SOH Prediction Using Impedance Pseudo-Image Features with Transformer-Based Data Enhancement

Data-driven state of health (SOH) estimation based on electrochemical impedance spectroscopy (EIS) is critical for battery management but currently faces two major challenges: the limited feature representation capability of one-dimensional impedance sequences and the training bias caused by uneven data distribution. To address these issues, this study proposes a novel SOH prediction framework utilizing impedance pseudo-image features with transformer-based data enhancement. First, the gramian angular field is employed to losslessly transform one-dimensional EIS data into two-dimensional pseudo-images, enabling the convolutional neural network (CNN) to capture deep spatial-temporal degradation features hidden in the frequency domain. Subsequently, a parametric generative regression model utilizing the Transformer architecture was constructed for dataset augmentation. Crucially, a physical consistency verification mechanism was incorporated within the generation workflow. Experimental results demonstrate that the proposed method significantly outperforms traditional benchmarks, taking the battery 45C01 as an example, the RMSE decreases from 0.0247 of the traditional CNN to 0.0054, with the error reduced by 78.1%. Furthermore, the MAPE drops from 2.3143% to 0.4853%, indicating a relative error reduction of approximately 79%. The proposed framework enhances both prediction stability and accuracy, offering a robust solution for battery health monitoring under data-imbalanced conditions.