A White-Box Equivalent Neural Network Ensemble for Health Estimation of Lithium-Ion Batteries

The recent data-driven solutions for state of health (SoH) estimation of lithium-ion batteries (LIBs) are black-box due to their inexplicable characteristics, thereby making the internal aging parameter variation like capacity or internal resistance difficult to characterize. Therefore, this article proposes a novel white-box equivalent approach for SoH estimation using a practical health indicator (HI) that can be extracted during battery relaxation phase. The main contribution consists of HI-based neural network (NN) design that combines system a priori knowledge with the nonlinear approximation capability. The NN ensemble is able to provide useful insights into cell physics powered with efficient nonlinear approximation capability during real on-board estimation. In proposed model, each circuit element of electrical equivalent circuit model (ECM) model is replaced with a dedicated NN that provides computational intelligence for flexible system identification. The model is validated experimentally for different batteries aged under realistic drive profiles at different temperatures and analyzed in terms of robustness, accuracy, and computational complexity. Furthermore, the estimation performance of the model is compared with conventional white-box ECM and conventional black-box NN along with eight state of art methods from the literature. The result shows that the proposed framework offers 21.94% improvement in mean squared error (mse) when compared to other discussed state of art methods and is more reliable to deploy in battery management system (BMS).