運用支撐向量機技術之FHMA/MFSK 訊號偵測

Abstract

本論文提供一個初步用來設計智慧型通訊處理器(intelligent communication processor)的整合架構。智慧型通訊處理器的一個重要特性是其可以利用離線式(off-line) 的訓練時期所獲得的知識(knowledge)來做及時性的訊號偵測。雖然離線式的訓練可能需要大量的運算，但所獲得的資訊卻是很簡潔的，因此智慧型通訊處理器只需要小量的運算就可以做訊號偵測。利用這樣的概念，我們從機械學習的觀點來探討在多重存取環境中跳頻(frequency hopping)訊號偵測的問題。

相對於直接序列展頻(DS-CDMA)技術，跳頻技術是另一種吸引人 的多重存取技術。除了通道統計特性，跳頻多重存取系統的效能會由兩項主要考量所決定：波形(waveform)的設計和接收機的架構。在給定跳頻波形，我們仍然很難設計最大相似性(ML)接收機，其主要原因是我們對於通道統計特性的瞭解是不完整的，即使是完整的，通常相對應的條件機率分佈(conditional pdf)是沒有簡潔的表示式。

藉由把接收到的訊號視為一個時間─頻率的圖形，我們可以把多使 用者偵測的問題轉換成圖形辨識的問題，然後運用支撐向量機的技術來解決所對應產生的圖形辨識問題。利用適當的核心函數(kernel function)，支撐向量機技術把收到的訊號轉換到高維度的特徵空間。利用了序列最小優化(Sequential Minimization Optimization, SMO)和有向無迴圈圖(Direct Acyclic Graph)演算法來找尋在特徵空間中的最佳分離平面，我們提出了基於支撐向量機技術的接收機。模擬的結果顯示我們的設計具有強軔性和令人滿意的效能。

This thesis documents an initial effort in establishing an uni‾ed framework for designing an intelligent communication processor (ICP). An important feature of a prototype ICP is its capability of applying the knowledge learned during an o□-line training period to real-time signal detection. Although the off-line training might require very intensive computing power, the extracted information does has a concise representation, which then enables the corresponding ICP to detect a signal using only simple and low-power operations. As an application of such a concept, we revisit the problem of detecting frequency-hopped (FH) signals in a multiple access (MA) environment from a machine learning perspective.

Frequency-hopping is an attractive alternative multiple access technique for direct sequence based code division multiple access (CDMA) schemes. Other than the com- munication channel statistic, the capacity of an FHMA system is determined by two major related design concerns: waveform design and receiver structure. Given the FH waveform, one still has di±culty in designing an FHMA ML receiver due to the facts that our knowledge about the channel statistics is often incomplete and even if it is complete the associated conditional probability density function (pdf) does not render a closed-form expression.

Regarding the FHMA/MFSK waveform as a time-frequency pattern, we convert the mutiuser detection problem into a pattern classi‾cation problem and then resolve to the Support Vector Machine (SVM) approach for solving the resulting multiple-class classi‾cation problem. By using an appropriate kernel function, the SVM essentially transforms the received signal space into a higher dimension feature space. We propose a SVM-based FHMA/MFSK receiver by applying the Sequential Minimization Optimization (SMO) and Directed Acyclic Graph (DAG) algorithms to ‾nd the optimal separating hyperplanes in the feature space. Simulation results indicate that our design does yield robust and satisfactory performance.