Research On Neural Network Inverse Modeling Methods Of Radio Frequency Modules

With the rapid development of wireless communication,the design speed of RF microwave circuit must be accelerated.The design cycle of circuits using traditional computer aided design(CAD)software is long.At the same time,with the emergence of new requirements for RF microwave modules,the design of new modules needs to be re-simulated and optimized,and the process is cumbersome.Therefore,it is urgent to find a more rapid and efficient method of module modeling and design.It is a research hotspot in recent years to use neural networks to model and optimize RF microwave modulesIn this paper,the main researches of the neural network inverse modeling methods for radio frequency modules are summarized as the following three points:Firstly,a bayesian regularized neural network inverse modeling method with L1/2 norm is proposed.L1/2 regularizer can make the network sparse,reduce the scale and accelerate the training speed.Bayesian regularization method can smooth the output curve,improve the stability and generalization of the network.This method is applied to inverse modeling of Doherty power amplifier,which obtains the output power,the f related to 511,and the f related to 521 comparing with direct inverse modeling method of mean square error is reduced by 8.83%,9.30%and 9%respectively,the running time is reduced by 99.34%,99.40%and 99.23%.It can solve multi-solution problems in designThen,a back-propagation(BP)neural network inverse modeling method based on improved ant colony optimization(IACO)and bayesian regularization(BR)is proposed.This method uses IACO algorithm to search the optimal weights of BRBP forward model quickly and effectively and keep them unchanged.Then the BR algorithm with L1/2 norm is used to iterate the input of the forward model inversely and make the network more stable.It is applied in reconfigurable power amplifier.Compared with the direct inverse model and the adaptive ? inverse model,the average modeling accuracy of the model is improved by 99.77%and 90.70%respectively,and the average running time is reduced by 35.76%and 2.05%respectively,which verifies its high efficiency.Finally,an improved Koch snowflake fractal ultra wideband(UWB)antenna is designed by using IACO-BRBP neural network inverse model and HFSS-MATLAB-API script file According to the return loss of the antenna,the dimension of the 0-order edge length is determined,which avoids the repetitive design and simulation of the antenna.At the same time,the bandwidth of the antenna is effectively expanded by using the fourth-order improved octagonal snowflake fractal structure and defective ground structure,and the antenna have good impedance matching characteristics.The antenna has good omnidirectionality in 3.2?12GHz band and its peak gain exceeds 2dBiThis paper has 38 figures,5 tables and 78 references.