A comparative study of battery state-of-health estimation based on empirical mode decomposition and neural network

Accurate estimation of state of health (SOH) in the battery management system furnishes powerful support for ensuring safe and reliable operation of lithium-ion batteries. Data-based neural networks have progressively developed into the most representative solution of SOH estimation. This paper systematically compares three typical neural networks and variants on the accuracy and robustness. Moreover, empirical mode decomposition is firstly adopted to withdraw efficacious health indicators from measurement data acquired during constant current and constant voltage charging. Secondly, Pearson correlation coefficient is applied to elect features with strong characterization from constant-current phase duration, constant-voltage phase duration, constant-current phase time proportion, constant-voltage phase time proportion, and total charge time. Finally, several neural network models such as simple recurrent neural networks, long and short-term memory neural networks (LSTM), gated recurrent units and opposite bidirectional structure are built and compared. Considering the universality and fairness of the results, the novel hyperband optimization seeks optimal configuration for deep learning models through dynamic resource allocation based on the early-stopping strategy and Successive Halving algorithm. Experimental results indicate that LSTM and bidirectional LSTM have higher charge–discharge conditions insensitivity and precision in battery SOH. • Different neural networks are compared and analyzed for battery health prognosis. • The empirical mode decomposition is employed to eliminate capacity regeneration. • Hyperband optimization is exploited to find optimal hyperparameter configurations.