Class-Incremental Learning (CIL) aims to develop a classification model where the number of classes increases gradually over time. CIL needs to address the stability-plasticity dilemma between learning new and existing classes: high plasticity can lead to catastrophic forgetting of old classes, while high stability can impair the model's ability to learn new classes. To alleviate this issue, unlike previous research that starts with a small number of classes, this paper explores a more realistic CIL scenario, beginning class-incremental learning with a model pre-trained on a large set of base classes. For this scenario, this paper proposes a Personalized Feature Combination Network (PFCN), which incorporates a three-stage training strategy. The first stage is feature expansion, where parts of the branches are cloned and expanded, and the network is fine-tuned to learn new data. The second stage is feature separation, where features are separated into global and personalized features using a conditional strategy. The final stage is knowledge fusion, where a unified probability distribution is generated through feature combination. Extensive experiments were conducted on image classification datasets, comparing the proposed model with current state-of-the-art CIL benchmark models. The results show that the proposed model achieves superior performance.