Nonlinear Model Optimization Of RF Power Amplifier

Digital predistortion is one of the most commonly used linearization techniques.With the enhancement of memory nonlinearity of RF Power Amplifiers(PA),the requirements for digital predistortion are getting higher and higher.With its strong adaptability and nonlinear fitting ability,neural networks have achieved better results in many fields than traditional algorithms.Based on neural network,this thesis models the memory nonlinearity of RF power amplifier and constructs predistorters.The main research work and innovations are as follows:Firstly,based on the BP neural network,the delay taps considering the memory effect are introduced in the input layer of the network,and the nonlinear series expansion is performed for each delay tap,a Modified Real-Valued Time-Delay Neural Network(MRVTDNN)model is proposed for RF power amplifiers with strong memory nonlinearity.The experimental results show that the modeling accuracy of the MRVTDNN model is higher than that of the BP neural network and the real-valued time-delay neural network model,and the normalized mean square error reaches-40.1306 d B.Then,based on the RBF neural network,the modules considering the memory nonlinearity are introduced in the input layer of the network,in which the envelope lag term and the envelope lead term considering the envelope lag effect and the envelope lead effect are introduced.A Generalized Modified Radial Basis Function Neural Network(GMRBFNN)model is proposed,which is more suitable for RF amplifiers with strong memory nonlinearity.When using GMRBFNN to model the power amplifier,the structure of the model is optimized by genetic algorithm,and the model becomes more streamlined.The experimental results show that the modeling accuracy of the GMRBFNN model is higher than that of the RBF neural network,the real-valued time-delay RBF neural network and the modified RBF neural network model,and the normalized mean square error reaches-40.8512 d B.Finally,based on the memoryless polynomial model,memory polynomial model,MRVTDNN model and GMRBFNN model,four predistorters are constructed.The experimental results show that the predistorter based on the GMRBFNN model has the best linearization effect,and the adjacent channel spectrum suppression can be up to about 16 d B,and the linearization effect of the predistorter based on the MRVTDNN model is second.In general,the linearization effect of the predistorters based on the neural network model is greatly improved compared with the linearization effect of the predistorters based on the traditional model.