Multimodal fusion dual-path neural network for state of health estimation of lithium-ion batteries using small samples

Abstract

Accurate state-of-health (SOH) estimation is important for battery management, but practical applications are often constrained by limited aging data and nonstationary operating conditions. This article proposes a multimodal fusion dual-path neural network (MFDPNN) for SOH estimation of lithium-ion batteries under small sample conditions. A physics-informed conditional generative adversarial network (PICGAN) is first employed to augment voltage–current degradation sequences. The measured voltage and current signals are then fused by principal component analysis (PCA) to obtain a compact health-related representation, which is further decomposed by variational mode decomposition (VMD) to extract multi-scale components associated with battery degradation. Finally, a dual-path network is designed to learn both local temporal variations and long-term aging trends from the processed signals, and the resulting representation is used for SOH regression. Experiments on the NASA, Oxford, CALCE, and Sino National real-vehicle datasets validate the effectiveness of the proposed method across different battery types and operating conditions. Specifically, across the four datasets, the proposed method achieves average MSE values ranging from 0.410% to 0.490%, average MAE values ranging from 0.410% to 0.610%, and average R^2 values ranging from 0.960 to 0.980. The proposed MFDPNN framework elevates SOH assessment from empirical prediction toward rigorous measurement science, with reduced estimation uncertainty, improved repeatability, and strong cross-dataset reproducibility.