arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2504.04410eess.SP
收藏

基于机器学习的自适应光无线网络功率分配

Machine Learning-Driven Adaptive Power Allocation for Optical Wireless Networks

Department of Engineering, Faculty of Natural, Mathematical

Walter Zibusiso Ncube, Ahmad Adnan Qidan, Taisir El-Gorashi, Jaafar M. H. Elmirghani

AI总结:

本文提出了一种结合低复杂度距离基用户关联算法与CNN的混合机器学习方案,用于自适应光无线网络功率分配,以提高性能并降低计算复杂性。

摘要:

垂直腔面发射激光器(VCSELs)由于其高调制带宽、窄谱宽和定向光束,在光无线通信(OWC)中受到青睐,相比发光二极管(LEDs)提供了更高的频谱效率和减少的多径色散。在本工作中,我们探讨了在室内环境下将VCSELs作为接入点(APs)部署,以应对移动性和时间变化的用户分布。为了提高性能,引入了合并接入点(MAP)拓扑结构以扩展每个小区的服务区域,同时采用零强迫(ZF)预编码技术来管理用户间干扰。然后提出一个求和速率最大化问题以维持动态环境下的高质量网络运行。尽管确定性方法可以解决所提出的问题,但它们在实时情况下由于计算复杂性变得不切实际,尤其是在高用户移动性和快速变化的信道条件之下。为此,我们提出了一种混合机器学习(ML)解决方案,结合低复杂度的距离基用户关联算法和卷积神经网络(CNN)用于自适应功率分配。仿真结果表明,所提出的混合关联CNN框架在接近最优性能的同时显著减少了与优化方案相比的计算复杂性。此外,它在实时运行,测量的中位数和P95推理延迟在毫秒范围内,并且保持了一个小的实证最坏情况差距到混合整数线性规划(MILP)最优解,证明了在移动性下的实用性和鲁棒性。

英文摘要:

Vertical Cavity Surface Emitting Lasers (VCSELs) have gained popularity in Optical Wireless Communication (OWC) due to their high modulation bandwidth, narrow spectral width, and directional beam, offering improved spectral efficiency and reduced multipath dispersion compared to Light Emitting Diodes (LEDs). In this work, we explore the deployment of VCSELs as Access Points (APs) in an indoor environment under mobility and time varying user distributions. To enhance performance, a Merged Access Point (MAP) topology is introduced to extend the serving area of each cell, whilst Zero Forcing (ZF) precoding is employed for inter user interference management. A sum rate maximisation problem is then formulated to maintain high quality network operation in the dynamic environment. Although deterministic methods can solve the formulated problem, they become impractical in real time due to computational complexity, particularly under high user mobility and rapidly changing channel conditions. To address this, we propose a hybrid Machine Learning (ML) based solution combining a low complexity distance based user association algorithm with a Convolutional Neural Network (CNN) for adaptive power allocation. Simulation results show that the proposed hybrid association CNN framework achieves near optimal performance while substantially reducing computation complexity relative to optimisation based schemes. Furthermore, it operates in real time, with measured median and P95 inference latencies in the millisecond range, and maintains a small empirical worst-case gap to the Mixed Integer Linear Programming (MILP) optimum, demonstrating both practicality and robustness under mobility.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