A High Precision DOA Estimation Method Based on Convolutional Neural Network

In recent years, deep learning (DL) techniques, such as convolutional neural networks (CNNs), have been widely applied to direction-of-arrival (DOA) estimation. However, existing DL-based methods face challenges in multi-target and dense-target scenarios, with limited improvement in accuracy as the signal-to-noise ratio (SNR) increases. To enhance estimation accuracy in complex wave scenarios, this paper proposes a CNN-based DOA estimation method with the following features: (1) By dividing the entire signal space into multiple subspaces, each CNN corresponding to a subspace treats in-subspace signals as targets and out-of-subspace signals as interference. (2) The proposed method enhances target resolution by narrowing subspace coverage and arranging overlapping networks. When the range is reduced, parameter transfer learning is introduced to lower the loss. (3) To address poor CNN performance at label boundaries, a label redefinition method based on grid interleaving is employed, reducing misjudgments from insufficient label weights. This method enables the CNNs to focus on capturing the features of targets within the current subspace. Through the collaboration of CNNs with different subspace coverage ranges, accurate DOA estimation is achieved for multi-target and dense-target scenarios. By comparing different training strategies, the relationship between signal numbers, subspace coverage, and CNN performance is analyzed. The proposed method demonstrates robustness to noise and strong generalization across various wave signal scenarios, highlighting its effectiveness and reliability in complex environments.