A deep learning approach for enhanced degradation diagnostics of NMC lithium-ion batteries via impedance spectra

Abstract

Electrochemical impedance spectroscopy (EIS) offers valuable insights into the dynamic behaviors of lithium-ion batteries, making it a powerful and non-invasive tool for evaluating battery health. However, EIS falls short in quantitatively determining the degree of specific degradation modes, which are essential for improving battery lifespan. This study introduces a novel approach employing deep neural networks enhanced by an attention mechanism to identify the degree of degradation modes. The proposed method can automatically determine the most relevant frequency ranges for each degradation mode, which can link impedance characteristics to battery degradation. To overcome the limitation of scarce labeled experimental data, simulation results derived from mechanistic models are incorporated into the model. Validation results demonstrate that the proposed method could achieve root mean square errors below 3% for estimating loss of lithium inventory and loss of active material of the positive electrode, and below 4% for estimating loss of active material of the negative electrode while requiring only 25% of early-stage experimental degradation data. By integrating simulation results, the proposed method achieves a reduction in maximum estimation error ranging from 42.92% to 66.30% across different temperatures and operating conditions compared to the baseline model trained solely on experimental data.