PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep Learning based Channel Estimation Algorithm over Time Selective Fading Channels

Qinbo Bai, Jintao Wang, Yue Zhang, Jian Song

arXiv:1908.11013v1eess.SP

TL;DR

This paper addresses channel estimation for time-selective fading, where receiver movement changes channel status and MLPs handle temporal relations poorly. It proposes a sliding bidirectional GRU estimator and reports improved comparisons with traditional algorithms and other neural networks, alongside robustness across pilot densities.

  • Problem

    Time-selective channels change with receiver movement, while memoryless MLPs handle temporal relations poorly and blockwise input may cause estimation discontinuities.

    Source: §I · p. 1

  • Method

    The paper proposes a sliding recurrent SBGRU channel estimator that handles arbitrary-length transmitted symbols and returns estimates as symbols arrive.

    Source: §I · p. 1

  • Results

    The estimator outperforms traditional algorithms and some differently structured neural networks, dynamically tracks channels, and is robust across pilot densities.

    Source: §V · p. 7

  • Takeaways & Limitations

    The proposed estimator does not require prior channel-model knowledge or the assumption that the channel is invariant within coherence time.

    Source: §V · p. 7

    The paper leaves open whether connecting a neural estimator and detector can outperform traditional systems in BER while remaining robust across pilot densities.

    Source: §V · p. 8

Abstract

from arXiv · show

The research about deep learning application for physical layer has been received much attention in recent years. In this paper, we propose a Deep Learning (DL) based channel estimator under time varying Rayleigh fading channel. We build up, train and test the channel estimator using Neural Network (NN). The proposed DL-based estimator can dynamically track the channel status without any prior knowledge about the channel model and statistic characteristics. The simulation results show the proposed NN estimator has better Mean Square Error (MSE) performance compared with the traditional algorithms and some other DL-based architectures. Furthermore, the proposed DL-based estimator also shows its robustness with the different pilot densities.

I. INTRODUCTION

Time-selective fading poses a challenge for channel estimation because receiver movement changes channel status over time, while memoryless MLPs handle temporal relations poorly. The paper proposes an SBGRU estimator and reports dynamic tracking, improved comparisons, and robustness across pilot densities.

I. INTRODUCTION

  • Time-selective channels change with receiver movement, and research applying deep learning to them is described as limited.Prior work includes recurrent detection and MLP-based channel estimation for time-selective channels.Source: §I · p. 1
  • MLPs lack temporal memory, and dividing data into blocks may cause discontinuities in channel estimates.The paper therefore motivates using an RNN for channel estimation.Source: §I · p. 1
  • The proposed SBGRU uses recurrent structure and a sliding approach to process arbitrary-length transmitted symbols and return estimates as symbols arrive.The estimator is designed for time-varying Rayleigh fading.Source: §I · p. 1
  • Simulations report dynamic channel tracking, better performance than traditional algorithms and other neural networks, and robustness across pilot densities.The supplied introduction summarizes these as the estimator's principal results.Source: §I · p. 1

II. SYSTEM MODEL

  • The system model describes a time-varying channel using a linear time-variant relation between transmitted and received signals, with additive white Gaussian noise.The channel and signal architecture are introduced as the system model.Source: §II · p. 2

A. Time Varying Rayleigh Fading Channel Model

The channel model focuses on time-selective Rayleigh fading, using Clarke’s model and the Jakes Doppler spectrum to describe temporal variation; multipath effects are left for future work.

A. Time Varying Rayleigh Fading Channel Model

  • The model considers time-selective fading alone as an initial exploration of rapidly varying channels, leaving multipath effects for future work.The text distinguishes time selectivity from frequency-selective fading caused by multipath.Source: A. Time Varying Rayleigh Fading… · p. 2
  • Clarke’s model and the Jakes Doppler spectrum are used to describe the time-varying channel.The passage identifies the Jakes spectrum as the adopted description of temporal channel variation.Source: A. Time Varying Rayleigh Fading… · p. 2
  • The channel is normalized to have expected gain E(|h[n]|^2) = 1 for the subsequent analysis.The text defines the normalized Doppler frequency using the sampling rate and sets the normalized channel gain.Source: A. Time Varying Rayleigh Fading… · p. 2

B. Signal Architecture

Frames contain repeated, equally spaced pilots among information symbols, and the receiver’s neural estimator uses pilots and the noisy received signal to estimate the channel.

