基於Inception編碼器的DOA估算在陣列缺陷下的應用

以作者查詢圖書館館藏

、以作者查詢臺灣博碩士

、以作者查詢全國書目

、勘誤回報

、線上人數：14

、訪客IP：3.145.171.144

姓名

楊書虹(Shu-Hung Yang) 查詢紙本館藏

畢業系所

通訊工程學系在職專班

論文名稱

基於Inception編碼器的DOA估算在陣列缺陷下的應用

相關論文

★ 利用手持式手機工具優化行動網路系統於特殊型活動環境	★ 穿戴裝置動態軌跡曲線演算法設計
★ 石英諧振器之電極面設計對振盪頻率擾動之溫度相依性研究	★ 股票開盤價漲跌預測
★ 感知無線電異質網路下以不完美頻譜偵測進行資源配置之探討	★ 大數量且有限天線之多輸入多輸出系統效能分析
★ 具有元學習分類權重轉移網路生成遮罩於少樣本圖像分割技術	★ 具有注意力機制之隱式表示於影像重建三維人體模型
★ 使用對抗式圖形神經網路之物件偵測張榮	★ 基於弱監督式學習可變形模型之三維人臉重建
★ 以非監督式表徵分離學習之邊緣運算裝置低延遲樂曲中人聲轉換架構	★ 基於序列至序列模型之 FMCW雷達估計人體姿勢
★ 基於多層次注意力機制之單目相機語意場景補全技術	★ 應用於3GPP WCDMA-FDD上傳鏈路系統的遞迴最小平方波束合成犛耙式接收機
★ 調適性遠時程瑞雷衰退通道預測演算法設計與性能比較	★ 智慧型天線之複合式到達方位-時間延遲估測演算法及Geo-location應用

檔案

[Endnote RIS 格式]

[Bibtex 格式]

[相關文章]

[文章引用]

[完整記錄]

[館藏目錄]

至系統瀏覽論文 (2030-3-15以後開放)

摘要(中)

方向到達（DOA）估算在無線通信、雷達監測和聲學定位等應用中至關重要。然而，傳統方法（如 MUSIC 和 ESPRIT）在面對陣列缺陷（例如增益/相位誤差、感測器互耦和位置偏移）以及低訊雜比（SNR）環境時，效能顯著下降。本文提出了一種基於 Inception 編碼器（IE）的 DOA 估算框架，並進一步優化為 IE+，通過結合多尺度特徵提取和注意力機制，實現了在雜訊和陣列不完善條件下的卓越性能。

IE 架構通過多尺度分析提取局部與全域特徵，能夠有效捕捉信號中的空間和時間相關性，同時濾除雜訊。與傳統卷積編碼器（CE）相比，IE 在面對複雜失真時展現出更高的準確性與穩健性，且其模組化設計和深度可分離卷積顯著降低了計算需求。IE+ 則通過加入注意力機制進一步提升了模型的適應能力，能動態關注關鍵特徵，有效緩解陣列缺陷的影響。

實驗結果表明，IE 和 IE+ 在不同場景中均優於 CE 和傳統方法。在不完善條件下，IE+ 的估算偏差明顯低於其他方法；在 SNR 降至 -15 dB 的極端情境下，IE+ 仍保持穩定的準確性。此外，IE+ 在大規模 MIMO 系統中展現了出色的可擴展性，其優化設計支持低延遲推理，適用於實時應用。

總之，基於 Inception 的編碼器為 DOA 估算在陣列缺陷下的應用提供了一個高效、穩健且可擴展的解決方案。該方法在處理現實世界中的失真和雜訊環境方面樹立了新的基準，並為現代通信和感測系統提供了可靠的技術支撐。

摘要(英)

