Blind Source Separation of Radar Signals in Time Domain Using Deep Learning

TL;DR

This work tackles the challenge of deinterleaving overlapping RF emitters in contested environments by reframing the problem as blind source separation in the time domain. The authors adapt a state-of-the-art audio separation architecture into RF territory, implementing a three-module network (Feature Extractor, Feature Transformer, Separator) that operates on time-frequency representations to estimate two masks for two unknown signals from a single-channel mix. The approach supports end-to-end differentiable training with iSTFT-based reconstruction, enabling direct time-domain outputs and leveraging powerful audio-loss paradigms. Experiments on simulated and real RF data show the method can separate two unknown waveforms within a defined band under difficult conditions, including simultaneous chirps and pulsed signals, highlighting its potential for RF deinterleaving where traditional methods struggle.

Abstract

Identification and further analysis of radar emitters in a contested environment requires detection and separation of incoming signals. If they arrive from the same direction and at similar frequencies, deinterleaving them remains challenging. A solution to overcome this limitation becomes increasingly important with the advancement of emitter capabilities. We propose treating the problem as blind source separation in time domain and apply supervisedly trained neural networks to extract the underlying signals from the received mixture. This allows us to handle highly overlapping and also continuous wave (CW) signals from both radar and communication emitters. We make use of advancements in the field of audio source separation and extend a current state-of-the-art model with the objective of deinterleaving arbitrary radio frequency (RF) signals. Results show, that our approach is capable of separating two unknown waveforms in a given frequency band with a single channel receiver.

Figures (5)

• Figure 1: Overview of the whole system: In the preprocessing, we convert the superposition of signals from time domain to time-frequency domain and apply transformations to handle the dynamic range of amplitudes. The processed input is then fed to our three trainable network modules. The predicted masks from the Separator are transformed again in the postprocessing stage, multiplied with the input mixture in time-frequency domain and the resulting separated signals are converted back to time domain.
• Figure 2: Description of the Feature Transformer block.
• Figure 3: Simple setup for real data generation: Test samples are automatically transmitted and received by a (USRP X310 by Ettus) as seen on the right. The channel is a cable with a 6 dB attenuator.
• Figure 4: Train and test losses with real and simulated data evaluated each training epoch.
• Figure 5: Separation under challenging conditions.