B. Signal Architecture

  • Each frame has K blocks, each containing Ns information symbols and Np equally spaced pilots, with the same pilot pattern repeated across blocks.The stated pilot density is Np/(Ns + Np).Source: B. Signal Architecture · p. 2

C. Signal Flow Model

  • The transmitter combines information and pilot bits, modulates and sends x, while the receiver estimator uses pilot p and noisy received signal y to estimate h.The signal-flow description places the neural estimator at the receiver.Source: C. Signal Flow Model · p. 2
  • The estimator requires no channel information, making the described communication system model-free; conventional channel coding can also be added at the transmitter.The text gives LDPC coding as an example of coding that can be added against noise.Source: C. Signal Flow Model · p. 2

D. Traditional Algorithms For Channel Estimation

The traditional baselines are LS with interpolation and MMSE estimation, whose required channel statistics are difficult to know in practice; the paper compares theory-based and simulation-based MMSE variants.

D. Traditional Algorithms For Channel Estimation

  • LS estimates channel values at pilot positions and uses interpolation for positions without pilots.The text gives an expected MSE expression for LS while omitting interpolation effects.Source: D. Traditional Algorithms For Channel… · p. 3
  • The MMSE estimator uses a channel correlation matrix, with its correlation values calculated from the channel autocorrelation function.The supplied text identifies the matrix as R_hh = E(hh^H).Source: D. Traditional Algorithms For Channel… · p. 3
  • MMSE estimation requires the channel autocorrelation form and Doppler speed in advance, quantities described as hard to know accurately in practical settings.The text motivates comparing two MMSE approaches in simulation.Source: D. Traditional Algorithms For Channel… · p. 3
  • “MMSE theory” directly uses known information, whereas “MMSE sim” estimates autocorrelation from LS estimates before calculating MMSE.The simulated method is named for deriving the computation from simulation results.Source: D. Traditional Algorithms For Channel… · p. 3

III. DL-BASED NN CHANNEL ESTIMATOR

The estimator uses recurrent processing to model channel time dependence, then adopts bidirectional GRU cells with a sliding structure after simple BGRU proved insufficient.

A. RNN structure

  • RNNs feed the previous time step’s output into the current input, allowing them to capture past information.Source: A. RNN structure · p. 3
  • Because channel states relate to both past and future, bidirectional RNNs process data in both directions and combine hidden states for a linear output layer.Source: A. RNN structure · p. 3
  • GRU cells replace basic RNN cells to address limited long-term information capture, using update and reset gates to control information flow.The paper notes GRUs have performance similar to LSTMs on many tasks and higher speed due to fewer gates.Source: A. RNN structure · p. 3
  • The estimator uses BGRU cells, but simple BGRU performance is insufficient, motivating the sliding BRNN structure for further improvement.The paper compares BGRU and SBGRU in Section IV.C.Source: A. RNN structure · p. 3

B. SBGRU structure

SBGRU applies bidirectional GRU computations over fixed-length windows that slide across the input sequence, then averages repeated estimates for each symbol.

B. SBGRU structure

  • Each BGRU block uses a fixed window length WL, and the paper notes that longer windows may improve performance because channel moments are correlated.The window-length effect is evaluated in Section IV.D.Source: B. SBGRU structure · p. 3
  • SBGRU computes over WL symbols and advances the window by one symbol after each computation.Source: B. SBGRU structure · p. 3
  • Because symbols are estimated multiple times as windows overlap, SBGRU averages their estimates to produce each final output.Source: B. SBGRU structure · p. 4

C. Train and test NN estimator

The estimator takes channel-distorted signals and pilot information as input, uses a two-layer BGRU followed by a linear layer, and trains its parameters with an MSE loss.

C. Train and test NN estimator

  • The SBGRU input includes the channel-distorted signal and pilot information, with complex-signal real and imaginary parts separated for the platform.Pilot positions carry the known pilot symbols, while information positions are set to zero.Source: C. Train and test NN estimator · p. 4
  • The estimator uses two BGRU layers followed by a linear layer to produce the final channel estimate.The linear layer uses weights W and bias b; θS and θL denote the SBGRU and linear-layer parameters.Source: C. Train and test NN estimator · p. 4
  • Training optimizes the estimator parameters using an MSE loss function.Source: C. Train and test NN estimator · pp. 4–5
  • The paper uses Adam for parameter optimization, and testing loads the trained parameters on data with the same structure and statistical characteristics as training data.Source: C. Train and test NN estimator · p. 5

IV. SIMULATION RESULTS

The simulation section describes the evaluation setup and presents four groups of results for the time-varying Rayleigh fading channel.

