Lithium-ion batteries have wide applications in the energy sector, and effective health prediction is crucial for ensuring their safety and performance. Most machine learning-based State of Health (SOH) estimation methods operate under the assumption that battery data follows the same distribution. However, the statistical characteristics of time-series data may change over time, affecting the accuracy of predictions. To address this issue, an adaptive time-series transfer learning method based on knee point is proposed in this paper. For different battery datasets, a knee method is employed to determine segmentation points of the time series, dividing the battery degradation curve into two stages. Then, healthy features are extracted from battery charge-discharge data for input into the transfer learning model, and the effectiveness of these features is validated. Finally, time-series transfer learning is used to train the health features, exploiting the shared information learned from different domains for SOH prediction. Experimental results demonstrate that the proposed method could achieve good performance across multiple battery data.