模糊類神經網路於震測圖形識別之研究

Abstract

我們提出三個模糊類神經網路模型，模糊Ｋ個最近鄰點法則網路 (Fuzzy K-NNR net)，二階段模糊類神經網路 (Two Learning Steps Fuzzy Neural Network)，及 Fuzzy Functional-Link 網路。我們將此三個模糊 類神經網路應用在 Seismic Trace Editing 和 Seismic First Arrival Picking 的問題上。第一個模糊類神經網路是模糊Ｋ個最近鄰點法則網路 。模糊Ｋ個最近鄰點法則網路由 Hamming net 發展而來的。在訓練網路 的階段，每一 pattern 都依據其所屬的類別而設定 fuzzy membership。 在測試的階段，測試的 pattern 通過神經網路而決定這個 pattern 是屬 於何種類別。第二個模糊類神經網路為採用模糊 C-means 理論的二階段 模糊類神經網路。這種網路的訓練過程可分為二個階段。第一個階段採用 模糊 C-means 理論是一種非監督式學習的方法而第二個學習段採用 perceptron learning 是一種監督式學習的方法。在在測試的階段，測試 的 pattern 被放進這個網路，然後轉換成 C 個 fuzzy degrees，輸出層 根據這 C 個 fuzzy degrees 得到最後的輸出。第三個模糊類神經網路 是 Fuzzy functional-link net，在學習的過程中加入了模糊的概念。這 三個模糊類神經網路都應用在震測的 Trace editing 和 First arrival picking 的問題上。所有的實驗都能獲得相當好的結果。 We propose three fuzzy neural network models, fuzzy K-nearest neighbor rule (fuzzy K-NNR) net, two learning steps fuzzy neural network, and fuzzy functional-link net. The three fuzzy neural networks are all applied to two important seismic pattern recognition problems, seismic trace editing and seismic first arrival picking. The first fuzzy neural network model is fuzzy K-neighbor rule neural network. Fuzzy K-nearest neighbor classification rule is implemented by neural network of the Hamming net. In the training stage of fuzzy K-nearest neighbor classification rule neural network, each pattern is assigned fuzzy membership. In the testing stage, testing patterns are through the neural network to determine which class the testing pattern belongs to. By adopting fuzzy C-means theorem the second fuzzy neural network model is two steps learning fuzzy neural network. The training stage of this neural network model are divided into two learning steps. The first step is applying unsupervised learning method using fuzzy C-means theorem as learning algorithm and second learning step is perceptron learning by gradient-descent method which is a supervised learning method. In the testing stage, each testing pattern is put in this network and transfer the pattern to C fuzzy degrees, output layer then get the output according to the C fuzzy degrees. Fuzzy functional-link net is 3rd fuzzy neural network model incorporated with fuzzy concept in learning procedure. The three fuzzy neural network are applied to seismic trace editing and first arrival picking. The experiments of seismic trace editing and seismic first arrival picking are quite encouraging.