Vision-powered generative paradigm for robust battery health prediction in real-world electric vehicles

Public summary

* A vision-powered generative paradigm predicts battery state of health (SOH) in real-world electric vehicles.

* Battery signals are transformed into visual representations to identify degradation patterns amid noise.

* Degradation knowledge learned from laboratory data is transferred to real-world electric vehicles.

* Interpretability reveals both visual and numerical modalities capture physics-consistent aging patterns.

* The paradigm enables accurate SOH predictions across different battery chemistries and vehicle types.

Abstract

Predicting the state of health (SOH) of electric vehicle (EV) batteries under real-world operating conditions remains challenging due to noisy measurements, irregular charging behaviors, and highly variable environments, which differ substantially from controlled laboratory settings. This study proposes a vision-powered, multi-modal generative paradigm for accurate EV battery SOH prediction, which reformulates numerical SOH signals into image-based visual representations and leverages pixel-level spatial correlations learned from clean data to reconstruct underlying degradation patterns from noisy EV measurements. The paradigm comprises four key stages: image synthesis, generative model construction, pixel mapping, and SOH prediction. The proposed generative framework integrates a U-Net-based visual channel with a numerical temporal channel, allowing local spatial degradation patterns to be effectively fused with multivariate temporal dynamics. Advanced attention mechanisms are further incorporated to selectively enhance degradation-related representations and to facilitate coherent cross-modal information fusion. The proposed method is evaluated on one laboratory dataset and three real-world EV datasets encompassing diverse battery chemistries, vehicle types, and operational conditions. Experimental results demonstrate that the proposed paradigm effectively mitigates data quality limitations inherent in field measurements and achieves highly accurate SOH prediction, attaining a minimum root mean-square error of 0.0013. Comprehensive benchmarking against 15 state-of-the-art methods further confirms its consistently superior generalization performance across heterogeneous datasets. Interpretability analyses further reveal that both visual and numerical modalities capture physics-consistent aging patterns. These findings highlight the potential of the proposed vision-powered paradigm to extract stable degradation signatures from noisy EV operational data and its applicability to broader health prognostics tasks.