Direction of Arrival (DOA) estimation is crucial in applications such as wireless communications, radar monitoring, and acoustic localization. However, traditional methods (e.g., MUSIC and ESPRIT) suffer significant performance degradation in the presence of array imperfections, such as gain/phase errors, mutual coupling, and sensor position deviations, as well as in low signal-to-noise ratio (SNR) environments. This paper proposes a DOA estimation framework based on the Inception Encoder (IE), further optimized as IE+, which integrates multi-scale feature extraction and attention mechanisms to achieve superior performance under noisy and imperfect array conditions. The IE architecture leverages multi-scale analysis to extract both local and global features, effectively capturing the spatial and temporal correlations within signals while filtering out noise. Compared to conventional convolutional encoders (CE), IE demonstrates higher accuracy and robustness against complex distortions, while its modular design and depthwise separable convolutions significantly reduce computational demands. IE+ further enhances model adaptability by incorporating attention mechanisms, dynamically focusing on key features to mitigate the impact of array imperfections. Experimental results show that both IE and IE+ outperform CE and traditional methods across various scenarios. Under imperfect conditions, IE+ exhibits significantly lower estimation bias than other approaches, and even in extreme cases where SNR drops to-15 dB, it maintains stable accuracy. Furthermore, IE+ demonstrates excellent scalability in large-scale MIMO systems, with an optimized design that supports low-latency inference, making it suitable for real-time applications. In summary, the Inception-based encoder provides an efficient, robust, and scalable solution for DOA estimation in the presence of array imperfections. This approach establishes a new benchmark for handling real-world distortions and noisy environments, offering reliable technological support for modern communication and sensing systems.

關鍵字(中)

★ 編碼器
★ 陣列缺陷

關鍵字(英)

★ Inception
★ DOA

論文目次

摘要 ............................................................................................................................................. i
目錄 ............................................................................................................................................ ii
圖目錄 ....................................................................................................................................... vi
表目錄 ...................................................................................................................................... vii
第1章介紹 .............................................................................................................................. 1
1.1 動機 ............................................................................................................................. 1
1.2 章節架構 ..................................................................................................................... 3
第 2 章卷積神經網絡 Convolutional Neural Networks ....................................................... 4
2.1 CNN的基本概念 ......................................................................................................... 4
2.2 CNN 的架構 ................................................................................................................ 5
2.2.1 CNN 的基本操作 ............................................................................................. 5
2.2.2 Convolution神經網路 (CNN) 操作在信號與陣列數據中的重要性 ........... 7
2.3 Inception 模組 The Inception Module ....................................................................... 8
2.3.1 CNN 架構的演變 ............................................................................................. 8
2.3.2 Inception 模組的結構與功能 .......................................................................... 8
2.3.3 Inception 模組的優勢 ...................................................................................... 9
2.3.4 Inception 模組在方向估計（DOA）中的應用 ........................................... 10
2.3.5 為何選擇 Inception 模組作為本研究的對象 ............................................. 11
2.4 與 DOA 估計的相關性 ........................................................................................... 11
第 3 章自編碼器 AutoEncoder ........................................................................................... 14
3.1 自編碼器（AutoEncoder）的基礎原理 .................................................................. 14
3.1.1 自編碼器（AutoEncoders）的歷史與演變 ................................................. 14
3.1.2 自編碼器（AutoEncoders）的應用 ............................................................. 15
3.1.3 為何自編碼器（AutoEncoders）適合高效數據表示 ................................. 15
3.1.4 自編碼器（AutoEncoders）與信號處理的相關性 ..................................... 16
3.2 自編碼器（AutoEncoder）的架構 .......................................................................... 17
3.2.1 自編碼器（AutoEncoder）的核心組成部分 ............................................... 18
3.2.2 Convolution自編碼器（Convolutional AutoEncoder）的運作 ................... 18
3.2.3 Convolution自編碼器（Convolutional AutoEncoders）的優勢 .................. 19
3.2.4 與信號處理的相關性 .................................................................................... 20
3.2.5 總結 ................................................................................................................ 20
3.3 自編碼器（AutoEncoders）在 DOA 估計中的應用 ............................................ 20
第４章 DOA 估算 .............................................................................................................. 22
4.1 DOA 估計的目的與重要性 ...................................................................................... 22
4.1.1 DOA 估計的目標 ........................................................................................... 22
4.1.2 在各領域中的重要性 .................................................................................... 22
4.1.4 DOA 估計的未來相關性 ............................................................................... 23
4.2 現有的 DOA 估計方法 ........................................................................................... 24
4.2.1 Classical Model-Based Methods ..................................................................... 24
4.2.2 Neural Network-Based Approaches ................................................................ 25
4.2.3 在陣列不完善情況下的有效性與局限性 .................................................... 26
第 5 章 The Proposed DOA 估算Method ........................................................................... 28
5.1 過去關於Convolution編碼器的研究概述 ............................................................. 28
5.1.1 Convolution編碼器在 DOA 估計中的優勢 ................................................ 28
5.1.2 Convolution編碼器的局限性 ........................................................................ 29
5.1.3 過去研究的貢獻 ............................................................................................ 29
5.1.4 Setting the Stage for the Proposed Method ..................................................... 30
5.2 Inception-Based Encoder (IE) .................................................................................... 30
5.2.1 IE 架構 ........................................................................................................... 30
5.2.2 IE 的運作細節 ............................................................................................... 31
5.2.3 IE 相較於 CE 的優勢 .................................................................................. 32
第 6 章 Experiment Setup ..................................................................................................... 35
6.1 實驗數據生成 ........................................................................................................... 35
6.1.1 Training Data ................................................................................................... 35
6.1.2 測試數據 ........................................................................................................ 36
6.2 Experimental Environment ......................................................................................... 37
6.2.1 Scenarios .......................................................................................................... 37
6.2.2 比較CE, IE和 IE+ 雜訊/偏差 .................................................................. 38
6.2.3 定量指標 ........................................................................................................ 45
6.2.4 Qualitative Insights .......................................................................................... 46
6.3 所提出方法與現有方法的比較總結 ....................................................................... 47
6.3.1 相對於原本方法的優勢 ................................................................................ 47
6.3.2 相對於 CE 的優勢 ....................................................................................... 48
6.3.3 Key Findings.................................................................................................... 49
第 7章結論與未來展望 ....................................................................................................... 50
7.1 實驗結果總結 ........................................................................................................... 50
7.1.1 Superior Accuracy in DOA 估算 .................................................................... 50
7.1.2 Robustness to Noise......................................................................................... 50
7.1.3 處理陣列不完美 ............................................................................................ 51
7.1.4 擴展性與效率 ................................................................................................ 51
7.1.5 跨多樣場景的泛化能力 ................................................................................ 51
7.1.6 對該領域的貢獻 ............................................................................................ 52
7.2 未來的研究方向 ....................................................................................................... 52
7.2.1 擴展至更大規模的陣列 ................................................................................ 52
7.2.2 Working with Real-World Data ....................................................................... 53
7.2.3 Further Refinements of the Model................................................................... 53
7.2.4 探索新應用 .................................................................................................... 53
7.3 結論 ........................................................................................................................... 54
參考文獻 ............................................................................................................................ 55

