Research On Electromagnetic Modeling Of Array Antennas Based On Artificial Neural Network

As wireless communication systems and cutting-edge applications continue to develop,future array antenna designs tend to become structurally complex,multifunctional,multiband,multi-polarized,and adaptive,posing high-dimensional nonlinear problems with huge solution spaces in modeling,design,and optimization.These complex characteristics make the modeling and design optimization of array antennas more difficult,and traditional optimization design requires exponentially increasing computational resources to handle such problems.In recent years,with the rapid development of neural networks and machine learning,new approaches and methods have been provided to solve complex electromagnetic problems.This dissertation mainly focuses on the modeling techniques based on neural network algorithms,proposing efficient neural network models applicable to various types of array antenna.The main research work of this dissertation is summarized as follows:1.A multi-branch neural network model is proposed for uniform linear arrays,combined with the active element pattern（AEP）technique for sub-arrays to process modeling methods with structure parameters as the input and array port/far-field performance as the output.The large scale array problem is transformed into solving small sub-arrays,reducing the dimension of the solution space and input variables by obtaining the radiation pattern results of small sub-arrays through vector fitting and transfer function.Sub-arrays constructed with different categories of elements are trained by independent and parallel multi-branch networks.The validation results of multiple numerical examples demonstrate that the model can effectively solve medium and large periodic array modeling and optimization problems,and accurately predict the electromagnetic response of array antennas in actual scenarios considering mutual coupling effects and array environment.Thus it is applicable to medium and large scale uniform linear array antenna modeling and design.2.A multi-level knowledge-based neural network（KBNN）model is proposed for one-dimensional thinned arrays to meet the modeling and design needs of large-scale thinned arrays with high design freedom and effectively reduce the sample demand.The model includes a coarse network model of complex mapping between element structure/spacing and transfer function coefficients,which outputs the deep electromagnetic behavior of the first-level network,such as element structure and array antenna deployment,as prior knowledge to the second-level fine network,thereby simplifying the network complexity and improving the network training speed.The verification results of numerical examples show that the proposed multi-level KBNN model can construct an accurate mapping from the element structure and spacing to the array radiation pattern.By calling the proposed model through the optimization algorithm,the thinned array design that requires low side lobes and phased beam scanning is realized,avoiding the time-consuming simulation iteration process of the entire array,and significantly reducing the model training cost.3.An outlier detection-aided supervised learning（ODASL）model is proposed for the thinned conformal array to solve the sensitivity to mutual coupling variation and data dependency problems in small sample datasets.The model adopts the outlier detection（OD）method based on data mining and outlier score mechanism（OSM）to quickly identify and process outlier and invalid samples in small sample datasets,improving the accuracy of the model.The validation results of the numerical example show that the proposed ODASL model can efficiently predict the electromagnetic response of thinned conformal arrays,with high prediction accuracy and stable generalization ability.The proposed model can be further extended to the design of other electromagnetic devices.4.A subarray characteristic matrix-aided supervised learning（SCMASL）model is proposed for sparse planar arrays to solve the problem that traditional sparse planar array designs require a large number of iterative optimizations and they are difficult to achieve optimal convergence.Taking into account the influence of mutual coupling effects,the model adopts the local sub-domain division technology based on cluster analysis to divide the array elements at different positions into local sub-arrays of different sizes to reduce the complexity of the model.At the same time,the introduction of the binarized feature matrix（BFM）representing the array element layout,input scaling factor K,and the mutual coupling evaluation coefficient M_e^sub reflecting the mutual coupling difference between sub-arrays enables the model to efficiently learn the mapping relationship between the array element structure/array information and radiation performance.The model reduces the impact of the performance difference between the sub-arrays on the model accuracy,and realizes the accurate prediction of the radiation pattern of the sparse planar array.The effectiveness of the proposed algorithm is illustrated through the verification results of two sparse planar array numerical examples.