RoSIP-Batt Improves Battery SOH and RUL Prediction

Accurate prediction of a lithium-ion battery's State of Health (SOH) and Remaining Useful Life (RUL) is vital for electric vehicles and other electrified systems. However, jointly predicting these two metrics is challenging due to "task heteroscedasticity"—SOH estimation has bounded, low-variance noise, while RUL prediction involves unbounded, nonlinearly expanding uncertainty. Traditional multi-task learning often struggles to balance these conflicting optimization goals. This research introduces the Rotary SOH-Injected Prior Battery Transformer (RoSIP-Batt), a unified framework designed to resolve these issues. RoSIP-Batt formulates joint prediction as a Bayesian multi-task objective, incorporating a homoscedastic uncertainty weighting mechanism. This mechanism dynamically scales task-specific gradients based on learned residual noise levels, ensuring that the optimization process appropriately handles the different uncertainty profiles of SOH and RUL. The architecture features decoupled dual classification tokens and a gated fusion mechanism, with a gradient-detachment operator preventing high-variance RUL updates from destabilizing the SOH representation. Crucially, it integrates Rotary Position Embedding (RoPE) into a shared Transformer backbone to model degradation patterns without relying on absolute cycle steps, and the intermediate SOH estimate is directly injected into the RUL regression head as a physical degradation prior. Evaluations across multiple datasets (NASA, MIT-Stanford, HUST) show RoSIP-Batt significantly outperforms state-of-the-art baselines, achieving substantially lower SOH and RUL prediction errors, making it a highly generalizable and computationally efficient solution for real-time battery management systems.