隨著全球人口高齡化趨勢加劇，獨居老人數量持續增長，傳統照護模式面臨人力不足與資源分配失衡的雙重挑戰。本研究提出KGVEC（Knowledge Graph-enhanced Voice-interactive Elderly Care）系統，一套基於知識圖譜增強的居家健康照護語音對話系統，專為獨居長者設計個人化智慧照護服務。系統採用MIAT方法論進行系統化設計，建立完整的IDEF0功能建模與GRAFCET行為建模框架，實現控制器與資料路徑分離的架構設計。技術層面整合Distil-Whisper語音辨識模型、SmolLM2小型語言模型與Kokoro-82M語音合成模型，構建本地化運行的智慧對話平台。核心創新在於建構基於Neo4j的個人化健康知識圖譜系統，透過知識索引層最佳化、檢索策略層最佳化與生成增強層最佳化，實現結構化健康知識與語言模型的有效融合。實驗結果顯示，相較於僅使用語言模型的基礎系統，知識圖譜增強系統在個人化程度方面提升，同時提供完全本地化的隱私保護機制，確保敏感醫療資訊不需上傳雲端。系統成功解決了傳統智慧照護系統缺乏記憶能力、過度依賴雲端運算與使用門檻偏高等關鍵問題，為高齡化社會的健康照護提供具實用性與安全性的技術解決方案。本研究展示了知識圖譜與小型語言模型結合在資源受限環境下的應用潛力，為相關領域技術發展提供重要參考依據。;With the accelerating global aging trend and the continuous growth of elderly living alone, traditional care models face dual challenges of insufficient human resources and imbalanced resource allocation. This study proposes KGVEC (Knowledge Graph-enhanced Voice-interactive Elderly Care), a knowledge graph-enhanced home healthcare voice dialogue system specifically designed to provide personalized intelligent care services for elderly individuals living alone. The system employs the MIAT methodology for systematic design, establishing comprehensive IDEF0 functional modeling and GRAFCET behavioral modeling frameworks, implementing a controller-datapath separation architecture. Technically, the system integrates the Distil-Whisper speech recognition model, SmolLM2 small language model, and Kokoro-82M speech synthesis model to construct a locally-operated intelligent dialogue platform. The core innovation lies in constructing a Neo4j-based personalized health knowledge graph system that achieves effective integration of structured health knowledge and language models through knowledge indexing layer optimization, retrieval strategy layer optimization, and generation enhancement layer optimization. Experimental results demonstrate that compared to baseline systems using only language models, the knowledge graph-enhanced system achieves improvement in personalization degree while providing completely localized privacy protection mechanisms, ensuring sensitive medical information does not require cloud uploading. The system successfully addresses key issues of traditional smart care systems including lack of memory capabilities, excessive dependence on cloud computing, and high usage thresholds, providing a practical and secure technical solution for healthcare in aging societies. This research demonstrates the application potential of combining knowledge graphs with small language models in resource-constrained environments, providing important reference for technological development in related fields.
