Consistence is a key metric for evaluating quality and performance of lithium battery packs, and grouping is a crucial means for improving consistence and overall performance of battery modules and packs. We introduce a novel framework that combines Knowledge Fusion-based Transformer (KFT) with an improved DPC clustering algorithm. The KFT serves as a data reconstruction model to extract features from multiple sources. In this model, we construct feature extractors based on expert knowledge, extracting features from both the original input samples and the reconstructed outputs of KFT. This achieves the solidification of knowledge, incorporating it into the data features. Finally, grouping is performed using the improved DPC clustering algorithm. After 250 charge-discharge cycles, the average State of Health (SOH) of the module is 92.98%, and the average inconsistence score of the pack is 0.0144, which outperforms the baselines, demonstrating the effectiveness.