In the evolving landscape of 5G new radio and related 6G evolution, achieving centimeter-level dynamic positioning is pivotal, especially in cooperative intelligent transportation system frameworks. With the challenges posed by higher path loss and blockages in the new frequency bands (i.e., millimeter waves), machine learning (ML) offers new approaches to draw location information from space-time wide-bandwidth radio signals and enable enhanced location-based services. This paper presents an approach to real-time 6G location tracking in urban settings with frequent signal blockages. We introduce a novel teacher-student Bayesian neural network (BNN) method, called Bayesian bright knowledge (BBK), that predicts both the location estimate and the associated uncertainty in real-time. Moreover, we propose a seamless integration of BNNs into a cellular multi-base station tracking system, where more complex channel measurements are taken into account. Our method employs a deep learning (DL)-based autoencoder structure that leverages the complete channel impulse response to deduce location-specific attributes in both line-of-sight and non-line-of-sight environments. Testing in 3GPP specification-compliant urban micro (UMi) scenario with ray-tracing and traffic simulations confirms the BBK’s superiority in estimating uncertainties and handling out-of-distribution testing positions. In dynamic conditions, our BNN-based tracking system surpasses geometric-based tracking techniques and state-of-the-art DL models, localizing a moving target with a median error of 46 cm.

中文翻译：

---

用于 6G 协作定位和跟踪的实时贝叶斯神经网络


在5G新无线电和相关6G演进的发展格局中，实现厘米级动态定位至关重要，特别是在协作智能交通系统框架中。面对新频段（即毫米波）更高的路径损耗和阻塞带来的挑战，机器学习 (ML) 提供了从时空宽带无线电信号中提取位置信息并实现增强的基于位置的服务的新方法。本文提出了一种在信号频繁阻塞的城市环境中进行实时 6G 位置跟踪的方法。我们引入了一种新颖的师生贝叶斯神经网络（BNN）方法，称为贝叶斯明亮知识（BBK），它可以实时预测位置估计和相关的不确定性。此外，我们提出将 BNN 无缝集成到蜂窝多基站跟踪系统中，其中考虑了更复杂的信道测量。我们的方法采用基于深度学习 (DL) 的自动编码器结构，利用完整的通道脉冲响应来推断视线和非视线环境中的特定位置属性。通过光线追踪和交通模拟在符合 3GPP 规范的城市微 (UMi) 场景中进行的测试证实了 BBK 在估计不确定性和处理分布外测试位置方面的优势。在动态条件下，我们基于 BNN 的跟踪系统超越了基于几何的跟踪技术和最先进的深度学习模型，以 46 厘米的中值误差定位移动目标。