基于FC-ED网络的无相位近场等效源重构方法OA

To address the issues of the highly ill-posed nature of inverse mapping equations,the tendency of traditional iterative algorithms to fall into local optima,and low computational efficiency in phaseless near-field antenna measurements,a single-plane equivalent Source Reconstruction Method(SRM)based on a Fully-Connected Encoder-Decoder(FC-ED)deep learning network architecture is proposed.In the data processing stage,a combined strategy of random Fourier contours and standard geometries is adopted to generate equivalent source training data with explicit electromagnetic diffraction physical guidance.In terms of network construction,a symmetric hourglass fully-connected topology is designed to achieve the capture of spatial electromagnetic features through dimensionality reduction and their reconstruction through dimensionality expansion.The convergence performance of the model,the accuracy of two-dimensional amplitude reconstruction on the source plane,and the one-dimensional far-field extrapolation are simulated and analyzed.The results indicate that the Mean Squared Error(MSE)of the two-dimensional equivalent source amplitude prediction of the model remains stable at the magnitude of 10-4,and the extrapolated cut-plane far-field radiation patterns are in high agreement with the theoretical true values with full amplitude and phase.Under the condition of requiring no iteration and relying solely on single-plane amplitude sampling,the proposed method effectively overcomes the challenge of the nonlinear inverse problem caused by near-field phase absence,achieving high-precision and highly robust field source reconstruction and far-field extrapolation.