Extremely Large-scale Array (ELAA) is considered a frontier technology for future communication systems, playing a crucial role in enhancing the rate and spectral efficiency of wireless networks. As ELAA employs a multitude of antennas operating at higher frequencies, users are typically situated in the near-field region where the spherical wavefront propagates. Near-field beam training requires information on both angle and distance, which inevitably leads to a significant increase in the beam training overhead. To address this challenge, we propose a near-field beam training method based on deep learning. Specifically, we employ a convolutional neural network (CNN) to efficiently extract channel characteristics from historical data by strategically selecting padding and kernel sizes. The negative value of the user average achievable rate is utilized as the loss function to optimize the beamformer, maximizing the achievable rate in multi-user networks without relying on predefined beam codebooks. Once deployed, the model requires only pre-estimated channel state information (CSI) to compute the optimal beamforming vector. Simulation results demonstrate that the proposed scheme achieves more stable beamforming gains and substantially outperforms traditional beam training approaches. Furthermore, owing to the inherent traits of deep learning methodologies, this approach substantially diminishes the near-field beam training overhead.

Extremely Large-scale Array (ELAA) is considered a frontier technology for future communication systems, pivotal in improving wireless systems’ rate and spectral efficiency. However, as ELAA employs a multitude of antennas operating at higher frequencies, users are typically situated in the near-field region. This inevitably leads to a significant increase in the overhead of beam training, requiring two-dimensional beam searching in both the angle and the distance domain. To address this problem, we propose a learning-based codebook-free near-field beamforming method. We strategically select padding and kernel size of convolutional neural network to efficiently extract complex channel state information features. We optimize the beamformers to maximize achievable rates in a multi-user network without predefined beam codebooks. Our solution requires only pre-estimated channel state information for optimal beamforming vector derivation during deployment. Simulation results demonstrate stable beamforming gain compared to baseline schemes, and the deep learning approach substantially reduces near-field beam training overhead.

In the realm of emergency disaster relief, it is paramount to attain a thorough comprehension of the semantic information associated with the local disaster scene for strategic rescue path planning and immediate rescue operations for affected individuals. Unmanned aerial vehicle (UAV) networks are widely utilized for rapid data collection in the aftermath of disasters due to their flexibility and maneuverability, assisting in rescue decision-making. However, some disasters, such as seismic events and floods have disrupted the initially structured ground shape information, leading to a disparate distribution of data collected by various UAV groups. This exposes traditional semantic segmentation models susceptible to shortcut bias, posing challenges in adapting to semantic segmentation tasks in disaster scenarios. Thus, this paper proposes a bias-compensation augmentation learning based semantic segmentation framework, which substantially enhances the extraction capability of semantic information. Initially, we exploit an artificial augmentation neural network for bias-awareness to determine the relative bias values of the collected image data. Subsequently, considering the limited computing power resources in UAV networks, we present a bias compensation computation offloading strategy to achieve a relatively balanced distribution of semantic information across UAV nodes, optimizing the trade-off between network scheduling efficiency and model accuracy. We conduct reconstruction validation on the FloodNet dataset, and a plethora of experimental results demonstrate that, compared to traditional methods, this approach greatly improves the accuracy of pixel-level semantic segmentation by over 86.5%. Moreover, the average combined processing time is also reduced by over 50%, enhancing the utilization efficiency of limited computational resources.