Research On Antenna Design Based On Support Vector Machine

Support vector machine(SVM) is a machine learning method based on statistical learning theory. SVM has been applied to many fields due to its advantages of excellent learning ability and generalization ability and small sample learning ability. However, the electromagnetic field involves such mathematical and physics problems like complicated model structure, nonlinear objective function and so on. SVM can not only reduce the amount of work and complicated mathematical calculation for the antenna designers, but it can also reduce the expensive measurement cost, so SVM is more suitable for the electromagnetic modelling problems. It is essential to select the kernel function because a single kernel cannot take into account both the interpolation ability and extrapolation ability. To overcome this drawback, a hybrid kernel function was proposed, and the particle swarm optimization(PSO) algorithm was used to optimize the parameters. Also, in order to further improve the prediction accuracy, the idea of SVM ensemble was proposed. The main researches were discussed as follows.(1) Considering respective advantage of each different single kernel, a hybrid kernel function combining the global kernel with local kernel function was proposed, in which the parameters of SVM and weight coefficient were optimized by PSO algorithm. The method was verified and analyzed by the data sets in the UCI database and test functions.(2) In order to improve the prediction accuracy and stability of SVM, SVM ensemble was proposed in which the parameters and weights were optimized by PSO algorithm. The result was better compared with ANN and single SVM when predicting the data sets in UCI database.(3) The designed PSO-SVM based on hybrid kernel function was used to fast model the direction of arrival(DOA) estimation, and the modeling error was analyzed and compared with that of ANN.(4) The designed PSO-SVM based on hybrid kernel function was used to model the resonant frequency of compact microstrip antenna(CMSA) including planar inverted-F antenna(PIFA), L-shaped MSA. HFSS was used to build simulation models to obtain the samples, and the modeling error of different methods was compared and analyzed.(5) The SVME was proposed to model resonant frequency of rectangle MSA, and the predicting accuracy was proved to be higher compared with the results by BP-NN and many other methods.(6) The SVME was proposed to model the circular polarized MSA, and the predicting accuracy was proved to be higher compared with the results by BP-NN and single SVM.