PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep Learning for mmWave Beam and Blockage Prediction Using Sub-6GHz Channels

Muhammad Alrabeiah, Ahmed Alkhateeb

arXiv:1910.02900v3cs.ITeess.SP

TL;DR

The paper asks whether sub-6GHz channels can directly predict mmWave beams and blockage status without relying on intermediate spatial-characteristic extraction. It proves suitable mappings and neural-network learnability, then evaluates an efficient model on DeepMIMO, reporting over 90% blockage-prediction success and near-optimal beam-based rates without beam-training overhead.

  • Problem

    The paper addresses whether sub-6GHz channels can directly predict mmWave beams and blockages, beyond prior approaches that first extract sub-6GHz spatial characteristics.

  • Method

    The paper establishes mapping conditions, uses universal approximation theory to prove neural-network learnability, and develops an efficient deep neural network for both predictions.

  • Results

    More than 90% success probability is achieved for predicting the LOS link status, while the model also predicts optimal mmWave beams and approaches optimal data rates.

  • Takeaways & Limitations

    Sub-6GHz channels can support direct mmWave beam and blockage prediction, with the evaluated deep-learning solution requiring no beam-training overhead.

Abstract

from arXiv · show

Predicting the millimeter wave (mmWave) beams and blockages using sub-6GHz channels has the potential of enabling mobility and reliability in scalable mmWave systems. These gains attracted increasing interest in the last few years. Prior work, however, has focused on extracting spatial channel characteristics at the sub-6GHz band first and then use them to reduce the mmWave beam training overhead. This approach has a number of limitations: (i) It still requires a beam search at mmWave, (ii) its performance is sensitive to the error associated with extracting the sub-6GHz channel characteristics, and (iii) it does not normally account for the different dielectric properties at the different bands. In this paper, we first prove that under certain conditions, there exist mapping functions that can predict the optimal mmWave beam and correct blockage status directly from the sub-6GHz channel, which overcome the limitations in prior work. These mapping functions, however, are hard to characterize analytically which motivates exploiting deep neural network models to learn them. For that, we prove that a large enough neural network can use the sub-6GHz channel to directly predict the optimal mmWave beam and the correct blockage status with success probabilities that can be made arbitrarily close to one. Then, we develop an efficient deep learning model and empirically evaluate its beam/blockage prediction performance using the publicly available dataset DeepMIMO. The results show that the proposed solution can predict the mmWave blockages with more than 90$\%$ success probability. Further, these results confirm the capability of the proposed deep learning model in predicting the optimal mmWave beams and approaching the optimal data rates, that assume perfect channel knowledge, while requiring no beam training overhead...

I. INTRODUCTION

mmWave systems face beam-training and blockage challenges, while sub-6GHz channels offer more robust propagation and low-overhead acquisition. This paper proposes directly predicting mmWave beams and blockages from sub-6GHz channels using theory and deep learning, with DeepMIMO evaluations supporting the approach.

  • Motivation: High mobility and reliability constraints make mmWave operation challenging because beam adjustment has high training overhead and propagation is sensitive to blockages.These challenges become harder with higher frequencies and larger antenna arrays.
  • Motivation: Sub-6GHz channels can generally be acquired with low training overhead and are spatially correlated with mmWave channels.This motivates using sub-6GHz information to reduce mmWave training overhead and maintain reliable links under blockages.
  • Contribution: This paper targets direct prediction of mmWave beams and blockages from sub-6GHz channels, unlike the cited prior work described as not addressing that task.The paper establishes conditions for such mappings and motivates deep learning because the mappings are hard to characterize analytically.
  • Related Work: Prior approaches estimate sub-6GHz spatial parameters to narrow mmWave beam search, leaving sensitivity to estimation error and often retaining mmWave training overhead.They also do not incorporate differences in materials’ dielectric coefficients between the two frequency bands.
  • Contribution: Large enough neural networks can learn mappings for optimal mmWave beams and blockage status with success probabilities arbitrarily close to one.The paper separately identifies conditions for blockage prediction and applies universal approximation theory to both tasks.
  • Evaluation: DeepMIMO evaluations confirm the capability of deep learning models to predict mmWave beams and blockages using sub-6GHz channels.DeepMIMO generates both channel types with a ray-tracing simulator that incorporates materials’ dielectric properties at the two bands.

