Statistics Learning Based Identification Of Water-Flooded Layers Of Oil

Identification and appraisal of water-flooded layers is one of the important facts in the description of water-controlled reservoir. The goal is to describe the reservoir more precisely, to reduce the extraction cost, and to enhance production and recovery.In order to identify and appraise the water-flooded, the petroleum technical personnel have used lots of methods, such as: micro fluorescence image analysis, geochemistry logging, geophysical well logging recognition and appraisal, but these techniques all have disadvantages like small applicable scope or high cost, so they can't be comprehensively used in the oil field.The geophysical well logging is an important method in identification and appraisal of water-flooded layers. With the obvious differences of logging response between non water-flooded and water-flooded layers,logging datum provides abundant information for the identification. In our country, logging workers have done lots of research and gained plentiful achievements in this field.For many years, lots of testing and production information has been accumulated for oil field. The petroleum technical personnel are looking for mapping relations between reservoir conditions among these data. Automatic identification of the water-flooded reservoir is the foundation of these research. The identification and appraisal is a complex, dynamic and territorial task, because water pouring and oil extraction are always in a dynamic change and different areas are sensitive to many factors. This dissertation mainly focuses on statistical techniques for automatic identification of the water-flooded layers. The dissertation has made some work in the following 4 aspects:First, neural network is one of the important statistical learning methods. An identification method for water-flooded layers is proposed, which based on normal fuzzy neural networks and image manipulation. We join the fuzzy logic analysis in neural networks and extract features of logging information using image manipulation, which has a stronger anti-noise ability.Second, another identification method is proposed,which based on signal disposal support vector machine and process neural networks. This method extracts features of logging information by B-spline transform or wavelet transform. And support vector machine has been used for experimental analysis,showing good performances. The results indicate high identification ability and strong generalization ability.Third, pattern classification algorithms should have calculation effectiveness and statistical stability (generalization ability). But the above-metioned model need firstly transform the logging data in practical application, thus the performance and generalization ability (statistical stability) tend to be influenced. Therefore, B-spline kernel has been introduced, which enhances the calculation effectiveness and statistical stability.Fourth, involving many subjects and functions, the identification system is very complicated , and its structure will change with the identification method changes. We need a flexible and extended system structure. In this dissertation, based on Agent theories, a systematic software model is proposed for water-flooded identification. It equips the system with flexibility and extension. Characterized of sociality, the system design has more human nature and can solve complex questions easily.