模糊類神經網路與專家系統於震測探勘之研究

Abstract

類神經綱路應用在Seismic Trace Editing 與 First Arrival Picking 的震測圖形識別工作上，目前以回傳式學習演算法（back-propagation lerning algorithm）為主，而且也擁有不錯的識別率，但是回傳式學習演算法本身卻有兩個主要的缺點存在：一是訓練時間（training time）過長，有時令人無法忍受；二是有區域最小值（local minima）的存在問題。 在本論文中，首先基於Keller [1985] 所提出的模糊Perceptron學習演算法（fuzzy perceptron learing algorithm）發展出模糊口袋學習演算法（fuzzy pocket learning algorithm）.模糊口袋學習演算法在訓練樣本非線性可分割時，能避免因不確定樣本（uncertain sample）所產生的的錯誤震盪行為，同時也免除了模糊Perceptron學習演算法其程式終止條件（stopping condition）不易設定的缺點，進一步地，當使用tensor forms做為模糊口袋學習演算法的訓練樣本時，我們可將模糊口袋學習演算法視為模糊Functional-Link綱路學習演算法，亦即模糊口袋學習演算法可直接擴展應用於 Functional-Link綱路的模糊學習演算法，此外，由於模糊 K 個最近鄰點法則（fuzzy K nearest neighbors rule）具有在平均辨識率以及辨識過程上，較之K個最近鄰點法則優越且穩健的特性；因此，基於模糊K個最近鄰點法則的方法觀念與辨識過程，我們嘗試發展出一個模糊類神經綱路模式：模糊K個最近鄰點法則綱路（ fuzzy K-NNR neural net）的演算法與架構。最後，將這些模糊類神經綱路模式應用在Seismic Trace Editing與 First Arrival Picking 的震測圖形識別工作上；我們發現，它們不但能免除回傳式學習演算法主要的缺點，而在平均辨識率上，模糊Functional-Link綱路學習演算法與模糊 K個最近鄰點法則綱路演算法不但分別優於Functional-Link綱路學習演算法與K個最近鄰點法則，而且均優於回傳式學習演算法。這也顯示出模糊類神經綱路未來在震測圖形識別工作上的良好發展潛力。 在本論文中亦就震測解釋專家系統（Seismic Interpretation Expert System）之推理方式做一探討。震測解釋專家系統在結合與傳遞確信因子（certainty factor）時，是依據機率理論的概念；但是最後結論的確信因子有時不能反應油氣是否存在的實際情況。因此，我們嘗試以模糊運算來結合與傳遞確信因子以使結論的確信因子能真正反應出實際的狀況。我們在一些實際範例中顯示出模糊推理方式的優越性與可行性。

The widely used model of neural network for seismic pattern recognition including seismic trace editing and first arrival picking is the back-propagation learning model, and it ration of recognitio is also good. But there are two major drawbacks of back-propagation learning model:one is that the training time is sometimes too long to accept, the other is that it cannot assure a successful learning due to the local minimum problem. In this thesis, we first propose the fuzzy pocket learning algorithm by using the weights updated rule of the fuzzy perceptron learning algorithm. The fuzzy pocket learning algorithm can reduce the influence of the uncrtain patterns in adjusting the decision boundary and free from the difficulty of setting the stopping condition of the fuzzy perceptron learning algorithm. Moreover, if the fuzzy pocket learning algorithm uses the tensor forms to be its training patterns, then we can view the fuzzy pocket learning algorithm as the fuzzy functional-link net learning algorithm, an expanded version of the fuzzy pocket learning algorithm. On the other hand, due to the fuzzy K nearest neighbors rule (fuzzy K-NNR) possesses superior average recognition rate and more robust recognition procedure than K-NNR, so we propose the fuzzy K-NNR neural net algorithm and its corresponding architecture based on the fuzzy K-NNR Finally, all these neural algorithms (fuzzy and nonfuzzy) are applied to process seismic trace editing and first arrival picking. We find that they can eliminate the drawbacks which BPNs suffer, moreover, in these nerual models including BPNs, the fuzzy functional-link net learning algorithm and fuzzy K-NNR neural net algorithm can give better average rates of correct classification in seismic trace editing and first arrival picking. This indicates the development potential of the fuzzy neural networks in seismic pattem recognition. We also discuss the inference procedure of Seismic Interpretation Expert System (SIES). In the inference procedure, the method used to combine and propagate certainty factors is probability-based. However, ever for the case of existence oil or gas, the certainty factor of conclusion may get a lower ralue (0.56, for example) by probability-based method. This seems that we cannot combine the input factors by probability-based method directly. Therefore, we focus on how to combine and propagate factors by fuzzy union and intersection operations in order to give a reasonable certainty factor which can esponse to a real situation.