Deep neural network battery impedance spectra prediction by only using constant-current curve

Electrochemical impedance spectroscopy (EIS) is an effective means for monitoring and diagnosing lithium ion batteries. However, its stringent test requirements hinder its wide adoption. In this paper, we propose a deep learning-based method to predict impedance spectra at the fully charged and fully discharged states over battery life using a convolutional neural network (CNN). The CNN only requires input data collected under constant-current charging, which is prevalent in battery applications. A battery degradation dataset that contains over 1500 impedance spectra collected from eight batteries over a wide lifespan is established to validate the proposed method. The results show that the impedance spectra can be accurately predicted with a root mean square error (RMSE)<1.5 mΩ. The effectiveness of the proposed method is also demonstrated by the distribution of relaxation times and the extracted ohmic resistance. Besides, the proposed method can give reliable predictions in the case of incomplete charging data. We demonstrate that using data collected in a 500 mV voltage window, our method can still give reliable predictions with most RMSEs less than 3mΩ. Our method makes EIS a more accessible tool and opens a new way to comprehensively monitor battery performances.