Complex Lithofacies Identification Using Improved Probabilistic Neural Networks

The probabilistic neural network (PNN) is functional in recognizing complex patterns without doing any pretraining of source data. However, for some data clusters, independence and colinearity characteristics of the variables in learning samples can seriously distort the window lengths of their probability density distributions, then leading to the incorrect or totally wrong calculated probability values and nal recognition results. In view of such drawbacks, an improved PNN that incorporates two techniques of mean impact value (MIV) and correlation analysis is proposed in order to perfect the original PNN's calculation mechanism by removing those interference and colinear variables from the source data. The data used to validate the method are from two wells in the Iara oil eld. Recognition accuracies of the improved network in four experiments are, 74.05%, 71.7%, 83.02% and 88.24%, respectively, each of which is the highest accuracy. The validation results demonstrate that the new network has the capability of recognizing complex carbonate lithofacies and the results are reliable enough to serve as the reference data for other geological efforts, such as analyzing sedimentary process and building a sequence framework.