Direction of Arrival (DOA) can be estimated through Sparse Recovery (SR) methods based on the sparsity of signals. However, conventional SR-DOA methods, such as the alternating direction method of multipliers (ADMM) encounter issues like difficulty in parameter setting and insufficient estimation accuracy. ADMM-Net is a neural network that combines ADMM with deep unfolding. In this letter, we propose a method for estimating DOA using ADMM-Net. The simulation results demonstrate that the proposed method achieves a better balance between DOA estimation performance and real-time constraints.

Range migration (RM) and Doppler frequency migration (DFM) make it challenging to achieve coherent integration for highly maneuvering targets. Generalized Radon-Fourier transform (GRFT) can compensate for RM and DFM, enabling coherent integration for targets, but its computational cost is significant. To address the aforementioned issue, this letter proposes a novel coherent integration method based on neural networks and GRFT. The proposed method utilizes a neural network to infer target trajectories from radar echoes. It narrows down the motion parameter search range of the GRFT method based on the inferred target trajectories, completing the coherent integration of target energy. Compared to GRFT and other existing methods, the proposed method achieves a better balance between detection performance and computational cost. At the same time, the proposed method possesses a certain degree of generalizability, maintaining good performance across different radar parameters. The results of simulation experiments confirm the aforementioned advantages.

Estimating the direction of arrival (DOA) for single snapshot signals is paramount in practical applications. In this letter, we present a method for single snapshot DOA estimation based on a neural network, namely the complex-valued residual angle estimation network (CVRAEN). Firstly, the CVRAEN network utilizes the initial feature extraction module including complex linear layer and complex convolutional layer to process the input signal and decompose it into multiple complex-valued feature signals for obtaining preliminary signal feature extraction. Subsequently, to extract depth features, the depth feature extraction module is employed to further process the extracted features. Finally, the mapping module combines and maps the extracted features for DOA estimation. This letter employs a series of training steps, including pre-training, signal-to-noise ratio (SNR) stratified training, and SNR adaptive training, to continuously optimize the network model. The main simulation results indicate that, within the simulated SNR range, the proposed method exhibits a decrease of approximately 30% in root mean square error compared to traditional methods.

In radar systems, high-resolution waveforms with favorable correlation properties are preferred. This letter addresses the challenge of designing unimodular radar waveform sets with low peak sidelobe level (PSL) in autocorrelation function (ACF). In contrast to conventional methods, this approach does not attempt to transform a non-convex problem into a convex one through relaxation. Inspired by neural network optimization techniques, an iterative neural network structure for minimizing PSL is proposed in this letter. By employing the Mellowmax operation and incorporating an additional penalty term into the loss function, the optimized ACF with low PSL is obtained. Corresponding simulation experiments demonstrate that our method achieves a superior PSL value of 2-3 dB lower than the state-of-the-art method.

Detection of highly maneuvering targets often suffers from the problem of the range migration (RM) and the Doppler frequency migration (DFM) within coherent processing interval, which results in performance degradation in coherent integration. In this letter, we propose a new coherent integration method which is based on neural networks and generalized Radon-Fourier transform (GRFT). Specifically, this method develops a neural network that can infer the target trajectory from the radar echo, and the trajectory reduces the search ranges of the GRFT method’s parameters according to some judgement criteria. After that, the RM and DFM are compensated via GRFT, and thereby it improves the performance of coherent integration. Besides, we introduce a customized regularization loss into the development of the neural network, which improves the detection performance. The superiority of the proposed method over GRFT is that it reduces the computational cost significantly with similar detection performance, which is confirmed by the results of experiments.