PRINCE: A Pruned AMP Integrated Deep CNN Method for Efficient Channel Estimation of Millimeter-Wave and Terahertz Ultra-Massive MIMO Systems

Zhengdong Hu, Yuhang Chen, Chong HanIEEE Trans. Wireless Communications 2023 · DOI

TL;DR

This paper introduces PRINCE, a novel method for efficient channel estimation in millimeter-wave and terahertz (mmWave/THz) ultra-massive MIMO systems. By integrating Approximate Message Passing (AMP) with a pruned Deep Convolutional Neural Network (DCNN), PRINCE significantly reduces computational complexity while maintaining high estimation accuracy, making it more practical for real-world deployment.

What Problem Does This Paper Address?

Modern wireless communication systems are pushing towards higher data rates, necessitating the use of millimeter-wave (mmWave) and terahertz (THz) frequency bands. These bands offer vast bandwidth but suffer from high propagation loss. To compensate, ultra-massive MIMO (UM-MIMO) systems with hybrid beamforming are employed. However, accurately estimating the channel state information (CSI) in these systems is a significant challenge. This is due to the extremely high dimensionality of the channels and the limited number of observations available from the few RF chains. Traditional methods struggle with accuracy and computational complexity, while existing deep learning approaches can be computationally intensive and prone to overfitting. The paper aims to bridge this gap by proposing a method that is both accurate and computationally efficient.

Methods

The proposed method, PRINCE, builds upon an AMP-DCNN architecture.

  • AMP (Approximate Message Passing): This iterative algorithm is used as a first step to obtain a coarse channel estimation. AMP is chosen for its low complexity and fast convergence in high-dimensional sparse signal recovery problems.
  • DCNN (Deep Convolutional Neural Network): Following AMP, a DCNN is employed to refine the coarse estimation from AMP, significantly improving accuracy. The DCNN learns to extract complex channel features.
  • Network Pruning (PRINCE): The core innovation is the PRINCE method, which addresses the computational burden of the DCNN. It involves:
  • Training with Regularization: A regularization term is added to the DCNN's training objective to encourage some feature maps to have near-zero scaling factors. This effectively creates "insignificant" feature maps.
  • Pruning: Feature maps with scaling factors below a certain threshold are identified and removed. This process is controlled by a "pruning ratio."
    • Refining: After pruning, the network is fine-tuned to recover any performance degradation caused by the removal of feature maps.

This pruning strategy aims to reduce the network's size and computational complexity without a significant loss in estimation accuracy.

Datasets

The paper utilizes two types of channel datasets for training and evaluation:

  • mmWave Dataset: A publicly accessible mmWave dataset at 60 GHz, selected from the Raymobtime datasets. This dataset considers 3D scenarios and receiver mobility.
  • THz Dataset: A self-generated THz dataset at 0.3 THz, constructed using Remcom Wireless Insite, simulating a typical street environment.

These datasets are generated using ray-tracing methods and incorporate realistic channel conditions.

Key Insights and Ideas

  • AMP as a powerful pre-processor: Using AMP to get an initial, albeit coarse, channel estimate provides a strong starting point for the DCNN, allowing the DCNN to focus on refinement rather than learning from scratch.
  • DCNNs excel at feature extraction for CSI: The DCNN architecture is well-suited to learning the complex spatial and frequency correlations inherent in mmWave/THz channels.
  • Network slimming for practical deployment: The key insight is that not all learned features are equally important. By identifying and removing less critical feature maps through pruning, significant computational savings can be achieved without drastically compromising performance. This addresses the practical challenge of deploying complex DL models in resource-constrained environments.
  • Regularization as a feature map generator: The regularization technique is crucial for creating the "insignificant" feature maps that can then be pruned, making the pruning process effective.

Key Results

  • AMP-DCNN Performance: The AMP-DCNN method achieves superior estimation accuracy compared to conventional methods (OMP, AMP, MMSE) and another DL-based method (SF-CNN). On the mmWave dataset, it reaches an NMSE of -10 dB at 10 dB SNR.
  • PRINCE Efficiency: The PRINCE method, by pruning up to 80% of feature maps, significantly reduces computational complexity and model size.
  • Accuracy-Complexity Trade-off: PRINCE demonstrates a good trade-off, achieving an NMSE of -10 dB at 10 dB SNR even with 80% feature maps pruned. This represents a substantial reduction in complexity with only a slight degradation in performance (e.g., ~2 dB degradation on average when pruning 80% of feature maps).
  • Robustness to Limited Data: The DCNN model shows robustness to limited training data, maintaining reasonable performance even with a training set size reduced by 83%.

Limitations and Future Work

  • Grid-Mismatch Problem: The paper acknowledges that the estimation result of AMP is not perfectly accurate due to the grid-mismatch problem, where physical parameters don't perfectly align with the predefined grid. The DCNN helps mitigate this.
  • Pruning Ratio Limits: The paper notes that pruning cannot be achieved if all feature maps in a CV layer are removed, as this would cause the DCNN to fail.
  • Future Work: While not explicitly detailed as future work, the success of PRINCE suggests that further exploration of pruning strategies and their application to other DL-based communication algorithms could be beneficial. The paper also implicitly highlights the ongoing need for efficient DL models in next-generation wireless systems.

Summary updated

BibTeX
@article{hu2023prince,
  author = {Zhengdong Hu and Yuhang Chen and Chong Han},
  title = {{PRINCE: A Pruned AMP Integrated Deep CNN Method for Efficient Channel Estimation of Millimeter-Wave and Terahertz Ultra-Massive MIMO Systems}},
  journal = {IEEE Trans. Wireless Communications},
  year = {2023},
  doi = {10.1109/twc.2023.3258405},
}
Figure