CARDNet: A Lightweight Regeneration-Aware Network for ElectricVehicle Battery State-of-Health Estimation from Weekly Telematics Data

Abstract Accurate state-of-health (SOH) estimation is essential for safe and economical electric vehicle (EV) operation, yet real-world fleet telematics typically provides short, weekly aggregated sequences with pronounced capacity regeneration and substantial cross-vehicle heterogeneity, challenging deep models developed for dense laboratory data. This paper proposes CARDNet, a lightweight network that isolates high-frequency input fluctuations, where regeneration dynamics are concentrated, through a parameter-free moving-average decomposition. The extracted component is encoded by a compact convolutional branch and complemented by a global feature-interaction path that captures cross-feature nonlinearities. We further provide an explicit operational definition of capacity regeneration, under which the studied fleet exhibits 117 regeneration episodes that transiently reverse capacity equal to 48.8% of the net decay. Experiments on 20 real EVs monitored for over two years use a unified one-week-ahead prediction setting and equal per-model hyperparameter search for all classical and deep baselines. CARDNet achieves the best held-out fleet accuracy with an RMSE of 1.040 Ah using only 53.7k parameters. Block-bootstrap confidence intervals separate CARDNet from all learned baselines on the held-out fleet, while leave-one-vehicle-out tests show statistically reliable gains over statistical forecasters and kernel regression and comparable mean accuracy to the strongest tree-ensemble, convolutional, and tuned attention baselines. Among this tied group, CARDNet provides superior stability and efficiency, achieving the lowest across-fold variance and worst-case error. On a single CPU core, it requires only 219 KiB of weights and 0.10 ms per inference, versus 3.5 MiB and 13.1 ms for the closest classical competitor. Error decomposition shows broad gains across week types, while a regeneration-weighted objective reduces regeneration-episode error by 5.2% at a 1.2% cost in overall accuracy.