PaperlayerView exact revision on arXiv

Source-linked AI summary

ComNet: Combination of Deep Learning and Expert Knowledge in OFDM Receivers

Xuanxuan Gao, Shi Jin, Chao-Kai Wen, Geoffrey Ye Li

arXiv:1810.09082v1eess.SPcs.IT

TL;DR

The paper addresses replacing an OFDM receiver while retaining expert wireless-communication knowledge rather than using a fully data-driven FC-DNN. It proposes block-by-block channel-estimation and signal-detection subnets initialized from traditional solutions, and reports improved accuracy, robustness, convergence, and efficiency relative to stated baselines.

  • Problem

    Existing FC-DNN OFDM receivers replace receiver modules without exploiting wireless-communication expert knowledge, motivating a model-driven alternative.

  • Method

    ComNet divides OFDM reception into channel-estimation and signal-detection subnets, using traditional communication solutions as initializations that deep networks refine.

  • Results

    ComNet offers more accurate channel estimation than LMMSE and higher data-recovery accuracy than existing methods and FC-DNN, with robustness to signal-to-noise ratio.

  • Takeaways & Limitations

    Combining deep learning with communication intelligence yields a ComNet receiver with faster convergence and fewer parameters than the FC-DNN OFDM receiver.

Abstract

from arXiv · show

In this article, we propose a model-driven deep learning (DL) approach that combines DL with the expert knowledge to replace the existing orthogonal frequency-division multiplexing (OFDM) receiver in wireless communications. Different from the data-driven fully connected deep neural network (FC-DNN) method, we adopt the block-by-block signal processing method that divides the receiver into channel estimation subnet and signal detection subnet. Each subnet is constructed by a DNN and uses the existing simple and traditional solution as initialization. The proposed model-driven DL receiver offers more accurate channel estimation comparing with the linear minimum mean-squared error (LMMSE) method and exhibits higher data recovery accuracy comparing with the existing methods and FC-DNN. Simulation results further demonstrate the robustness of the proposed approach in terms of signal-to-noise ratio and its superiority to the FC-DNN approach in the computational complexities or the memory usage.

I. INTRODUCTION

ComNet proposes a model-driven OFDM receiver that combines deep learning with conventional wireless-communication knowledge, addressing limitations of fully connected data-driven receivers. It divides processing into channel-estimation and signal-detection subnets, initialized from traditional solutions, and reports improved performance with fewer parameters and faster convergence than FC-DNN.

  • Motivation and contribution: FC-DNN receivers replace all OFDM receiver modules but do not exploit expert knowledge and require many parameters and large training datasets.The paper associates this data-driven design with unexplainability, unpredictability, slow convergence, and high computational complexity.
  • Motivation and contribution: ComNet replaces conventional or FC-DNN OFDM receivers with a model-driven architecture combining deep learning and wireless-communication expert knowledge.In this paper, expert knowledge means block-based architectures and algorithms from conventional wireless communications.
  • Architecture: The ComNet receiver processes signals block by block through channel-estimation and signal-detection modules rather than one entire deep-learning architecture.Its subnets use existing receiver models and traditional solutions as initializations, while the networks refine coarse inputs.
  • Reported findings: ComNet reports better performance than LMMSE and FC-DNN, with faster convergence and fewer parameters than the FC-DNN OFDM receiver.The reported comparison concerns channel-estimation or receiver performance and computational efficiency, as described in the introduction.
  • Architecture: A dotted short path provides a relatively robust candidate for binary-symbol recovery within the receiver architecture.This path is described alongside the subnet initializations and coarse-to-fine processing design.

II. COMNET

The ComNet section presents an OFDM receiver with channel-estimation and signal-detection subnets and describes its training and configuration choices.

  • Architecture: The proposed ComNet receiver contains channel-estimation and signal-detection subnets for OFDM reception.The section introduces the architecture and elaborates the DL-based subnets in Section II A.
  • Training configuration: The section explains network-weight initialization, cost function, optimizer, and hyperparameter configurations.These choices are presented as part of the receiver’s design and training setup.

A. ComNet Architecture

ComNet replaces the conventional OFDM receiver with two cascaded DL-based subnets for channel estimation and signal detection. These subnets initialize traditional solutions and refine them to support robust data recovery across scenarios.

