Abstract
Direction of Arrival (DOA) estimation in passive radar systems is challenging due to the reliance on non-cooperative "illuminators of opportunity" and the inherent physical imperfections of antenna arrays. This paper proposes a robust deep learning framework, termed LCN-CNN, which utilizes a Linear Classification Network (LCN) for coarse spatial localization followed by parallel Convolutional Neural Networks (CNNs) for fine-grained DOA estimation. By learning the mapping between received signal covariance and spatial angles, the framework demonstrates superior resilience to antenna gain/phase errors and mutual coupling compared to traditional subspace-based methods like MUSIC.
Introduction
Passive radar systems leverage existing electromagnetic signals—such as FM radio, digital television (DVB-T), and satellite transmissions—to detect and track targets. Unlike active radar, these systems do not emit their own signals, making them covert and cost-effective. However, estimating the Direction of Arrival (DOA) of target echoes is complicated by the diverse nature of these illuminators and the non-ideal characteristics of practical antenna arrays.
Traditional DOA estimation algorithms, such as MUSIC and ESPRIT, are mathematically elegant but physically brittle. They rely on precise mathematical models of the antenna array. In real-world deployments, factors like sensor position inaccuracies, gain/phase imbalances, and mutual coupling between elements cause these models to deviate from reality, leading to significant performance degradation for traditional algorithms.
The LCN-CNN Framework
To address these challenges, the authors propose a "divide and conquer" deep learning architecture. The core idea is to reduce the generalization burden on any single neural network by decomposing the 180-degree spatial horizon into smaller sub-regions.
The framework operates in several stages:
- Preprocessing: The system computes the normalized covariance matrix of the received signals, extracting essential spatial correlation features while reducing input dimensionality.
- Coarse Classification (LCN): A Linear Classification Network identifies which of the $P$ predefined spatial sub-regions contains the signal.
- Fine Estimation (Parallel CNNs): The signal data is routed to a specific CNN dedicated to that sub-region. Because each CNN is specialized for a narrow angular range, it can achieve much higher precision than a monolithic network.
- Output Fusion: A weighted SoftMax function integrates the outputs to provide a final high-resolution DOA estimate.
Experimental Results and Robustness
The framework was evaluated using simulated data for four types of illuminators: FM, DVB-T, GPS, and GSM. A 10-element antenna array was modeled with various levels of induced defects, including amplitude and phase errors.
Key findings include:
- Robustness to Defects: While the performance of the traditional MUSIC algorithm collapsed under significant antenna array errors, the LCN-CNN framework maintained a consistently low Root Mean Square Error (RMSE). This is because the network learns the actual response of the "imperfect" array during training.
- SNR Performance: The system achieves stable and accurate estimation once the Signal-to-Noise Ratio (SNR) exceeds 4dB, which is well within the operational range of passive radar systems.
- Full Angular Coverage: The architecture successfully covers the entire [0, 180] degree range, overcoming the limited field-of-view constraints seen in some previous deep learning approaches.
Conclusion
The LCN-CNN framework represents a significant step forward in making passive radar systems more practical and resilient. By shifting from rigid mathematical models to adaptive deep learning architectures, DOA estimation can remain accurate even in the presence of hardware imperfections. Future research will focus on extending this model to handle multiple simultaneous targets and validating the performance with real-world field data.
Summary updated