Remaining useful life prediction of lithium-ion batteries via an EIS based deep learning approach

Reliable life prediction technology is of great significance in ensuring a safe and efficient lifetime of lithium batteries. However, the traditional health factors such as battery capacity are the effects of battery aging rather than the direct causes, which cannot directly reflect the internal degradation mechanism information of the battery, and the prediction accuracy is easily affected by the working environment of the battery. Electro-chemical impedance spectroscopy (EIS) data can more directly reflect the internal mechanism information of the battery, which includes a wealth of battery aging information. In order to deeply investigate the mapping relationship between impedance spectrum and remaining useful life (RUL) of lithium batteries, EIS method is employed to obtain the impedance and phase of lithium batteries under different health states and temperatures, as well as explore the visualization and quantification of impedance frequency response of lithium battery. Furthermore, the mapping relationship between RUL and the lithium battery impedance is investigated in a full impedance spectrum at different temperatures. It is found that, as lithium battery aging, its negative imaginary parts impedance increases significantly, especially in the middle of the frequency band, and has no significant dependence on temperature. While the real part impedance shows an obvious dependence on temperature. Therefore, it is found that the negative imaginary parts impedance of impedance spectrum has a well-fit characterization ability for RUL of a lithium battery. In this paper, a fusion neural network model of Conv1d-SAM (one-dimensional convolutional neural network-self-attention mechanism) was established with negative imaginary part impedance as input factor to predict battery RUL. The predicted results show that Conv1d-SAM has improved accuracy and stability in RUL prediction, and the mean absolute error function of the proposed model is increased by 72% compared with the latest published method.