TL;DR
This paper introduces a novel method for identifying radar emitters, even in noisy environments with low signal-to-noise ratios (SNRs). By combining a new feature extraction technique called Energy Cumulant of Short Time Fourier Transform (EC-STFT) with a Reinforced Deep Belief Network (RDBN), the authors achieve robust and accurate radar emitter recognition.
What Problem Does This Paper Address?
Radar emitter recognition is crucial for intelligence gathering and defense systems. Traditional methods often rely on basic radar parameters (like frequency, pulse width, etc.), which are insufficient in today's complex electromagnetic environments with numerous overlapping signals. Pulse compression radars offer better performance, but recognizing their intra-pulse signals, especially at low SNRs where noise can obscure critical features, remains a significant challenge. Existing methods struggle to extract discriminative features from these noisy signals, leading to lower recognition rates.
Methods
The proposed method involves two main components:
- Feature Extraction: Energy Cumulant of Short Time Fourier Transform (EC-STFT)
- Short Time Fourier Transform (STFT): First, the intra-pulse radar signal is transformed into a time-frequency distribution (TFD) using STFT. This captures how the signal's frequency changes over time.
- Noise Reduction: Before feature extraction, a preprocessing step is applied to reduce noise in the STFT TFD. This involves a zero-mean scaling and a thresholding operation, repeated multiple times to enhance the signal's clarity.
- Energy Cumulant: The core of the new feature is the "energy cumulant." This is calculated by accumulating the absolute values of the STFT TFD across the time dimension for each frequency bin. This process essentially sums up the energy at each frequency over the entire pulse duration, creating a more robust feature that is less susceptible to noise and preserves key signal characteristics.
- Recognition Model: Reinforced Deep Belief Network (RDBN)
- Deep Belief Network (DBN): A DBN is a type of deep learning model composed of stacked Restricted Boltzmann Machines (RBMs). It's used here for its ability to learn hierarchical representations of data.
- Reinforcement Learning Integration: The authors propose a "reinforced" DBN. After the initial unsupervised training of the RBMs, a reinforcement learning concept is integrated to refine the weights within the RBMs. This reinforcement processing focuses on modifying the weight matrix based on the significance of connections, aiming to improve the network's ability to distinguish between different radar signals.
- Training: The RDBN is trained using the extracted EC-STFT features. The network learns to map these features to specific radar emitter types.
Datasets
The paper does not explicitly mention using pre-existing public datasets. Instead, the authors generated their own dataset of eight types of radar intra-pulse signals (NS, BFSK, QFSK, BPSK, QPSK, LFM, NLFM, and FRANK). These signals were simulated with varying SNRs, including both high SNRs (30-40 dB) for training and low SNRs (-10 to 5 dB) for testing. The training set consisted of 400 samples per type, and the testing set had 200 samples per type at each SNR.
Key Insights and Ideas
- EC-STFT as a Robust Feature: The EC-STFT feature is shown to be more resilient to noise than traditional STFT or Fourier Transform (FT) features, particularly at low SNRs. By accumulating energy across time for each frequency, it effectively "smooths out" noise while preserving discriminative signal characteristics.
- Reinforcement Learning Enhances DBN: Integrating reinforcement learning into the DBN architecture significantly improves its recognition performance, especially when dealing with subtle differences in radar signals. This suggests that fine-tuning the network's weights based on signal importance can lead to better classification.
- Low SNR Performance is Key: The primary focus on achieving high recognition rates at low SNRs is critical for practical applications where signals are often weak and corrupted by noise. The proposed method demonstrates a notable advantage in this challenging scenario.
- Network Architecture Matters: The number of hidden layers in the RDBN has a significant impact on performance. The study found that two hidden layers often provided the best balance of performance and computational cost, with more layers not always leading to better results due to accumulating losses.
Key Results
- High Recognition Rates at Low SNR: The proposed EC-STFT + RDBN method achieved a recognition rate of 95% at -5 dB SNR.
- Comparison with Other Methods: The EC-STFT feature, when used with RDBN, outperformed other feature sets (FT, FT-WMD) and other recognition networks (DBN, SDAE, SVM, NN) in various simulations.
- Parameter Tuning: Optimal network parameters were identified through traversal search, with a 2-hidden-layer architecture (70-110 nodes) achieving a 95.06% recognition rate at -5 dB.
- LFM Slope Impact: For Linear Frequency Modulated (LFM) signals, variations in slope and bandwidth (especially between LFM2 and NLFM) could lead to confusion, but the proposed method still maintained good recognition ability.
Limitations and Future Work
- Dataset Size and Format: The paper acknowledges that the format of datasets fed to deep learning models needs consideration, suggesting that input features should ideally be vectors. They also mention that unlimited conditions (e.g., longer time-domain signals, more TFD pixels) could lead to even higher performance.
- Single Feature Approach: The current work focuses on a single feature (EC-STFT). The authors suggest that exploring multiple features in conjunction with deep learning models could further enhance recognition performance.
- Pre-training Algorithms: Finding effective pre-training algorithms to generate desired input formats for deep learning models is identified as an area for future research.
- Online Learning: The current method relies on offline training with a pre-existing database. Future work could explore online learning capabilities for real-time adaptation.
Summary updated