Computation-light AI models for Robust Battery Capacity Estimation based on Electrochemical Impedance Spectroscopy

Electrochemical impedance spectroscopy (EIS) holds great promise for assessing battery degradation. Nevertheless, many present methods for extracting battery aging features from EIS data are unsuitable for cells that have very different aging behaviors. Another challenge is that the time complexity of machine learning (ML) models based on full-frequency EIS significantly increases. To tackle these issues, this article investigates two feature extraction methods together with suitable regression algorithms. First, manual feature extraction methods are studied that involve feature extraction via feature correlation and consistency analysis, enhancing the robustness of battery capacity estimation within Gaussian process regression (GPR) and support vector machine (SVM) algorithms. On the other hand, automatic feature extraction methods extract latent features using convolutional neural networks (CNNs) as inputs to the model. Best results are achieved with the CNN-based methods that achieve robust capacity estimation with 2.88% accuracy for different battery cells at varying temperatures while remaining insensitive to EIS measurement uncertainty. This is a factor of 2-3 more accurate than the existing published approaches on the same dataset. By combining the advantages of manual and automatic feature extraction approaches, a reduced EIS dataset is used to train a lightweight CNN model, resulting in an inference time of less than 21.48 ms on STM32L476RG, whereas the existing approaches take up to 50 ms for achieving a similar accuracy.