隨著整合式感測與通訊（Integrated Sensing and Communication， ISAC）、智慧物聯網（Internet of Things， IoT）以及分散式感測網路的快速發展，多個感測裝置共同觀測同一目標環境已成為重要的感測架構。然而，在實際多視角感測（Multi-View Sensing）環境中，各感測裝置通常僅能觀測全域稀疏訊號中的部分成分，且量測資料可能受到遮蔽效應、通訊限制以及異常值（Outliers）干擾。此外，在許多光學成像、雷達感測與無線通訊系統中，接收端通常只能取得振幅或強度量測，而無法獲得完整相位資訊，使訊號恢復問題進一步形成具有高度非凸特性的稀疏相位恢復（Sparse Phase Retrieval）問題。 為解決上述問題，本文研究多視角分散式稀疏相位恢復架構，並結合仿射量測（Affine Sampling）模型與異常值污染環境，建立一套多視角訊號重建框架。首先，利用多個感測裝置所提供之局部量測資訊，建立全域稀疏訊號與局部遮罩矩陣之關聯模型。接著，提出一套三階段重建流程：第一階段利用計數統計法估計全域訊號支援集合；第二階基於此估計結果並透過仿射量測模型中的偏項向量(Bias Vector)完成全域訊號之非零值重建；第三階段則根據前兩階段所獲得之全域訊號資訊，恢復各感測裝置對應之局部遮罩結構，進而完成多視角訊號重建。 為提升異常值環境下之重建效能，本文進一步將局部遮罩恢復問題轉換為基於最小絕對誤差準則之最佳化問題，並提出次梯度重建演算法以提高對異常量測之魯棒性。此外，為改善傳統迭代式演算法收斂速度較慢以及參數需人工調整之缺點，本文進一步利用深度展開（Deep Unfolding）技術，將次梯度更新流程映射為可訓練之神經網路架構，透過資料驅動方式學習最佳更新參數，以提升有限迭代次數下之訊號恢復效能。 最後，本文分別於集中式與非集中式兩種系統架構下進行模擬驗證，並比較不同遮罩機率、裝置數量以及異常值機率對系統效能之影響。模擬結果顯示，所提出之方法能有效恢復全域稀疏訊號及局部遮罩結構，且深度展開架構相較於傳統次梯度方法具有更快的收斂速度與更佳的重建效能，驗證了所提方法在多視角分散式感測環境中的可行性與優越性。 關鍵字：多視角感測、稀疏相位恢復、仿射取樣、異常值、次梯度法、深度展開、分散式感測網路;With the rapid development of Integrated Sensing and Communication (ISAC)， the Internet of Things (IoT)， and distributed sensing networks， collaborative sensing using multiple sensing devices has become an important paradigm in modern sensing systems. However， in practical multi-view sensing environments， each sensing device can only observe a subset of the global sparse signal due to occlusion effects， communication constraints， and sensing limitations. Furthermore， measurement data may be contaminated by outliers. In addition， many optical imaging， radar sensing， and wireless communication systems can only acquire amplitude or intensity measurements without access to phase information， resulting in a sparse phase retrieval problem that is highly nonconvex and challenging to solve. To address these challenges， this thesis investigates a multi-view distributed sparse phase retrieval framework under affine measurements and outlier-corrupted environments. A unified multi-view signal reconstruction framework is developed by exploiting the relationship between the global sparse signal and the local masking structures associated with different sensing devices. First， local measurements collected from multiple sensing devices are utilized to establish a model that characterizes the interaction between the global sparse signal and local masking matrices. Subsequently， a three-stage reconstruction scheme is proposed. In the first stage， a counting-rule-based approach is employed to estimate the support set of the global sparse signal. In the second stage， the nonzero entries of the global sparse signal are recovered through a one-dimensional phase retrieval relationship derived from the affine measurement model. In the third stage， the local masking structure of each sensing device is reconstructed based on the global signal information obtained from the previous stages， thereby completing the multi-view signal reconstruction process. To improve reconstruction performance in the presence of outliers， the local mask recovery problem is further formulated as an optimization problem based on the least absolute deviation criterion. A subgradient-based reconstruction algorithm is then developed to enhance robustness against abnormal measurements. Moreover， to overcome the slow convergence and manually tuned parameters associated with conventional iterative algorithms， a Deep Unfolding framework is proposed. By mapping the subgradient update procedure into a trainable neural network architecture， the proposed method learns optimal update parameters in a data-driven manner， leading to improved reconstruction performance under a limited number of iterations. Finally， extensive simulations are conducted under both centralized and non-centralized architectures. The effects of different masking ratios， numbers of sensing devices， and outlier levels are investigated and compared. Simulation results demonstrate that the proposed framework can accurately recover both the global sparse signal and local masking structures. Furthermore， the proposed Deep Unfolding approach achieves faster convergence and superior reconstruction performance compared with the conventional subgradient method， validating its effectiveness and applicability in multi-view distributed sensing environments. Keywords: Multi-View Sensing， Sparse Phase Retrieval， Affine Sampling， Outliers， Subgradient Method， Deep Unfolding， Distributed Sensing Networks
