Named Entity Recognition (NER) in classical Chinese is challenging due to ambiguous word boundaries, variant-standard character coexistence, and contextual polysemy. To address these issues, we propose a NER method based on the GuwenBERT-BiGRU-FOCAL-CRF framework enhanced with PGD adversarial training. GuwenBERT, pre-trained on classical Chinese corpora, provides character-level contextual representations, improving the capture of linguistic features. A Bidirectional Gated Recurrent Unit (BiGRU) models long-range dependencies, while a Conditional Random Field (CRF) captures label relations. Focal Loss optimizes training by emphasizing hard-to-classify samples. Robustness is further improved through adversarial training, where learnable perturbations are added to embeddings or intermediate features, enabling stable predictions under minor perturbations. Experiments show that our model outperforms the BERT-BiLSTM-CRF baseline on classical Chinese NER, with F1-score gains of 12.40% and 15.11%. These results confirm the effectiveness and superiority of our approach for named entity recognition in ancient Chinese texts.