Deep neural network based battery impedance spectrum prediction using only impedance at characteristic frequencies

Electrochemical impedance spectroscopy can be used for characterizing and monitoring the state of batteries. However, the difficulty in the onboard acquisition limits its wide applications. This work proposes a new method to obtain the impedance spectrum by using convolutional neural network, which uses the impedance measured at several characteristic frequencies as input. The characteristic frequencies are determined according to the time constants corresponding to the characteristic peaks and valleys of contact polarization and solid electrolyte interphase growth processes from the distribution of relaxation time. The proposed method is validated based on the dataset which contains the impedance spectra of eight batteries over the whole life cycle. The predictions coincide with the ground truth, with a maximum root mean square error of 0.93 mΩ. The developed method can also be quickly adapted to acquire the impedance spectrum of other batteries with different chemistries and be used for predictions of various battery states based on the transfer learning approach.