隨著全球人口高齡化加劇，居家安全監測已成為智慧長照系統中不可或缺的關鍵功能。雖然現行多數監測系統採用視覺型技術以提供即時監控，但卻常引發隱私侵擾與監視壓力，進而降低使用接受度與系統實用性。為解決此問題，本研究提出一種基於聲音感測的非視覺監控系統——聲音異常居家活動偵測系統（SAHAD），結合深度學習模型與分散式通訊架構，實現具隱私保障、可擴展且具實用性的異常監測機制。系統依據 MIAT 方法論進行模組化設計，包含三大功能模組：聲音採集子系統負責擷取與傳輸環境聲訊，異常辨識子系統運用深度學習模型進行危險聲音識別，監控介面子系統提供設備管理與即時通報功能。系統採用 DDS（Data Distribution Service）通訊協定，確保多節點間資料傳輸的可靠性與即時性，同時支援動態裝置識別與擴展部署。實驗結果顯示，SAHAD 能準確辨識多種居家與異常聲音，具備高偵測效能與低安裝門檻，特別適用於獨居長者或長照機構的安全需求。相較傳統視覺監控技術，本系統在隱私保護、可用性與擴充性方面具顯著優勢，為智慧長照應用提供具潛力的創新解決方案。;With the rapid growth of the aging population, home safety monitoring has become a critical component of smart elderly care systems. While vision-based monitoring technologies offer intuitive surveillance, they often raise concerns about privacy and intrusiveness, limiting their long-term acceptance. This paper proposes the Sound-based Abnormal Home Activity Detection (SAHAD) system as a non-visual alternative that leverages audio sensing, deep learning, and a decentralized communication framework. The proposed system is developed using the MIAT methodology and consists of three functional modules: an audio acquisition subsystem for environmental sound collection and transmission, an abnormal sound recognition subsystem utilizing deep learning models, and a monitoring interface subsystem for device management and real-time alerting. The system is built on a Data Distribution Service (DDS)-based communication protocol, enabling reliable multi-node data exchange and dynamic device discovery, thereby addressing the scalability and latency limitations of centralized architectures. Experimental results demonstrate the system’s ability to accurately detect a variety of household and abnormal sounds with low deployment complexity. SAHAD provides a privacy-preserving and scalable solution suited for both independent elderly living and institutional care, contributing to enhanced safety without compromising user acceptance.
