在 5G 邁向 6G 的演進中，毫米波 (mmWave) 與太赫茲 (THz) 通訊已成為支撐高傳輸速率與低延遲的關鍵技術。為克服高頻訊號嚴重的路徑損耗與遮蔽效應，系統必須結合混合式波束合成技術(Hybrid Beamforming) ；由於通道呈現稀疏結構，使壓縮感知 (Compressive Sensing， CS) 技術成為通道估測的有效途徑。 本論文針對配備均勻線性陣列 (Uniform Linear Array， ULA) 之下行鏈路 大規模 MIMO-OTFS 系統之三維延遲-都卜勒-角度 (Delay-Doppler-Angle，DDA) 通道估測問題，於演算法層面提出三維結構化正交匹配追蹤結合網格旋轉演算法 (Three-Dimensional Structured Orthogonal Matching Pursuit with Rotation，3DSOMPR)。所提之 3DSOMPR 演算法以階層式 結構結合三步流程：(1) 確定式導頻設計與張量初始化；(2) 三維原子生成，於時域或延遲-都卜勒域以 2D 循環位移、空間波束合成、局部擷取、CP 與相位補償等運算合成稀疏字典原子；(3) 三維離格參數估測，於延遲、都卜勒與角度三個維度分別進行分數式二分逼近 以達到子網格精度的參數細化，並以干擾消除 (Interference Cancellation， IC) 機制提升多路徑下的估測準確度。在模擬實驗中，所提之 3DSOMPR 演算法可在延遲、都卜勒與角度三個維度同時為 Off-grid 的條件下，正確估測多路徑通道參數；與未進行離格細化之標準 3D-SOMP 基準方法相比，於相同訊號雜訊比 (SNR) 下展現顯著改善的歸一化均方誤差 (NMSE) 表現。於硬體實現層面，本論文延續實驗室前期成果之 T-OMP-R-IC 演算法以及RTL 設計，接續完成於 TSMC 28nm 製程下進行後端自動佈局繞線 (APR) 實體設計，並透過 V93000 自動測試機台 (ATE) 完成晶片之功能驗證與量測，量測結果成功驗證了 T-OMP-R-IC 加速器的功能正確性，並取得時序與功耗等實測指標，完成了以 TSMC 28nm 製程實現毫米波環境下 MIMO-OFDM 五維通道張量估測加速器的全流程驗證。;In the evolution from 5G toward 6G， millimeter-wave (mmWave) and terahertz (THz) communications have become key enablers of high data rate and low latency. To combat the severe path loss at these frequencies， systems usually employ hybrid beamforming antenna arrays. Such systems render the channel sparsity and make compressed sensing (CS) an effective approach to channel estimation. This thesis targets the three-dimensional delay-Doppler-angle (DDA) channel estimation problem of a downlink OTFS system equipped with a uniform linear array (ULA). At the algorithm level， a Three-Dimensional Structured Orthogonal Matching Pursuit with Rotation (3DSOMPR) is proposed. The 3DSOMPR algorithm adopts a hierarchical， three-step procedure: (1) deterministic pilot design and tensor initialization; (2) three-dimensional atom generation， in which sparse dictionary atoms are synthesized through 2-D cyclic shift， spatial beam combining， local windowing， CP insertion， and phase compensation; and (3) three-dimensional off-grid estimation， in which fractional bisection is carried out independently along the delay， Doppler， and angular dimensions to refine the parameters to sub-grid accuracy， together with an interference-cancellation (IC) mechanism that improves accuracy under multipath. Simulation results show that the proposed 3DSOMPR can correctly estimate multipath channel parameters when the delay， Doppler， and angle are simultaneously off-grid， achieving a significantly lower normalized mean square error (NMSE) than the standard 3D-SOMP baseline at the same signal-to-noise ratio (SNR). On the hardware side， continuing the laboratory′s prior work on T-OMP-R-IC accelerator， the back-end automatic place-and-route (APR) in a TSMC 28 nm process is completed. Its functional correctness is verified and measured on V93000 automated test equipment (ATE). The measurements confirm the functional correctness of the accelerator and provide timing and power metrics， completing a full-flow demonstration of an accelerator for five-dimensional MIMO-OFDM channel tensor estimation in TSMC 28 nm.
