PaperlayerView exact revision on arXiv

Source-linked AI summary

Radar Aided 6G Beam Prediction: Deep Learning Algorithms and Real-World Demonstration

Umut Demirhan, Ahmed Alkhateeb

arXiv:2111.09676v1eess.SPcs.IT

TL;DR

The paper addresses beam-training overhead in mobile mmWave and THz communication by learning beam selection from radar observations. It combines domain-knowledge radar preprocessing with deep learning and evaluates the approach on real-world vehicular measurements, achieving high top-K prediction accuracy and reduced training overhead.

  • Problem

    High-overhead narrow-beam training limits mobile mmWave and THz applications, while practical radar measurements do not map straightforwardly to the best communication beam.

  • Method

    Deep-learning models learn radar-to-beam mappings from preprocessed radar features and are evaluated using the real-world DeepSense 6G dataset.

  • Results

    93.5% top-5 beam prediction accuracy is achieved by the deep-learning solutions, compared with 63% for the classical baseline.

  • Takeaways & Limitations

    The real-world evaluation demonstrates the feasibility of radar-aided mmWave beam prediction for reducing beam-management overhead in highly mobile communication scenarios.

Abstract

from arXiv · show

This paper presents the first machine learning based real-world demonstration for radar-aided beam prediction in a practical vehicular communication scenario. Leveraging radar sensory data at the communication terminals provides important awareness about the transmitter/receiver locations and the surrounding environment. This awareness could be utilized to reduce or even eliminate the beam training overhead in millimeter wave (mmWave) and sub-terahertz (THz) MIMO communication systems, which enables a wide range of highly-mobile low-latency applications. In this paper, we develop deep learning based radar-aided beam prediction approaches for mmWave/sub-THz systems. The developed solutions leverage domain knowledge for radar signal processing to extract the relevant features fed to the learning models. This optimizes their performance, complexity, and inference time. The proposed radar-aided beam prediction solutions are evaluated using the large-scale real-world dataset DeepSense 6G, which comprises co-existing mmWave beam training and radar measurements. In addition to completely eliminating the radar/communication calibration overhead, the experimental results showed that the proposed algorithms are able to achieve around $90\%$ top-5 beam prediction accuracy while saving $93\%$ of the beam training overhead. This highlights a promising direction for addressing the beam management overhead challenges in mmWave/THz communication systems.

I. INTRODUCTION

mmWave and THz systems need narrow-beam gains but face costly beam training in mobile settings. The paper introduces radar-aided machine-learning beam prediction using practical radar and communication measurements.

  • Motivation: Narrow beams provide sufficient receive power, but finding the best beam or beam pair creates high training overhead for mobile applications.The motivation includes vehicular, drone, and augmented/virtual reality communications.
  • Motivation: Radar or other sensory information can help beam selection by revealing transmitter/receiver locations and surrounding-environment geometry.The paper contrasts radar awareness with sub-6GHz channels and user-position information as alternative guidance sources.
  • Contributions: The paper formulates radar-aided beam prediction and develops machine-learning algorithms using classical signal preprocessing to extract relevant features.The study evaluates range-velocity, range-angle, and range-velocity-angle representations.
  • Contributions: The evaluation uses DeepSense 6G, a large-scale real-world dataset containing co-existing mmWave beam-training and radar measurements.The dataset supports evaluation in a realistic vehicular communication scenario.
  • System Model: An FMCW radar at the base station observes the environment while a phased-array mmWave terminal communicates with a mobile user.The radar frame consists of transmitted chirps whose reflections are received, mixed, filtered, sampled, and collected as raw measurements.

B. Communication Model

The communication model uses a narrowband mmWave channel and restricts transmission to a finite beamforming codebook. The optimal beam is the codebook vector maximizing received SNR, while the proposed pipeline uses radar-derived predictions to select it.

