Research On Digital Predistortion Of Power Amplifier Based On Neural Network

In recent years,with the continuous increase of bandwidth and the use of efficient topology of power amplifier,the linearity of power amplifier has deteriorated seriously,and the resulting nonlinear distortion has seriously affected the signal quality of the system.Because of its high flexibility and excellent linearization performance,digital predistortion technology has become a basic component of current and next generation wireless communication systems.The neural network predistortion model has achieved better linearization effect than polynomial predistortion model in many application scenarios,but the current neural network model has too many coefficients,and the coefficients are nonlinear relative to the model,and the training time is long.These reasons make it difficult to widely use the digital predistortion model of power amplifier based on neural network.This paper is devoted to reducing the complexity of digital predistortion algorithm of power amplifier based on neural network.The main work contents and innovations are as follows:1.A two-stage pruning algorithm for feedforward neural network model is proposed.The redundant components of feedforward neural network model are pruned by evaluating the importance of the input components,and the structure of the feedforward neural network model is simplified.The hyperparameters optimized by the algorithm focus on each input term itself,which leads to smoother performance changes and can find a more suitable feedforward neural network model.The pruning algorithm is divided into two stages.In the first stage,the importance of the input components is roughly sorted based on the hypothesis testing principle,and in the second stage,the importance of the input components is finely sorted based on the performance fluctuation principle,and finally a more accurate ranking of the importance of the input components can be obtained.By pruning unimportant components,a feedforward neural network model with only a few input components can be obtained.Verified by the built test platform,the proposed pruned FNN model reduces the model coefficients from 842 to 262,and the reduction ratio of model coefficients reaches 69%on the premise of ensuring the performance.2.A power adaptive neural network digital predistortion algorithm is proposed,which divides the neural network predistortion model coefficients into common behavior characteristic coefficients and specific power behavior characteristic coefficients.The power adaptive neural network digital predistortion model is used to extract the common behavior characteristic coefficient matrix under different input signal power conditions off-line,which does not change with the change of input power conditions,and the least square method is used to extract the specific power behavior characteristic coefficient matrix under the current input signal power condition.This algorithm greatly reduces the number of online updating neural network model coefficients,and changes online updating neural network model coefficients from nonlinear to linear.This algorithm is suitable for the application scenarios where the input signal power of power amplifier changes rapidly.It can reduce the adaptive complexity of digital predistortion neural network model and improve the adaptive speed of digital predistortton neural network model while retaining the excellent performance of digital predistortion neural network model.Verified on the built test platform,the proposed model only needs to update 62 coefficients online,which is 95%lower than the adaptive ARVTDNN model.Comparing the time of updating model coefficients based on MATLAB,the proposed model only takes 0.037 s,which is only 0.5‰ of the time required for adaptive ARVTDNN model.The two-stage pruning algorithm of feedforward neural network model and power adaptive neural network digital predistortion algorithm proposed in this paper reduce the computational complexity of digital predistortion based on neural network from different angles,which lays a foundation for the application of neural network model in broadband and power dynamic change scenarios.