Research On Imbalanced Data Classification For Lithologic Identification Of Complex Reservoirs

In recent years,the issue of imbalanced data is widely occurred in real life,and has attracted more and more attention.The phenomenon of imbalanced data also occur in the field of geological prospecting,along with the difficulty of exploration in complex reservoirs,the traditional classification algorithm is difficult to do the detailed reservoir description and lithology identification in complex reservoirs.Through the analyzing,we find that this kind of rock data are all imbalanced data,and some of them show high similarity.Consequently,after analyzing the achievements and problems of imbalanced data classification algorithm,binary classification and multi classification algorithms for such rock data has been proposed by the thesis.The specific researches and works of the thesis are as follows:Firstly,by analyzing the characteristics of rock data that collected from complex reservoir,a data balanced algorithm C&SM(cluster and samples move based on feature distance)is proposed.The algorithm would obtain several sub samples firstly by clustering majority class,then compares the capacity of minority sample with those sub samples,if the imbalanced phenomenon show again among the sub samples,use the samples move method based on feature distance for secondary processing,and finally multi balanced sub samples would be got.The algorithm not only can avoid the change of original data distribution and data missing by use the resampling algorithm,but also can solve the problem of imbalanced phenomenon show again among the sub samples.Secondly,a novel ensemble rule based on feature weight of formation elements is proposed,which could solve the problem of binary classification of imbalanced rock data with high similarity.After processing by the feature selection algorithm,different weights are assigned to formation elements,which can guide the voting rule of base classifiers.Experiments results show that the AUC value of the novel ensemble rule can reach at least 0.93.Thirdly,after studying the multi classification algorithm of imbalanced rock data,the thesis has proposed a rock sample fusion strategy and multi-layer funnel classification model,MICMF.The multi imbalanced classification problem could be transformed into multi balanced classification problem at first by fusion strategy,and then the MICMF model would be used to do the lithology identification.Experiments show that the accuracy of the MICMF model can reach 92%,and the precession of the model can reach 84%.