Research Of RF Power Amplifiers’ Behavior Model With Neural Network

In the new generation of mobile communication system，there will be a strong chanllenge forRadio Frequency subsystem of mobile communication, especially for the bandwidth and linearityindex of base station RF power amplifiers. DPD (Digital PreDistriction) is a great way to improve thelinearity of PA while the key is to establish a fitable model of PA. This paper mainly makes a study ofIBPNN (Improved BP Neural Network) and used it to establish a PA model.First of all, this paper discusses the reason of the nonlinear characteristics of PA and memoryeffects. The causes of non-linear of power amplifiers are analysed and index of linearity of poweramplifier is given. At the same time, infroduce the definition of memory effects and give a detailanysis of causes of PA’s memory effects and how to reduce them.Secondly, the principle of neural network is elaborated. It mainly introduces the ANFIS neuralnetwork’s and BP neural network’s network structure, learning algorithm and learning process, theanvantages and disadvantages of them and the application fields of the are also been analysed.Analyse the models of PA, such as Saleh model, polynomial model without memory, Voterra model,Wiener&Hammerstein model and neural network model. The PA’s linearization technique, such asnegative feedback linearization technique, feed-forward linearization technology, LINC linearizationtechnology as well as the predistortion technology, is also discussed. At the same time, the detailedanalysis and comparison between those linearization technologies are made and the applications ofthose technologies are also summarized.Then, the PA’s models are built according to ANFIS and Improved BP neural network. Specificmodeling steps, specific parameter setting and some notices in building model with the two kinds ofneural network are given. Also, this paper gives the main code when modeling.At last, according to the models which have been trained with the date from test platform, theprecision and convergent rate of calculation of models are compared. The result show that theImproved BP Neural Network (IBPNN) resolves the disadvatages of the condional BP neural networkwhich has slow convergent rate and easily trapped in the local minimum value. At the same time, theIBPNN model has better precision than ANFIS model. In the end, this paper makes a conclusion thatthe IBPNN has better performance in convergence speed and prediction accuracy.