應用類神經網路辨識摩斯碼

Abstract

近年來，由於類神經網路在各方面應用漸漸地愈來愈受重視，因此依據一 個較有系統，有層次的方法來設計及製作類神經網路是相當重要的。在本 論文中，將階層式設計的理論應用於摩斯碼的辨識上。而在訂定方法及研 製實用系統的過程中，我們深切了解到一個高性能的類神經網路應用系統 ，應當是傳統的方式與類神經網路相搭配而成的系統。因此，在本論文中 提出混合式的類神經網路。階層式的設計，是由問題的分析，將問題分成 數個子部分，並且針對每個子部分，分析其主要目標。對於每個目標，研 析出數種方法，再對每個方法加以評估，找出較好的方法，最後將之整合 。摩斯碼的辨識系統藉由階層式的設計，將系統分成三部分，分別是前處 理，特徵擷取，以及辨識。在辨識的過程中，希望能有較佳的辨識結果， 因此採用混合式類神經網路。混合式類神經網路在訓練時間上較單一之神 經網路省時間；前者約為後者的三分之一。而在辨識的效果上，也較單一 種的方法好；混合式類神經網路的辨識率為 90%，傳統方法與單一之類神 經網路的辨識率分別為 85.7%， 82%。但在輸入值方面，需選擇適合的輸 入，才能有上述之優點。藉由混合式類神經網路以及階層式的設計，摩斯 碼辨識系統能有較好的結果。 Recently, the studies of artificial neural network have attracted the attention of more and more researchers from various fields. It is important to design and implement neural networks on a systematic and hierarchical methodology. In this thesis, we develop the hierarchical design flow and follow this flow to design, analyze, simulate and implement to solve Morse Codes Recognition System. In the recognition system design, we were aware that a high performance system which contains both neural networks and traditional method system. Therefore, we propose a hybrid neural network system in this thesis. The design of hierarchical approaches is to analyze problem and then divided the problem into several parts. The major function of each part should be clearly defined in advance. Then, we find some methods to achieve its function and choose the best one. Finally, integration of each part is required. With hierarchical design flow, Morse Codes recognition system is divided into three subsystems, preprocessing, feature extraction and code recognition. In order to achieve high performance, we use hybrid neural network to design this system. There are two advantages in hybrid neural network, training time reduced and the better results than only one method is used. In our experiment, the training time in MLP network is about three times the time in the hybrid neural network. The final correct rates in the hybrid neural network and MLP network are 90\% and 82\% respectively. But we should note that only the suitable input values result in better performance and training time reducing. Consequently, The hierarchical design and hybrid neural network lead to the better results in Morse Codes Recognition System.