Bridging Neural Networks and Wireless Systems with MIMO-OFDM Semantic Communications

Technical Summary: Bridging Neural Networks and Wireless Systems with MIMO-OFDM Semantic Communications

Problem Statement
While semantic communications—specifically Deep Joint Source-Channel Coding (DeepJSCC)—have demonstrated superior performance in simulations compared to conventional modular systems (separate source coding, channel coding, and modulation), significant performance gaps persist when these systems are deployed in real-world wireless environments. Prior research has identified discrepancies between simulated results and prototype measurements, yet the specific causes related to hardware limitations and practical channel characteristics remain underexplored. This article investigates why theoretical gains often fail to materialize in practice, focusing on the challenges introduced by Power Amplifier (PA) nonlinearity, Peak-to-Average Power Ratio (PAPR) variations, and frequency-selective noise in Multiple-Input Multiple-Output (MIMO) Orthogonal Frequency-Division Multiplexing (OFDM) systems.

Methodology
The authors developed a functional MIMO-OFDM semantic communication prototype to validate performance in a real-world setting. The system architecture employs an end-to-end deep learning approach where a CNN-Transformer encoder maps image patches directly to complex-valued wireless symbols, bypassing traditional bit-stream processing.

• Hardware Setup: The prototype utilizes NI USRP-2953R Software-Defined Radios (SDRs) configured for 2×2 and 4×4 MIMO in a static indoor Line-of-Sight (LoS) environment. A Mini-Circuits ZVA-183W-S+ amplifier is integrated to introduce controlled PA nonlinearity for testing.
• Signal Processing: The system transmits 32×32 image patches as 512 complex symbols using OFDM (128-point FFT, 72 active subcarriers). The receiver employs Zero-Forcing (ZF) channel estimation and combining.
• Mitigation Strategies: To address practical impairments, the authors implemented:
  1. Symbol Shuffling: Randomizing symbol allocation across subcarriers and spatial streams to average out frequency-selective fading and SNR variations, mimicking Additive White Gaussian Noise (AWGN) conditions.
  2. PAPR Regulation: Incorporating a PAPR penalty term into the loss function during training to constrain symbol power levels, thereby reducing distortion in the PA's nonlinear region.
  3. Normalization: Scaling symbols to prevent clipping within the fixed-point dynamic range of the DAC/ADC hardware.

Key Contributions

1. Prototype Validation: This work presents the first validation of semantic communication performance using a MIMO-OFDM prototype, moving beyond single-input single-output (SISO) or purely simulation-based studies.
2. Diagnosis of Performance Gaps: The study identifies that variations in Signal-to-Noise Ratio (SNR) across different subcarriers and spatial streams are the primary cause of the discrepancy between simulation and prototype results.
3. Demonstration of PAPR Benefits: The authors empirically demonstrate that while PAPR-restricted models may underperform in ideal (AWGN) simulations due to reduced representation power, they outperform baseline models in practical scenarios involving PA nonlinearity.
4. Open-Source Foundation: The authors have made their source code and hardware setup publicly available to facilitate further research into real-world semantic communication systems.

Experimental Results

• Linear Power Region: In the linear operating region, semantic systems outperformed conventional BPG+LDPC+QAM baselines across all tested SNR levels. Crucially, the prototype results with symbol shuffling closely aligned with simulated AWGN performance, whereas unshuffled prototypes showed significant degradation, particularly at lower SNRs. This confirms that SNR variance across resources is the main culprit for performance gaps.
• Nonlinear Power Region: When operating in the PA's nonlinear region (high input power), models trained with PAPR reduction techniques showed approximately a 1 dB improvement in Rx SNR compared to baseline models. While strong PAPR regulation reduced the Peak-to-Average Power Ratio significantly (from ~7.6 dB to ~4.3 dB), it incurred a trade-off in Peak Signal-to-Noise Ratio (PSNR) due to reduced symbol representation power. However, a "weak" PAPR regulation model achieved the best balance, outperforming the baseline in PSNR at higher power levels by mitigating nonlinear distortion without excessive representation loss.

Significance and Claims
The paper claims to establish a foundational bridge between theoretical semantic communication models and real-world deployment. It argues that the theoretical advantages of semantic communications can be realized in practical wireless systems if specific hardware and channel constraints are addressed. Specifically, the work highlights that:

• Performance gaps are not inherent to the semantic approach but are artifacts of unaddressed channel variations and hardware nonlinearity.
• Targeted mitigation strategies, such as symbol shuffling and PAPR-aware training, are essential for aligning prototype performance with theoretical expectations.
• Future system designs must move beyond idealized AWGN assumptions to incorporate channel-adaptive mechanisms and hardware-aware training to ensure robustness in dynamic wireless environments.

The authors maintain a modest stance, acknowledging that their current setup (static LoS, ZF equalization, AWGN-trained models) does not cover all deployment scenarios (e.g., non-line-of-sight, high mobility, massive MIMO). They position this work as a step toward understanding practical constraints rather than a complete solution for all future wireless generations.