Multimodal Feature Fusion for Lithium Battery Remaining Capacity Prediction

Accurate prediction of the remaining capacity of lithium batteries is a core challenge in battery health management systems. To address the limitations of existing approaches, which are often sensitive to noise in capacity sequences and insufficient in modeling complex temporal dependencies, this paper proposes a novel multimodal feature fusion method that integrates timedomain, frequency-domain, and state transition representations for remaining capacity prediction. First, Singular Spectrum Analysis is employed to capture the trend components of the original capacity signals. The Transformer encoder is then used to extract time-domain features from the raw capacity sequences. Subsequently, the original signal is transformed into Continuous Wavelet Transform and Markov Transition Field representations to capture frequency characteristics and state transition patterns of the time series, respectively. Convolutional Neural Networks are applied to the CWT spectrogram to extract deep spectral features and to the MTF image to learn temporal correlation features. Finally, the features from different modalities are fused and passed through fully connected layers for final prediction of battery remaining capacity. Experimental results demonstrate that the proposed multimodal fusion model outperforms traditional models, validating its accuracy and effectiveness for battery health management.