Abstract
The limited over-the-air (OTA) pilot symbols in multiple-input-multiple-output orthogonal-frequency-division-multiplexing (MIMO-OFDM) systems presents a major challenge for detecting transmitted data symbols at the receiver, especially for machine learning-based approaches. While it is crucial to explore effective ways to exploit pilots, one can also take advantage of the data symbols to improve detection performance. Thus, this paper introduces an online attention-based approach, namely RC-AttStructNet-DF, that can efficiently utilize pilot symbols and be dynamically updated with the detected payload data using the decision feedback (DF) mechanism. Reservoir computing (RC) is employed in the time domain network to facilitate efficient online training. The frequency domain network adopts the novel 2D multi-head attention (MHA) module to capture the time and frequency correlations, and the structural-based StructNet to facilitate the DF mechanism. The attention loss is designed to learn the frequency domain network. The DF mechanism further enhances detection performance by dynamically tracking the channel changes through detected data symbols. The effectiveness of the RC-AttStructNet-DF approach is demonstrated through extensive experiments in MIMO-OFDM and massive MIMO-OFDM systems with different modulation orders and under various scenarios.
1. Introduction
In modern wireless communication systems, Multiple-Input-Multiple-Output Orthogonal Frequency Division Multiplexing (MIMO-OFDM) is the workhorse technology. A critical challenge for these receivers is the rapid variation of the wireless channel due to user mobility and environmental interference. Traditionally, receivers rely on "pilot symbols"—known reference signals—to estimate the channel. However, pilots are overhead; increasing their number reduces the bandwidth available for actual data transmission.
While Deep Learning (DL) has shown promise for symbol detection, many models are data-hungry and struggle to adapt to dynamic environments after offline training. This paper proposes RC-AttStructNet-DF, a hybrid architecture designed to address pilot scarcity and dynamic channel tracking by "learning as it detects" through a decision feedback mechanism.
2. Proposed Architecture: RC-AttStructNet-DF
The proposed receiver architecture integrates time-domain and frequency-domain processing to maximize efficiency and accuracy.
2.1 Time-Domain Reservoir Computing (RC)
The initial stage utilizes Reservoir Computing (RC) in the time domain. Unlike traditional Recurrent Neural Networks (RNNs) that require extensive backpropagation through time, RC keeps the internal reservoir weights fixed and only trains the output layer. This significantly reduces the computational burden for online adaptation, allowing the system to perform initial stream decoupling and channel equalization efficiently.
2.2 2D Multi-Head Attention (MHA)
Following the Fast Fourier Transform (FFT), the signal enters the frequency-domain network. A key innovation is the 2D MHA module. Standard attention mechanisms typically focus on either temporal or spectral correlations. In contrast, the 2D MHA captures correlations across both dimensions simultaneously, which is vital for tracking fading channels where Doppler shifts and multi-path delays create a complex "checkerboard" of correlations.
2.3 StructNet and Decision Feedback
The final stage is StructNet, a frequency-domain network that embeds domain knowledge regarding the repetitive geometric structure of QAM constellations. It includes a Parameter Estimation (PE) layer that acts as a learnable channel estimator.
The Decision Feedback (DF) mechanism allows the receiver to use detected data symbols as pseudo-pilots for online weight updates. To mitigate "error propagation" (learning from a wrong guess), the authors employ an Attention Loss function. This function weights training samples based on the model's prediction confidence, effectively filtering out noise in the feedback loop by prioritizing "sure" detections.
3. Experimental Evaluation
The RC-AttStructNet-DF approach was evaluated using high-fidelity simulations based on the 3GPP 3D Channel Model via the QuaDRiGa simulator. Tests included standard MIMO (4x4) and massive MIMO (up to 64 receive antennas) with user speeds of 30 km/h.
Key Results
- Superior Accuracy: The model consistently achieved lower Bit Error Rates (BER) than conventional LMMSE (Linear Minimum Mean Square Error) and specialized DL models like MMNet, especially in high-mobility and high-modulation (64-QAM) scenarios.
- Computational Efficiency: RC-AttStructNet-DF is approximately 100 times faster than MMNet in total processing time per subframe, making it a viable candidate for real-time implementation.
- Nonlinear Resilience: The system showed significant gains in scenarios with Power Amplifier (PA) distortion, where traditional linear receivers typically fail.
- Pilot Efficiency: It performs better with just 4 pilot symbols than other methods do with significantly more overhead, effectively increasing the data throughput of the link.
4. Conclusion
RC-AttStructNet-DF represents a significant step toward adaptive, real-time MIMO-OFDM receivers. By integrating structural domain knowledge with efficient online learning mechanisms like RC and DF, the system effectively overcomes the limitations of static DL models. Future work will explore even higher mobility scenarios (e.g., high-speed rail at 500 km/h) and further model compression for deployment on low-power edge devices.
Summary updated