Source-linked AI summary
Data-Driven Deep Learning to Design Pilot and Channel Estimator For Massive MIMO
Xisuo Ma, Zhen Gao
TL;DR
Massive MIMO channel estimation must recover high-dimensional downlink channels from underdetermined, low-overhead measurements. The paper jointly learns pilot signals and a channel estimator with an end-to-end DNN, achieving better NMSE than SOMP in the evaluated one-ring and cluster-sparse settings.
Problem
FDD massive MIMO needs accurate high-dimensional downlink CSI despite users having few antennas and limited training overhead.
Method
An end-to-end DNN jointly learns pilot signals through dimensionality reduction and channel reconstruction through fully connected and cascaded convolutional layers.
Results
The proposed scheme outperforms SOMP in NMSE for one-ring and cluster-sparse channels, including settings with lower pilot overhead.
Takeaways & Limitations
Jointly optimizing pilots and reconstruction enables effective high-dimensional channel estimation from low-dimensional measurements in the evaluated massive MIMO scenarios.
Abstract
from arXiv · showhide
In this paper, we propose a data-driven deep learning (DL) approach to jointly design the pilot signals and channel estimator for wideband massive multiple-input multiple-output (MIMO) systems. By exploiting the angular-domain compressibility of massive MIMO channels, the conceived DL framework can reliably reconstruct the high-dimensional channels from the under-determined measurements. Specifically, we design an end-to-end deep neural network (DNN) architecture composed of dimensionality reduction network and reconstruction network to respectively mimic the pilot signals and channel estimator, which can be acquired by data-driven deep learning. For the dimensionality reduction network, we design a fully-connected layer by compressing the high-dimensional massive MIMO channel vector as input to low-dimensional received measurements, where the weights are regarded as the pilot signals. For the reconstruction network, we design a fully-connected layer followed by multiple cascaded convolutional layers, which will reconstruct the high-dimensional channel as the output. By defining the mean square error between input and output as loss function, we leverage Adam algorithm to train the end-to-end DNN aforementioned with extensive channel samples. In this way, both the pilot signals and channel estimator can be simultaneously obtained. The simulation results demonstrate that the superiority of the proposed solution over state-of-the-art compressive sensing approaches.
I. INTRODUCTION
Massive MIMO channel estimation requires recovering high-dimensional CSI from low-overhead measurements, especially in FDD systems. The paper addresses limitations of separate pilot and estimator design with an end-to-end DNN that jointly learns both.
- Motivation: Accurate high-dimensional CSI is essential for beamforming and detection, but FDD users must estimate downlink channels with few antennas and low training overhead.The paper motivates exploiting angular- and delay-domain channel compressibility to address this underdetermined estimation problem.
- Motivation: Existing deep-learning channel-estimation methods generally design transmitter pilots and receiver estimators separately, potentially causing performance loss.
- Proposed approach: The proposed end-to-end DNN jointly learns pilot signals and channel estimation through dimensionality reduction and reconstruction networks.The dimensionality reduction weights represent pilots, while the reconstruction network estimates the high-dimensional channel.
- Proposed approach: The framework trains both components with an MSE loss and the Adam algorithm using extensive channel samples.
II. SYSTEM MODEL
The system models wideband downlink channel estimation for hybrid-beamforming massive MIMO across low-frequency and mmWave bands. It represents channels with cluster-based propagation models and collects measurements over multiple OFDM time slots.
- System configuration: The downlink system uses OFDM with K subcarriers, a UPA of NBS = Nh×Nv antennas, NRF RF chains, and U single-antenna users.The channel-estimation phase considers hybrid beamforming at the base station.
- Estimation setting: The proposed scheme targets high-dimensional channel recovery from much-reduced pilot overhead, making estimation underdetermined unlike conventional small-scale MIMO LS estimation.
- Measurement model: Received signals are collected across M successive time slots, while analog and digital precoders characterize the hybrid-beamforming measurements.
- Channel models: The channel includes path gains, delays, sampling rate, and azimuth/elevation AoDs, with the UPA represented through its antenna response vector.
- Channel models: For mmWave channels, the model uses Nc clusters with Np paths per cluster; the low-frequency one-ring model is the Nc = 1 special case.
III. DNN-BASED CHANNEL ESTIMATION SCHEME
The scheme models wideband massive-MIMO channel estimation using angular-domain compressibility and an end-to-end architecture that jointly learns dimensionality reduction and reconstruction.
- The simplified same-signal assumption across subcarriers can cause high PAPR, but scrambling and descrambling codes can relax it.The paper notes this assumption may be impractical in OFDM systems.