II. SYSTEM AND CHANNEL MODELS

The paper models dual-band communication in which sub-6GHz uplink channels support mmWave downlink beam selection and blockage prediction. The mmWave beam is analog-only, codebook-constrained, and otherwise requires exhaustive search with substantial training overhead.

  • The base station uses separate transceivers for sub-6GHz and mmWave operation, with sub-6GHz supporting digital channel estimation.
  • The system uses co-located sub-6GHz and mmWave arrays at the base station, while the mobile user has one antenna at both bands.
  • The mmWave transceiver uses one RF chain with an analog beamforming vector selected from a finite codebook.
  • The geometric channel model represents propagation using path gains, delays, and arrival angles, while capturing dependence on environment geometry, materials, and frequency band.
  • Because one analog beam applies across subcarriers and must come from the codebook, achievable-rate optimization is non-convex and requires exhaustive search.
  • Estimating the mmWave channel or performing online exhaustive beam training creates substantial training overhead, motivating sub-6GHz-based beam and blockage prediction.

IV. PREDICTING MMWAVE BEAMS USING SUB-6GHZ CHANNELS

Under a bijective position-to-sub-6GHz-channel mapping, the paper establishes deterministic mappings from sub-6GHz channels to optimal mmWave beams. These mappings are theoretically valid but difficult to characterize analytically.

  • The paper seeks to predict optimal mmWave beams from sub-6GHz channels instead of searching a large mmWave beam codebook.
  • For a given environment, a bijective position-to-sub-6GHz-channel mapping enables a deterministic mapping from sub-6GHz channels to mmWave channels.
  • The bijectiveness condition depends on antennas, array geometry, channel paths, and the surrounding environment, but a few antennas may satisfy it with high probability.
  • Under the bijectiveness assumption, a sequence of sub-6GHz-to-achievable-rate mappings exists for the candidate mmWave beams.
  • The optimal mmWave beam is obtained by selecting the candidate beam associated with the largest mapped achievable rate.
  • Although the mappings exist, their analytical characterization is difficult, motivating deep learning to learn them.

B. Deep Learning Based Beam Prediction

The paper uses neural networks to approximate sub-6GHz-to-achievable-rate mappings and select mmWave beams directly. Training reuses conventional beam-training data, while deployment removes exhaustive mmWave beam-training overhead.

  • A sufficiently large neural network can approximate the achievable rate associated with every candidate mmWave beam with arbitrarily small error.
  • The predicted beam is selected by taking the codebook beam whose neural-network output is largest.
  • Deep Learning Training Phase: During training, one sub-6GHz uplink pilot and exhaustive mmWave beam training provide data points pairing sub-6GHz channels with achievable rates.
  • Deep Learning Deployment Phase: During deployment, the base station uses the sub-6GHz channel directly to predict the mmWave downlink beam.
  • Deep Learning Deployment Phase: The deployment operation saves the overhead associated with exhaustive mmWave beam training.
  • Dataset collection and neural-network training can occur without affecting classical mmWave system operation because beam training would typically occur anyway.
  • Practical Challenges: Practical errors can arise from measurement noise, phase noise, and dynamic scatterers that disrupt the bijective channel mapping or change observed sub-6GHz channels.

V. PREDICTING MMWAVE BLOCKAGES USING SUB-6GHZ CHANNELS

The paper investigates whether sub-6GHz channels can reveal mmWave line-of-sight blockage status. Accurate blockage prediction is motivated by mmWave reliability needs and could support proactive transmission adaptation or handoff.

  • MmWave blockage prediction addresses reliability challenges caused by the sensitivity of high-frequency signals to obstructions.
  • A blocked LOS path can cause a sudden SNR drop, motivating prediction before the link is obstructed.
  • The paper asks whether sub-6GHz channels can determine whether the mmWave LOS link is blocked.
  • Blockage knowledge can support adaptive transmission choices such as changing power or modulation and coding, or handing the session to sub-6GHz.
  • The paper considers deep neural networks as an implementation for inferring LOS blockage status from sub-6GHz channels.

