High-Fidelity State-of-Health Estimation for Li-Ion EV Batteries Using EMA-based Neural Networks

Accurate State-of-Health estimation for lithium-ion batteries is essential for the safety and reliability of electric vehicles and grid storage systems. Existing deep learning architectures, such as CNN-BiLSTM hybrids, face an inherent performance ceiling of approximately 99.96% due to sequential lag and sensitivity to capacity regeneration noise. To address these limitations, this paper introduces the Polynomial-enhanced neural network framework, a novel paradigm that replaces stochastic sequence learning with deterministic manifold reconstruction via high-order topological projection and recursive stochastic noise decoupling. The framework utilizes a Recursive Exponential Moving Average to isolate underlying aging trends from transient electrochemical noise, followed by a 5th-order polynomial projection to linearize the non-linear degradation manifold. Experimental results on the NASA Prognostics Center of Excellence dataset demonstrate that the proposed framework achieves a record-breaking R² of 99.9919% with a Mean Absolute Error (MAE) of 0.0012 and a Root Mean Square Error (RMSE) of 0.0015, effectively eliminating temporal lag and residual oscillations common in recurrent models. This breakthrough in estimation fidelity provides a robust solution for high-stakes Battery Management Systems, ensuring precise End-of-Life predictions and facilitating safe second-life battery utilization.