Deep Learning-based mmWave MIMO Channel Estimation using sub-6 GHz Channel Information: CNN and UNet Approaches

Basic Information

• Paper ID: 2506.11714
• Title: Deep Learning-based mmWave MIMO Channel Estimation using sub-6 GHz Channel Information: CNN and UNet Approaches
• Authors: Faruk Pasic, Lukas Eller, Stefan Schwarz, Markus Rupp, Christoph F. Mecklenbräuker
• Institution: Institute of Telecommunications, TU Wien, Vienna, Austria
• Classification: eess.SP cs.IT math.IT
• Publication Date: October 15, 2025 (arXiv v2)
• Paper Link: https://arxiv.org/abs/2506.11714

Abstract

Future wireless multiple-input multiple-output (MIMO) systems will integrate sub-6 GHz and millimeter-wave (mmWave) frequency bands to meet growing high data rate demands. MIMO link establishment typically requires accurate channel estimation, which is particularly challenging at mmWave frequencies due to low signal-to-noise ratio (SNR). This paper proposes two novel deep learning-based approaches for estimating mmWave MIMO channels by leveraging out-of-band information from the sub-6 GHz band. The first approach employs a convolutional neural network (CNN), while the second utilizes a UNet architecture. Simulation results demonstrate that the proposed out-of-band assisted deep learning methods outperform existing alternatives in terms of achievable spectral efficiency.

Research Background and Motivation

Problem Definition

1. Core Challenge: Channel estimation problem in mmWave MIMO systems, particularly the estimation difficulties caused by low pre-beamforming SNR
2. Technical Requirements: Future 5G/6G systems need to simultaneously utilize sub-6 GHz and mmWave bands to achieve high data rate transmission
3. Existing Limitations: Traditional methods rely solely on in-band mmWave information, constrained by low SNR, resulting in limited performance

Research Significance

• mmWave communication offers tremendous bandwidth potential but suffers from high propagation loss, requiring large antenna arrays and MIMO techniques
• Accurate channel estimation is critical for link establishment in mmWave MIMO systems
• Multi-band measurements indicate that sub-6 GHz and mmWave bands exhibit similar multipath characteristics, but sub-6 GHz provides higher pre-beamforming SNR

Limitations of Existing Methods

1. Pure In-Band Methods: Use only mmWave channel information, constrained by low SNR
2. Traditional Out-of-Band Methods: Such as pilot-assisted approaches in references [12]13, which do not fully exploit machine learning to discover inherent relationships between in-band and out-of-band channel estimation
3. Deep Learning Methods: Existing CNN, FFDNet, ResNet-UNet and other methods rely solely on in-band information

Core Contributions

1. Proposes Two Novel Deep Learning Methods: CNN and UNet-based mmWave MIMO channel estimation methods that effectively utilize sub-6 GHz out-of-band information
2. Improves Existing Out-of-Band Assisted Methods: Incorporates machine learning into the MRC method from reference 13, better exploiting the relationship between in-band and out-of-band channel estimation
3. Comprehensive Performance Evaluation: Compares with pure in-band methods and existing out-of-band methods using normalized channel MSE and spectral efficiency metrics
4. Significant Performance Improvement: Achieves 3.35% and 4.03% median spectral efficiency improvements over the MRC method for CNN and UNet, respectively

Methodology Details

Task Definition

• Input: In-band mmWave channel estimate H̃^(m)n, out-of-band auxiliary channel estimate Ĥ^(m)n, estimated sub-6 GHz K-factor K̃^(s), noise variance Γ^(m)
• Output: Improved mmWave channel estimate H̄^(m)n
• Objective: Minimize normalized channel estimation MSE, maximize achievable spectral efficiency

System Model

Channel Model

Employs a frequency-selective Rician fading channel model:

H^(b)[n] = √(η^(b)) * [√(K^(b)/(1+K^(b))) * H_fs^(b) + √(1/(1+K^(b))) * H_rp^(b)[n]]

which includes deterministic line-of-sight (LOS) components and Rayleigh scattering components.

