Deep domain adaptation for cross-chemistry battery SOH prediction with relaxation voltage features

Cross-chemistry state-of-health (SOH) prediction for batteries is hampered by disparate electrochemical systems, because prevailing feature engineering methods are tailored for specific chemistry, with limited cross-system adaptability. This paper is the first to present a cross-chemistry SOH prediction framework to enable knowledge transfer from lithium-ion batteries to sodium-ion ones using limited labelled data. To begin with, a comprehensive cross-chemistry dataset is acquired by integrating the Tongji University dataset with our laboratory experimental data from eight lithium iron phosphate batteries and eight sodium-ion batteries, providing five distinct chemical systems for cross-chemistry transfer learning. Relaxation voltage (RV) characteristics are analysed to facilitate domain adaptation learning. Then, a deep neural network model is devised to that leverages convolutional layers to automatically extract RV features, thereby integrating temporal-dynamics modelling with global contextual awareness. Furthermore, a multi-layer maximum mean discrepancy alignment and fine-tuning strategy is proposed to simultaneously align feature distributions across multiple network layers, and to reduce the domain shift between heterogeneous battery chemistry systems by combining an entropy minimisation regularisation and a cosine warm-up scheme. The experimental results show that the proposed method achieves effective cross-domain knowledge transfer through 5 % of the target domain label data, thereby attaining average root mean square error values of 0.65 %, 0.98 %, and 1.55 % across three target domain datasets.