神經網路於無線通訊系統非線性失真補償之應用

Compensation of nonlinear distortion in wireless communication system using neural networks

曾昭俊(Chao-Chun Tseng)

指導教授 ： 闕志達

國立臺灣大學/電機資訊學院/電機工程學研究所/碩士(2007年)

https://doi.org/10.6342/NTU.2007.00254

摘要

實際的無線通道和有線通道不同之處在於，無線通道充滿各樣式的干擾和不確定性，這使得無線通訊的發展充滿著挑戰。影響傳輸品質的原因可能來自於傳送端的功率放大器，也可能來自於通道上的符號間干擾，再加上無線頻譜資源的有限，使得這些失真效應日趨嚴重。所以我們必須尋求改善這些失真的方法。
本論文將探討造成非線性失真的一個主要原因，功率放大器的非線性。並結合神經網路理論的優點，發展出神經網路類型的預失真架構，利用誤差倒傳遞演算法，進行神經網路的訓練過程。本論文提出的兩種預失真技巧中，第一種技巧的神經網路僅有一個輸入節點，其為振幅輸入；第二種技巧的神經網路有兩個輸入節點，其分別為複數訊號的實部和虛部。經過浮點數模擬驗證兩個輸入節點的神經網路表現優於僅有一個輸入節點的神經網路。同時在硬體實現上，利用2次冪的群組符號(Group signed power-of-two)來表示的突觸權值(synaptic weight)，簡稱為GSPT權值，更可以簡化預失真電路的設計。
除了功率放大器非線性外，通道傳輸時的符號間干擾也一併討論，我們可以在接收端使用非線性等化器來補償這些失真現象。本論文提出了兩種類型的神經網路等化器架構，分別為多層感知器等化器和Chebyshev 等化器。浮點數模擬驗證了兩種神經網路類型的等化器在失真補償上皆有不錯的表現。在硬體設計上，我們選擇了多重感知器等化器，配合GSTP誤差倒傳遞演算法，以簡化電路架構。系統模擬結果顯示，本論文提出的方法可以有效的對抗非線性功率放大器和符號間干擾等失真效應。

關鍵字

神經網路 ； 預失真 ； 非線性等化器

並列摘要

The channel ideality make it a big challenge in developing wireless communication system. The distortion of received signals may result from nonlinear power amplifier or iter-symbol interference. Besides, the distortion is getting worse due to the limted bandwidth for transmission.

The dissertation is mainly aimed at the compensation from the power amplifier nonlinearity which is one of the nonlinear effects of the channel. The power amplifier pre-distortion is used to compensate for distortion based on the error back-propagation algorithm in neural networks. Two techniques of pre-distortion are proposed. The neural network of first technique includes only one node in input layer; while the neural network of the second technique includes two nodes in input layer. The floating point simulation shows that the performance of the first technique is better than that of the second technique. In hardware implementation, the group signed power-of-two (GSPT) number system is used to reduce hardware complexity in error back-propagation algorithm.

Next, the iter-symbol interference is also considered in the channel. It is necessary to seek appropriate equalizer. In the dissertation, there are two proposed equalizers available. They are multilayer perceptron (MLP) equalizer and Chebyshev equalizer. The floating point simulations of these two equalizer show good performance in compensation from nonlinear distortion. In hardware implementation, GSPT-based error back-propagation algorithm is also applied to MLP equalizer. The simulation results show the proposed methods can combat with distortions, such as power amplifier nonlinearity and inter-symbol interference.

並列關鍵字

Neural networks ； Pre-distortion ； Nonlinear equalizer

參考文獻

[1] McCulloch, W.S., and W. Pitts, 1943. “A logical calculus of the ideas immanent in nervous activity,” Bulletin of Mathematical Biophysics, vol. 5, pp.115-133.

[3] Rosenblatt, F., 1958. “The Perceptron: A probabilistic model for information storage and organization in the brain,”Psychological Review”, vol. 65, pp386-408.

[4] Hopfield, J.J., 1982. “Neuron networks and physical systems with emergent collective computational abilities,”Proceedings of the National Academy of Sciences, USA, vol.79, pp. 2554-2558.

[5] Rumelhart, D.E. and J. L. McClelland, eds 1986. Parallel Distributed Processing: Explorations in the Microstructure of Cognition, vol.1, Cambridge, MA:MIT Press.

[6] S Haykin, Neural Networks:A Comprehensive Foundation, Prentice Hall,1999.