Link Establishment Process

1. Training Phase:
   • Step 1: Transmit non-precoded pilot symbols on both frequency bands
   • Step 2: Utilize sub-6 GHz channel estimates for angle estimation, transmit precoded pilots on mmWave band
2. Data Transmission Phase: Use improved channel estimates for SVD precoding and combining

Model Architecture

1. MRC Baseline Method

Maximum ratio combining method:

H̄^(m)[n] = ŵ * Ĥ^(m)[n] + (1-ŵ) * H̃^(m)[n]

where ŵ is an adaptive combining factor determined by K-factor and noise variance.

2. CNN Architecture

• Input: 6 real-valued matrices (real and imaginary parts of H̃^(m) and Ĥ^(m), K̃^(s), Γ^(m))
• Structure: 9 convolutional layers, each with 64 3×3 filters, ReLU activation, batch normalization, zero padding
• Output Layer: 2 channels (real and imaginary parts), Tanh activation function
• Mapping Relationship: H̄^(m)n = f_ΦC(H̃^(m)n, Ĥ^(m)n, K̃^(s), Γ^(m); ΦC)

3. UNet Architecture

• Encoder-Decoder Structure: Two encoders (contracting path) and two decoders (expanding path)
• Distinctive Design: Skip connections directly link encoder and decoder layers, transmitting feature information
• Advantages: Capable of capturing global information and fine-grained details, more robust to noise
• Layer Configuration: Convolutional layers with 64→128→256 channels, coupled with max pooling and upsampling

Technical Innovations

1. Multi-Modal Information Fusion: First combines sub-6 GHz out-of-band information with deep learning for mmWave channel estimation
2. Adaptive Architecture Design: Network adapts to different K-factor and SNR conditions
3. Skip Connection Mechanism: UNet's skip connection design better preserves spatial localization information
4. End-to-End Optimization: Directly optimizes channel estimation quality related to spectral efficiency

Experimental Setup

Dataset

• Training Samples: 500,000 channel realizations
• Test Samples: 500,000 channel realizations
• Parameter Distribution:
  • SNR γ^(m): Uniformly distributed -20, 10 dB
  • K-factor K^(m): Uniformly distributed -20, 30 dB
  • AoD ϑ, AoA φ: Uniformly distributed -90, 90 degrees

Simulation Parameters

Evaluation Metrics

1. Normalized Channel MSE:

   NMSE = (1/L_r N^(m)) * Σ ||H_l^(m)[n] - H̄_l^(m)[n]||_F^2 / ||H_l^(m)[n]||_F^2
   

2. Achievable Spectral Efficiency:

   SE = (1/N^(m)) * Σ Σ log2(1 + SINR_μ[n])
   

Comparison Methods

• Pure In-Band Methods: Traditional estimation, in-band CNN, in-band UNet
• Out-of-Band Assisted Methods: MRC, proposed CNN, proposed UNet
• Ideal Upper Bound: Perfect channel state information (CSI)

Implementation Details

• Loss Function: Normalized channel estimation MSE
• Optimizer: Adam optimizer
• Batch Size: 64
• Training Epochs: 200
• Single Subcarrier Training: Training uses only the center subcarrier, deployed across all subcarriers

Experimental Results

Main Results

Performance Comparison Under Different K-Factors

1. K = -20 dB: In-band and out-of-band assisted ML methods perform comparably, both superior to MRC
2. K = 10 dB: MRC outperforms in-band CNN/UNet, but out-of-band assisted CNN/UNet exceed MRC at all SNR levels
3. K = 20 dB: Out-of-band assisted ML methods significantly outperform MRC and in-band methods

Spectral Efficiency Improvements

• In-Band CNN vs Traditional Method: 86% median spectral efficiency improvement
• In-Band UNet vs Traditional Method: 88% median spectral efficiency improvement
• MRC vs Traditional Method: 151% median spectral efficiency improvement
• Out-of-Band CNN vs MRC: 3.35% median spectral efficiency improvement (159% total)
• Out-of-Band UNet vs MRC: 4.03% median spectral efficiency improvement (161% total)

