類神經網路與樹自動機於震測圖形之辨認

Abstract

在複雜的地震計記錄圖(synthetic seismogram)，存有相當多結構的震測圖型(structural seismic patterns)。為了能有效率的來偵測這些圖形，本研究將提出一個樹狀結構的自動機系統(tree automaton system)架構，用來辨認出這些震測圖形。 這個系統將包含兩部份，第一部分是將震測圖形轉換為相對應的樹狀結構資料，此部分包含四個步驟，首先定義每種震測圖形的子樣本的樹狀表示，第二步驟再找出震測圖形中這些已知的子樣本的位置，由於震測圖形有形狀失真或尺寸大小易於變化等特點，所以為了能有效的找出這些子樣本，吾人使用多層認知器(Multilayer perceptron)類神經網路，將每個子樣本轉成一組七個不變矩(Moments)當做特徵給輸入層來訓練。第三步驟是計算所有找出的子樣本之間的距離與其位置間的相對關係。第四步驟將根據這些子樣本的相對關係資料，利用先廣後深搜尋(Breadth-First Search)方法發展出一個可轉換為所對應的樹狀結構資料的演算法。根據這包含四個步驟的方法，將可有效的來處理一些失真不連續點的震測圖形，使它可以轉換為所對應的樹狀結構資料。 第二部分是將所建立出的樹狀結構資料，使用樹狀結構文法推導演算法，建構出其相對應的樹狀結構文法(tree grammar)與樹狀結構自動機(tree automaton)，然後再使用由下往上(bottom-up)結構保留可校正錯誤樹自動機(SPECTA)，與由上往下(top-down)可校正錯誤樹自動機(ECTA)等方法來分析樹狀結構文法(tree grammar)，而產生辨識結果，決定震測圖形所屬於的類別。由實驗結果我們將比較這些可容錯的樹自動機之優缺點，我們將以震測亮點圖形(seismic bright spot pattern)來當成例子分析。

In a synthetic seismogram, there exist certain structural seismic patterns. In order to recognize seismic patterns efficiently by computer and improve seismic interpretation, We propose a block diagram of tree automaton system. The system includes two parts. The first part transforms the seismic patterns into their corresponding tree representations. This part includes four steps. In step 1, we assign primitives for each subpattern. In step2, we want to find the position of these known subpatterns in the seismic pattern. The properties of seismic patterns such as deformation in shape or change in size. In order to find these subpatterns effectively, we use the multilayer perceptron neural network and transform each subpatterns into a set of seven invariant moments as the features for training in neural network. In step 3, we calculate the distances and describe the relations among the position of all subpatterns. In step 4, according to the relations, we use breadth-first search method to develop the tree construction algorithm. Using this method that include four steps, we will deal with corrupted seismic patterns effectively. The second part, we propose the tree grammar inference algorithm, the tree grammar and tree automata are inferred from tree representations. To obtain the recognition results of seismic patterns, we use bottom-up structure preserved error-correcting tree automata (SPECTA) and top-down error-correcting tree automata (ECTA). In the experiments we discuss the performance of these fault tolerance tree automata. We use the fundamental tree grammar and automaton on the analysis of seismic bright spot pattern as the example.