State-of-Health Estimation of Electric Vehicle Batteries From Multisource Onboard Sensor Data Using RGMamba

Abstract

Accurate state-of-health (SOH) estimation is essential for the safe operation and maintenance of electric vehicles. Battery management systems (BMSs) continuously record multisource onboard sensor signals, but measurement noise and varying operating conditions make it difficult to extract reliable indicators of battery degradation. This article proposes an SOH estimation framework that integrates cross-scale features extracted from multisource onboard data with a residual gating Mamba (RGMamba) network. During charging, incremental capacity features are extracted from cell-voltage and pack-current measurements to characterize battery aging. During discharging, vehicle-speed data are used to derive acceleration-based user-behavior features while charging temperature and accumulated mileage are incorporated to describe thermal and usage conditions. A feature-gating module adaptively weights the cross-scale features, the Mamba module captures long-term dependencies in battery degradation, and a residual module refines the nonlinear relationship between the learned features and SOH. The framework is evaluated using real-world operating data from seven electric vehicles, with the reference SOH derived from onboard state-of-charge (SOC) and current measurements. For the benchmark vehicle, RGMamba achieves a root-mean-square error (RMSE) of 0.0474 and a mean absolute percentage error (MAPE) of 4.40%. After fine-tuning with the first 20% of each target vehicle’s data, the MAPE values for the other six vehicles range from 3.17% to 6.54%. These results indicate that feature-level fusion of multisource onboard data, combined with target-vehicle fine-tuning, can support reference SOH estimation across vehicles equipped with the same battery system.