Research On Multi-state Generalization Model Based On Deep Learning For Power Amplifiers

As a power gain module,power amplifiers(PAs)are indispensable parts of the wireless communication system.Whether power amplifiers can work with high efficiency and linearity is related to the operation quality of the communication system.However,its nonlinear characteristics lead to the conflict between linearity and efficiency.To better analyze and evaluate the nonlinear characteristics of PAs,the research of PAs behavior modeling has been widely paid attention to,and many achievements have been made.Neural networks are widely used in the modeling of PAs because of their strong ability to fit arbitrary nonlinear functions.However,with the rapid increase of communication system bandwidth,the nonlinear characteristics of PAs become more serious,and the modeling becomes more difficult.For this reason,the deep neural network(DNN)containing more hidden layers is gradually used in PA modeling.Its performance will be improved with the increase of hidden layers.However,the traditional neural network models are usually modeled for PAs under fixed working states;PAs’ signal bandwidth and input power remain unchanged.When the working states change,the performance of the PA model will decline sharply,so it is necessary to retrain and update the model’s coefficients.This thesis proposes a multi-state model design method based on deep learning to meet the actual needs of modeling PAs in various working states.When there are PA signals in different working states,the signals in different states can be modeled with only one training.This effectively alleviates the problem that traditional models must be retrained for performance degradation caused by state change.In this paper,one PA model is established based on the deep neural network,and code vectors representing corresponding states of PAs under different working states are established.Then,the corresponding dimension-reduced embedding vector is obtained based on the state coding vector,denoted as the state representation vector,and combined with the state representation vector to realize the modeling of PAs with different states.Compared with the traditional DNN model,the proposed method can guarantee the modeling performance and establish one PA model aiming at various states,effectively alleviate the increase of modeling time caused by frequent training,and improve the model’s generalization ability to a certain extent.