臺灣博碩士論文加值系統

由於類神經網路在各種問題上的廣泛適應能力以及水下聲波信號應用的未
來發展潛力, 所以本篇論文中主要探討的領域是, 結合類神經網路與水下
訊號處理--特別是在於被動式聲波訊號的偵測與分類識別的應用上,以類
神經網路作為分類辨識的處理工具, 所表現出來的能力與效應, 做一較深
入的研究探討.利用幾種一般較常用之不同模式類型的類神經網路模式，
如倒傳遞理論(Backward-Propagation), 模糊諧調共振理論映射(Fuzzy
ARTMap)及自我組織映射(Self-Organization Map)等三種類神經網路系
統.在對實際的水下聲波訊號做前置信號處理及頻普轉換後, 以萃取出較
具有代表性的頻譜信號作為特徵信號參數, 將此特徵信號參數輸入至各類
神經網路系統,以作為分類辨識程序中的輸入訊號來源, 而後進行各個類
神經網路系統的訓練與學習過程, 最後在每一個類神經網路系統訓練完成
後, 依據每一個類神經網路系統在水下聲波訊號分類辨識的正確率上所表
現出的結果, 及達到合理之辨識分類結果時所需要之訓練資源, 探討何種
網路模式為最適合於被動式水下聲波訊號之辨識分類處理.最後藉由比較
分類辨識率高低, 找出在水下聲波信號的分類辨識應用中, 表現最好的類
神經網路系統後, 改變其類神經網路系統中學習演算法則之學習參數值與
變化情形, 或是在此類神經網路的架構上做一適度的調整, 進一步加強其
辨識分類之正確率, 縮短網路樣本學習之時間。

For the widely adaptability of the Neural Networks in various
applications, and the bright future of the development of
appli- cation in Underwater Acoustic Signal. So, In this
thesis, by combine the Neural Network and the Processing of the
Underwater Acoustic Signal together, we try to discuss the
ability of the Neural Network applied on the detection and
recognition of the Passive Acoustic Signal, and the fitness for
the Neural Network work in Underwater Acoustic Signal
recognition. Use several different Neural Network Models-Back-
Propagation Network, Fuzzy Adaptive Resonance Theory(Fuzzy
ARTMAP) Neural Network and the Self-organization Map(SOM)
Neural Network, after the processed of the Underwater Acoustic
Signal, and use Fourier Transform to convert the signal to
frequency domain then get the Feature Vector for this Signal.
Use this Feature Vector as the input training data for the
Neural Network. After the Neural Network trained, compare each
Neural Network for the correct rate of the classification, then
can decide which is the proper modal for Passive Sonar Signal
recognition and classification. Finally, we promoting the rate
of the classification for Passive Acoustic Signal, trying to
accelerate the speed of the classification, reduce the learning
times for the Neural Network by adjust the learning parameter
or the activity function on the neuron unit and modify the
architecture of the model of the Neural. Network we chosen in
the research of this thesis.