对无线观测信号的建模研究，为无线系统的有效信息挖掘与高性能算法设计提供统一的理论支撑和技术框架，在无线定位和场景感知等领域具有重要的研究价值。随着采集方法的优化以及存储技术的演进，海量、高分辨率的观测信号为无线系统的性能提升与智能化提供了巨大的发展机遇，也催生了信号的理解与表达、推断与生成的新挑战。由于复杂场景下无线观测信号的分布及其传播过程难以显式表达，从中推断出距离和环境等关键参数，以及生成高度仿真的观测信号，都具有很大的挑战性。传统模型驱动的信号处理方法受限于简化的假设，而新型数据驱动的方法泛化性有限，难以推广到不同场景和多样的信号处理任务。为此，本文研究无线观测信号的统一视角模型，围绕信号的推断与生成任务，构建结合变分法和深度学习的有效方法框架。主要的研究内容和贡献总结为如下三个方面：首先，针对基于无线信号的距离推断任务，建立隐变量模型，提出基于深度变分网络的距离分布估计算法。通过分析环境对距离估计的影响，设计双路神经网络学习距离分布的参数，并将算法推广到弱监督学习情形。实验验证了所提方法在复杂环境的距离估计任务上的准确性和在弱监督数据集上的可扩展性。其次，针对基于无线信号的环境推断任务，建立包含环境特征的层次隐变量模型，提出基于深度变分自编码器的联合距离估计与环境识别算法。通过引入距离和环境特征的低层隐变量及其分布的深度高斯假设，从信号中解耦复杂环境特征，进而同时推断出距离和环境标签。实验验证所提方法在复杂环境数据集上的准确性以及在不同数据集上的鲁棒性和泛化能力。最后，针对面向距离和环境推断的无线信号的生成任务，建立了带鉴别变量的层次隐变量模型，提出基于生成式对抗网络的无线信号生成算法。通过引入观测变量的隐式分布假设，学习从标签到复杂观测变量的生成过程。实验验证所提方法所生成的信号能有效模拟真实信号的物理特征，且能为处理距离和环境推断任务的机器学习模型提供有效的训练数据集。综上所述，本文提出了一套无线观测信号的推断与生成任务的深度变分模型与算法框架。所述信号模型通过结合统计理论和深度学习技术，为无线定位与感知的技术研究提供了一种统一、可扩展的建模指导，也为统计推断与深度学习的结合，以及未来无线信号处理的研究提供了算法框架支撑。
The inference and generation of wireless signals play a crucial role in wireless localization and sensing applications. However, inferring positional and environmental features as well as generating realistic observed signals remain challenging due to the intractable distribution of observed signals in complex environments. Traditional model-driven methods are limited due to the simplified assumptions, while data-driven methods lack generalization and scalability. To address these limitations, this paper introduces a unified modeling perspective for wireless signal processing, combining variational inference (VI) and deep learning (DL) to construct effective algorithms.First, a parallel latent variable model (LVM) with environmental variable is established for distance distribution estimation from wireless signals. A deep variational network-based latent variable inference algorithm is proposed, considering the environment‘s impact on observed signals and distance estimation. Experimental results demonstrate the accuracy of the proposed method on distance estimation and scalability in weakly supervised learning schemes.Second, a hierarchical LVM with variable for abstract environmental features is established for concurrent distance and environment inference from wireless signals. A deep variational autoencoder (VAE)-based latent variable inference algorithm is proposed, considering the influence of distance features and refined environment features in signals. Experimental results demonstrate the superiority of the proposed method on both tasks and generalization across different datasets.Lastly, a hierarchical LVM with global discriminative variable is established for wireless signal generation. A deep adversarial generative network (GAN)-based observation variable inference algorithm is proposed, with implicit distribution assumption. Experimental results illustrate that the realism of generated signals in waveform and key physical features, and effectiveness in training practical models.In summary, this paper presents a deep variational model and algorithm framework for the inference and generation problems of wireless signals. It offers a unified and scalable modeling perspective for the complex observed signal processing and provides new algorithm support for future wireless applications.
