随着元宇宙、人工智能、云会议和云计算等新应用的出现，海量数据的激增导致了对高速和低延迟通信系统的需求不断增长。然而，目前的无线频谱资源不足以满足这些需求。可见光通信技术具有带宽大、速度快、抗干扰强的优点，在解决这一问题上具有很大的潜力。与此同时，机器学习领域近年来蓬勃发展，其应用遍布各个领域。机器学习与可见光通信技术的交叉研究具有巨大的潜力，可以解决可见光通信中面临的挑战，并推动该领域的创新发展。此研究主要面向可见光通信的自动调制分类和信号质量估计领域，通过在自动调制分类过程中应用主动学习和迁移学习技术来解决现实世界的可见光通信系统中训练数据不足的问题。此外，采用先进的深度学习模型来探索可见光通信系统中少有研究的信噪比估计领域。所提出的方案利用机器学习算法来降低了训练相关的成本，强化了可见光通信系统的智能性和灵活性，并提高了系统的监测能力。首先，本文对可见光通信的整体发展进行了概述，然后全面回顾了该领域中机器学习应用的研究现状和进展，介绍了机器学习的发展和卷积神经网络的原理。其次，本文提出了应用于可见光正交频分复用通信系统的新型自动调制分类方法。该方法采用迁移学习和主动学习来提高分类准确度，仿真和实验共同证明了其相对于现有方法在数据量较小的情况下改善了识别精度，减少了实际可见光通信系统中数据收集和标记工作所需的成本。此外，当训练样本稀缺时，引入的数据增强技术能够进一步提高分类精度。本文提出的新颖的自动调制分类方法在训练深度学习模型的数据有限的现实场景中展示了巨大的潜力。再次，本文还设计了新型的基于深度学习的方案来估计水下和自由空间可见光通信系统的信噪比，采用了卷积神经网络和视觉变压器模型分别捕获星座图的局部特征和全局特征。实验结果表明，该方法在水下环境下对二进制正交幅度调制(2QAM)、4QAM和8QAM的信噪比估计精度分别达到99.70%、98.00%和94.70%。在自由空间可见光通信中，2QAM、4QAM、8QAM和16QAM的信噪比最高估计精度分别为95.00%、79.67%、58.33%和50.33%。这些结果证明了深度学习模型在精确监测可见光通信网络方面的潜力。最后，本文总结了提出的基于机器学习的方案对未来智能可见光通信发展的重要贡献，并对未来的研究方向进行了展望。
The increasing demand for high-speed and low-latency communication systems has resulted from a surge of massive data, driven by the emergence of new applications such as metaverse, artificial intelligence, cloud conferences, and cloud computing. However, the current wireless spectrum resources prove inadequate to meet these demands. As a solution, visible light communication (VLC) technology presents great potential due to its large bandwidth, high speed, and low interference. Meanwhile, the field of machine learning (ML) has witnessed significant developments in recent years and its applications have expanded across various domains. The incorporation of ML into VLC technology has tremendous potential to address the challenges faced by VLC and drive innovations in the field. In this study, the focus is on two specific aspects of VLC, namely, automatic modulation classification (AMC) and signal quality estimation. The primary objective is to resolve the issue of insufficient training data in real-world VLC systems by implementing active learning (AL) and transfer learning (TL) techniques for the AMC process. Furthermore, the thesis aim to explore the less-studied field of SNR estimation in VLC systems by employing advanced deep learning (DL) models. The proposed methods utilize the potential of ML algorithms to decrease the costs associated with training, enhance the system‘s intelligence and flexibility, and advance the system‘s monitoring capabilities.Firstly, the thesis provide an overview of the development of VLC, followed by a comprehensive review of the latest ML applications in this field, highlighting significant progress made in recent years. Additionally, the development of ML and the principle of convolutional neural network (CNN) are presented.Then, the thesis introduce novel methods for AMC in orthogonal frequency-division multiplexing VLC systems. The proposed approaches employ TL and AL to enhance the classification accuracy, with both simulations and experiments demonstrating its effectiveness compared to existing methods. The proposed method has the potential to reduce the need for data collection and labeling efforts in practical VLC systems. Furthermore, data augmentation techniques are employed to improve the classification accuracy when training samples are scarce. Our proposed AMC methods exhibit promising results for use in real-world scenarios with limited data for training DL models.Moreover, the thesis introduce novel DL-based schemes for estimating SNR in both underwater and free space VLC systems. CNNs and vision transformer models are employed to capture local and global features, respectively. Experimental results demonstrate the efficacy of the proposed technique, achieving high SNR estimation accuracies of 99.70%, 98.00%, and 94.70% for 2 quadrature amplitude modulation (2QAM), 4QAM, and 8QAM in underwater environments, respectively. In free space VLC, the highest achieved accuracies for 2QAM, 4QAM, 8QAM, and 16QAM are 95.00%, 79.67%, 58.33%, and 50.33%, respectively. These findings highlight the potential of DL models to accurately monitor impairments in VLC networks.Finally, the thesis highlights the significant contributions of the proposed ML-based schemes to the development of future intelligent VLC and provides a promising outlook for future research.