IV. SIMULATION RESULTS

  • The section first describes the NN estimator’s simulation setting, then presents and analyzes four groups of simulation results.Source: §IV · p. 5

A. Simulation Setting

The simulations use randomized QPSK sequences and separate training, validation, and test data to evaluate channel estimation. SBGRU tracks the channel well and approaches MMSE theory performance without channel knowledge, outperforming LS and MMSE simulation estimates.

A. Simulation Setting

  • The simulation generates 1,200 channels—800 for training, 200 for validation, and 200 for testing—and 120,000 sequences split 100,000/10,000/10,000.Sequences use randomly generated i.i.d. bits mapped with QPSK; each sequence randomly selects a channel.Source: A. Simulation Setting · p. 5
  • The implementation uses an Intel i7-6700K CPU, NVIDIA GeForce GTX 1080 GPU, 16 GB memory, PyTorch 1.0.0, and Python 3.6.Source: A. Simulation Setting · p. 5

B. Performance Comparison with the traditional algorithm

  • SBGRU approaches MMSE theory performance without channel knowledge and outperforms LS and MMSE sim across the tested SNR range.MMSE theory performs best; LS performs worst, and MMSE sim improves only slightly over LS, especially at high SNR.Source: B. Performance Comparison with the… · p. 5, Figure 6 · p. 5
  • At 20 dB SNR over a 4,000-symbol channel, SBGRU tracks most linear channel segments with slight oscillation in nonlinear segments.LS and MMSE sim estimates vibrate heavily in the comparison.Source: B. Performance Comparison with the… · p. 5

C. Performance Comparison with different structures of NN

Sliding BGRU uses channel information over a time window and performs better than blockwise BGRU as SNR increases. SBGRU outperforms the tested MLP estimators above 5 dB, while window-length results indicate a balance between estimation performance and training and testing speed.

C. Performance Comparison with different structures of NN

  • Sliding BGRU performs better than blockwise BGRU as SNR increases, with sliding using average channel information within a time window.Source: C. Performance Comparison with… · p. 5, Figure 8 · p. 7
  • The MLP with block length 16, matching the design in, performs poorly; length 32 improves slightly, while length 40 reduces performance.The authors suggest insufficient model parameters may explain the 16-length result; the 40-length setting is not an integral multiple of the original 16-symbol data block and cannot fully exploit repeated pilot information.Source: C. Performance Comparison with… · p. 6
  • Above 5 dB SNR, SBGRU outperforms all tested MLP estimators with block lengths of 16, 32, and 40.The recurrent structure captures previous channel estimates without requiring them as network inputs.Source: C. Performance Comparison with… · p. 6, Figure 9 · p. 7

D. Performance vs window length

  • Among the tested window lengths, performance increases monotonically with length; the three lengths other than 16 symbols perform nearly the same.A window that is too short may not provide enough estimation information, while a longer window brings little further improvement; a suitable length balances accuracy with training and testing speed and depends on channel characteristics.Source: D. Performance vs window length · pp. 6–7

E. Performance vs pilot density

With lower pilot density, the SBGRU estimator’s MSE worsens only slightly and remains better than LS and “MMSE sim” estimation. The conclusion reports that the estimator tracks the channel dynamically, performs better than traditional algorithms and some NN estimators, and does not require prior channel-model knowledge or a channel-invariance assumption during training and testing.

E. Performance vs pilot density

  • As pilot density decreases, MSE worsens only slightly, while SBGRU remains much better than LS and “MMSE sim” estimation.The paper presents this result as evidence of robustness across pilot densities.Source: E. Performance vs pilot density · p. 7

V. CONCLUSION

  • The estimator is designed for a time-varying Rayleigh fading channel.Source: §V · p. 7
  • The proposed DL-based estimator performs better than traditional algorithms and some NN estimators, dynamically tracks the channel, and is robust to pilot density.The conclusion summarizes these results without specifying numerical values.Source: §V · p. 7
  • DL-based estimation requires neither prior channel-model knowledge nor the assumption that the channel is invariant in coherence time during training and testing.The paper presents this as a potential advantage under time-varying channels.Source: §V · p. 7
  • Combining the estimator with high-performance channel coding before modulation is described as a convenient way to protect performance against Gaussian noise and further improve MSE.Source: §V · pp. 7–8
  • The paper proposes exploring a DL-based detector for equalization and demodulation alongside the estimator, including whether the resulting system improves BER and remains robust across pilot densities.Source: §V · p. 8
Loading 1908.11013v1…