A. Mapping Sub-6GHz Channels to Link Blockages

Under a bijective channel mapping, sub-6GHz channels can distinguish blocked from unblocked LOS links through a continuous discriminant; neural networks can learn this mapping with arbitrarily high success probability under the stated conditions.

  • Assumption: The analysis assumes that a blockage obstructing the mmWave LOS path also obstructs the sub-6GHz LOS path.Co-located arrays make this assumption typically applicable, although sub-6GHz obstruction may reduce power without fully blocking the ray.
  • Theoretical condition: A bijective mapping Ψ yields a continuous discriminant f that assigns blocked and unblocked channel sets to separate classes.The proof uses disjoint channel sets and the Urysohn Lemma.
  • Theoretical condition: The bijectiveness condition requires distinct blocked and unblocked channels at every user position and across users.
  • Learning motivation: Deep neural networks are proposed because the blockage discriminant is difficult to characterize analytically and may be highly nonlinear.The decision depends on spatial and power properties of the rays forming the channel.
  • Practical challenge: Obtaining ground-truth blocked/unblocked labels is a practical challenge, motivating a labeling strategy based on mmWave beam-training results.The paper evaluates this strategy against settings with available blockage-status knowledge.

VI. DEEP LEARNING MODEL

The proposed model treats beam and blockage prediction as classification from shared sub-6GHz inputs, using a common deep network with task-specific output layers and transfer learning.

  • Problem formulation: Beam and blockage prediction are formulated as classification because each selects an option from a beam codebook or binary blockage set.Beam prediction maps each sub-6GHz channel to one of D = |F| classes, while blockage has D = 2 classes.
  • Architecture: A shared base network processes both tasks to reduce training computation and enable transfer learning.
  • Architecture: The base network uses LNN stacks of fully connected, ReLU, and dropout layers, with MNN neurons per fully connected layer.
  • Task-specific outputs: Task-specific output stacks use Softmax classification layers whose dimensions match the beam-codebook size or the two blockage classes.The beam layer has D = |F| outputs; the blockage layer has D = 2 outputs for blocked and unblocked.
  • Transfer learning: Training first fits beam prediction, then replaces the final stack and fine-tunes the shared network for blockage prediction.This strategy offers faster convergence and improved blockage-task performance compared with training from scratch.

B. Learning Model

The learning pipeline normalizes and real-encodes sub-6GHz channels, constructs one-hot labels, trains with cross-entropy, and deploys the trained classifier for direct beam or blockage prediction.

  • Pre-processing: Sub-6GHz complex channels are normalized, split into real and imaginary parts, and stacked across subcarriers into real-valued neural-network inputs.The resulting input dimension is 2 × K × Msub-6.
  • Labels: Beam labels are one-hot vectors marking the optimal codebook beam, while blockage labels use [1, 0] for blocked and [0, 1] for unblocked.
  • Training objective: The supervised model minimizes cross-entropy between target one-hot vectors and predicted class probabilities.
  • Operating modes: Background training collects sub-6GHz channels with beam and, when available, blockage labels; deployment predicts both outputs directly from sub-6GHz channels.
  • Evaluation setup: The evaluation uses two publicly available DeepMIMO ray-tracing scenarios covering LOS beam prediction and blocked/unblocked blockage prediction.Wireless InSite captures channel dependence on frequency, with co-located arrays at 28GHz and 3.5GHz.

B. Dataset Generation

The study builds separate DeepMIMO datasets for optimal-beam and blockage classification, trains a five-stack MLP, and evaluates accuracy and achievable rates under noisy sub-6GHz inputs.

