Abstract
A novel low-complexity wavenumber-domain method is proposed for near-field sensing (NISE). By exploiting the non-linear relationship between target positions and power-concentrated regions in wavenumber-domain channels, a bi-directional convolutional neural network (BiCNN) is developed for efficient target localization. This approach enables gridless localization using a single antenna array and limited bandwidth. Simulation results show that the BiCNN method achieves a 100x speedup over conventional 2D-MUSIC algorithms while maintaining comparable accuracy and requiring fewer trainable parameters than standard neural architectures.
Introduction
The transition to 6G wireless systems involves the adoption of extremely large aperture arrays (ELAAs) and higher frequency bands. This shift significantly expands the "near-field" region, where the spherical curvature of wavefronts becomes prominent. While this curvature allows for simultaneous estimation of a target's angle and distance using a single array, traditional localization methods like the 2D Multiple Signal Classification (MUSIC) algorithm are computationally prohibitive. These methods rely on exhaustive grid searches, leading to delays that are unacceptable for real-time applications such as autonomous driving.
Wavenumber-Domain Methodology
To address these computational challenges, the proposed method bypasses the grid search by transforming spatial antenna data into the wavenumber domain.
Wavenumber-Domain Transform
In the wavenumber domain, the target's spatial coordinates are encoded in the signal's power distribution. Specifically, the target's angular position determines the location of the power concentration, while its distance from the array is reflected in the width of this power region. This transformation converts a complex geometric estimation problem into a more manageable pattern recognition task.
Bi-directional CNN (BiCNN) Architecture
A specialized Bi-directional Convolutional Neural Network (BiCNN) is employed to map these wavenumber-domain patterns to exact target coordinates. Convolutional layers are used to capture local features of the power distribution. The "bi-directional" aspect involves feeding the network both the original observation vector and its flipped counterpart. This architectural choice allows the model to learn positional distributions more effectively with reduced complexity.
The training process utilizes the Huber Loss function, which provides robustness by combining the benefits of quadratic loss for small errors and linear loss for large outliers, ensuring stable convergence even in the presence of noise.
Performance Evaluation
The proposed BiCNN was evaluated using a synthetic dataset of 28 GHz signals received by a 511-antenna base station.
Localization Accuracy
The model demonstrates high precision in predicting target coordinates across various distances and angles. As shown in the comparison with ground truth data, the BiCNN accurately tracks target positions even under varying transmit power levels.
Computational Efficiency
The most significant advantage of the BiCNN method is its speed. During online implementation, the BiCNN requires only 0.0037 seconds to localize a target, representing a 100x speedup over a high-resolution 2D MUSIC search (0.5565 seconds). Despite this efficiency, the BiCNN maintains a Root Mean Square Error (RMSE) of approximately 0.24 meters, which is highly competitive with the gold-standard MUSIC algorithm.
Conclusion
The wavenumber-domain BiCNN approach offers a promising solution for real-time 6G near-field sensing. By eliminating the need for grid searches and leveraging the inherent symmetry of the signal space, it achieves high-precision, gridless localization with minimal computational overhead. Future work will focus on extending this framework to multi-target scenarios and complex non-line-of-sight environments.
Summary updated