A GPT-powered automated feature extraction framework for state of health estimation of fast-charging batteries

Efficient and reliable feature extraction engineering plays a crucial role in improving the accuracy and real-time operational capability for data-driven state-of-health (SOH) estimation of lithium-ion batteries. However, conventional feature extraction methods are time-intensive and prone to manual bias, particularly under the multi-step constant current fast-charging protocols prevalent in practical scenarios. To address this problem, this study introduces a generative pre-trained transformer (GPT)-powered automated feature extraction (AFE) method that identifies a series of health features from voltage and capacity aging data in both the time domain and frequency domain. Observation and analysis reveal that with increasing temperature, capacity-related features exhibit a more stable correlation with the SOH compared to voltage-related features. Based on this insight, we propose a two-stage kernel extreme learning machine-autoencoder (TS-KELM-AE) model, which integrates deep feature extraction and nonlinear mapping capabilities to estimate SOH under varying temperatures. Compared to manual feature extraction (MFE) and mainstream deep learning as a feature extractor, the proposed AFE method is feasible and efficient, and the TS-KELM-AE model achieves higher accuracy. Furthermore, analyses of feature importance, model multicollinearity, and cross-battery generalization under diverse operating conditions demonstrate that integrating AFE with the TS-KELM-AE framework establishes a robust and scalable solution for practical applications.