LW Rui-Lin, Zhou Cheng-Dang and Jin Zhen-Wu
Dept. of Geophysical Exploration, Jianghan Petroleum Institute, Jiangling, Hubei, P.R. CHINA
Abstract
Three—layer perceptrons and statistical pattern recognition techniques are usually used for lithofacies recognitions from well logs which recognize facies as isolated patterns, with no consideration of the inter-relationship between lithofacies. Thus, the performance of three-layer perceptrons in recognizing facies is similar to those of existing statistical pattern recognition techniques,
In this paper, a Time-Delay Neural Network(TDNN) is utilized to recognize lithofacies sequence from well logs. This approach is based on two points: (1) in a formation, lithofacies order Is a varying sequence with depth, the Time(Depth)—Delay architecture in TDNN enable the network to discover the varying features of lithofacies sequence and the dependance between lithofacies in depth, that is, the TDNN has the ability to represent relationship between lithofacies in depth; (2) the TDNN trained by error backpropagation learning algorithm allows to generate arbitrary nonlinear decision surfaces. The result obtained from an oil field example shows that the performance of TDNN is better than that of a three-layer perceptron. The TDNN can recognize reservoir facies in our test case with an accuracy of 94.3 percent, while the three-layer perceptron has only an accuracy of 89.2 percent. And the TDNN recognizes non-reservoir facies also with a higher accuracy than the three-layer perceptron.