Year: 2026
Type: article
Abstract: Recently, one of the critical challenges in beamforming is beam selection in modern communication systems, especially for large antenna arrays, due to the significant training overhead for exhaustive beam sweeping results.This paper presents a proposed modeldriven convolutional neural network (CNN)-based beam selection framework that depends on physical channel modeling and data-driven learning together to reduce the requirements for full channel estimation and lower the system overhead.CNN is trained offline in a supervised manner to converge the mapping between beam power patterns and optimal beam index.This enables fast beam selection during inference with minimal computational complexity.Monte Carlo simulations are employed to evaluate the performance of the proposed framework.The results confirmed that the proposed system achieves accuracies exceeding 99% and 93.8% under ideal and realistic channel conditions, respectively and 93.65% for the multi-user framework.Furthermore, the loss curve converges rapidly to zero, which indicates that the degradation in physical layer performance is insignificant, while Top-K accuracy demonstrates that the optimal beam is included in predicted candidates.Additional analyses such as reproducibility assessment over ten independent executions, sensitivity to measurement error, and multi-user performance validation were performed.All these metric analyses validate that the proposed framework is reliable, robust and effective under noise-free and multipath noisy channels.The system supports scalability due to its stable performance without noticeable degradation in the multi-user framework.
DOI: 10.18280/i2m.250408
Author: M. F. Ghanim
Institution: University of Mosul
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204257189
