Patent analysis plays a vital role in innovation statistics and strategic planning, but faces challenges in semantic understanding and structured knowledge extraction. This study proposes a statistical and multimodal framework that integrates BERTopic-based topic modeling, International Patent Classification (IPC) codes, and TF-IDF keywords to capture both semantic and structural features of patent data. An adaptive clustering model generates coherent topic groups, linked to IPC codes and keywords in a multilayer knowledge graph. Experiments show that the proposed method achieves the highest Silhouette and Calinski–Harabasz Index scores, and the lowest Davies–Bouldin Index, outperforming traditional baselines. These findings highlight the effectiveness of multimodal integration and knowledge graph construction in revealing meaningful patterns in complex patent corpora, thus enhancing decision support in innovation analysis.