A novel sparse behavioral model design method based on the global representative point selection and the randomized SVD algorithm

Digital pre-distortion (DPD) technology is currently the dominant technology employed to compensate for the nonlinearity of power amplifiers (PAs). Recently, PA modeling methods based on machine learning have attracted much attention. However, the traditional model still suffers from the defects of long modeling time or insufficiently high modeling accuracy. To solve this problem, the global representative point selection (GRPS) algorithm and the randomized SVD (RSVD) algorithm based on the least squares twin support vector regression (LSTSVR) model are introduced. The GRPS algorithm is first used to select a specific number of globally representative points from all the data to construct the support vector set. Then the RSVD algorithm is used to perform a low-rank approximation to the target kernel matrix. The modeling performance of the proposed approach is compared with the existing model using different communication signals, and the results show that the proposed model can improve the modeling accuracy and reduce the modeling time. Further, a pre-distortion experimental platform is established, and the proposed model is used to pre-distort the Class F PA and the Doherty PA respectively, which proves that the proposed model has a good nonlinear correction effect.