  • Communication Model: The base station uses a phased-array mmWave transceiver to communicate with a single-antenna mobile user over a narrowband channel.The channel comprises path gains and transmit azimuth and elevation angles.
  • Beam Codebook: The transmitted data symbol is precoded by a beamforming vector selected from an N-vector beamforming codebook.The codebook restricts the feasible beamforming vectors to its entries.
  • Beam Selection: The optimal beam index is obtained by maximizing the received SNR over the available codebook beams.The paper states that this optimum can be found by exhaustive search.
  • Radar-Aided Selection: Radar observations are preprocessed into useful features and fed to a deep neural network that predicts the base-station beamforming vector.This provides a learned alternative to directly searching the beam codebook.

III. MACHINE LEARNING FOR RADAR AIDED BEAM PREDICTION: PROBLEM FORMULATION

This section frames radar-aided beamforming as a machine-learning problem built on the system model. It introduces the mapping from radar observations to the optimal communication beam.

  • Problem Formulation: The paper defines the radar-based beamforming problem using the previously described communication and radar system models.The section then presents the proposed solution and defines the associated machine-learning task.

A. Problem Definition

The paper learns a mapping from radar measurements to the optimal beam index, using deep models and radar-domain preprocessing to address practical mismatches between sensing and communication data.

  • Problem Definition: Radar measurements X_l are used to infer the optimal communication beamforming vector f_n⋆ for each frame.The formulation uses target position and orientation information available in radar observations when a user is in the base station’s line of sight.
  • Problem Definition: The learning objective is to choose a mapping Ψ_Θ and parameters Θ that maximize accuracy in predicting the optimal beam.The objective is expressed through an indicator of correct beam-index prediction.
  • Motivation for Machine Learning: Direct radar-to-channel mapping is nontrivial because radar measurements differ from communication channel state information.Lookup-table approaches are also affected by radar imperfections, surrounding objects, and nonideal beam patterns.
  • Motivation for Machine Learning: Deep learning is proposed to learn the radar-to-beam mapping while domain-knowledge radar preprocessing reduces the learning problem’s complexity in realistic environments.The approach integrates machine-learning models with radar feature extraction.

C. Machine Learning Task: Radar aided Beam Prediction

The paper formulates radar-aided beam prediction as a multi-class task that returns the most promising communication beams from radar observations. It supports top-K evaluation to assess whether the optimal beam is included among several predictions.

  • Given radar observation X_l, the model returns the index of the optimal beam in the beamforming codebook.
  • Top-K prediction returns an ordered set of the K most likely beam indices rather than only one prediction.
  • The beam prediction task is formulated as a multi-class classification problem.
  • Top-K accuracy evaluates whether the optimal beam is included among the model's K predictions.

IV. MACHINE LEARNING FOR RADAR AIDED BEAM PREDICTION: PROPOSED FRAMEWORK

The proposed framework decomposes radar-to-beam prediction into preprocessing, neural-network, and evaluation functions. It compares radar representations that retain different combinations of range, velocity, and angle information before neural-network inference.

  • The framework integrates radar preprocessing with deep neural networks to reduce learning complexity and support efficient training.
  • Preprocessing: Radar measurements are transformed into range-angle maps, range-velocity maps, or a radar cube containing range, velocity, and angle information.
  • Preprocessing: The extracted radar modalities are standardized and then fed into neural networks.
  • Preprocessing: The radar cube uses a 3D FFT and preserves more dimensions, whereas range-angle and range-velocity inputs use 2D FFTs with smaller input sizes.

2) Neural Network Modeling:

The neural-network component uses compact convolutional models adapted to each radar-input modality and trains them with cross-entropy for beam classification. Their outputs are converted into beam predictions by selecting the maximum-scoring beam.

  • The DNNs contain eight layers, including five convolutional layers with ReLU activations and average pooling to reduce data size.
  • Network input, output, and kernel sizes are adjusted for each radar modality so model size remains reasonable and comparable.
  • The networks are trained with cross-entropy loss because beam prediction is a multi-class classification problem.
  • The target beam is represented by a one-hot vector, while the network produces soft beam scores that are converted into beam indices.

