Feature Selection And Rule Extraction In Oil-bearing Reservoir Recognition

With the rapid economic development,the demands of china for oil and other energies are increasing gradually.However,the production of oil cannot meet the demands for various factors,such as the limitation of technical level and industrial structure.Our foreign dependence ratio of oil has been above danger level for nine consecutive years since 2008,indicating a contradiction between oil supply and demand.As effective measures,the growth in oil reserve and the improvement in oil output can make great contributions in easing the tensions in oil consumption,and the accurate identification of oil reservoir can guarantee the growth in oil reserve and the improvement in oil output.Oil reservoir identification mainly includes oil-bearing reservoir recognition and predicition of key attributes.Oil-bearing reservoir recognition is an important content of reservoir recognition,which restores logging information to geological information,and it will directly affect the efficiency and success rate of oil exploration.In the process of oil-bearing reservoir recognition,the key logging attributes such as whether the description information of logging attributes is complete,whether the attributes is redundant,whether the attributes include the information related to reservoir classification,have great effects on the recognition rate.At the same time,the managers' knowledge on oil-bearing reservoir classification will directly affect the rationality of their decisions.Thus,it is particularly important to select the key logging attributes from many logging attributes,and transform the key logging attributes into easily understandable knowledge.Feature selection is an important technique which is used to remove irrelevant and redundant attibutes,and it is often adopted in data mining and machine learning.Feature selection performs well on obtaining the key attributes.If feature selection is applied on logging attributes,the logging attributes which are high correlate to classification will be picked out and the logging attributes which are irrelevant to classification will be deleted.On the basis of guaranteeing the classification accuracy,the number of logging attributes is reduced,in order to reduce the data in the oil exploration.As a result,the cost will be reduced and the requirements of exploration technical level will be reduced in the oil exploration.And rule extraction is an important technique which converts the hidden information in data into knowledge.The result of rule extraction can form rules as “IF-Then” which is similar to expertise,and it relates the logging attributes information to the results of reservoir classification in a simple and straightforward way.The result of rule extraction can help managers who are not professional understand the classification information of reservoir easily,thus the managers can make better decision.Based on the relevant literatures,this paper has studied the feature selection and rule extraction,and a new feature selection algorithm and a new rule extraction algorithm are proposed by considering the advantages and disadvantages of various evaluation algorithms,classification algorithms,and search algorithms in feature selection,and the advantages and disadvantages of rules conflict resolution,rule set evaluation criteria,and search algorithms in rule extraction.In the proposed algorithms,differential evolution algorithm is selected as the search algorithms,and this paper improves the traditional differential evolution algorithm to solve problems on the sensitivity of control parameters and the poor ability of local searching.In the improved differential evolution algorithm,different control parameters have been used to make mutation and crossover.Here is the detailed process: Firstly,it initializes two groups of population.Secondly,it generates a set of control parameters for one of the two populations and then further creates another new series of control parameters for the other population through mutating the initial control parameters.Thirdly,once the control parameters are generated,the two populations are mutated and crossed to produce two groups of trial vectors.Finally,the target vectors are selected from the two groups of trial vectors.In order to enhance its global search capabilities,simulated annealing is involved in the selecting operation and the control parameters with better performance are chosen as the initial control parameters of the next generation.To test the performance of the improved differential evolution algorithm,we compared the improved differential evolution with several other differential evolution algorithms using 17 benchmark functions.The results of comparison between improved differential evolution algorithm and several other differential evolution algorithms show that the improved differential evolution algorithm is superior to other algorithms on the total performance.In the new feature selection algorithm,Relief F,BIF,FCBF and randomly selected algorithm are selected as the pool of evaluation standard;SOM neural network,fuzzy c-means,k-means and k-Nearest Neighbor algorithm are selected as the pool of classifier,and the improved differential algorithm is selected as the search strategy.In order to test the performance of the new feature selection algorithm,this paper will test the new feature selection algorithm on five benchmark datasets,and will compare its result with the result of other feature selection algorithms.The results of these analyses show that the new feature selection algorithm can extract the key feature and ensure a high recognition rate at the same time,thus the new algorithm performs better than other feature selection algorithms.To verify the accuracy is not reduced when the new feature selection algorithm is applied,the test result will be compared with the classification results of the SOM neural network,fuzzy c-means,k-means and k-Nearest Neighbor algorithm on original datasets.Then,the new feature selection will be applied on the datasets of 5 wells in Jianghan oilfield to extract the key logging attributes.In the new rule extraction algorithm,“first-come,first-served” and “maximum membership” are selected as the principle to deal with the rules' conflict;the accuracy rate of classified samples of the rule set,rules' number in the rule set,the total number of rules' antecedents in the rule set,and samples' number that are not covered in the rule set are selected as the criteria of rule set;the improved differential algorithm is as the search strategy.The new rule extraction algorithm adopts IF-THEN as the rule's form,AND as connection word for the rule's antecedent,class label as the rule's consequence.Rules are encoded as individuals in population of the improved differential evolution algorithm,and each individual represents a rule set consisting of three parts: the rule's parameters(including the controls of the rule,antecedents and class labels),the control parameters(including scaling factors and crossover rates),and the fitness value.In order to test the performance of the new rule extraction algorithm,this paper will test the new rule extraction algorithm on ten benchmark datasets,and will compare its result with the result of other rule extraction algorithms.And the results of analyses show that the new rule extraction algorithm performs better than other algorithms in terms of the interpretability and the recognition rate of rule set.Then,the new rule extraction will be applied on the key logging attributes generated by the feature selection algorithm.It gives the relationships between the key logging attributes and the classification of reservoir in the form of “IF-THEN”.There are three main innovations in this paper,firstly,an enhanced self-adaptive differential evolution is proposed,namely ESADE;secondly,a feature selection algorithm which has embedded ESADE as a search strategy is presented;thirdly,a rule extraction method based on ESADE is proposed when the ESADE is applied on rule extraction.