類神經網路於震測圖型分析之研究

Abstract

在這篇論文中包含三個主要的章節, 第二章. 自我組織神經網路用於震測水平層連接 第三章. 循序分類於網路節點成長之研究 第四章. 類神經網路於震測圖型之強健性辨識 在第二章中, 我們基於自我組織映射的原理設計了一個演算法來連接地震圖中的震測水平層. 神經元的拓撲是一個線性圖形用來表示一條水平層. 神經元的權重係數是被地震圖的點 (Peak Point) 以自我組織方式來調整的. 一個 Peak Point 抓取一個神經元. Moment 被用來決定新的神經元建立的位置. 我們用這個演算法用再模擬的以及真實的地震圖上面, 得到相當好的結果. 在第三章中, 我們結合了兩個重要的技術, the approximating posteriori probability functions of the classes in the outputs of the trained multilayer perceptron 以及 sequential classification technique, 我們可以從 perceptron 和 two-layer perceptron 中得到最小的節點數目. 我們將這個技術用於典型的 exclusive OR 的問題和震測圖型識別上面. 網路節點降低的比率非常好. 在第四章中, multilayer perceptron 類神經網路被訓練當作一個分類器, 用於辨識震測圖型. 在平移, 旋轉, 大小可以維持不變的 Seven moments 被運用在每一個震測圖型的特徵產生. 在這個系統裡面, 有訓練圖型組和測試圖型組. 在測試圖型組裡面包含了不同等級的雜訊. Multilayer perceptron 在初始的時候先用無雜訊的震測圖型進行訓練. 在訓練收斂之後, 神經網路用於辨識包含有雜訊的測試圖型組. 接著在訓練圖型組中加入一些雜訊較高辨識錯誤的圖型用來重新訓練. 這樣子的訓練和辨識過程反覆幾次的階段. 在每一個訓練階段, 把收斂好的網路用於辨識在 Mississippi Canyon 的真實震測資料. 結果在使用到較高的雜訊訓練階段, Bright spot 圖型可以被偵測出來.

This thesis contains three major chapters, Chapter 2. Self-Organizing Neural Network for Seismic Horizon Linking. Chapter 3. Node Growing of Perceptrons By Sequential Classification Technique. Chapter 4. Neural Network for Robust Recognition of Seismic Patterns. In Chapter 2, we design an algorithm based on self-organizing feature maps to link the seismic horizon in the seismogram. The topology of the neurons is a linear pattern that forms a horizon. The weighting coefficients of the neurons are self-organized by the peaks of the seismogram. One peak catches one neuron. Moment is used in the determination of new neuron creation position. We have applied the algorithm on simulated and real seismograms and the results are quite well. In Chapter 3, combining the important property of the approximating a posteriori probability functions of the classes in the outputs of the trained multilayer perceptron and sequential classification technique, we can get the minimum number of nodes in perceptron and two-layer perceptron. We apply the technique to the typical exclusive OR problem and seismic pattern recognition. The reduction rate of nodes is quite good. In Chapter 4, the multilayer perceptron neural network is trained as a classifier and is applied to the recognition of seismic patterns. Seven moments that are invariant to translation, rotation, and scale, are employed for feature generation of each seismic pattern. In the system, there are training and testing pattern sets. The testing pattern set includes different noise level. The multilayer perceptron is initially trained with the training set of noise-free seismic patterns. After convergence of the training, the network is applied to the classification of the testing set of noisy seismic patterns. Some misclassified patterns with higher noise level are added to the training set for retraining. The training and classification process is repeated through several stages. The converged network at each training stage is applied to the real seismic data at Mississippi Canyon, the bright spot pattern can be detected when the stage is using higher level noisy patterns in the training.