3) Evaluation:

The evaluation selects a single predicted beam by taking the maximum value in the neural network's soft output.

  • The predicted beam is the index corresponding to the maximum neural-network output, implemented as ΨE(.) = arg max(.).

V. REAL-WORLD DATASET

The real-world dataset combines radar and wireless measurements from a stationary 60 GHz testbed as a car passes by in a realistic traffic environment. It contains cleaned target-car samples with optimal beams labeled from measured received power.

  • Testbed configuration: The testbed pairs an FMCW radar and 60 GHz mmWave receiver at a stationary unit with a 60 GHz transmitter on a mobile car unit.The stationary receiver uses a 16-element uniform linear array, while the mobile transmitter remains oriented toward it.
  • Beam and radar measurements: The receiver uses a 64-vector oversampled beamforming codebook, labeling the highest-power combiner as the optimal beam.The radar configuration supports a 45 m maximum range and 56 km/h maximum velocity.
  • Collection scenario: Scenario 9 records radar measurements and received power for every communication beam as a passenger-held transmitter-equipped car passes the stationary unit.Other cars, pedestrians, and bikers were present during collection, preserving realistic environmental conditions.
  • Dataset construction: 6319 samples are retained after manually keeping target-car-in-sight frames, with other environmental elements preserved and a 70/30% training/testing split.The optimal beam is defined as the beam providing the highest measured power.

VI. RESULTS

The evaluation compares radar representations, a lookup-table baseline, and their accuracy, complexity, and data-size behavior. Range-angle deep learning performs best overall, while higher-resolution representations improve accuracy at added preprocessing cost.

  • Evaluation setup: The evaluation measures prediction accuracy, complexity and inference time, and required dataset size across radar-cube, range-velocity, range-angle, and baseline solutions.The DNNs run on the GPU while preprocessing runs on the CPU, so their durations are not directly one-to-one comparable.
  • Beam prediction accuracy: Range-angle deep learning outperforms range-velocity and radar-cube solutions, while deep learning exceeds the 33% baseline top-1 accuracy by at least 8%.Range-velocity remains close to the other deep learning approaches in this scenario.
  • Beam prediction accuracy: 93.5% top-5 accuracy is achieved by range-angle deep learning, versus 63% for the baseline; top-3 range-angle accuracy reaches 79.7%.The other deep learning approaches also improve as K increases, but the baseline gains less.
  • Complexity: The radar-cube, range-velocity, and range-angle DNNs each use approximately 175k parameters, compared with 1024 or 16384 for the lookup-table baselines.The baseline parameter counts correspond to 4-point and 64-point angle FFTs.
  • Complexity: Higher-resolution angle FFTs add substantial preprocessing time, creating a trade-off between beam prediction accuracy and solution complexity.Without angle FFT oversampling, the approaches show similar preprocessing durations.
  • Impact of dataset size: The 64-point range-angle solution begins saturating after 10–20% of the training data, while other solutions require around 20–30%.Range-angle and range-velocity solutions generalize well with smaller datasets, whereas the radar cube saturates later.

VII. CONCLUSION AND TAKEAWAYS

The real-world evaluation demonstrates radar-aided beam prediction feasibility, with deep learning outperforming the classical baseline and range-angle preprocessing offering a performance–complexity trade-off.

  • 45% and 33% top-1 beam prediction accuracy are achieved by the proposed deep learning and lookup table solutions, respectively, over a 64-beam codebook.The comparison emphasizes the promise of radar for mmWave beam prediction and management.
  • 93.5% top-5 beam prediction accuracy is achieved by the deep learning solutions, compared with 63% for the classical baseline.For top-3 prediction, the corresponding accuracies are around 80% and 56%.
  • 92% top-5 beam prediction accuracy is achieved by the range-angle-map solution with 4-point angular FFT processing and 15ms preprocessing and inference time.This operating point illustrates the reported trade-off between performance and complexity/inference time.
Loading 2111.09676v1…