Deep Learning Enabled Beam Tracking for Non-Line of Sight Millimeter Wave Communications

Ruiyu Wang, Paulo Valente Klaine, Oluwakayode Onireti, Yao Sun, Muhammad Ali Imran, Lei ZhangIEEE Open J. Commun. Soc 2021 · DOI

Abstract

Millimeter-wave (mmWave) communications are essential for 5G and beyond, but they suffer from high path loss and susceptibility to blockage. In Non-Line-of-Sight (NLOS) environments, beam alignment becomes a complex "needle-in-a-haystack" problem. This paper presents a deep learning-based framework that integrates a Deep Neural Network (DNN) with a trajectory prediction algorithm to proactively track beams. By mapping user locations to optimal reflection paths, the proposed system achieves a 42,000-fold reduction in computational complexity compared to exhaustive search methods while maintaining high angular precision.

Introduction

The transition to mmWave frequencies offers the massive bandwidth required for modern mobile networks. However, the short wavelength of mmWave signals makes them easily obstructed by urban infrastructure. To maintain connectivity in NLOS scenarios, base stations must rely on directional beams that reflect off environmental surfaces. Traditional beam alignment relies on exhaustive search—scanning all possible Angle of Arrival (AOA) and Angle of Departure (AOD) pairs. This reactive approach is often too slow and power-hungry for mobile users.

Spherical coordinate system defining the azimuth (phi) and elevation (theta) angles used for beam tracking.

This research proposes a proactive solution: using the user's location as a proxy for the complex geometry of the urban environment. If the system can predict the user's trajectory and map their location to the best reflection paths, it can maintain a stable connection without the overhead of constant scanning. The core insight is that in a static urban environment, optimal reflection paths are essentially a function of geometry.

Methodology and Dataset

The proposed framework consists of two primary components: a predictive DNN and a trajectory tracking algorithm.

Deep Learning Framework

A multi-layer perceptron with four hidden layers was developed to predict the optimal AOA and AOD. The model utilizes Xavier initialization to stabilize training and a 40% Dropout rate to ensure generalization. To enhance performance, the authors employed K-means clustering on raw channel data, adding a "cluster ID" as a feature. This allows the DNN to categorize propagation environments more effectively.

Trajectory Prediction

To account for real-time mobility, the system integrates the Dynamic Window Approach (DWA). DWA is a robotics-inspired algorithm that predicts a user's future path by considering physical constraints like velocity and acceleration, ensuring the beam is ready for the user's next position before they arrive.

Simulation Environment

The system was validated using a high-fidelity synthetic dataset generated via Wireless InSite ray-tracing software. The simulation modeled the University of Glasgow’s Gilmorehill campus, covering over 73,000 grid positions to provide a robust training set for NLOS scenarios.

3D ray-tracing model of the University of Glasgow campus used to generate the NLOS training dataset.

Results and Discussion

The deep learning approach demonstrated significant improvements over traditional methods:

  • Efficiency: The inference-based prediction is approximately 42,000 times faster than an exhaustive search.
  • Accuracy: The DNN predicted beam angles with an average error of only ±2°, well within the ±7° tolerance required for stable mmWave links.
  • Reliability: Connection loss occurred in only 0.1% to 0.5% of tested positions, even in complex NLOS environments.

The study found that standard DNNs outperformed Convolutional Neural Networks (CNNs) for this specific task because the input features (coordinates and power levels) are tabular rather than spatial/pixel-based, where CNNs typically excel.

Conclusion and Future Work

This research demonstrates that deep learning can effectively solve the beam tracking challenge in NLOS mmWave communications. By leveraging environmental geometry through location-based prediction, the system drastically reduces latency and power consumption.

However, limitations remain. The current model was trained on a static campus map; real-world dynamic obstacles like large vehicles could alter reflection geometry. Future work will focus on making the model more adaptive to environmental changes and extending the framework to multi-user, multi-base station scenarios.

Summary updated

BibTeX
@article{wang2021deep,
  author = {Ruiyu Wang and Paulo Valente Klaine and Oluwakayode Onireti and Yao Sun and Muhammad Ali Imran and Lei Zhang},
  title = {{Deep Learning Enabled Beam Tracking for Non-Line of Sight Millimeter Wave Communications}},
  journal = {IEEE Open J. Commun. Soc.},
  year = {2021},
  doi = {10.1109/ojcoms.2021.3096118},
}
Figure