Overview
Traditional radar detectors struggle with high-speed targets because rapid movement creates "Doppler mismatch," causing significant signal loss. This paper introduces the Sliding Residual Network Detector (SRND), a deep-learning approach that bypasses traditional matched filtering to detect hypersonic targets with high sensitivity and waveform flexibility.
The Challenge of High-Speed Detection
In radar signal processing, the standard way to find a target is through "coherent integration," usually via a matched filter. This works perfectly if the target is stationary or moving slowly. However, when a target moves at high speeds (like hypersonic missiles or aircraft), it introduces a massive Doppler shift.
For phase-coded waveforms—which are popular for their complexity and security—this shift is a dealbreaker. These waveforms are "Doppler-sensitive," meaning even a slight mismatch between the expected and received signal causes the signal-to-noise ratio (SNR) to plummet. Traditional fixes involve using a "bank" of many filters tuned to different speeds, but this increases the risk of false alarms and requires more computational power.
The Sliding Residual Network Detector (SRND)
The authors propose the Sliding Residual Network Detector (SRND) as a practical approximation of the Likelihood Ratio Test (LRT), the mathematically optimal statistical test for detection. Its architecture is built on several key innovations:
- Waveform Adapter: This preprocessing step "decouples" the specific radar waveform from the detection problem. By compensating the received signal with the transmitted code, the residual network only needs to learn the representation of a target echo, making the model compatible with different phase codes without retraining.
- 1D Residual Network (ResNet): The core is an 8-layer 1D ResNet. Residual connections (skip connections) allow the network to be deep enough to capture complex features without suffering from vanishing gradients.
- Sliding Window Approach: To handle real-world data, the network "slides" across range cells, processing each one individually to produce a binary "target present/absent" result.
Experimental Validation
The study uses two distinct types of data:
- Simulated Data: Computer-generated radar echoes in Additive White Gaussian Noise (AWGN). This included targets moving at velocities ranging from -3400 m/s to 3400 m/s (roughly Mach 10).
- Measured Data: Real-world data recorded by a 16-element MIMO (Multiple-Input Multiple-Output) radar. This data included actual ground clutter and moving targets, which were pre-processed to remove clutter using the CLEAN technique.
Key Results
- Superior Detection: In simulations, the SRND consistently outperformed the standard CA-CFAR (Cell-Averaging Constant False Alarm Rate) detector and even multi-filter banks, especially in low-SNR environments.
- Velocity Robustness: While a single traditional filter's performance drops to near zero as target speed increases, the SRND maintains a high probability of detection across the entire $\pm 3400$ m/s range.
- Real-World Performance: On measured data, the SRND achieved a significantly higher detection count than the CA-CFAR detector, proving its practical utility and ability to generalize from synthetic training data.
- Interpretability: Using Guided Backpropagation (GBP), the authors demonstrated that the network effectively reconstructs the target echo, with deeper layers capturing the representation of the entire target pulse.
Limitations and Future Directions
The current SRND processes a single radar echo pulse. In many radar systems, multiple pulses are integrated over time to improve sensitivity. Additionally, determining the exact threshold for extremely low false alarm rates (e.g., $10^{-6}$) requires extensive Monte Carlo trials. Future work will explore extending this model to "coherent bursts" (sequences of pulses) to further improve detection performance in even noisier environments.
Summary updated