Broad Neural Networks for Sparse Signal Retrieval and Array Interpolation in Automotive Radar

Sparse antenna arrays are widely used in automotive radar systems to achieve high angular resolution with reduced hardware complexity and mutual coupling. However, the high sidelobes inherent in sparse configurations can degrade angle estimation accuracy, motivating the need for effective interpolation methods. Although deep neural networks (DNNs) have shown strong performance in sparse signal recovery, their deployment is hindered by high computational cost, complex activation functions, and limited interpretability due to deep-layer architectures. To address these challenges, we propose a novel framework based on broad neural networks (BNNs) for efficient and accurate sparse signal retrieval. Unlike conventional DNNs, BNNs utilize parallel network structures with simplified, custom-designed activation functions, eliminating hidden layers to reduce computational overhead and improve transparency. We further enhance this design with an iterative BNN approach that incorporates a scalar mask and a phase-enhanced layer, enabling high-accuracy data recovery with fewer iterations. The experimental results on simulated and real-world radar datasets demonstrate that the proposed BNN framework significantly outperforms baseline methods in angle spectrum estimation and data interpolation while maintaining low computational complexity. These findings highlight the potential of BNNs as a practical, interpretable, and scalable alternative for advanced automotive radar applications.