智慧感測器節能的信息年齡感知資源分配

Abstract

整合感測與通信是國際電信聯盟無線電通信部門與第三代合作夥伴計畫推薦的第六代及更高世代的潛在應用場景。本論文研究了兩種整合感測與通信環境監測場景：智慧工廠和災害救援。在智慧工廠中，業主可以通過減少能源使用降低運營成本；而在災害救援中，可以通過節約能源延長感測器的壽命。然而，系統效能（如資訊的新鮮度）不應因此受到影響。在此背景下，本論文的三個研究工作的最終目標是，在滿足資訊陳舊度要求的前提下，實現數據傳輸能耗的最小化。

本論文的第一個研究工作探討了一個實時環境監測場景，其中感測器主動收集環境數據並將其傳輸至控制器。第一個工作採用熵風險測度，並結合李亞普諾夫優化的概念，在滿足陳舊度違反概率和超過預定閾值的平均極端陳舊度約束下，實現能耗的最小化。為應對極端陳舊度，本研究融入聯邦學習的框架以生成全局統計數據。數值結果表明，基於聯邦學習的解決方案在建模能耗方面優於集中式基線方法。繼承第一項工作的基礎，第二項研究進一步探討感測器數據採樣頻率與數據傳輸能耗之間的權衡，同時保持資訊的新鮮度。同樣地，第二項工作利用熵風險測度實現全球傳輸能耗的最小化，並滿足陳舊度約束。為尋找最合適的平衡，該研究提出了一個基於聯邦學習的兩階段優化框架，通過李亞普諾夫優化迭代學習最佳採樣頻率與相應的全球傳輸能耗。數值結果顯示，該基於聯邦學習的框架在輕微性能損失的情況下，比集中式基線方案節省了更多計算能耗。

考慮到地球表面活動的不斷增加，第三項研究著重於為災害救援整合臨時地面與非地面感測網絡。為了保持採樣數據的新鮮度，本研究提出了一種方法其包括數據流量安排機制和資源分配最佳化，目的在考慮兩種數據時效性要求（即端到端延遲和條件監測間隔）的情況下，最佳化感測數據傳輸的能耗。具體而言，感測數據流量以動態輪詢方式進行管理，而功率和頻寬分配則採用交替方向乘子法進行最佳化，以滿足上述的的兩種數據時效性要求。除了多場景分析外，數值結果顯示，從簡化提出之方法中節省能量對於系統成本（即目標能量、處理能耗和約束損失總和）而言並不有效。

Integrated sensing and communication (ISAC) is a promising sixth generation (6G) and beyond usage scenarios recommended by the International Telecommunication Union Radiocommunication Sector (ITU-R) and the 3rd Generation Partnership Project (3GPP). This dissertation studies two ISAC environment monitoring scenarios: the smart factory and disaster relief. In the smart factory, the owner can lower the operation costs by reducing energy usage, whereas in disaster relief, the sensor's life cycle can be prolonged by conserving energy. However, the system performance, e.g., the freshness of information, should not be compromised. In this regard, the ultimate objective of the three works in this dissertation is to minimize data transmission energy subject to some staleness requirements.

In this dissertation, the first work studies a real-time environment monitoring scenario, where the sensors proactively collect environmental data and transmit it to the controller. The first work adopts the entropic risk measure (ERM) to minimize the energy consumption subject to the constraints on the staleness violation probability and the average extreme staleness exceedances over a pre-defined threshold by leveraging the concept of Lyapunov optimization. To deal with extreme staleness, this work weaves the framework of federated learning (FL) to create global statistics. Numerical results demonstrate that the proposed FL-based solution outperforms the centralized baseline regarding modeling energy consumption. Inheriting the first work, the second work further investigates the trade-off between the sensor's data-sampling frequency and the data transmission energy while maintaining information freshness. Similarly, the second work minimizes global transmission energy by leveraging ERM and is subject to staleness constraints. To find the most appropriate trade-off, the second work proposes an FL-based two-stage optimization framework to iteratively learn the optimal sampling frequency with the corresponding global transmission energy via Lyapunov optimization. Numerical results show that the proposed FL-based framework saves more computing energy than the centralized baseline with slight performance degradation.

Considering Earth's increasing surface activity, the third work focuses on integrated Ad Hoc terrestrial and non-terrestrial sensing networks for disaster relief. To keep the sampling data fresh, this work proposes an approach, which includes a data traffic orchestration mechanism and a resource allocation optimization scheme, to optimize the sensing data transmission energy while taking into account two staleness requirements, i.e., the End-to-End Delay (EED) and the Condition Monitoring Interval (CMI). Therein, the traffic of sensing data is managed in a dynamic round-robin manner, while the power and bandwidth allocation is optimized for sensing data transmission in an alternating direction approach of multipliers (ADMM) manner regarding the EED and CMI requirements. In addition to the multi-scenario analysis, the numerical results reveal that saving energy by simplifying the approach is inefficient regarding the system cost as the sum of objective energy, processing energy, and constraint loss.