addition, this method demonstrates excellent robustness at different discharge rates. Even at a high discharge rate like 3C, it can still maintain high accuracy.