Principal Feature Extraction Based Low-complexity Convolutional Neural Network for Concurrent Multiband Digital Predistortion of RF Power Amplifiers

With the advent of the fifth generation (5G) New Radio (NR), the Non-Terrestrial Network (NTN) stands out as a solution to enable wider coverage of broadcast satellites. NTN systems require higher data rates and bandwidth. RF power amplifiers (PAs) are critical components of NTN system transmitters, but power amplifiers fed with large bandwidth multi-band signals present more complex nonlinear characteristics, bringing severe in-band distortion and out-of-band spectral leakage and placing higher demands on the efficiency of power amplifiers. The behavior modeling of power amplifiers is an effective method for nonlinear analysis. Volterra-based models and neural network-based models have been demonstrated to express the nonlinearity of PAs. However, the computational cost of these models increases, and their implementation demands more processing resources as the signal bandwidth gets wider or the number of carrier aggregations increases. Previous work has demonstrated that the convolutional neural network (CNN) based behavior model has low model complexity when modeling single-carrier PAs. However, due to the coupling between multiple carriers, the CNN model cannot be directly applied to the case of multiple carriers. To address this issue, a concurrent multi-band wireless transmitter DPD structure based on principal component extraction with CNN is proposed in this paper. The method uses principal component extraction to simplify the fully connected layer of the CNN to reduce the complexity of the model. Compared with the conventional CNN DPD method, the experimental results show that the proposed method can achieve low computational complexity while ensuring modeling accuracy.