Research On Neural Network Modeling Based On Image Domain In Electromagnetic Fields

In recent years,the neural network has been widely applied to the modeling of microwave components.The mapping relation between the component geometry and the electromagnetic(EM)response is learned by the neural network model.The trained model can predict the response quickly and accurately,and substitute the full-wave simulation to speed up the design especially for complex and electrically large components.However,there are several limitations that need to be addressed.First,the traditional model usually is a parametric one,where the shape change and topology change of components are not involved.It only focuses on the relationship between geometric parameters and EM responses.Second,the neural network model used in the computer vision is an end-to-end model,where the EM knowledge is not involved.The neural network learns the relationship and suitable features from the given sample by itself.Third,to meet the increasingly strict accuracy requirements of indicators,the samples are generated by full-wave EM simulation,which is time-consuming.Thus,it is significant to make full advantage of the limited samples.The main contribution of this dissertation is introduced as follows:For the modeling of the component shape,an effective model based on the convolutional neural network(CNN)is proposed.The input of the model is the crosssectional image of components instead of the geometric parameters.To define the training samples,a one-to-one relation between the component contour and the knot positions(and the slope of the tangent line to the interpolation curve at endpoint)is built with a shape-changing technique based on cubic spline interpolation.CNN is employed to map cross-sectional images into EM responses in the model.The proposed model is confirmed with several filters and antennas.The shape modeling method shows the ability to expand the search area of the solution and it can reach some specified performance indices which are difficult or even impossible for parametric modeling.For the lack of EM knowledge,two model based on knowledge based neural network(KBNN)are proposed.Based on the prior knowledge input method,one model takes EM responses of the original components as the prior knowledge.Based on the space mapping method,another model takes the parametric model as the prior knowledge.Due to use of prior knowledge,the relationship that remains to be learned is simplified.The effectiveness of two proposed models is confirmed with an example of a microstrip/coplanar waveguide filter.Compared with the shape modeling based on CNN and the parametric-like model based on the artificial neural network(whose inputs are the knots of component contour,and outputs are the EM responses),the two KBNN models are trained by fewer samples with no significant deterioration in model accuracy.For the efficient utilization of microwave component samples,a least square support vector machine(SVM)model based on the histogram of oriented gradient is proposed.In the model,the histogram of oriented gradient feature is extracted from the component image to show the appearance and shape of the component.The relationship between the histogram of oriented gradient features and the EM responses is preliminarily built on SVM and the transfer function.Then a radial basis function network is used for error correction.The proposed model is confirmed with a tri-band slot antenna and a bandpass filter with quarter wavelength short-circuited stubs.Compared with shape modeling based on CNN,the feature extracted by CNN is substituted by the histogram of oriented gradients feature,and CNN with many hyper-parameters is substituted by SVM,which can be implemented easily with several hyper-parameters.The SVM model shows the same accuracy and the improvement of training efficiency.