A U-Net Architecture for Time-Frequency Interference Signal Separation of RF Waveforms

This paper presents a data-driven approach to solve the challenge of separating co-channel mixture signals in the radio spectrum. The main aim is to extract the signal-of-interest with high fidelity from the mixture signal, allowing improved performance in demodulation and decoding tasks. We have developed a U-Net architecture specifically designed for the separation of interference signals within the time-frequency domain. This architecture integrates elements of OFDM signal resource grid configurations, like the cyclic prefix, ensuring a tailored and effective approach to signal processing. This approach has demonstrated a significant improvement, with an average 63% enhancement in MSE performance over the baseline model on four different interference types.