  • Dataset construction: The beam dataset pairs each sub-6GHz channel with a one-hot label identifying the optimal mmWave beam in codebook F.
  • Dataset construction: The blockage dataset combines blocked and LOS channels, selecting users in the marked scenario region and labeling whether the LOS ray is obstructed.
  • Training: The adopted network has LNN = 5 layer stacks and MNN = 2048 neurons per layer, with noisy training and testing samples at target SNRs.
  • Training: Beam prediction is trained from scratch, whereas blockage prediction initializes the shared weights from the best beam model and retrains only the end stack initially.The transfer-learning approach generally converges faster than random initialization.
  • Metrics: The evaluation reports Top-1 and Top-n classification accuracy and achievable rates using predicted beams.
  • Beam results: 85% and 99% are the reported beam-prediction Top-1 and Top-3 accuracies for the adopted setup.Using 30% of the total training subset gives a success probability approximately 12% below the upper bound.
  • Beam results: 81% is the reported Top-3 beam accuracy at 0 dB SNR, while the Top-1 prediction is around 50% at the same SNR.The Top-3 achievable rate is about 6% below the perfect-mmWave-channel upper bound.
  • Beam results: At approximately 15 dB SNR, the Top-1 rate gap falls to slightly below 5%.

F. Blockage Prediction

The blockage-prediction experiments compare ground-truth and power-rule labeling for mixed LOS and blocked users. Transfer learning yields excellent ground-truth classification and exceeds 90% accuracy at high SNRs despite label contamination.

  • Experimental setup: The experiments evaluate blockage prediction on a dataset mixing LOS and blocked users, using ground-truth and power-rule labels.Transfer learning is used to train the deep neural network.
  • Labeling approaches: Fig. 7 shows that the majority of blocked users have power ratios close to one, supporting threshold-based power-rule labeling.The figure presents separate histograms for blocked users and LOS users.
  • Labeling approaches: Ground-truth labeling assumes accurate user labels, for example from simultaneous localization and mapping, but provides an upper bound for other labeling techniques.The paper notes that this approach may not be practical.
  • Labeling approaches: Power-rule labeling uses the strongest-to-second-strongest beam power ratio, which is expected to be large for LOS users and close to one for blocked users.A threshold can therefore be used to create training labels.
  • Results: The model has excellent classification ability across a wide range of SNRs with ground-truth labels, while power-rule labels still produce accuracy exceeding 90% at high SNRs despite label contamination.The comparison uses the best-performing beam-prediction model for blockage prediction.

VIII. CONCLUSION

The paper establishes that sub-6GHz channels can map directly to optimal mmWave beams and blockage status under certain conditions, and uses neural networks to learn these mappings. Experiments show high-fidelity prediction, including more than 90% LOS-link-status success at high SNRs, while identifying practical labeling and environmental-dynamics challenges.

  • Conclusion: Under certain conditions, mapping functions from sub-6GHz channels to the optimal mmWave beam and blockage status exist.These mappings avoid relying on an intermediate extraction of spatial channel characteristics.
  • Conclusion: Large enough neural networks can learn both mappings with success probabilities arbitrarily close to one.This follows from universal approximation theory.
  • Conclusion: The proposed neural network uses sub-6GHz channels to predict optimal mmWave beams and blockage status, with evaluation datasets developed using accurate 3D ray tracing.The model is designed to perform both prediction tasks.
  • Conclusion: The network performs both tasks with relatively high fidelity when trained with enough data, even with noisy sub-6GHz channels.The beam-prediction experiments also reveal a tendency to learn the correct beam direction.
  • Conclusion: More than 90% success probability is achieved for predicting LOS link status at high SNRs, while future work must address environmental dynamics and practical blockage-data labeling.Beam mispredictions often select a beam near the optimum.

APPENDIX A

The appendix proves the beam-prediction success-probability result by expressing the predicted beam through neural-network outputs and bounding the aggregate approximation error. Universal approximation then establishes arbitrarily small prediction error under the stated assumptions.

  • Proof: The proof begins by expressing the success probability of predicting the optimal mmWave beam from sub-6GHz channels.The predicted beam is denoted f̂ and the optimal beam f⋆.
  • Proof: The predicted beam is selected from the outputs of |F| neural networks by applying N(.) and choosing the corresponding beam in codebook F.This connects neural-network outputs to the beam decision.
  • Proof: Defining the maximum approximation error across all beams and applying Proposition 2 and Assumption 2 yields an arbitrarily small error bound for any ϵ > 0.This completes the proof.
Loading 1910.02900v3…