  • Architecture: The ComNet receiver uses cascaded channel-estimation and signal-detection subnets instead of a conventional OFDM receiver.Its inputs and outputs remain similar to those of the conventional receiver.
  • Architecture: Unlike brute-force FC-DNN estimation, ComNet combines traditional communication solutions with DL networks to refine coarse inputs.The architecture connects conventional methods with DL-based processing for relatively robust recovery.
  • Channel estimation: The CE subnet takes LS channel estimation from pilot symbols as input and uses LS_RefineNet to generate more accurate channel estimates.LS_RefineNet is a one-layer DNN with a linear output layer of 128 neurons.
  • Signal detection: The SD subnet uses zero-forcing symbol detection, refined with estimated channels and received data to improve transmit-data estimation.For 64 subcarriers, eight independent SD subnets process groups of eight consecutive subcarriers.

B. Training Specification

The networks use expert-informed initialization and sequential training. The CE subnet is trained first and fixed before training the SD subnet, reducing parameter requirements while preserving block-level optimization.

  • Initialization: The CE subnet is initialized with the real-valued LMMSE channel-estimation weight matrix.The initialization maps the real and imaginary parts of LS and LMMSE channel estimates.
  • Initialization: The SD subnet’s fully connected-layer weights are initialized using an existing method before DL optimization.After initialization, network parameters are adjusted by minimizing the cost between estimated and raw transmit binary symbols.
  • Training procedure: Both subnets use mean-squared-error cost and the Adam optimizer, with sequential training in TensorFlow.The CE subnet is trained for 2,000 epochs, followed by 5,000 epochs for the SD subnet.
  • Training procedure: Sequential training can guarantee optimality of each block and accelerate training with fewer network parameters.Each epoch uses 50 mini-batches with a total batch size of 1,000.
  • Training procedure: The learning rate starts at 0.001 and decreases by scheduled factors during CE and SD training.It decreases 10-fold every 1,000 CE epochs and 5-fold every 2,000 SD epochs.

III. NUMERICAL RESULTS

The simulations evaluate ComNet against FC-DNN and traditional receivers across linear, CP-removal, and other configured cases, comparing accuracy and complexity.

  • The experiments consider linear, CP-removal, and clipping cases, with “SameSNR” marking training and deployment at SNR = 5 dB.The postfixes no postfix, “_CP,” and “_CR” correspond to the respective cases.
  • The study focuses on a model-driven OFDM receiver architecture rather than resolving nonlinear distortion through traditional compensatory methods.Traditional nonlinearity compensatory methods are not further discussed because the stated contribution is the receiver architecture.
  • Simulations compare ComNet, FC-DNN, and traditional communication methods in accuracy and complexity.The evaluated traditional baseline is LMMSE-MMSE, while ComNet is represented by ComNet-BiLSTM and ComNet-FC.
  • The OFDM configuration uses 64 subcarriers, 16 CP samples, one pilot symbol, and one data symbol per frame.The modulation is 64-QAM under the LTE standard, with a WINNER II C1-scenario NLOS channel at 2.6 GHz.
  • ComNet uses traditional receiver solutions to initialize its channel-estimation and signal-detection subnetworks.The ComNet variants include BiLSTM-Detection and FC-Detection ZF_RefineNet designs.

A. CE Subnet

The channel-estimation subnet improves channel estimation under CP removal relative to traditional LMMSE channel estimation by learning refinements from an initialized solution.

  • Accurate channel-state information from ComNet is useful for channel analysis and downlink CSI feedback.
  • ComNet’s channel-estimation subnet better rectifies CP-removal effects than traditional LMMSE channel estimation.Figure 4 compares MSE performance under linear and CP-removal cases.
  • The LS_RefineNet training process adjusts initialized multiplicative weights by minimizing channel MSE with Adam.

B. SD Subnet

In the linear case, ComNet achieves better BER performance than FC-DNN and LMMSE-MMSE, while ComNet-FC also converges faster and uses fewer parameters.

