類神經網路方法於井測資料反推之研究

Abstract

井測反推是一種利用儀器來蒐集地下的資訊並藉此推測地層特性的技術，已經被廣泛地使用在油氣探測的工業上。然而，井測視導電率與地層真實導電率之間是一種非線性的映對關係，對於井測的反推來說是一項困難的工作。類神經網路的架構提供了非線性輸入與輸出間的映對關係，而且可以在反推的問題上獲得不錯的結果。我們提出三種類神經網路網路架構來解決井測反推的問題。首先，多層感知器(Multilayer Perceptron, MLP) 網路利用梯度坡降法經由訓練過程來調整權重值，以逼近輸入輸出值之間的映對關係。除此之外，為了加強網路的非線性映對能力，我們提出高階特徵值類神經網路(Higher Order Feature Neural Network, HONN)，其特性為使用次方函數來反應出輸入特徵值的高階項。另外，我們也提出修改後的兩層輻射半徑基底函數(Radial Basis Function Network, RBF) 網路與改良式三層的RBF，應用於井測資料的反推。針對我們所提出的三種網路架構用於井測的模擬資料。由初步的實驗結果顯示，MLP網路在輸入為十個特徵值，隱藏節點個數為二十七個以及輸出節點為十個的時候，可以得到與模擬資料間的最小平均絕對值誤差。HONN在輸入為十個特徵值且擴展到三階，隱藏節點為八個，以及輸出節點為十個的時候，可以得到與模擬資料間的最小平均絕對值誤差。而改良式三層的RBF在輸入為十個特徵值，第一層節點為廿七個，第二層節點為九個，以及第三層(輸出)節點為十個的時候，可以得到與模擬資料間的最小平均絕對值誤差。在實際應用的方面，我們也將所提出的三種網路架構MLP，HONN與三層的RBF應用在真實的井測資料上，以測試網路的反推結果。對井測的反推問題有不錯的結果，將表示我們所提出的方法對於井測資料的反推是可行且有效率的。

Well log is a technique of using instruments to gather and identify subsurface information. It has been widely used in oil and gas exploration industry. However, inversion of non-linear mapping from apparent conductivity to the true formation conductivity is difficult. Neural network provides a way for non-linear input-output mapping and a robustness result. We propose three neural network methods to solve the inversion problem. First, multilayer perceptron neural network (MLP) is able to approximate the input-output mapping through training process with iterative adjustments of connection weights by gradient descent. To consider higher order mapping from apparent conductivity (Ca) to true formation conductivity (Ct), we propose higher order feature neural network (HONN) where the input features is composed of its expanding second order and third order. Besides, we propose the modified radial basis function network (RBF) and three-layered RBF to perform the inversion task of well log data. Simulated data are used in our proposed three models. From the preliminary results, the multilayer perceptron network with 10 input features, 20 hidden nodes, and 10 output nodes can get the smallest average mean absolute error. And the higher-order multilayer perceptron network with 10 input features, the expanding input features to the third order, 8 hidden nodes, and 10 output nodes can get the smallest average mean absolute error on simulated well log data. The three-layered RBF network with 10 input data, first layer with 27 nodes, second layer with 9 hidden nodes, and 10 output nodes can get the smallest average mean absolute error. Also, we will apply our proposed network on the real field well log data to get the inverted Ct. From the good result, it will show that our methods provide a feasible and effective way to do the well log data inversion.