Abstract
We consider the problem of classifying radar pulses given raw I/Q waveforms in the presence of noise and absence of synchronization. We also consider the problem of classifying multiple superimposed radar pulses. For both, we design deep neural networks (DNNs) that are robust to synchronization, pulse width, and SNR. Our designs yield more than 100x reduction in error-rate over the previous state-of-the-art.
Introduction
Classifying radar pulses is a cornerstone of cognitive radar and electronic warfare, yet it presents significant challenges. Unlike standard image classification, radar signals (raw I/Q waveforms) exhibit extreme scale variance in pulse widths and are often asynchronous, meaning the pulse's start time is unknown. Furthermore, these systems must often identify signals buried in noise, sometimes far below 0 dB. Traditional methods often rely on hand-crafted features or real-valued neural networks that fail to fully exploit the complex-valued nature (In-phase and Quadrature) of radar data.
Methods
The researchers developed a Complex-Valued Deep Residual Network (CResNet) to process raw time-domain samples directly. Unlike standard approaches that treat I and Q components as separate real-valued channels, CResNet utilizes complex-valued arithmetic within its convolutional layers. This approach better captures the mathematical relationship between phase and amplitude while reducing the number of learnable parameters to prevent overfitting.
To handle asynchronicity, the team employed asynchronous data augmentation, training the model with "noise padding" and random delays. This forces the network to identify the signal regardless of its position in the observation window. Additionally, they identified that input length acts as a critical hyperparameter, where an optimal window size prevents the signal from being diluted by excessive noise.
The study utilized the SIDLE dataset, focusing on 17 classes of phase-modulated radar pulses with Signal-to-Noise Ratios (SNRs) uniformly distributed between -12 dB and +12 dB. For multi-pulse scenarios, the model was adapted to use Binary Cross-Entropy (BCE) loss for multi-label classification.
Results
The transition from shallow DNNs to deep complex-valued architectures yielded dramatic improvements. A previous state-of-the-art 9-layer DNN achieved an 18.29% error rate on asynchronous data. In contrast, the CResNet-30 architecture achieved significantly higher accuracy with fewer parameters than its real-valued counterparts.
Key findings include:
- Architecture Impact: Moving to a 30-layer ResNet using only the real component reduced error to 2.14%. Incorporating the full complex signal via CResNet further reduced this to 0.36%.
- Optimization: Final hyperparameter tuning led to a test error of 0.14%, representing a 100x improvement over the baseline.
- Multi-Pulse Robustness: The model maintained a 4.0% absolute error rate even when four different pulses overlapped simultaneously.
Conclusion
This work demonstrates that deep complex-valued networks, combined with proper data augmentation and input length optimization, can achieve unprecedented accuracy in radar waveform classification. While the error rate grows approximately linearly with the number of overlapping pulses, the current results set a new benchmark. Future research will explore time-domain object detection techniques to better localize and identify pulses in increasingly dense signal environments.
Summary updated