標題: 倒傳遞網路應用於油氣儲集層亮點偵測
Detection of Oil Reservoir Bright Spot Using Back-propagation
作者: 詹世偉
Shih Wei Chan
曾仁杰
Ren Jye Dzeng
土木工程學系
關鍵字: 油氣儲集;圖型辨識;倒傳遞網路;亮點偵測;Oil Reservoir;Pattern Recognition;Back-propagation;Bright Spot Detection
公開日期: 2007
摘要: 震測剖面圖圖資的複雜程度常常引起人員的誤判,然而因誤判而進行鑽探的結果往往造成三千萬成本的損失。震測資料解釋人員憑其經驗法則判圖,若此經驗法能以類神經網路取代,以類神經網路辨識的結果做為解釋人員判圖的參考,屏除震測解釋人員情緒化的主觀判斷或因眼誤造成誤判機會。 過去的油氣亮點圖形辨識,尚無法實際運用在震測資料解釋上,一為所使用的震測剖面圖測試樣本為最完美情況下之圖形,與實際震測剖面圖有相當的落差,二為抽取之特徵值並非實際油公司常用的特徵,故應用上有困難。 本研究主要探討如何使用類神經網路於震測剖面圖的亮點圖形辨識,經由實地訪談油公司取得震測剖面圖資,並提出了五項常用特徵,如下:震波原始訊號(Seismic Signal)、震幅強度(Envelope)、瞬間頻率(Instantaneous Frequency)、瞬間相位(Instantaneous Phase)、逆推阻抗(Inversion Impedance),再將特徵做前處理(Pre-processing),如:除去異常極大值、除去極端異常值、正規化,最後輸入網路,反覆測試出使辨識率最佳的可能特徵組合、訓練函數(倒傳遞網路之變形Levenberg-Marquardt、Conjugate Gradient)、網路層數與神經元數,共12480組,最後提出最佳化辨識率的儲集層亮點辨識法。
he complexity of two-dimensional seismic data often leads to mistakes in discriminating oil reservoir. However, mis-drilling caused by these mistakes brings about thirty million loss in cost each time. Seismic data interpreters do recognition by rules and experiences. If we can use neural network in place of rules and experiences, then we can get rid of some chances of seismic data interpreters’ emotional discrimination or mislook. The past oil reservoir bright spot detections still can't be applied to practice for two reasons. First, target seismic data in past research to be detected is in perfect condition. Interpreters seldom deal with these kinds of seismic data in reality. There is a big difference between seismic data in perfect condition and seismic data in practice. Second, the seismic attributes that past research used differs from that interpreters used in oil companies. The primary goal of this research is to apply back-propagation neural network to pattern recognition of oil reservoir bright spot. By interviewing with Seismic data interpreters in the oil company, we propose five seismic attributes in common use including seismic signal, evelope, instantaneous frequency, instantaneous phase and inversion impedance. After five seismic attributes of feature extraction, we do pre-processing on extracted features including transformating .segy file into .mat file, elimination of blunder, elimination of outlier, normalization and building feature set matrixes. Then, we import feature set into neural network and train matrix by matrix. By tuning any possible neural network layer, hidden layer node, training function(Levenberg-Marquardt, Conjugate Gradient), we have summed up to 12480 times of neural network training. Finally, we propose a method of optimized-recognition rated oil reservoir bright spot detection.
URI: http://140.113.39.130/cdrfb3/record/nctu/#GT009516535
http://hdl.handle.net/11536/38692
Appears in Collections:Thesis


Files in This Item:

  1. 653501.pdf