Research On Fracture Identification Methods Based On Sequence Features In CT Images Of Coal-rock

During long-term geological movements,fractures of various sizes were widely distributed inside coal-rock.These fractures are exactly where gas and groundwater are stored and transported.Underground coal mining activities will break the internal mechanical balance of the original coal and rock,which will lead to changes in the fracture network such as extension and initiation,which will cause gas and groundwater to protrude,resulting in major casualties.Therefore,it is of great significance to study the evolution of fractures when the stress field of coal-rock changes,and the accurate identification of fractures in CT images of coal-rock can provide a strong support for related research.As a natural body,coal has complex backgrounds in CT images of coal-rock,which makes it difficult to accurately identify fractures.This thesis originates from a school-selected project.Based on the sequence characteristics of fractures in the sequence coal-ock images,this paper studies the method of accurately identifying fractures.The main difficulties in identifying fractures in coal-rock images are:(1)Noises such as artifacts and impurities are serious,and the background is complicated:As a naturally occurring mineral in the geological evolution,coal-rock has various impurities in it and appears as dense noise in the images,which seriously affects the identification of coal-rock fractures.In addition,the gray value range of the fractures and coal matrix are close,and the contrast of the image is low.It is difficult to effectively remove image noise and improve image contrast using only the traditional image processing methods based on grayscale.(2)Weak boundary fractures are widespread:In the coal matrix,weak boundary fractures are often widely distributed.When underground coal mining causes changes in the stress field of coal-rock,these weak boundary fractures are often the most likely locations for geological changes such as extension and fracture.Therefore,weak boundary fractures are very important for subsequent fracture mechanics research.Accurate identification of weak boundary fractures is the focus of this study.(3)Complex fracture features:Fractures have no fixed shape characteristics,and their shape and location distribution are random.It is difficult to accurately define the coal-rock fractures with complex features,making it difficult to propose algorithms that can accurately describe the characteristics of the fractures,which limits the application of traditional image processing methods in coal-rock fracture identification.(4)It is difficult to accurately label a large number of fractures:The characteristics of coal-rock fractures are complex,and it is very difficult to accurately label weak boundary fractures in coal-rock images.This poses a great challenge for deep neural networks to learn fracture features from a small amount of labeled data.As a data-driven technology,the lack of training data will lead to inaccurate identification models.Because of these difficulties,the accurate identification of fractures in coal-rock images has great challenges.Through in-depth research,in addition to the two-dimensional features,the fractures in the sequence coal-rock images also have three-dimensional gray features and three-dimensional shape evolution features.Based on these two characteristics,this topic researches the method to realize fracture identification.The main research contents and innovations are as follows:(1)A method of fracture identification based on curve evolution:The traditional spatial filtering method is extended by using the three-dimensional gray features of the fractures to denoise and enhance coal-rock images and improve the quality of coal-rock images.A gradual change model of the fracture shape is proposed to realize preliminary prediction and identification of the fracture boundary,and then the gradient direction consistency optimization model is used to optimize the fracture boundary location to obtain a more accurate fracture location.The method is simple in calculation and can obtain good fracture identification results for small batches of coal-rocks with a single fracture.However,it cannot cope with the topology change of the fracture contour.(2)A fracture identification method based on surface evolution is proposed:Based on the level set method,the research raises the problem of two-dimensional identification to three-dimensional space to consider,and effectively responds to the topological change of the fracture boundary during the evolution.Aiming at the difficulty of arbitrary fracture shape,the prior shape of the fracture is obtained based on the progressive changes of the three-dimensional shape of the fracture.The prior shape of the fracture is used to constrain the evolution of the level set surface function,and finally the fracture identification is realized.The level set function can easily fuse various features,but it is difficult to define other features of the fracture artificially,which limits the identification effect of this method.(3)A method of convolution neural network fracture identification based on multi-scale features is proposed:Constructing a deep neural network automatically learns the fracture features from a small amount of fracture annotation data.In order to improve the feature learning ability of the fracture,two different feature extraction networks are used in parallel.The weighted feature fusion layer is used in the network to better fuse multi-scale fracture features.At the same time,data augmentation technology and transfer learning are used to alleviate the shortcomings of the small amount of fracture annotation data.The identification results of the convolutional network are further optimized based on the gradual change characteristics of the fracture shape.This method realizes the automatic learning of fracture features,but cannot directly learn the three-dimensional features of fractures.(4)A 3D convolution neural network fracture identification method based on sequence difference is proposed:Sequential coal-rock images are sequentially superimposed to form a discrete three-dimensional data field.The study uses a three-dimensional deep convolutional neural network to directly learn the three-dimensional features of fractures from this three-dimensional data field.Replacing the pooling operation with expansion convolution,while expanding the receptive field,does not increase the amount of calculation,and also plays a role in reducing the dimensionality of the image.The network uses weighted feature fusion to make the model more accurate.Based on the sequence difference method to locate the change position of the fracture boundary between adjacent images,construct a weighted cross-entropy loss function,increase the weight of the change position of the fracture boundary,except for the change position of the fracture and the rest of the pixel points in the label image.Significantly reduce the training time of the 3D deep network,speed up the network convergence speed,and improve the accuracy of the final fracture model.After in-depth research on the characteristics of fractures in serial coal-rock images,four fracture identification methods based on different theoretical foundations are proposed.These four methods are related and supported to achieve accurate identification of coal-rock fractures.