RC-Struct: A Structure-Based Neural Network Approach for MIMO-OFDM Detection

Jiarui Xu, Zhou Zhou, Lianjun Li, Lizhong Zheng, Lingjia Liu

Abstract

In this paper, we introduce a structure-based neural network architecture, namely RC-Struct, for MIMO-OFDM symbol detection. The RC-Struct exploits the temporal structure of the MIMO-OFDM signals through reservoir computing (RC). A binary classifier leverages the repetitive constellation structure in the system to perform multi-class detection. The incorporation of RC allows the RC-Struct to be learned in a purely online fashion with extremely limited pilot symbols in each OFDM subframe. The binary classifier enables the efficient utilization of the precious online training symbols and allows an easy extension to high-order modulations without a substantial increase in complexity. Experiments show that the introduced RC-Struct outperforms both the conventional model-based symbol detection approaches and the state-of-the-art learning-based strategies in terms of bit error rate (BER). The advantages of RC-Struct over existing methods become more significant when rank and link adaptation are adopted. The introduced RC-Struct sheds light on combining communication domain knowledge and learning-based receive processing for 5G/5G-Advanced and Beyond.

Introduction

Modern wireless communication standards like 5G and 6G rely on MIMO-OFDM for high-speed data transmission. Traditionally, symbol detection is handled by mathematical models. However, real-world hardware often introduces non-linearities, such as power amplifier distortion, which are difficult to model analytically. While Deep Learning (DL) offers a way to learn these complex behaviors, it faces two major hurdles in wireless systems: data scarcity (limited pilot symbols) and high training complexity.

RC-Struct addresses these challenges by combining the temporal processing power of Reservoir Computing (RC) with domain-specific structural knowledge. It achieves state-of-the-art performance in a purely online fashion using extremely limited training data.

The RC-Struct Architecture

RC-Struct is a two-stage architecture that splits the detection task between the time and frequency domains to maximize efficiency and accuracy.

  • Time Domain (Reservoir Computing): The system uses an Echo State Network (ESN). Unlike standard Recurrent Neural Networks (RNNs), the internal weights of the "reservoir" are fixed and random; only the output layer is trained. This allows for incredibly fast, stable online learning that captures the temporal dynamics of the wireless channel.
  • Frequency Domain (Structure-based NN): Instead of a generic multi-class classifier, the authors designed a binary classifier that exploits the symmetry of the modulation constellation. By "shifting" the received signal based on estimated channel information, the network only needs to learn a simple "positive vs. negative" decision.
The RC-Struct architecture, illustrating the transition from time-domain reservoir computing to frequency-domain binary classification.

This architecture enables Online Learning, where the entire system is trained subframe-by-subframe. It doesn't require a massive offline dataset; it learns the specific channel conditions of the moment using only the few pilot symbols available in that specific millisecond.

Performance Evaluation

The researchers evaluated RC-Struct using the QuaDRiGa channel simulator, which follows 3GPP 3D MIMO models. The model was tested against traditional methods (LMMSE, Sphere Decoding) and modern DL approaches (MMNet, RCNet) in a 4x4 MIMO system.

Experimental results demonstrate that RC-Struct consistently achieves lower Bit Error Rates (BER) compared to both conventional model-based detectors and state-of-the-art DL models like MMNet across various signal-to-noise ratios (Eb/No).

Bit Error Rate (BER) performance comparison showing RC-Struct outperforming several baseline methods, including MMNet and RCNet.

Key Insights and Discussion

  • Domain Knowledge as a Regularizer: By forcing the neural network to follow the known structure of a QAM constellation, the search space for the model is significantly reduced. This allows high performance even with very little data.
  • Efficiency through Binary Tasks: Converting complex multi-class classification into parallel binary tasks allows the network to reuse training samples more effectively. One pair of data points can generate multiple binary training samples.
  • Robustness and Adaptation: RC-Struct is particularly effective during link and rank adaptation. Because it learns online and leverages constellation symmetry, it adapts to changes in data streams or modulation types almost instantly, whereas other DL models may suffer from model mismatch.

Conclusion

RC-Struct meets the 3GPP requirement of using less than 20% pilot overhead, making it a viable candidate for actual 5G/6G implementation. Future work may extend this to high-mobility scenarios and explore even more efficient pilot patterns to further reduce overhead.

Summary updated

Abstract

In this paper, we introduce a structure-based neural network architecture, namely RC-Struct, for MIMO-OFDM symbol detection. The RC-Struct exploits the temporal structure of the MIMO-OFDM signals through reservoir computing (RC). A binary classifier leverages the repetitive constellation structure in the system to perform multi-class detection. The incorporation of RC allows the RC-Struct to be learned in a purely online fashion with extremely limited pilot symbols in each OFDM subframe. The binary classifier enables the efficient utilization of the precious online training symbols and allows an easy extension to high-order modulations without a substantial increase in complexity. Experiments show that the introduced RC-Struct outperforms both the conventional model-based symbol detection approaches and the state-of-the-art learning-based strategies in terms of bit error rate (BER). The advantages of RC-Struct over existing methods become more significant when rank and link adaptation are adopted. The introduced RC-Struct sheds light on combining communication domain knowledge and learning-based receive processing for 5G/5G-Advanced and Beyond.
BibTeX
@article{xu2022rc,
  author = {Jiarui Xu and Zhou Zhou and Lianjun Li and Lizhong Zheng and Lingjia Liu},
  title = {{RC-Struct: A Structure-Based Neural Network Approach for MIMO-OFDM Detection}},
  journal = {IEEE Transactions on Wireless Communications},
  year = {2022},
  doi = {10.1109/twc.2022.3155945},
}
Figure