Multipoint channel charting-based radio resource management for V2V communications

Hanan Al-Tous, Tushara Ponnada, Christoph Studer, Olav TirkkonenEURASIP J. Wirel. Commun. Netw 2020 · DOI

Abstract

This paper introduces a novel framework for managing vehicle-to-vehicle (V2V) communications using Multipoint Channel Charting (MPCC). By applying unsupervised machine learning to radio signals already received by base stations, the system constructs a logical map of vehicles that predicts link quality and outage probability. This approach effectively eliminates the need for GPS-based location reporting or high-overhead signaling, providing a robust solution for resource management in millimeter-wave (mm-Wave) networks.

Introduction

As the automotive industry moves toward fully autonomous driving, the demand for high-bandwidth vehicle-to-vehicle (V2V) and vehicle-to-infrastructure (V2I) communication has surged. Millimeter-wave (mm-Wave) frequencies are ideal for these data-intensive tasks but present significant challenges: signals are easily blocked by urban structures or other vehicles, and channel conditions can fluctuate within milliseconds.

System model illustrating V2V and V2I communication links between vehicles (UEs) and multiple base stations (BSs).

Traditionally, managing these radio resources requires precise knowledge of vehicle locations via Global Navigation Satellite Systems (GNSS). However, GNSS can be power-hungry, inaccurate in "urban canyons," and requires constant reporting that consumes valuable bandwidth. The authors propose solving this by using Channel Charting (CC), a method that allows the network to "locate" vehicles in a logical radio space using only the Channel State Information (CSI) it already collects.

Methodology

The core of the proposed system is the Multipoint Channel Charting (MPCC) algorithm and its out-of-sample extension (EMPCC). The process involves several key steps:

  1. Feature Extraction: Massive MIMO base stations use the MUSIC (Multiple Signal Classification) algorithm to extract multipath components, such as power and phase, from the received CSI.
  2. Dissimilarity Mapping: The system calculates the difference between vehicles' signals. Similar multipath features indicate physical proximity.
  3. Dimensionality Reduction: High-dimensional CSI data is compressed into a 2D or 3D "logical map" using Laplacian Eigenmaps (LE). This preserves the neighborhood structure, ensuring vehicles near each other in reality remain close on the chart.
  4. EMPCC: Unlike standard transductive charting, the developed out-of-sample extension allows new vehicles to be mapped to the existing chart instantly without retraining the entire model.
  5. Link Quality Prediction (LQP): By correlating distance on the logical chart with measured V2V signal-to-noise ratios (SNR), the network builds a model to predict the probability of V2V link outages based solely on chart positions.
Visualization of physical x-y coordinates (top) mapped into logical radio space coordinates (v1-v4) using dimensionality reduction.

Experimental Setup and Results

The study utilized a simulated mm-Wave cellular channel model based on ray-tracing in a Manhattan-style grid with 10 massive MIMO base stations. The training set included 5,000 vehicle locations, while evaluation for V2V prediction involved 1,000,000 random vehicle pairs.

Key Findings

  • Mapping Accuracy: The logical map proved to be a highly reliable proxy for physical proximity, with "Continuity" and "Trustworthiness" metrics consistently exceeding 0.97.
  • Outage Prediction: The LQP was remarkably accurate, with a relative mean square error for predicting V2V outage probability of less than 2% in most scenarios, and only 3% in the most challenging cases.
  • Infrastructure-Led Intelligence: By shifting the computational burden of machine learning to the base station, the system reduces battery drain and processing requirements for the vehicles.

Discussion and Conclusion

The MPCC framework demonstrates that physical coordinates are not strictly necessary for effective radio resource management; a "radio neighborhood" is a sufficient and often more efficient metric. The introduction of EMPCC marks a practical advancement, proving that radio maps remain valid as new vehicles enter the network.

While the simulation results are promising, future work is needed to address real-world dynamics like dynamic shadowing from large moving objects (e.g., trucks) and long-term changes in the physical environment. The authors suggest that deep learning-based architectures, such as autoencoders, could further enhance the robustness of link quality predictions in these complex scenarios.

Summary updated

BibTeX
@article{altous2020multipoint,
  author = {Hanan Al-Tous and Tushara Ponnada and Christoph Studer and Olav Tirkkonen},
  title = {{Multipoint channel charting-based radio resource management for V2V communications}},
  journal = {EURASIP J. Wirel. Commun. Netw.},
  year = {2020},
  doi = {10.1186/s13638-020-01723-3},
}
Figure