Automatic Radar Signal Detection and FFT Estimation using Deep Learning

Akila Pemasiri, Zi Huang, Fraser Williams, Ethan Goan, Simon Denman, Terrence Martin, Clinton FookesICSPCS 2024 · OpenReview

TL;DR

This paper introduces a novel deep learning approach for simultaneously detecting radar signals and estimating their frequency content in challenging Electronic Warfare (EW) scenarios. By employing a compound loss function that combines time and frequency domain information, their method achieves improved accuracy and robustness compared to existing techniques.

What Problem Does This Paper Address?

Accurately detecting radar signals and estimating their characteristics, like bandwidth, is a fundamental step in radar signal processing. This is particularly crucial in Electronic Warfare (EW) environments where signals are often low probability of intercept (LPI) and lack prior information about their modulation. Traditional statistical methods struggle in these scenarios, and while deep learning has shown promise, existing methods often treat signal detection as a simple binary classification problem. This paper tackles the limitation of existing deep learning approaches by proposing a method that not only detects a signal but also precisely localizes it in time and estimates its operating frequency band.

Methods

The core of the proposed method is a deep learning architecture based on UNet, a well-established network for segmentation tasks. This UNet architecture is adapted for 1D signal processing using convolutional layers. What's novel is the compound loss function used during training. This loss function combines two components:

  • Segmentation Loss (Lseg): A Binary Cross-Entropy (BCE) loss to accurately identify the presence or absence of a signal in the time domain.
  • FFT Estimation Loss (Lfft): A Mean Squared Error (MSE) loss to estimate the frequency content (represented by the Fast Fourier Transform or FFT) of the detected signal.

By jointly optimizing for both segmentation and FFT estimation, the model learns complementary information from both time and frequency domain representations, leading to more robust performance. The network takes raw In-phase and Quadrature (IQ) signal data as input and outputs a segmentation mask (indicating signal presence) and an estimated FFT representation.

Datasets

The authors utilize a newly created synthetic radar signal detection dataset. This dataset builds upon previous benchmarks and includes five radar signal classes: coherent unmodulated pulses (CPT), Barker codes, polyphase Barker codes, Frank codes, and Linear frequency-modulated (LFM) pulses. Crucially, this dataset incorporates channel impairments beyond just Additive White Gaussian Noise (AWGN), including frequency and phase offsets, to better simulate realistic EW environments. The dataset is split into training (120,000 signals), validation, and testing sets (20,000 signals each).

Key Insights and Ideas

  • The Power of Compound Loss: The most significant insight is that combining segmentation and FFT estimation losses leads to substantial improvements in both tasks. This suggests that the time-domain localization and frequency-domain estimation are not independent but rather mutually reinforcing.
  • UNet's Adaptability: The UNet architecture, traditionally used for image segmentation, proves effective when adapted with 1D convolutions for time-series signal processing tasks like radar signal segmentation.
  • Robustness in Low SNR: The proposed method demonstrates improved performance even at low Signal-to-Noise Ratios (SNRs), which is a critical requirement for LPI signal detection in EW.
  • Beyond Binary Classification: Moving beyond simple signal detection to simultaneous localization and parameter estimation (like frequency band) is essential for more comprehensive radar signal analysis.

Key Results

  • The proposed compound loss function significantly improves FFT estimation accuracy compared to using only the segmentation loss, especially at lower SNRs. For instance, at -20 dB SNR, the cosine similarity for FFT estimation jumps from around 0.3 to 0.54 when the FFT loss is included.
  • The F1 score for segmentation also sees a notable improvement with the compound loss, indicating better overall detection accuracy.
  • The model achieves high probability of correct detection (Pa) of 98.36% and a low probability of false alarm (Pfa) of 1.34%, outperforming traditional energy detection methods.
  • Performance remains consistent across different sequence lengths (4096, 8192, 16384 samples).

Limitations and Future Work

  • Synthetic Dataset: While the dataset is enhanced with realistic impairments, it is still synthetic. Future work could involve testing the model on real-world radar data.
  • Dataset Extension: The authors suggest extending the dataset to include more challenging and realistic channel effects.
  • Multi-stage Models: Exploring other segmentation models, potentially multi-stage architectures, could further enhance performance.
  • Real-world Deployment: Adapting the model for real-time processing and deployment in actual EW systems would be a logical next step.

Summary updated

Abstract

This paper addresses a critical preliminary step in radar signal processing: detecting the presence of a radar signal and robustly estimating its bandwidth. Existing methods which are largely statistical feature-based approaches face challenges in electronic warfare (EW) settings where prior information about signals is lacking. While alternate deep learning based methods focus on more challenging environments, they primarily formulate this as a binary classification problem. In this research, we propose a novel methodology that not only detects the presence of a signal, but also localises it in the time domain and estimates its operating frequency band at that point in time. To achieve robust estimation, we introduce a compound loss function that leverages complementary information from both time-domain and frequency-domain representations. By integrating these approaches, we aim to improve the efficiency and accuracy of radar signal detection and parameter estimation, reducing both unnecessary resource consumption and human effort in downstream tasks.
BibTeX
@inproceedings{pemasiri2024automatic,
  author = {Akila Pemasiri and Zi Huang and Fraser Williams and Ethan Goan and Simon Denman and Terrence Martin and Clinton Fookes},
  title = {{Automatic Radar Signal Detection and FFT Estimation using Deep Learning}},
  booktitle = {ICSPCS},
  year = {2024},
  url = {https://openreview.net/forum?id=3b50kAakX2},
}
Figure