摘要: | 隨著延展實境、自動駕駛、遠距醫療等新應用出現,在有限的頻譜資源下實現更大的資料量與低延遲,傳統通訊系統著重於符號(Symbol) 或位元 (Bit) 正確傳輸已無法滿足需求。傳統通訊系統發送端利用通道編碼 (如 Turbo Coding、Polar Coding、LDPC Coding等),使接收端得以使用通道解碼 (Channel Decoding) 修正錯誤之位元。然而,當位元錯誤率超過容錯能力時,通道編解碼技術可能無法正確還原資訊。 語意通訊 (Semantic Communication) 本質為使用AI/ML模型訓練語意模型,使其能透過前後文還原訊息之語意。若將語意通訊應用於通訊網路,發送端與接收端須同步訓練語意模型,使其具相同知識條件下,當傳輸過程中發生位元錯誤時,接收端仍有能力透過前後文解碼還原發送端之語意訊息。因此,我們構想如何採行語意編碼取代現行通道編碼技術,透過語意編解碼模型之語意與特徵對抗通道雜訊。換言之,於存在位元錯誤環境下,驗證其區塊錯誤率 (Block Error Rate,BLER) 是否優於現行通道編碼技術。 於5G New Radio (NR) 行動通訊系統內,基站傳送下行控制訊息 (Downlink Control Information,DCI) 至終端設備,其攜帶排程與參數調整等重要控制訊務。因時效性,此控制訊務傳輸須為高可靠傳輸,以確保終端設備正確接收與觸發相對應之行為;也因此,我們選擇下行控制訊息傳輸作為導入語意通訊之可行性評估。目標為利用實現聯合來源通道編碼 (Joint Source Channel Coding,JSCC),達成高可靠傳輸之要求,同時控制資源使用率。 AI-native communication為6G關鍵應用之一,AI/ML模型在訓練與推理階段普遍使用浮點數 (Floating-Point,FP),以提取資料特徵的變化。我們構想在有限的星座圖上有效映射與傳輸浮點數,確保即使在傳輸過程中受到干擾,資料也能被恢復。 最後,為了驗證基於語意通訊控制訊務傳輸之可行性,本研究將上述語意通訊系統使用開源軟體OpenAirInterface5G (OAI) 進行實際驗證,觀察其容錯能力與資源使用率。 ;With the emergence of new applications such as extended reality, autonomous driving, and telemedicine, which achieves a large amount of data and low latency within limited spectrum resources. Traditional communication system, which focus on the accurate transmission of symbols and/or bits, cannot fulfill the demands of these emerging applications. Specifically, it utilizes channel coding (e.g., Turbo Coding, Polar Coding, LDPC Coding, etc.,) to allow the receiver to recover erroneous bits via channel decoder. However, channel coding techniques may fail to recover bit information when the bit error rate (BER) exceeds the tolerance limitation. Semantic communication (SC) fundamentally uses AI/ML model(s) to train semantic model(s) that can reconstruct the semantics of messages through context. If applied to communication networks, both transmitter and receiver must synchronously train semantic models under the same knowledge base. This allows the receiver to decode and restore the semantic message sent from transmitter when bit errors occur during transmission. Thus, we attempt to replace current channel coding with semantic communication against noisy channel through the semantics and features of semantic decoding models. In other words, we develop the experiment in channel noise environment to observe whether the proposed SC outperforms existing channel coding techniques in term of Block Error Rate (BLER). In 5G New Radio (NR), gNB transmits downlink control information (DCI) to UE, it carries critical control information for scheduling and parameter adjustments. This control information transmission must be highly reliable to ensure that UE receives it correctly and triggers the appropriate actions. Therefore, we chose DCI transmission to evaluate the feasibility of introducing semantic communication. The goal is to fulfill the requirements of high-reliability transmission through Joint Source-Channel Coding (JSCC) while controlling resource utilization rates. AI-native communication is one of the key applications for 6G, which utilizes floating-point numbers (FP) during the training and inference phases of AI/ML models to capture semantic features. The new challenge involves how to efficiently mapping and transmitting FP on a limited constellation diagram, which ensures data recovery despite transmission distortions. Finally, this study is built on OpenAirInterface5G (OAI) open-source software to validate the actual performance of the proposed semantics-based control information transmission mechanism. |