Improved SVM Applied In Water-flooded Zone Identification

At present, the various oil fields have been entered the high water cut stage and will enter tertiary recovery stage. Because of complex ground conditions, in order to improve oil recovery efficiency, and to extend the period of oil field development, the establishment of detailed geological interpretation model, exact model on the distribution of remaining oil field has become an urgent task. It is necessary to improve the accuracy of water-flooded zone identifying, to study the geological interpretation model and the distribution of remaining oil.SVM is the newest achievement in the field of machine learning. It uses the kernel function method in the data mapping process to avoid the "dimension explosion" problem. Meanwhile, support vector machine problem into a quadratic programming problem of convex set, and avoid the "local minimum" problem of neural network effectively. In this paper, the improved support vector machine for Water-flooded Zone Identifying is suggested. In this paper, I focus on improving the learning speed of support vector machines by the pre-extract support vectors algorithm and incremental learning algorithm, in the situation of large-capacity data set and improving the accuracy of multi-classification algorithm of support vector machine. The algorithms proposed are based on the fact that:the SVs is the samples close to the plane of classification. Based on this fact, I put forward the symmetric pre-fetch SVs algorithms, a SMO based on the symmetric pre-fetch SVs algorithms, an incremental learning based on symmetric pre-fetch algorithms, and an improved DAGSVM algorithm.The main idea of Symmetric pre-fetch SVs algorithms is that:first, compute the Symmetric set of the learning samples set., then find the samples that everyone of them is the nearest to one sample of the sample set, and belong to the Symmetric set, finally, the set of these samples is the approximates support vector set-Boundary vector set. This algorithm could avoid the disadvantage of the center distance ratio method that needs user to set threshold. Through two experiments, we can find that The Symmetry pre-extraction SVs algorithm can get a close approximates support vector set. On the basis of this result, the SMO based on the Symmetric pre-fetch SVs algorithms is proposed. The algorithm run on the border vector set that is get by the Symmetric pre-fetch SVs algorithms. So you can get an approximates Lagrange multipliers, and on the results, the algorithm trains of all samples, in the further, avoids the early blind training, and makes the training speed faster.When data cannot be prepared one time in practice, or amount of data is too large to be read by pc memory, incremental learning based on symmetric pre-fetch SVs algorithms is proposed in this paper. This method retains the current SVs and the current boundary vectors that has chance to become SV,discard the samples those unlikely to be a vector and combines incremental sample to form a new sample set, finally complete the incremental learning on the new sample set.SVM is essentially a binary algorithm. Improve DAGSVM algorithm is proposed in this paper to resolve the problem that SVM do not have multi- classification function. First, the algorithm judges the degree of that the input belongs to every class, then the dynamic test path is arranged by descending sequence of degree. The algorithm avoids the disadvantage of the DAGSVM algorithm that has a fixed test path for all of the samples, increases the classification accuracy. The improved DAGSVM algorithm is applied to determine the degree of Water-flooded Zone, and we get better results.