本篇論文提出以數據驅動的方式用於適應均勻線性陣列天線(Uniform Linear Array, ULA) 中存在的各種陣列缺陷(Array imperfection) 並估計訊號到達的方向(Direction of Arrival, DOA) 獲得良好的陣列缺陷適應性與增強通用性。在深度學習演算法提出摺積自動編碼器的架構比一般的深度自動編碼器強調了學習局部的結構特徵，自動編碼器的概念類似空間濾波器，將入射訊號的空間範圍分解成多個較小的子空間範圍，在此架構下比起原始的入射訊號每個子空間涵蓋的入射範圍更窄使得入射訊號的的分佈更加集中，有助於減輕後續DOA 估計的負擔。在機器學習演算法中支撐向量機(Support Vector Machine, SVM) 利用核函數將數據映射至高維度的空間進行分類，在對DOA 估計進行分類時，我們使 用有向無環圖(Directed Acyclic Graph, DAG) 改善一般Multiclass SVM 處生的不可辨識區域的問題。在結果分析與比較中，可以看出以數據驅動的方法在各種陣列缺陷下均有令人滿意的性能。;This thesis studies a datadriven approach adapted to various imperfections in uniform linear array (ULA), which can estimate the direction of arrival (DOA) to obtain a good adaptability and enhance versatility for imperfect arrays. The architecture of the convolutional autoencoder in deep learning substantially focuses on the learning of local features of the structure than that of the general deep autoencoder. The autoencoder acts like a group of spatial filters, decomposing the input into multiple small spatial sub-regions. The range of the input covered by each spatial sub-region is narrower than that of the original input, and hence, the distribution of the input is better centralized. In machine learning algorithms, support vector machines (SVM) use kernel functions to map data in a high-dimensional space for classification. For the application of DOA estimation, we apply Directed Acyclic Graph (DAG) to solve the problem of unidentifiable regions with the general multi-class SVM. From simulation analysis and comparison, it can be seen that the data-driven method has satisfactory performance for different cases of imperfect arrays.