參考文獻

[1] Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). ImageNet classification with deep convolutional neural networks. Advances in Neural Information Processing Systems, vol. 25, pp. 1097-1105. [2] Alzubaidi, L., Zhang, J., Humaidi, A.J. et al. Review of deep learning: concepts, CNN architectures, challenges, applications, future directions. J Big Data, vol. 8, no.53, pp. 2021. [3] Zheng, H., Shi, Z., Zhou, C., Vorobyov, S. A., & Gu, Y. (2024). Deep Tensor 2-D DOA estimation for URA. IEEE Transactions on Signal Processing, vol. 72, pp. 4065–4080. [4] Agarap, A. F. (2018). Deep Learning using Rectified Linear Units (ReLU). International Journal of Machine Learning, 45(3), 210-230. [5] Klambauer, G., Unterthiner, T., Mayr, A., & Hochreiter, S. (2017). Self-normalizing neural networks. In Advances in Neural Information Processing Systems, pp. 972-981. [6] Chang, D.-C., & Liu, Y.-T. (2023). DOA estimation based on convolutional autoencoder in the presence of array imperfections. Electronics, vol. 12, pp. 771. [7] Liu, Z.-M., Zhang, C., & Yu, P. S. (2018). Direction-of-Arrival estimation based on deep neural networks with robustness to array imperfections. IEEE Transactions on Antennas and Propagation, vol.66, no.12, pp. 7315–7327. [8] Yu, J, Howard, W. W, Xu, Y, & Buehrer, R. M. (2023). Model order estimation in the presence of multipath interference using residual convolutional neural networks. IEEE Transactions on Wireless Communications, vol. 23, no. 7, pp.7349–7361.

指導教授

陳永芳(Yung-Fang Chen)

審核日期

2025-3-13

推文