Neural network-based impedance profile reconstruction of nonuniform microstrip transmission lines from broadband frequency-domain reflectometry

Impedance discontinuities in nonuniform transmission lines (TLs) produce reflections that correspond to the spatial variation of the impedance. Reconstructing this variation from broadband reflection data is a challenging inverse problem of critical importance for fault diagnosis and integrity monitoring. In this work, a neural network (NN) based inverse modeling approach is proposed to estimate the impedance profile of nonuniform microstrip TLs from frequency-domain reflectometry responses. To this end, a large synthetic dataset is generated using a physics-based forward TL model, including rectangular and Gaussian profiles with single and double variations. Each sample consists of the broadband complex reflection response (10 MHz–7 GHz) as input and the spatial characteristic impedance profile at the center frequency (3.5 GHz) as the output. Three NN architectures (a multilayer perceptron (MLP), a one-dimensional convolutional NN, and a transformer encoder (TE)) are trained on the synthetic dataset and evaluated using (i) synthetic data, (ii) computer simulation technology (CST) based full-wave microstrip simulations, and (iii) measurements from fabricated FR-4 microstrip prototypes. The proposed method accurately reconstructs the impedance profile based on experimental measurements, with the best performance achieved by the TE ( (Formula presented) (Formula presented) up to 0.824 and root mean square error (RMSE) of (Formula presented) (Formula presented) ), while the MLP achieves comparable reconstruction accuracy for profiles with sharp impedance transitions ( (Formula presented) (Formula presented) up to 0.897 and RMSE of (Formula presented) (Formula presented) ). These results demonstrate that NN–based models enable direct reconstruction of impedance profiles from broadband reflection data.