A Memory Term Reduction Approach for Digital Pre-Distortion Using the Attention Mechanism

Over recent years, the memory effect compensated digital pre-distortion (DPD) has been widely used for linearizing wideband radio frequency power amplifiers (RF-PAs). However, with the increase of memory depth, the DPD model becomes more accurate, but the computational complexity of the DPD is significantly increased. In this paper, an approach using the attention mechanism is proposed, which can be used for locating and eliminating the memory terms with a small contribution to the performance of the DPD model so that the complexity of the DPD model is significantly reduced. The attention mechanism is employed to obtain the weighted correlation coefficient matrix of the memory effect. And the memory terms in the DPD model will be retained only if the contributions are high, which are evaluated by ensemble averaging over each diagonal of the weighted correlation coefficient matrix. To verify the applicability of the approach, a three-carrier wideband code-division multiple access signals with a bandwidth of 15 MHz and a single carrier long-term evaluation signal with a bandwidth of 20 MHz are employed for testing two Doherty RF-PAs with an operating frequency of 460 and 1900 MHz. Moreover, the generalized memory polynomial model is used to verify the effectiveness of the proposed approach. The simulation and experimental results show that the modeling accuracy and the DPD linearization performance of the DPD model with and without the memory term reduction are all almost the same, which indicates the effectiveness of the proposed approach.