Research On Rock Composition Based On Deep Learning

Rock composition is a complex problem,such as mineral particles,clay materials and pores,which all have its own particularities.The author of this dissertation takes the pores that are important for the exploration and development of oil and gas fields as the component research targets and analyzes the types of pores to realize the classification of intergranular pores,intercrystalline pores and dissolved pores.Rock pore identification is an important research work in geological exploration.A deep understanding of the pore structure is a key factor to improve oil recovery.The traditional identification of rock pores is mainly based on visual observation,which has problems such as long period,difficult quantitative,and low recognition efficiency.To solve this problem,this dissertation applies deep learning algorithm to the classification of pore components.The convolutional neural network and deep belief network are selected as the research objects in this dissertation.The dataset of rock casting thin slices collected from Ordos Basin are used to realize the automatic classification of rock pores.Firstly,using the sample image to adjust the network structure to obtain the most suitable convolutional neural network structure and deep belief network structure to solve the rock pore classification problem.Then the sample images were transformed into HSV and YCbCr color space,tested and compared separately.Finally,the classification performance of convolutional neural network and deep belief network in different color spaces was analyzed and compared with other classification methods.It was concluded that convolutional neural network and deep belief network performed better in solving the problem of rock pore image classification.In contrast,CNN and DBN can achieve higher accuracy in the YCbCr color space.In this dissertation,the deep learning algorithm is applied to the classification of rock pores,and its feasibility and applicability are verified by experiments.The algorithm not only overcomes the subjective influence of traditional manual recognition,but also improves the classification accuracy.The solution to the problem of analysis in this dissertation puts forward a new ideas for works about rock analysis and has certain engineering application value.