Key Findings

1. Importance of Out-of-Band Information: Out-of-band information significantly improves performance in high K-factor scenarios
2. Advantages of Deep Learning: ML methods better exploit complex relationships between in-band and out-of-band channel estimates
3. Marginal UNet Advantage: Although computationally more complex, UNet slightly outperforms CNN on large-scale data
4. Robustness: Proposed methods demonstrate stable performance across different channel conditions

Related Work

mmWave Channel Estimation

• Traditional ML Methods: Multi-layer perceptron, CNN, FFDNet, ResNet-UNet, etc., but using only in-band information
• Out-of-Band Assisted Methods: References [12]13 propose pilot-assisted methods utilizing sub-6 GHz LOS components

Multi-Band Communication

• Measurement Validation: Multiple measurement campaigns confirm similar multipath characteristics between sub-6 GHz and mmWave bands
• System Integration: Trend toward joint deployment of mmWave systems with sub-6 GHz systems

Deep Learning Applications in Channel Estimation

• CNN Applications: Exploit antenna spatial arrangement characteristics
• UNet Advantages: Encoder-decoder structure and skip connections provide noise robustness

Conclusions and Discussion

Main Conclusions

1. The proposed out-of-band assisted deep learning method significantly outperforms existing solutions
2. UNet architecture slightly outperforms CNN in capturing fine-grained details
3. Deep learning effectively exploits inherent relationships between multi-band channel information
4. The method demonstrates good adaptability and robustness across different channel conditions

Limitations

1. MIMO Configuration Constraints: Networks require retraining for different antenna configurations
2. Computational Complexity: UNet's complex design increases computational overhead
3. Single Subcarrier Training: While scalable to all subcarriers, potential frequency correlation loss may occur
4. Simulation Environment: Results based on 3GPP channel models; real environments may differ

Future Directions

1. Multi-Configuration Adaptation: Develop universal network architectures adaptable to different MIMO configurations
2. Real-World Validation: Verify method effectiveness in actual environments
3. Computational Optimization: Reduce network complexity and improve real-time deployment feasibility
4. Multi-Band Extension: Extend to joint channel estimation across more frequency bands

In-Depth Evaluation

Strengths

1. Strong Innovation: First to combine deep learning with multi-band channel estimation, novel approach
2. Complete Methodology: Forms a complete technical chain from system modeling to network design to performance evaluation
3. Comprehensive Experiments: Thorough comparisons across multiple K-factors and SNR conditions, credible results
4. Engineering Value: Significant performance improvements make the method practically applicable

Weaknesses

1. Insufficient Theoretical Analysis: Lacks theoretical explanation for why deep learning effectively fuses multi-band information
2. Computational Complexity Analysis: Missing detailed computational complexity comparison and real-time performance analysis
3. Generalization Capability: Validation limited to specific channel models; generalization to other environments unknown
4. Practical Deployment Considerations: Lacks discussion of synchronization, calibration, and other practical system deployment issues

Impact

1. Academic Contribution: Provides new research direction for channel estimation in multi-band MIMO systems
2. Industrial Value: Important reference for multi-band deployment in 5G/6G systems
3. Technology Promotion: Method extensible to other multi-band communication scenarios
4. Reproducibility: Detailed implementation parameters facilitate reproduction and improvement

Applicable Scenarios

1. 5G/6G Base Stations: Base station systems supporting dual sub-6 GHz and mmWave bands
2. High-Speed Mobile Communication: High-speed scenarios like vehicular networks requiring reliable channel estimation
3. Dense Deployment Environments: Complex propagation environments such as urban macrocells
4. Large-Scale MIMO Systems: Large antenna array systems requiring precise channel information

References

The paper cites 18 relevant references, primarily including:

• Fundamental mmWave communication literature 1-3
• Deep learning channel estimation methods 4-6
• Multi-band measurements and modeling 8-11
• Out-of-band assisted channel estimation 12-13
• 3GPP standards and channel models 15

Overall Assessment: This is a high-quality paper with strong technical innovation and well-designed experiments, providing an effective deep learning solution for channel estimation in multi-band MIMO systems with significant academic value and engineering application prospects.