A Review of State-of-Health Estimation and Prediction Methods for Lithium-Ion Traction Batteries in Vehicular Applications: Task Definition, Benchmark Evaluation, and Deployment Challenges

The state of health (SOH) of traction batteries is a common basis for safety warning, residual value assessment, and second-life utilization of electric vehicles, but a gap exists between the laboratory accuracy of existing deep learning methods and their on-board usability. This paper reviews the relevant research progress using on-board usability as the screening criterion. SOH estimation, SOH prediction, and remaining useful life (RUL) prediction are distinguished according to label source and evaluation protocol, and the limits on the comparability of error values are explained. Existing methods are grouped into six categories by modeling mechanism and evaluated category by category using four on-board questions: input availability, cross-domain generalization validation, uncertainty output, and edge-computing resource constraints. Six models are compared on the CALCE CS2 dataset under random splitting, temporal extrapolation, and cell-level holdout protocols. Under random splitting, the mean absolute errors of all six models do not exceed 0.0150. Under temporal extrapolation, they range from 0.0204 to 0.1392, and the R² values of three neural networks become negative. Under cell-level holdout, they range from 0.0254 to 0.0378. Under the CALCE CS2 dataset and the settings of this paper, protocol selection changes both the error level and the model ranking.