Lithium-ion power battery state of health evaluation for electric vehicles based on multi-source transfer learning

Estimating state of health (SOH) of a battery for electric vehicles is constrained by the scarcity of labeled data under real-world operating conditions. While transfer learning addresses data shortages, challenges regarding domain selection and negative transfer remain. This study presents a multi-source transfer-learning framework designed for SOH estimation using unlabeled or limited samples. By implementing a systematic domain selection process based on similarity metrics, the approach identifies relevant source data to mitigate negative transfer. Furthermore, a deep domain adaptation network minimizes discrepancy across multiple sources to enhance generalization. The model achieves SOH estimation errors of 1.64% under unsupervised conditions and 1.58% with limited samples. These findings demonstrate that integrating systematic domain selection with multi-source adaptation significantly improves estimation accuracy, offering a robust solution for battery-health monitoring in practical applications.