適用於智慧物聯網應用之高效能邊緣雲共同計算技術

Abstract

近年來，人工智慧 (AI) 技術快速的興起，強大的 AI 模型逐漸成為我們生活中不可或缺的一部份。人工智慧物聯網 (AIoT) 藉由結合邊緣運算和 AI 模型，在系統上達成傳感器-邊緣-雲的各層級協作，戰略性的將運算任務分配於不同層集中。這種階層式的系統，不僅能善用稀少的硬體資源，還能兼顧複雜問題的處理性，達成最佳的使用者體驗。然而，此議題上有許多潛在問題等待解決，包含在如今面臨的頻寬不足問題下，如何提高整體系統的運作效率和拓展性，是維持智慧物聯網生態系不可或缺的議題。 本論文就「IoT-邊緣端」和「邊緣-雲」兩項不同硬體等級的領域進行探討。此二領域同有著傳輸頻寬不足和運算資源限制的問題，而這也是過去文獻較為專注的部分。然而兩項研究還是有相異之處，在於前者有著 IoT 傳感器的訊號不穩定問題，而後者更多的是 AI模型-系統共同優化的問題。過去的文獻大多只專注於架構的設計，而缺乏特殊問題點的優化、整合性的分析、以及系統層面的優化。這會造成許多運算和網路資料傳輸資源的浪費，在 AIoT 系統內硬體資源非常珍貴的情況下，使的系統運作效率低落，無法完全的發揮系統協作的潛能。 本論文基於上述議題，分別對兩項研究領域進行問題整合，並提出嶄新的架構優化方法。對於IoT-邊緣端的雜訊干擾議題，而本論文提出一個動態調整的架構，先進行訊號品質評估，在進行最佳化的運算，進而使判斷變得更準確，從而提升硬體資源使用效率；本論文還提出一個輕量的訊號品質評估方法，不僅能精準的偵測雜訊大小，具有高度的可更新性，為系統的可擴展性加值。而對於邊緣-雲的共同優化問題，本論文提出一個兩階段的架構優化演算法，使的邊緣-雲中的 AI 模型能夠共同優化；此外，本論文更對網路資料的傳輸做更進一步的分析，使系統能更精確的傳輸真正關鍵的訊息，藉此極大化邊緣-雲的共同合作機制。基於這四項的深耕研究，本論文提及的架構和演算法打破過往的框架，有效的提升 AIoT 系統的效能。

In recent years, the advent of Artificial Intelligence (AI) has rapidly became a part of our daily lives. Artificial Intelligence of Things (AIoT) systems deftly allocate computing tasks across various layers, integrating edge computing with sophisticated AI models to facilitate seamless sensor-edge-cloud collaboration. This structured hierarchy not only optimizes limited hardware resources but also adeptly manages complex challenges to enhance user experiences. Nevertheless, numerous issues remain unresolved, including the enhancement of operational efficiency and scalability within the constraints of limited bandwidth, a critical factor for sustaining the AIoT ecosystem. This dissertation examines the "IoT-edge" and "edge-cloud" domains. Both sectors confront common challenges such as insufficient bandwidth and restricted computational resources, themes prevalent in prior studies. However, distinctions emerge in the focus of these studies: IoT sensors frequently suffer from signal instability, whereas Edge-Cloud interactions primarily involve AI model-system co-optimization. Previous research has concentrated on architectural design yet often overlooks the optimization system-level enhancements. This oversight leads to inefficient utilization of computing capabilities and network data transfers, thereby undermining system efficacy and failing to leverage the full potential of collaborative operations in resource-constrained AIoT environments. To address the deficiencies, this dissertation synthesizes issues from both research domains and introduces an innovative architectural optimization strategy. First, to counteract IoT-edge noise interference, we propose a dynamic architecture that initiates with signal quality assessment to select the best model. We also introduce a lightweight signal assessment technique that not only accurately gauges noise levels but is highly adaptable. Second, regarding the edge-cloud co-optimization challenge, we advocate for a two-stage architectural optimization algorithm that facilitates the joint optimization of AI models among edge and cloud. In addition, we further delve into data transmission optimizations, enabling the system to more effectively relay critical information and maximize the cooperative mechanisms between edge and cloud components. Through these comprehensive studies, the proposed architectures and algorithms transcend traditional frameworks, significantly improving the performance of AIoT systems.