在物聯網 (IoT) 應用領域中，基於通道狀態信息 (CSI) 的活動識別因其在實現精確的無線定位、行為識別和健康監測系統方面的潛力而引起了極大關注。 CSI 利用無線信號在物理環境中的傳播細節，為開發先進的識別系統提供了基 礎。然而，收集涵蓋各種環境條件和特定行動（如跌倒）的全面數據集具有相當大的挑戰。這一難度在嘗試捕捉罕見或複雜事件時更為顯著，此時需要多樣且廣泛的數據集變得尤為重要。為應對這些挑戰，本研究倡導採用擴散生成模型作為合成原始 CSI 數據的新方法。我們的實驗結果表明，這種方法顯著減少了傳統數據收集所需的時間和資源，同時提升了活動識別系統的性能。本研究的主要貢獻包括： (1) 引入一種創新技術來生成原始 CSI 數據，該技術能夠產生多樣化的合成樣本，促進訓練穩健的分類器；以及 (2) 開發了一種獨特的分類器架構，通過整合來自不同時間段的數據並使用 Transformer 編碼器，克服了傳統時間卷積網絡 (TCN) 的限制。這種整合，結合優化的 TCN 核大小，確保了在活動識別任務中的卓越準確性。;In the realm of Internet of Things (IoT) applications, activity recognition based on Channel State Information (CSI) has garnered significant interest due to its potential for enabling precise wireless localization, behavior recognition, and health monitoring systems. CSI leverages the intricacies of wireless signal propagation in physical environments, providing a foundation for developing advanced recognition systems. Nonetheless, amassing a comprehensive dataset covering a wide array of environmental conditions and specific actions, such as falls, poses substantial challenges. This difficulty is exacerbated when attempting to capture rare or complex events, where the need for a diverse and extensive dataset becomes paramount. In response to these challenges, this study advocates for the adoption of diffusion generative models as a novel methodology for synthesizing raw CSI data. Our experimental findings reveal that this approach markedly mitigates the time and resources traditionally required for data collection, while simultaneously enhancing the performance of activity recognition systems. Main contributions of this research include:(1) The introduction of an innovative technique for generating raw CSI data, capable of producing a varied set of synthetic samples to facilitate the training of robust classifiers; and (2) The development of a unique classifier architecture that surpasses the limitations of conventional Temporal Convolutional Networks (TCNs) by amalgamating data from disparate time segments through a Transformer encoder. This integration, when combined with an optimized TCN kernel size, ensures exceptional accuracy in activity recognition tasks.
