廣義空間調變多工系統（Generalized Spatial Modulation with Spatial Multiplexing， GSMx）可透過啟用天線組合與調變符元共同傳遞資訊，以提升頻譜效率與資料傳輸率。然而，接收端除了需偵測傳送符元外，亦須判斷實際啟用之天線組合，使候選解數量隨發射天線數、啟用天線數與調變階數增加而快速成長。若採用最大概似偵測器雖可獲得最佳錯誤率效能，但其計算複雜度與硬體成本皆相當龐大。為降低 GSMx 接收端之偵測複雜度，本論文採用兩階段偵測架構，先進行天線組合候選篩選，再針對保留之候選天線組合進行符元偵測。在符元偵測階段中，本文提出適合硬體實作之實數分解方式，將複數上三角系統轉換為實數上三角模型，並利用其相鄰層間的結構特性設計分散式 K-Best（Distributed K-Best， DKB）偵測器。所提出之架構結合 Schnorr-Euchner 列舉概念，在實作上透過查表方式，簡化候選點產生、節點比較與部分歐幾里得距離計算流程。此外，本文亦設計軟性輸出產生器，利用候選列表之距離資訊產生對數概似比，以支援後端維特比解碼器進行軟性決策解碼。模擬結果顯示，在 Nt = 8、Na = 3、Nr = 5 且採用 64-QAM 調變之 GSMx 系統中，所提出之 DKB 偵測器可達到接近 ML 偵測器之錯誤率效能。在 BER 為 1e−5 時，未編碼系統下 DKB 與 ML 約相差 0.85 dB；於軟性決策解碼系統下，兩者差距約為 0.4 dB，且相較於硬性決策解碼約可獲得 0.6 dB 之效能增益。最後，本論文以 TSMC 40 nm 製程完成晶片實現，操作頻率可達 487 MHz，正規化吞吐量為 2045.4 Mbps。結果顯示，本文所提出之軟性輸出 DKB 偵測器能在維持近似 ML 效能的同時，兼具高吞吐量與硬體實作可行性。;Generalized spatial modulation with spatial multiplexing (GSMx) improves spectral efficiency and data transmission rate by jointly conveying information through transmit antenna combinations and modulation symbols. However， the receiver must detect not only the transmitted symbols but also the actual active transmit antenna combination. As a result， the number of candidate solutions increases rapidly with the number of transmit antennas， active antennas，and modulation order. Although maximum likelihood (ML) detection can achieve the optimal error-rate performance， its computational complexity and hardware cost are considerably high.
To reduce the detection complexity of the GSMx receiver， this thesis adopts a two-stage detection architecture. The transmit antenna combination candidates are first selected， and symbol detection is then performed for the retained antenna combinations. In the symbol detection stage， this thesis proposes a hardware-oriented real-valued decomposition method， which transforms the complex upper triangular system into a real-valued upper triangular model. Based on the structural property between adjacent layers， a distributed K-Best (DKB) detector is designed. The proposed architecture combines the concept of Schnorr-Euchner enumeration with a table-based implementation， thereby simplifying candidate generation， node comparison， and partial Euclidean distance calculation. In addition， a soft-output generator is designed to produce log-likelihood ratios using the distance information of the candidate list， enabling soft-decision Viterbi decoding.
Simulation results show that， for a GSMx system with Nt = 8， Na = 3， Nr = 5， and 64-QAM modulation， the proposed DKB detector achieves near-ML BER performance. At a BER of 1e-5， the proposed DKB detector is approximately 0.85 dB away from ML detection in the uncoded system. In the soft-decision coded system， the performance gap is reduced to about 0.4 dB， and approximately 0.6 dB gain can be achieved compared with hard-decision decoding. Finally， the proposed design is implemented using TSMC 40 nm CMOS technology. The operating frequency reaches 487 MHz， and the normalized throughput is 2045.4 Mbps. These results demonstrate that the proposed soft-output DKB detector can maintain near-ML performance while achieving high throughput and practical hardware feasibility.
