Machine-learning-assisted Mobile Communication Base Station Array Antenna Design

With the continuous and rapid development of mobile communications,the requirements for base station antennas are also getting higher.Therefore,after completing the initial design of the antenna,in order to achieve a balance between numerous structural parameters and design goals,antenna engineers often need to spend a lot of time on parameter analysis and optimization,which requires a lot of time and cost.The introduction of machine learning into the design and optimization of antennas can greatly save the number and time required for full-wave simulations,and can also help explore the physical upper limit of the antenna structure.In this paper,the main work and innovation points of the machine-learning-assisted base station array antenna design are as follows:Firstly,according to the requirements of the low profile of the base station antenna,three antenna units with different feeding methods are designed,using coplanar waveguide feeding,electromagnetic coupling feeding and T-shaped proximity coupling feeding respectively.The-10d B impedance bandwidth of the three antenna units covers the N41 frequency band（2.496GHz～2.69GHz）,and the gain at the center frequency of 2.6GHz reaches 5d Bi.For the electromagnetic coupling feeding scheme and the T-shaped proximity coupling feeding scheme,one-to-three power division networks with equal output port amplitudes are designed,which are combined with the previously designed antenna units to form two three-element base station array antennas.The-10d B impedance bandwidth of the coupled feed base station array antenna is 2.51～2.81GHz（300MHz）,the isolation|₂₁|in the N41 frequency band floats between-20d B～-10d B,and the gain at 2.6GHz reaches 9.55d Bi.The-10d B impedance bandwidth of the T-shaped proximity coupled feed base station array antenna is 2.31～2.97GHz（660MHz）,and the-14d B impedance bandwidth is 2.41～2.77GHz（360MHz）,the isolation|₂₁|in the N41frequency band fluctuates around-14d B and the gain at 2.6GHz reaches 10.05d Bi.Secondly,in view of the problem that antenna optimization requires a lot of computational resources and time costs,the machine-learning-assisted optimization（MLAO）method is proposed.The core idea is to fit the relationship between the structural parameters and the response of the antenna through the Gaussian process regression surrogate model,to predict the solution given by the genetic algorithm,and at the same time a comprehensive analysis of the group of solutions is performed using full-wave simulations.The predicted value is compared with the simulated value,the new input and response are updated to the database of the surrogate model,and a process of updating the surrogate model online is completed.The two base station array antennas proposed above are optimized by the MLAO method.As for the electromagnetic coupling feeding base station array antenna,the part above-10d B of|₁₁|in the previous 2.48～2.51GHz frequency band was successfully optimized to below-10d B,and the gain at 2.6GHz was increased from 9.55d Bi to 9.78d Bi.And as for the T-shaped proximity coupling feeding base station array antenna,the part above-10d B of|₂₁|is successfully optimized to be below-14d B,while ensuring that the-14d B impedance bandwidth and 9.5 d Bi gain bandwidth of the antenna completely cover the N41 frequency band.Finally,aiming at the problems of insufficient degrees of freedom in parameter optimization and asymmetric feed port design,machine-learning-assisted antenna shape optimization is proposed.On the basis of the machine-learning-assisted antenna optimization（MLAO）,the edge or center of the antenna is meshed,and the length of the mesh is set as an optimization parameter,so as to optimize the shape of the edge or center of the antenna.The example optimization and verification of the T-shaped adjacently coupled feed base station array antenna proposed above are carried out.Through the edge shape optimization,the|₂₁|in the N41 frequency band is optimized from-14d B to-20d B,and the cross-polarization ratio of the two ports is higher than 18 d B,while ensuring that the-14d B impedance bandwidth and 9.5 d Bi gain bandwidth of the antenna completely cover the N41 frequency band.