  • 1) Linear Case:: The ComNet receiver’s BER is closest to the ideal Y/H_true bound compared with FC-DNN and LMMSE-MMSE.The comparison concerns the linear case without nonlinear effects.
  • 1) Linear Case:: 1 dB lower SNR reaches BER = 10^-3 for ComNet than for FC-DNN and LMMSE-MMSE.
  • 1) Linear Case:: ComNet outperforms FC-DNN under longer delay spread.
  • 1) Linear Case:: ComNet-FC uses one-eighth the number of parameters of the compared FC-DNN approach.
  • 1) Linear Case:: ComNet-FC needs 200 epochs to converge, whereas FC-DNN needs approximately 2,000 epochs to reach the same BER level.The result is presented as evidence of faster convergence and lower parameter requirements for the model-driven approach.

2) CP Removal:

Without a cyclic prefix, traditional LMMSE-MMSE saturates at high SNR, while DL-based approaches resolve inter-symbol interference; ComNet-BiLSTM performs best among the compared learned methods.

  • 2) CP Removal:: At SNR = 20 dB, traditional LMMSE-MMSE becomes saturated when the CP is omitted.The CP normally mitigates inter-symbol interference caused by multipath channels.
  • 2) CP Removal:: DL-based FC-DNN and ComNet-BiLSTM perform better than LMMSE-MMSE in resolving inter-symbol interference without CP.
  • 2) CP Removal:: Above SNR 25 dB, ComNet-BiLSTM has about 50% of FC-DNN’s BER.
  • 2) CP Removal:: ComNet-BiLSTM recovers transmitted symbols more accurately than the other approaches in OFDM systems without CP.The passage attributes this ability to the BiLSTM recurrent neural network design.

3) Clipping:

Under nonlinear clipping and related cases, ComNet-BiLSTM achieves the lowest reported BER and remains robust to SNR mismatching, while complexity varies across ComNet configurations.

  • Clipping: Clipping at a ratio of 1.6 introduces nonlinear distortion, and ComNet-BiLSTM obtains the lowest BER among all methods.The comparison concerns BER curves under clipping nonlinear distortion.
  • Robustness of SNR Mismatching: Training offline at SNR = 40 dB while deploying online at arbitrary SNRs produces SNR-mismatched results.The reported experiments compare SNR-matched and SNR-mismatched conditions.
  • Robustness of SNR Mismatching: At SNR = 5 dB, SNR mismatching causes approximately 3 dB MSE loss but only slight BER performance loss for binary-symbol recovery.This indicates robustness of the ComNet receiver against SNR mismatching in the reported setting.
  • Complexity Analysis: To achieve better BER performance than FC-DNN within the same time period, ComNet-BiLSTM needs more than twice FC-DNN’s FLOPs but approximately one-fourth its memory.The comparison is reported as a complexity trade-off between computation and memory usage.
  • Complexity Analysis: ComNet-FC requires 0.37 million FLOPs and 1.22 MBytes of memory, and consumes fewer FLOPs than LMMSE-MMSE.Its settled post-training parameters avoid recomputing the LMMSE channel-estimation matrix as conditions vary.

IV. CONCLUSION

The conclusion presents ComNet as a model-driven OFDM receiver that combines deep learning with communication expertise for linear and nonlinear distortions. It highlights manually created features, nonlinear signal detection, and efficient deployment as implications of this design.

  • Conclusion: ComNet combines deep learning with expert knowledge to recover transmitted data in OFDM systems with linear and nonlinear distortions.The conclusion frames this architecture as a model-driven approach for physical-layer communications.
  • Receiver Implications: The coarse-to-fine idea in ComNet is described as intuitive and as offering deeper insight into its implications.The conclusion identifies this design perspective as an interpretive insight.
  • Receiver Implications: Nonlinear activation functions introduce nonlinearity into the signal-detection module, forming a nonlinear signal detector.This connects the neural-network operation to the receiver’s signal-processing role.
  • Receiver Implications: Communication intelligence enables manually created novel features that can accelerate training and result in efficient deployment performance.The conclusion presents this as a benefit of model-driven deep learning.
  • Future Work: Combining deep learning with expert knowledge in ComNet motivates future applications of model-driven deep learning to wireless physical layers.The stated implication is inspiration for future work.
Loading 1810.09082v1…