UPCNN: CNN Framework for DOA Estimation in Presence of Unequal-Power Sources
UPCNN: CNN Framework for DOA Estimation in Presence of Unequal-Power Sources
Yuan Zhou Arnatovich Yauhen 1Aifei Liu 1Fengyu Chen 1Si Ouyang Xinyu Lu
作者信息
- 1. Xi'an Jiaotong Liverpool University; Northwestern Polytechnical University
- 折叠
摘要
Most existing deep neural network-based direction-of-arrival (DOA) estimation methods cannot be well generalized to scenarios involving unequal-power sources, as the DOA estimation of the weak signal may become inaccurate or completely lost. This issue is critical in practical applications such as automotive radar and intelligent transportation systems, where targets such as vehicles and pedestrians possess different radar cross-sections (RCS) due to their material composition and varying degrees of occlusion, resulting in substantial differences in received signal power. This paper proposes a new DOA estimation framework based on Convolutional Neural Network (CNN), named CNN for Unequal-Power sources (shortened as UPCNN). The proposed UPCNN fully explores the array signal information by constructing a 4-channel input data for the CNN, comprising the real parts, imaginary parts, magnitudes, and phases of the upper off-main diagonal elements of the array covariance matrix. Thus, the unequal-power source information can be better embedded in the input data. Additionally, zero-padding is used to ensure that each input channel has a square shape. Moreover, we specifically design the UPCNN label vector involving the power ratios of multiple source signals. Furthermore, we design a power-aware enhancement (PAE) layer to adaptively amplify the weak features and retain the strong features.With these domain knowledge-embedded designs and the PAE layer, the UPCNN is able to accurately estimate the DOAs of sources, even in the presence of a large source-power ratio, thereby improving the detection of weak signals. Ablation study is conducted to analyze the contribution of each novel component of the UPCNN. It is illustrated that 4-channel input data and the label vector with source-power ratios work together to improve the robust generalization of the UPCNN in multi-source scenarios. On the other hand, the PAE layer further enhances the robustness and increases the estimation accuracy. Finally, numerical simulation results verify the effectiveness of the UPCNN in different source-power ratios, signal-to-noise ratios (SNRs), numbers of snapshots, DOA separations, correlated signals, and non-uniform noise.
Abstract
Most existing deep neural network-based direction-of-arrival (DOA) estimation methods cannot be well generalized to scenarios involving unequal-power sources, as the DOA estimation of the weak signal may become inaccurate or completely lost. This issue is critical in practical applications such as automotive radar and intelligent transportation systems, where targets such as vehicles and pedestrians possess different radar cross-sections (RCS) due to their material composition and varying degrees of occlusion, resulting in substantial differences in received signal power. This paper proposes a new DOA estimation framework based on Convolutional Neural Network (CNN), named CNN for Unequal-Power sources (shortened as UPCNN). The proposed UPCNN fully explores the array signal information by constructing a 4-channel input data for the CNN, comprising the real parts, imaginary parts, magnitudes, and phases of the upper off-main diagonal elements of the array covariance matrix. Thus, the unequal-power source information can be better embedded in the input data. Additionally, zero-padding is used to ensure that each input channel has a square shape. Moreover, we specifically design the UPCNN label vector involving the power ratios of multiple source signals. Furthermore, we design a power-aware enhancement (PAE) layer to adaptively amplify the weak features and retain the strong features.With these domain knowledge-embedded designs and the PAE layer, the UPCNN is able to accurately estimate the DOAs of sources, even in the presence of a large source-power ratio, thereby improving the detection of weak signals. Ablation study is conducted to analyze the contribution of each novel component of the UPCNN. It is illustrated that 4-channel input data and the label vector with source-power ratios work together to improve the robust generalization of the UPCNN in multi-source scenarios. On the other hand, the PAE layer further enhances the robustness and increases the estimation accuracy. Finally, numerical simulation results verify the effectiveness of the UPCNN in different source-power ratios, signal-to-noise ratios (SNRs), numbers of snapshots, DOA separations, correlated signals, and non-uniform noise.关键词
Array signal processing/DOA/CNN/Unequal-power sourcesKey words
Array signal processing/DOA/CNN/Unequal-power sources引用本文复制引用
Yuan Zhou,Arnatovich Yauhen,Aifei Liu,Fengyu Chen,Si Ouyang,Xinyu Lu.UPCNN: CNN Framework for DOA Estimation in Presence of Unequal-Power Sources[EB/OL].(2026-07-27)[2026-07-30].https://chinaxiv.org/abs/202607.00267.学科分类
通信/无线通信/雷达
