Applications of artificial neural networks to RF and microwave measurements

Abstract

This article describes how artificial neural networks (ANNs) can be used to benefit a number of RF and microwave measurement areas including vector network analysis (VNA). We apply ANNs to model a variety of on-wafer and coaxial VNA calibrations, including open-short-load-thru (OSLT) and line-reflect-match (LRM), and assess the accuracy of the calibrations using these ANN-modeled standards. We find that the ANN models compare favorably to benchmark calibrations throughout the frequencies they were trained for. We summarize other current applications of ANNs, including the determination of permittivities of liquids from the reflection coefficient measurements of an open-ended coaxial probe and the determination of moisture content of wheat from free-space transmission coefficient measurements. We also discuss some potential applications of ANN models related to power measurements, material characterization, and the comparison of nonlinear vector network analyzers.

Water-to-Cement Ratio Prediction Using Anns From Non-Destructive and Contactless Microwave Measurements

“…The difference between the output and target is calculated using a certain error function in order to give the prediction error made by the network. Then, the training algorithm is used to adjust the network's weights and thresholds in order to minimize this error [28,29].…”

Levenberg-Marquardt and Conjugate Gradient Neuro-Modeling of Simulated Miniature Rectangular Microstrip Antenna

“…NNs are used in impedance matching [16,17], inverse modeling [18], measurements [19], and synthesis [20]. Multilayer perceptron (MLP), radial basis function (RBF), knowledge based neural network (KBNN), wavelet network, and recurrent neural network (RNN) are commonly used as ANN structures.…”

Knowledge-Based Support Vector Synthesis of the Microstrip Lines

“…Both of these nonlinear learning machines are once trained, they are capable of responding almost instantly to any input variable set, thus they are as fast as the approximate (coarse) models and can be as accurate as the detailed electromagnetic (fine) models [6][7][8][9]. Here, the key problem is to train these machines with the accurate data which may be measured or simulated data.…”

“…Recently, there is a new trend in the Electromagnetic (EM)-ANN area for searching the techniques using reduced number of accurate training data, thus resulting in lessened CPU and human time together with faster model development. The pioneering techniques for reducing the need for accurate traning data can be given as follows: Neural networks with knowledge such as the knowledge-based neural networks (KBNN) [7], difference method (DM) [8], prior-knowledge input (PKI) network [9] and space mapped neural networks (SMNN) [10][11][12][13]. Furhermore an efficient knowledgebased automatic model generation (KAMG) technique is proposed [14] combining automatic model generation, knowledge neural networks, and space mapping, where the two data generators -coarse and fine generators-are simultaneously employed for the first time.…”