應用於OFDM 發射機之增強實值時滯神經網路預失真補償設計

Abstract

傳統射頻功率放大器數位預失真方式可分為多項式法、查表(LUT)法…等。近年來因為5G與物聯網的發展許多產品都必須以低功耗低面積去做設計，傳統數位預失真無法處理嚴重的非線性失真因此造成效能上的嚴重損失，而近年來類神經網路(Deep Neural Network, DNN)和機器學習(Machine Learning, ML)的發展，將其與數位預失真上做結合得到很好的效果。本論文主要研究神經網路演算法應用於數位預失真上，首先我們先分析不同神經網路演算法在數位預失真上的ACPR,EVM,再來我們選擇最適當的神經網路演算法並根據所制定的發射機規格模擬出ACPR,NMSE,EVM結果，在模擬過程中為了以後寫硬體做準備引入量化流程將浮點數運算轉換成定點數運算，在不大量損失精確度的同時，大幅減少儲存空間與運算複雜度，並利用訓練序列最佳化shuffle演算法加快訓練序列收斂速度，進而減少訓練時間，我們發現加了shuffle演算法將序列打亂後得到優化序列後所需要的training時間比原先tranining少了40%。 我們的ARVTDNN演算法應用於數位預失真上在5G規格下並使用Saleh PA與Doherty PA Model當作模型，我們可以發現與其他幾篇DPD相關做數據比較可以發現我們的ACPR,NMSE,EVM數值皆有較佳的表現。

The digital predistortion (DPD) methods for traditional RF power amplifiers can be categorized into polynomial methods, lookup table (LUT) methods, and others. In recent years, due to the development of 5G and the Internet of Things (IoT), many products must be designed with low power consumption and minimal area. Traditional digital predistortion techniques struggle to handle severe nonlinear distortions, leading to significant performance degradation. However, with the recent advancements in deep neural networks (DNN) and machine learning (ML), integrating these techniques into digital predistortion has shown promising results. This paper primarily investigates the application of neural network algorithms in digital predistortion. First, we analyze different neural network algorithms in terms of Adjacent Channel Power Ratio (ACPR) and Error Vector Magnitude (EVM). Next, we select the most suitable neural network algorithm and simulate the ACPR, Normalized Mean Square Error (NMSE), and EVM results based on the defined transmitter specifications. During the simulation process, in preparation for future hardware implementation, we introduce a quantization process that converts floating-point operations into fixed-point operations. This significantly reduces storage space and computational complexity while maintaining accuracy. Additionally, we optimize the training sequence using a shuffle algorithm to accelerate sequence convergence, thereby reducing training time. We found that incorporating the shuffle algorithm to randomize the sequence resulted in a 40% reduction in training time compared to the original training process. Our ARVTDNN algorithm is applied to digital predistortion under 5G specifications, using the Saleh PA and Doherty PA models. Compared to other DPD-related studies, our results demonstrate superior ACPR, NMSE, and EVM performance.