Multi-stage Learning for Radar Pulse Activity Segmentation

Zi Huang, Akila Pemasiri, Simon Denman, Clinton Fookes, Terrence Martin

Abstract

Radio signal recognition is a crucial function in electronic warfare. Precise identification and localisation of radar pulse activities are required by electronic warfare systems to produce effective countermeasures. Despite the importance of these tasks, deep learning-based radar pulse activity recognition methods have remained largely underexplored. While deep learning for radar modulation recognition has been explored previously, classification tasks are generally limited to short and non-interleaved IQ signals, limiting their applicability to military applications. To address this gap, we introduce an end-to-end multi-stage learning approach to detect and localise pulse activities of interleaved radar signals across an extended time horizon. We propose a simple, yet highly effective multi-stage architecture for incrementally predicting fine-grained segmentation masks that localise radar pulse activities across multiple channels. We demonstrate the performance of our approach against several reference models on a novel radar dataset, while also providing a first-of-its-kind benchmark for radar pulse activity segmentation.

Introduction

In the high-stakes domain of Electronic Warfare (EW), the ability to identify and localize radar pulses in real-time is survival-critical. Systems must generate Pulse Descriptor Words (PDWs)—detailed data structures describing a pulse's width, frequency, and repetition interval—to deploy effective countermeasures. However, modern contested environments are increasingly complex, characterized by interleaved signals from multiple emitters and low Signal-to-Noise Ratios (SNR). Traditional digital signal processing methods often struggle with these conditions. While deep learning has transformed fields like medical imaging and audio processing, its application to raw radar IQ (In-phase and Quadrature) data has been hindered by a lack of large-scale, sample-wise annotated datasets.

Proposed Architecture: MS-UNet1D

To address these challenges, researchers from the Queensland University of Technology have introduced MS-UNet1D, a multi-stage deep learning architecture designed for precise pulse segmentation.

The MS-UNet1D architecture consists of sequential UNet1D stages that incrementally refine segmentation masks.

The model is built upon a 1D adaptation of the UNet architecture, traditionally used in medical imaging. MS-UNet1D operates directly on raw IQ samples, eliminating the need for manual feature engineering. The architecture consists of a sequential stack of UNet1D stages: the initial stage ($S_0$) generates a preliminary segmentation mask from the raw signal, while subsequent stages take both the previous stage's output and the original features to incrementally "clean up" and refine the predictions. This multi-stage loss approach forces the network to improve the mask progressively, effectively acting as a series of learned matched filters.

The RadSeg Dataset

A significant contribution of this research is the release of RadSeg, a first-of-its-kind open-source dataset providing a benchmark for the EW community. RadSeg contains 80,000 unique radar signals (60k training, 10k validation, 10k test) featuring five distinct signal classes: Coherent Unmodulated Pulses (CPT), Barker codes, Polyphase Barker codes, Frank codes, and Linear Frequency-Modulated (LFM) pulses. Unlike previous datasets, RadSeg provides binary segmentation masks for every sample over long time horizons (32,768 samples), including complex interleaving and noise levels ranging from -20 dB to +20 dB SNR.

Performance and Refinement

Experimental results demonstrate that MS-UNet1D excels in low-SNR environments. At -20 dB SNR, the model achieved a 15.5% increase in Intersection-over-Union (IoU) compared to a single-stage baseline. Furthermore, MS-UNet1D is computationally efficient; it achieves competitive performance with approximately 10.8M parameters, significantly fewer than the 23.1M parameters required by MS-TCN++, making it a strong candidate for real-time hardware deployment.

Qualitative results showing how predictions are refined from Stage 0 to Stage 3, aligning more closely with ground truth masks.

Qualitative analysis shows that the multi-stage process effectively eliminates "over-segmentation" errors—where a single pulse is incorrectly fragmented—and smooths boundaries to align closely with ground truth. This refinement is particularly crucial for maintaining the temporal continuity of radar pulses in high-noise scenarios.

Conclusion

MS-UNet1D represents a significant step forward in radar pulse activity recognition, offering a robust solution for deinterleaving signals in contested environments. While the model shows great promise, future research will focus on expanding the dataset to include more complex modulations and testing the system in real-world hardware-in-the-loop simulations to further validate its utility in electronic warfare applications.

Summary updated

Abstract

Radio signal recognition is a crucial function in electronic warfare. Precise identification and localisation of radar pulse activities are required by electronic warfare systems to produce effective countermeasures. Despite the importance of these tasks, deep learning-based radar pulse activity recognition methods have remained largely underexplored. While deep learning for radar modulation recognition has been explored previously, classification tasks are generally limited to short and non-interleaved IQ signals, limiting their applicability to military applications. To address this gap, we introduce an end-to-end multi-stage learning approach to detect and localise pulse activities of interleaved radar signals across an extended time horizon. We propose a simple, yet highly effective multi-stage architecture for incrementally predicting fine-grained segmentation masks that localise radar pulse activities across multiple channels. We demonstrate the performance of our approach against several reference models on a novel radar dataset, while also providing a first-of-its-kind benchmark for radar pulse activity segmentation.
BibTeX
@inproceedings{huang2024multi,
  author = {Zi Huang and Akila Pemasiri and Simon Denman and Clinton Fookes and Terrence Martin},
  title = {{Multi-stage Learning for Radar Pulse Activity Segmentation}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2024},
  url = {https://openreview.net/forum?id=Ublc5VsHh0},
}
Figure