This paper is devoted to the development of a mathematical model for the analysis of electrochemical impedance spectra of electrochemical power sources. Due to the complex internal structure of modern electrochemical batteries, a new modeling approach, not reliant on manual expertise, is required to analyze large volumes of battery data. While the existing mathematical models are predominantly tailored for online battery management, the battery sorting task is an increasingly important challenge: choosing the best batteries with similar impedance characteristics for assembly into battery packs. Therefore, a mathematical model based on an autoencoder neural network is proposed as an effective tool for dimensionality reduction and removal of outlier batteries. This paper shows the advantages of the developed neural network architecture, composed of recurrent and convolutional layers, over the basic convolutional architecture for formalizing and analyzing impedance spectra. The impact of dataset augmentation by equivalent electrical circuits and utilizing class labels on the accuracy of the autoencoder and its latent space is evaluated. Implementation of the developed mathematical model is compared with other solutions for impedance-based management and sorting of lithium-ion power sources.
Artem Popov (Wed,) studied this question.