Federated domain generalization aims to learn a global model that generalizes well to unseen domains. Feature alignment with prototypes enhances domain-invariant representation learning, aligning inconsistent feature spaces caused by domain skew. However, we identify two key problems that limits feature alignment: local prototypes become unreliable due to mismatched batch normalization statistics and insufficient inter-class separation, while conventional prototype aggregation biases the global feature space toward overrepresented domains. Building on these findings, we propose FedCoda, a cross-domain feature alignment framework that calibrates local prototype geometry and debiases global prototype aggregation. FedCoda first performs calibrated feature alignment by transitioning to global batch normalization statistics in the later training stage, while introducing a prototype calibration loss that encourages well-separated class prototypes. It then applies fair prototype aggregation, which optimizes global prototypes by minimizing the variance of their distances to local prototypes, reducing dominance from domains with more clients. Extensive experiments show that FedCoda consistently improves generalization ability and representation learning under various settings, demonstrating the importance of jointly addressing prototype reliability and aggregation fairness for federated domain generalization.

The continuous evolution of information technology and network expansion pose significant challenges to traditional network operation and maintenance (O\&M), where fragmented, multimodal knowledge and isolated knowledge silos hinder intelligent automation. From the artificial intelligence perspective, we propose a novel entity alignment diagram construction method that semantically associates visual entities to enhance knowledge fusion and retrieval efficiency. This method is integrated with retrieval-augmented generation (RAG) and large language models (LLMs) to build an intelligent question-answering system tailored for network operations. From the engineering application perspective, we develop two specialized datasets to evaluate RAG performance and multimodal question answering in real-world network O\&M scenarios. Experimental results demonstrate that our method outperforms traditional multimodal RAG and pure LLM baselines in answer quality, effectively enabling knowledge fusion and vertical-domain multimodal alignment for practical network O\&M deployment.