Robust 3-D Near-Field Source Localization Method Based on Deep Feature Fusion Extraction Network

A deep learning (DL) based approach for near-field (NF) source localization is introduced to enhance the accuracy and robustness in both direction of arrival (DOA) and range estimation. The received data is first preprocessed by leveraging temporal and spatial correlation of the incident signals and symmetry of the employed cross array. This preprocessing step achieves parameter decoupling, reduces mutual interference between different features and simplifies the model training process. Next, a feature fusion extraction network (FFEN) is proposed, which mainly consists of a feature fusion network (FFN) and a feature extraction network (FEN). Adopting a multiple-input-multiple-output (MIMO) model structure, FFEN transforms the NF parameter estimation problem into a multi-task learning problem. The network adopts a dual-branch strategy for processing input data, which enables the extraction and learning of diverse parameter features from the data stream, thereby acquiring comprehensive information of all parameters of interest. Utilizing residual connections and introducing attention mechanisms allows the network to focus on important features within the input data, avoiding issues like gradient vanishing and network degradation, and subsequently, the three branches output labels for different parameters separately. This design ensures the network's efficiency and accuracy when dealing with complex multi-task learning problems. Experimental results demonstrate that the proposed method achieves high-precision parameter estimation with good generalization capabilities, and it also outperforms traditional model-driven approaches.