基於能源效益之聯邦式學習應用於無人機通訊之優化;Optimization of Energy-Efficient Federated Learning over UAV Wireless Communication System

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Communication Engineering > Electronic Thesis & Dissertation > Item 987654321/93529

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/93529

Title:	基於能源效益之聯邦式學習應用於無人機通訊之優化;Optimization of Energy-Efficient Federated Learning over UAV Wireless Communication System
Authors:	柯羽軒;Ke, Yu-Xuan
Contributors:	通訊工程學系
Keywords:	無人機通訊;聯邦式學習;軌跡設計;功率控制;連續凸逼近;凸優化;聯合優化設計
Date:	2024-01-25
Issue Date:	2024-03-05 17:42:48 (UTC+8)
Publisher:	國立中央大學
Abstract:	在現代的無人機系統中，結合機器學習技術的應用變得越來越重要，以實現自主性、智能性和高效性，其應用在生活中發揮著關鍵作用，包括監控、搜索和救援、農業、環境監測等。然而，機器學習模型的訓練需要大量的數據，而無人機可能會分散在不同的地點，這些地點可能無法輕易地將數據傳輸到中央訓練伺服器。於是無人機系統通常面臨數據隱私和頻寬限制的挑戰。聯邦式學習（Federated Learning，FL）是一項創新技術，可以有效地應對這些挑戰，利用分散式學習的特性，將無人機系統的性能提升到新的水平。本研究考慮在單一無人機與多個地面基站的場景中，設計一個時槽模型來約束無人機與地面基站的行為，且地面基站作為FL的用戶端進行本地模型的訓練，而無人機做為資料聚合的中心。在一定的準確度下，以最小化整體系統的能耗為優化目標函數，進行聯合優化無人機飛行軌跡、地面基站發射功率、無人機懸停時間與資料大小分配，而此聯合設計是一個高難度的非凸問題。為克服這些設計的難題，本論文利用連續凸逼近的方法，將此問題轉化為一凸問題，再利用凸優化求解。此外，本論文提出無人機飛行軌跡漸進解，簡化了最佳化問題，證明了在長時間的任務時間下的能耗逼近本論文所提出的最佳解。運用聯邦式學習結合單一無人機系統進行聯合設計凸優化，除了能夠解決資料隱私性的問題外，也能夠減少在訓練模型時所產生的能耗，進一步減少整體性統的能耗。 ;In modern unmanned aerial vehicle (UAV) systems, the integration of machine learning technology has become increasingly crucial for achieving autonomy, intelligence, and efficiency. Its applications play a critical role in various aspects of life, including surveillance, search and rescue, agriculture, environmental monitoring, and more. However, the training of machine learning models requires a substantial amount of data, and UAVs may be dispersed across different locations, making it challenging to easily transmit data to central training servers. As a result, UAV systems often face challenges related to data privacy and bandwidth limitations. Federated Learning (FL) is an innovative technology that effectively addresses these challenges by leveraging the characteristics of distributed learning. It elevates the performance of UAV systems to new levels, overcoming issues related to data privacy and bandwidth constraints. This research considers designing a time-slotted model to constrain the behavior of UAV and multiple base stations in a scenario involving a single UAV and multiple base stations. The base stations act as FL clients for training local models, while the UAV serves as the data aggregation center. The objective function is to minimize the overall system energy consumption, subject to a certain level of model accuracy. The optimization involves joint design of UAV flight trajectories, UAV-to-ground station channel gains, ground station transmission power control, and data size allocation. This joint design poses a highly challenging non-convex problem. To overcome the difficulties in this design, this paper employs a method of successive convex approximation (SCA), transforming the problem into a convex one and subsequently utilizing convex optimization for solution. Additionally, the paper proposes an asymptotic solution for UAV flight trajectories, simplifying the optimization problem. It proves that the energy consumption over extended mission times approximates the optimal solution proposed in this paper. The application of FL combined with a single UAV system for joint design convex optimization not only addresses data privacy concerns but also reduces energy consumption during model training, thereby further minimizing the overall system energy consumption.
Appears in Collections:	[Graduate Institute of Communication Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	157	View/Open

社群 sharing

Loading...