Deep Learning-Based Channel Estimation Algorithm Over Time Selective Fading Channels

Qinbo Bai, Jintao Wang, Yue Zhang, Jian SongIEEE Trans. Cogn. Commun. Netw 2020 · OpenReview

TL;DR

This paper introduces a novel deep learning-based channel estimator, SBGRU, designed to accurately track time-varying Rayleigh fading channels without prior knowledge of channel statistics. The proposed method significantly outperforms traditional algorithms and other deep learning approaches in terms of Mean Square Error (MSE) and demonstrates robustness to varying pilot densities.

What Problem Does This Paper Address?

In wireless communication systems, accurately estimating the channel's state is crucial for reliable data transmission. Traditional channel estimation methods often rely on pre-defined channel models and statistical characteristics, which can be difficult to obtain or may not accurately reflect real-world conditions, especially in dynamic environments. Time-selective fading channels, where the channel characteristics change rapidly over time due to factors like receiver movement, pose a particular challenge. Existing deep learning approaches have shown promise but often struggle with the complexity and dynamic nature of these channels, or they require extensive prior knowledge. This paper aims to bridge this gap by developing a deep learning solution that can dynamically adapt to time-varying channels without needing explicit channel models.

Methods

The core of this paper is the Sliding Bidirectional Gated Recurrent Unit (SBGRU) channel estimator. The authors leverage Recurrent Neural Networks (RNNs) due to their ability to process sequential data and capture temporal dependencies, which is essential for time-varying channels.

Here's a breakdown of their approach:

  • Bidirectional GRU (BGRU): Instead of a standard RNN, they use a Bidirectional GRU. This allows the network to process the input sequence in both forward and backward directions, capturing context from both past and future channel states, which is crucial for time-varying channels.
  • Sliding Window Mechanism: To further enhance performance, they introduce a "sliding" operation. The SBGRU processes data in blocks, but instead of processing each block independently, it slides the processing window by one symbol at a time. This allows for a more continuous estimation and leverages information from overlapping windows, effectively averaging estimations over a temporal window.
  • Input Representation: The input to the SBGRU is the received signal (channel-distorted data) and the known pilot signals. Since the current deep learning platforms primarily handle real numbers, complex signals are separated into their real and imaginary parts before being fed into the network.
  • Architecture: The final SBGRU estimator consists of two layers of BGRU cells followed by a linear layer for the final channel estimation.

Datasets

The paper does not explicitly mention using pre-existing public datasets. Instead, they generate their own simulation data based on the described channel model and parameters.

  • Training Data: 100,000 sequences.
  • Validation Data: 10,000 sequences.
  • Testing Data: 10,000 sequences.

The channel parameters and pilot densities used for simulation are chosen to be comparable to those in related work [13] for fair comparison.

Key Insights and Ideas

  • Dynamic Tracking without Prior Knowledge: The most significant insight is that a deep learning model, specifically the SBGRU, can effectively track time-varying channels without needing explicit knowledge of the channel's statistical properties (like Doppler shift). This is a major advantage over traditional methods.
  • Bidirectionality is Key: The use of a bidirectional RNN structure (BGRU) is crucial for capturing the full temporal context of a time-varying channel, outperforming unidirectional RNNs.
  • Sliding Window Enhances Robustness: The sliding window mechanism, rather than block-wise processing, leads to a more continuous and accurate channel estimation, especially at higher SNRs. It effectively averages information over a temporal window, smoothing out estimations.
  • SBGRU Outperforms Traditional and Simpler DL Methods: The SBGRU consistently shows superior performance compared to Least Squares (LS) and simulated Minimum Mean Square Error (MMSE) estimators, as well as simpler deep learning architectures like MLP.

Key Results

  • MSE Performance: The SBGRU estimator achieves performance comparable to "MMSE theory" (which assumes perfect channel knowledge) and significantly outperforms LS and "MMSE sim" (MMSE estimated from LS) estimators across various Signal-to-Noise Ratios (SNRs). For example, at 20dB SNR, the SBGRU's MSE is significantly lower than LS and "MMSE sim."
  • Channel Tracking: Visualizations show that the SBGRU estimator tracks the real channel very well, with only slight oscillations in non-linear parts, whereas LS and "MMSE sim" estimators exhibit heavy vibrations.
  • Robustness to Pilot Density: The SBGRU estimator demonstrates robustness to decreasing pilot densities. While MSE performance decreases slightly with fewer pilots, it remains substantially better than LS and "MMSE sim" estimators.
  • Window Length Impact: Increasing the sliding window length generally improves performance, but the gains diminish beyond a certain point, suggesting a balance between accuracy and computational efficiency.

Limitations and Future Work

  • MLP Limitations: The paper notes that simpler MLP networks struggle with time-varying channels, especially when the estimation block length is not optimally chosen or when the network lacks sufficient capacity.
  • Window Length Optimization: While the sliding window improves performance, selecting the optimal window length requires balancing accuracy and training/testing speed, and it may be related to specific channel characteristics.
  • Future Work: The authors suggest several promising directions:
  • Joint Detection and Estimation: Extending the deep learning approach to jointly perform channel estimation, equalization, and demodulation.
  • End-to-End System Construction: Building a complete communication system using deep learning for both transmitter and receiver.
  • Bit Error Rate (BER) Performance: Investigating whether such DL-based systems can achieve better BER performance than traditional systems under time-varying channels.
  • Robustness to Pilot Densities: Further exploring the robustness of DL-based systems with varying pilot densities.
  • Integration with Channel Coding: Combining the DL estimator with high-performance channel coding techniques to further improve MSE performance.

Summary updated

Abstract

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.
BibTeX
@article{bai2020deep,
  author = {Qinbo Bai and Jintao Wang and Yue Zhang and Jian Song},
  title = {{Deep Learning-Based Channel Estimation Algorithm Over Time Selective Fading Channels}},
  journal = {IEEE Trans. Cogn. Commun. Netw.},
  year = {2020},
  url = {https://openreview.net/forum?id=wt9HOujefcE},
}
Figure