FlowSep-RF: Rectified Flow Matching for Single-Channel Radio-Frequency Signal Separation

• Matched filtering and LMMSE can be limited when structured, non-Gaussian communication interference violates the additive-Gaussian-noise model.Source: §1 · p. 1
• Direct neural regression does not explicitly model mixture-to-SOI transport, while diffusion and score-based separators commonly require iterative sampling or optimization.Source: §1 · p. 1
• Flow matching learns a velocity field between source and target distributions, avoiding long reverse-diffusion chains.Source: §1 · p. 1
• FlowSep-RF learns clean-endpoint transport from the received mixture to the SOI, combining time-domain IQ features with frequency-domain structure for one-step separation.Source: §1 · p. 1
• FlowSep-RF improves BER and MSE over classical, discriminative, and generative baselines on QPSK mixtures with OFDM-BPSK, OFDM-QPSK, and CommSignal2 interference.Source: §1 · pp. 1–2

• The formulation names s as the SOI, b as the co-channel interferer, and y as the received mixture; κ controls interference scale.Source: §2 · p. 2
• Because the true interference scale is unavailable in blind receivers, FlowSep-RF uses a power-based estimate κ̂(y).Source: §2 · p. 2
• Training jitters the conditioning scale by ξ sampled uniformly from 0.9 to 1.1 to improve robustness to estimation error.Source: §2 · p. 2
• BER is the primary metric for transmitted-information recovery, while MSE measures waveform reconstruction accuracy.Source: §2 · p. 2

3. FLOWSEP-RF

• FlowSep-RF learns direct transport from the mixture to the clean SOI, rather than predicting and subtracting the interference.This clean-endpoint formulation aligns the transport target with the separation objective.Source: §3 · p. 2

3.1. Clean-endpoint rectified flow

• Training interpolates between mixture and SOI; inference evaluates the learned field at the mixture endpoint to produce the separated signal in one neural evaluation.The mixture remains conditioning information, with jittered scale during training and a power-based estimate at inference.Source: §3.1 · p. 2

3.2. Dual-domain architecture

• The dual-domain architecture processes IQ and Fourier features in parallel, then maps frequency-domain velocity back to time and combines it with time-domain velocity.A learned scalar gate controls the frequency-domain contribution.Source: §3.2 · p. 2

3.3. Training objective

• Training combines clean-endpoint velocity prediction with waveform, symbol-proxy, and spectral-magnitude fidelity losses, up-weighting low- and mid-SIR examples.The symbol projection is a differentiable proxy for demodulation quality; the SIR weighting emphasizes harder mixtures.Source: §3.3 · pp. 2–3

4.1. Datasets

• Three settings pair an RRC-filtered QPSK SOI with OFDM-BPSK, OFDM-QPSK, or RFChallenge CommSignal2 interference.The test grid spans −24 to −3 dB SIR, with 400 held-out examples per SIR level for each interference type.Source: §4.1 · p. 3

4.2. Baselines

• Baselines include matched filtering and LMMSE, one-pass U-Net-RF and DirectComm-RF, and iterative DDPM, reverse diffusion, and α-RGS.These represent linear receivers, fast supervised separators, and iterative generative separation methods.Source: §4.2 · p. 3

4.3. Implementation details

• Evaluation reports BER, MSE, and inference efficiency; Fig. 2 plots BER and waveform MSE versus SIR, while FlowSep-RF uses one endpoint evaluation.Models train for 20 epochs using the best validation checkpoint; U-Net-RF and DirectComm-RF also use one-pass inference, unlike iterative generative baselines.Source: §4.3 · pp. 3–4, Figure 2 · p. 4

5. RESULTS

• FlowSep-RF achieves the lowest average BER and MSE across all three settings; on OFDM-BPSK, BER is 6.45 × 10−4 and MSE is 1.47 × 10−3.On OFDM-QPSK and CommSignal2, it also outperforms DDPM, U-Net-RF, DirectComm-RF, Reverse Diffusion, and α-RGS.Source: §5 · p. 4
• One endpoint evaluation matches U-Net-RF and DirectComm-RF, while DDPM uses 25 denoising steps and α-RGS uses 2000 test-time optimization steps.Reverse diffusion is iterative.Source: §5 · p. 4
• Removing the rectified-flow objective or any auxiliary component degrades BER and MSE across all three settings.The symbol-proxy, spectral, and SIR-weighting terms improve receiver accuracy, frequency agreement, and low-SIR robustness, respectively.Source: §5 · p. 4, Table 2 · p. 4

6. CONCLUSION

• FlowSep-RF learns a clean-endpoint transport from the received mixture to the SOI, combines time- and frequency-domain features, and performs one-step inference.Experiments use QPSK mixtures with synthetic OFDM and RFChallenge CommSignal2 interference.Source: §6 · p. 4
• Across the evaluated settings, FlowSep-RF achieves the lowest average BER and MSE among evaluated baselines.The conclusion reports experiments with synthetic OFDM and RFChallenge CommSignal2 interference.Source: §6 · p. 4
• Ablations confirm the value of the rectified-flow objective and receiver-aware, spectral, and SIR-aware training terms.The authors conclude that these results support clean-endpoint rectified flow for structured co-channel interference.Source: §6 · p. 4

Compliance with Ethical Standards

• No ethical approval was required because this was a numerical simulation study.The authors report no funding for the study.Source: Compliance with Ethical Standards · p. 5
• The study received no funding.Source: Compliance with Ethical Standards · p. 5
• The authors disclosed no relevant financial or nonfinancial interests.Source: Compliance with Ethical Standards · p. 5