5G新無線電接收機之基於神經網路的通道估測和秩指標及預編碼矩陣指標選擇之設計

Design of Neural Network Based Channel Estimation and RI/PMI Selection for 5G New Radio Receivers

王鈺凱(Yu-Kai Wang)

指導教授 ： 闕志達

國立臺灣大學/電機資訊學院/電子工程學研究所/碩士(2019年)

https://doi.org/10.6342/NTU202000337

摘要

近年來因為電腦科技進步，神經網路的各種應用也再次蓬勃發展，很多研究者已經將神經網路的技術擴展到各自的領域，當然無線通訊領域也不例外，雖然已經有應用在無線通訊系統的上層(Upper layers)，但實體層(Physical layer)的應用會因為複雜通道環境的阻礙而有所限制，實作起來相對困難，儘管如此還是相信神經網路能夠提出有用且有見解的解決方案，並且有望能在難以用數學模型描述的通訊場景中有所突破。而本論文針對兩個實體層上和通道相關的模組，分別為通道估測與秩指標及預編碼矩陣指標選擇，嘗試使用神經網路的解決方案來處理。
本論文的第一個主題是基於神經網路的通道估測，將通道頻率響應視為一張二維影像，利用影像上超解析度(Super-resolution)的技術，藉由一個統一的神經網路(Unified neural network)，優化傳統使用內插得到的通道，並可以獲得更平滑的通道頻率響應，且降低通道估測誤差以及提升位元錯誤率(bit error rate)品質。最後發現在長延遲擴展造成嚴重的頻率選擇性衰減的通道能有顯著的效能增加，但是所需要的複雜度比起傳統通道估測卻是較高的，而這在卷積神經網路的解決方案當中是一個很難避免的問題。
本論文的第二個主題是基於自組織特徵映射圖的秩指標及預編碼指標選擇，因為需要對多個預編碼矩陣計算複數矩陣乘法與矩陣反矩陣或是矩陣行列式，所以傳統搜尋的複雜度非常大，而隨著天線數的增加或是天線的擺放方式不同，預編碼矩陣的碼簿大小也會急遽增加，所以本論文提出一低複雜度的解決方案，使用和以往完全不同的通道共變異數分群做法，對不同的多輸入多輸出相關性通道分群，並建立秩指標與預編碼矩陣指標的查找表，來完成秩指標與預編碼矩陣指標選擇，且在可容忍範圍內的效能降低，來達到降低運算複雜度的目的。

關鍵字

卷積神經網路 ； 超解析度 ； 通道估測 ； 自組織特徵映射圖 ； 秩指標及預編碼矩陣指標選擇

並列摘要

Recently, due to the advancement of computer technology, various applications of neural networks have flourished again. Researchers from different fields have extended neural networks to their respective fields, and of course, wireless communication is no exception. Although neural networks have been applied to the upper layer of the wireless communication system, applying them to the physical layer is quite challenging due to the sophistication of channel environments, which renders implementation tougher. Nevertheless, we still believe that neural networks can provide useful and insightful solutions, and they are expected to make a breakthrough in the communication scenarios that can hardly be expressed by mathematical models. In this thesis, we focused on the two channel-related modules of the physical layer processing, which are channel estimation and RI/PMI selection, respectively, and tried to process them with neural network solutions.
The first topic in this thesis is neural network-based channel estimation. We regard the channel frequency response(CFR) as a 2D image and utilize the super-resolution technique, which was originally used on images, and optimize the CFR obtained by traditional interpolation methods with a unified neural network to obtain a smoother CFR. This technique helps reduce channel estimation error and improves bit error rate quality. In the end, it was found that decoding performance for channels with highly frequency selective fading caused by long delay spread has improved significantly, but the required complexity is higher than traditional channel estimation approaches. However, this is usually an inevitable issue for CNN type solutions.
The second topic in this thesis is SOFM-based RI/PMI selection. The traditional RI/PMI selection approach utilizes all precoding matrices to calculate the complex matrix multiplications and matrix inversion or matrix determinant, which will lead to huge complexity. As the number of antennas increases or the arrangement of antennas varies, the size of the precoding matrix codebook will increase drastically. Therefore, we proposed a low complexity solution, which is the channel covariance matrix clustering. This approach is completely different from past approaches in that it groups different MIMO correlation channels and builds RI/PMI look-up tables for RI/PMI selection. In summary, this approach is a low-complexity solution with very similar performance for RI/PMI selection.

並列關鍵字

convolution neural network (CNN) ； super-resolution ； channel estimation ； self-organizing feature map (SOFM) ； RI/PMI selection

參考文獻

[1] Jeff Dean. 2017. AIFrontiers: Trends and Developments in DeepLearning Research [Online]. Available: https://www.slideshare.net/AIFrontiers/jeff-dean-trends-and-developments-in-deep-learning-research

[2] T. Wang, C. Wen, H. Wang, F. Gao, T. Jiang, and S. Jin, “Deep learning for wireless physical layer: Opportunities and challenges,” China Communications, vol. 14, no. 11, pp. 92-111, Nov. 2017.

[3] T. O’Shea and J. Hoydis, “An introduction to deep learning for the physical layer,” IEEE Transactions on Cognitive Communications and Networking, vol. 3, no. 4, pp. 563-575, Dec. 2017.

[4] E. Nachmani, Y. Be'ery, and D. Burshtein, “Learning to decode linear codes using deep learning,” in Proc. of 54th Annual Allerton Conference on Communication, Control, and Computing (Allerton), Monticello, IL, USA, Sept. 2016, pp. 341-346.

[5] E. Nachmani, E. Marciano, D. Burshtein, and Y. Be’ery, “RNN decoding of linear block codes,” arXiv preprint arXiv:1702.07560, 2017.