An end-to-end neural network framework for state-of-health estimation and remaining useful life prediction of electric vehicle lithium batteries

This study proposes an end-to-end prognostic framework for state-of-health (SOH) estimation and remaining useful life (RUL) prediction. In such a framework, a hybrid neural network (NN), i.e., the concatenation of onedimensional convolutional NN and active-state-tracking long-short-term memory NN, is designed to capture the hierarchical features between several variables affecting battery degeneration, as well as the temporal dependencies embedded in those features. The prior distribution over hyperparameters, specified to the popular NNs applied in SOH or RUL tasks, is built through the Kolmogorov-Smirnov test. Such prior distribution is regarded as a surrogate to investigate the degeneration data's impact on modeling such NNs. Based on such a surrogate, a Bayesian optimization algorithm is proposed to build SOH and RUL models, selecting the most promising configuration automatically in the sequential evolution progress of hyperparameters. Compared with the existing NNs, the experiments indicate that our method hits a lower average RMSE 0.0072 and global average RMSE 0.0269 for SOH and RUL tasks. Code and models are available at https://github.com/LipenghuaCQ/CNN-ASTLSTM.

Automated summary

This study proposes an end-to-end prognostic framework for state-of-health (SOH) estimation and remaining useful life (RUL) prediction. A hybrid neural network is used to capture hierarchical features and temporal dependencies, with a Bayesian optimization algorithm for automatic configuration selection. The experiments show superior performance in accuracy compared to existing methods.