A dual-polarised array antenna using AEP-DNN technique for beam-steering in millimeter wave applications

Abstract

Nowadays, the evolution of wireless networks has become increasingly robust due to rising demands. Notably, one key feature of 5G is its capacity to offer broader coverage of received signals. Polarization diversity can be used to improve received signal coverage and mitigate link drop caused by polarization mismatch. In this study, a novel approach uses a single feeding port to simultaneously excite dual linearly polarised patch antenna array (DLPAA) at 28 GHz. Antenna array optimization involves pattern synthesis to achieve the desired radiation pattern of the proposed DLPAA. Currently, full-wave and conventional optimization tools are usually used to synthesise an antenna array, which makes the design slow and computationally expensive. This could lead to high inaccuracy in array synthesis especially when the antenna is highly sensitive in millimeter-waves. Deep learning algorithms map input features and target variables, automatically adjusting parameters based on known pairs. The project aims to develop active element pattern and deep neural network (AEP-DNN) algorithms to improve antenna array synthesis efficiency and reliability. Therefore, this study has led to three research objectives; the work begins with the design of dual-polarised antennas operating at 28 GHz. This is followed by the synthesis of a proposed DLPAA using the AEP-DNN method to enable beam-steering capabilities. This method is proposed because it demonstrates faster convergence, improved pattern prediction accuracy for large input datasets, reduces simulation costs and computational complexity while increasing overall learning efficiency. Finally, the antennas are fabricated and the algorithm is verified through testing. The proposed DLPAA is a 1×4 linear array (62.45 × 23.08 × 1.575 mm³) with 15.60 mm spacing. Simulated in CST, it achieves S₁₁ of –13 dB per port, with measured values ranging from –11.57 dB to –35.23 dB. The proposed DLPAA achieves these values through optimized impedance matching, dual-polarised design, and careful array configuration. The simulated co-polarised gains are improved from 12.10 dBi to 14.40 dBi (elevation) and 12.00 dBi to 13.10 dBi (azimuth). Although some discrepancies exist between simulated and measured results is due to fabrication or instrument errors, the performance remains acceptable. The AEP-DNN method, implemented in MATLAB, successfully steered beams to 0°, 5°, 10°, 15°, and 20°, with the predicted DNN patterns aligning well with the desired main lobe directions. RMSE values converged to approximately 1.5 (training) and 1.3 (validation), while training and validation losses reached minimums of 2.1–2.3 and 1.7–1.9, respectively. The method has been effectively applied for pattern synthesis and verified through both CST simulations and measurements on the fabricated antenna.