- The received signal is stacked across subcarriers, and the angular-domain transformation produces sparse channel rows that enable compressed reconstruction.The transformed channel uses DFT matrices to expose angular-domain sparsity.
- The end-to-end DNN jointly optimizes pilot signals and channel estimation through dimensionality reduction and reconstruction networks trained together.The dimensionality-reduction weights represent pilots, while the reconstruction network estimates the high-dimensional channel.
1) Pilot Signals:
The pilot-signal network compresses channel inputs into low-dimensional measurements through a bias-free, non-linear-activation-free fully connected layer.
- 1) Pilot Signals:: Pilot signals are represented by the weights of a fully connected layer that maps high-dimensional channels to low-dimensional measurements.The input channel is multiplied by the learned pilot signals to generate real and imaginary measurement components.
- 1) Pilot Signals:: Each measurement component is a weighted linear combination of the corresponding real and imaginary channel values.
- 1) Pilot Signals:: The compression layer omits both bias values and nonlinear activation functions.
2) Channel Estimator:
The channel estimator reconstructs the high-dimensional channel from compressed measurements using an initial fully connected estimate followed by cascaded convolutional reconstruction.
- 2) Channel Estimator:: The reconstruction network first produces a coarse channel estimate and then performs iterative-style reconstruction with cascaded convolutional layers.The coarse estimate imitates the matching operation used by greedy compressive-sensing methods such as OMP.
- 2) Channel Estimator:: The estimator is trained by minimizing mean squared error between input channel samples and reconstructed outputs.The loss is computed over all training samples in a batch.
- 2) Channel Estimator:: Adam jointly trains the dimensionality-reduction and reconstruction parameters with learning rate 0.001, batch size 128, and 300 epochs.
IV. SIMULATION RESULTS
The simulations evaluate the proposed estimator against SOMP using a 256-antenna, 256-subcarrier system and channel datasets generated for training, validation, and testing.
- IV. SIMULATION RESULTS: The evaluation uses a 256-antenna UPA, 256 OFDM subcarriers, and compression ratio ρ = M/NBS, with SOMP as the comparison method.SOMP uses a quantized angular dictionary with 64 grids in each antenna dimension.
- IV. SIMULATION RESULTS: The data comprise 2,000 training groups, 1,000 validation groups, and 500 test groups generated from specified channel scenarios.Each group contains channel data across subcarriers and antenna dimensions.
A. Low-Frequency Massive MIMO over One-Ring Channel Model
Under the one-ring channel model, the proposed DNN estimator achieves better NMSE than SOMP while using less pilot overhead.
- A. Low-Frequency Massive MIMO over One-Ring Channel Model: The DNN scheme with compression ratio ρ = 0.25 outperforms SOMP with ρ = 0.5 across the evaluated SNRs and angle spreads.The comparison uses angle spreads of ±7.5° and ±15° under the one-ring model.
B. MmWave Massive MIMO over Cluster Sparse Channel Model
For cluster-sparse mmWave channels, the proposed DNN estimator outperforms SOMP across compression ratios and benefits from multi-subcarrier training samples.
- B. MmWave Massive MIMO over Cluster Sparse Channel Model: Multi-subcarrier channel samples produce better DNN channel-estimation performance than single-carrier samples in the OFDM evaluation.The multi-carrier samples include channels from all subcarriers across different channel realizations.
- B. MmWave Massive MIMO over Cluster Sparse Channel Model: Fig. 4 evaluates NMSE against SNR for the cluster-sparse model and separately examines multi-subcarrier processing effectiveness.The evaluations use angle spread ±3.75° for the SNR comparison.
- B. MmWave Massive MIMO over Cluster Sparse Channel Model: The trained DNN performs online estimation using two fully connected operations and multiple cascaded convolutional operations after offline training.The paper compares this testing-stage complexity with SOMP-based estimation.
V. CONCLUSIONS
The paper proposes an end-to-end DNN that jointly designs pilot signals and channel estimation for massive MIMO. Simulations show improved estimation performance over conventional schemes in both low-frequency one-ring and mmWave cluster-sparse models.
- V. CONCLUSIONS: The proposed end-to-end DNN jointly designs pilot signals and a channel estimator through dimensionality reduction and channel reconstruction networks.The dimensionality-reduction weights represent pilot signals, while the reconstruction network uses a fully connected layer and cascaded convolutional layers.
- V. CONCLUSIONS: Simulation results show significant channel-estimation improvement over the conventional scheme in low-frequency one-ring and mmWave cluster-sparse channel models.