Self-evolving multimodal digital twin for health-aware and safety-aware intelligence in electric vehicle lithium-ion batteries

Battery intelligence is a vital component for the safe, energy-efficient, and sustainable adoption of electric vehicles. This study proposes a self-evolving multimodal digital twin for EV lithium-ion batteries that enables concurrent electro-thermal state estimation, health-dependent parameter adaptation, uncertainty-based inference, and anomaly-oriented safety intelligence. Compared with existing battery monitoring systems that mainly depend on voltage, current, and temperature sensing channels, the proposed approach integrates an additional impedance-derived measurement channel with a second-order electrical network, a two-node thermal network, stress-dependent health evolution rules, and an Unscented Kalman Filter for nonlinear state inference. A self-evolving trust-allocation strategy is employed for dynamic redistribution of trust among voltage, temperature, and impedance-derived sensing channels based on innovation consistency, and bounded process-noise inflation is used to enhance estimator adaptability under persistent model–measurement mismatch. An anomaly and thermal-risk accumulation layer is also adopted to generate fault-oriented safety decisions. Four configurations of increasing complexity are considered to evaluate the relationship between estimation accuracy and anomaly sensitivity, and the final v4 formulation is selected as the optimal configuration. The simulation results indicate that the proposed framework delivers an RMSE of 0.0525 for SOC, 0.4049 K for core temperature, 234.92 C for capacity, and 0.01078 Ω for ohmic resistance, in addition to a detection accuracy of 0.9017, recall of 1, and F1-score of 0.8716. Furthermore, the framework is validated using a lithium-ion battery aging dataset, which produces RMSE values of 0.081 V for voltage, 0.6458 K for temperature, 0.0297 for SOC, 0.0231 for SOH, and 1.4 × 10⁻⁷ Ω for the internal-resistance-based impedance proxy. These results demonstrate that the proposed digital twin provides integrated electro-thermal-health estimation, adaptive multimodal fusion, and safety-aware abnormality detection while remaining compatible with real battery cycling data.