Graph representation learning driven electric imaging logging crack online identification method and system

Abstract

The invention discloses an electric imaging logging crack online identification method and system driven by graph representation learning, and relates to the field of geological resource exploration and mining, and the method comprises the steps: segmenting obtained real-time electric imaging logging crack data into a plurality of sequence electric imaging fragment graphs; performing binarization processing and de-noising processing on each sequence of electric imaging fragment images to obtain binarization fragment images; determining an original path diagram according to the binarization fragment diagram; according to engineering-geology priori knowledge, determining spatial features and morphological features of each node in the original path diagram and embedding corresponding nodes to obtain a reconstructed knowledge path diagram, projecting the reconstructed knowledge path diagram to a high-dimensional feature space, and calculating to obtain high-dimensional spatial distances of different segments; connecting the fragments of which the distances are lower than a threshold value to obtain a recombined meta-crack knowledge graph; and inputting the recombined element crack knowledge graph into an intelligent crack identification model driven by graph representation learning to obtain a crack type. According to the invention, on-site electric imaging-based incomplete crack intelligent on-line identification can be realized.

Description

Technical Field

[0001] The present invention relates to the technical field of geological resource exploration and exploitation, and in particular to an online identification method and system for electrical imaging logging fractures driven by graph representation learning. Background Art

[0002] The exploitation of underground energy in reservoirs plays an important role in global economic activities, including oil resources, natural gas resources, geothermal resources, etc., and it is highly dependent on the continuous development of advanced exploration technology. In this process, natural fractures, as an important medium for underground energy storage and seepage, can effectively describe and evaluate the properties of underground reservoirs. Therefore, how to accurately extract, identify and interpret the natural fractures in reservoirs is a key engineering problem.

[0003] The electrical imaging logging technology based on the difference in reservoir rock background and pore microresistance response can effectively and intuitively describe the reservoir fracture-pore situation, and has the advantages of wide borehole coverage and high vertical resolution. On this basis, the fracture detection technology that performs electrical imaging from discrete micro-electrical scanning data and further extracts and interprets the distribution of natural fractures is a key step in the quantitative description and evaluation of reservoir properties. However, unlike the ideal laboratory simulation test environment, the field electrical imaging logging data inevitably has incomplete problems, which are mainly reflected in the data missing caused by sensor failure, the strong background noise caused by the widespread existence of pores-holes-sand, and the incomplete natural fracture segments with broken and missing items. This incomplete data problem is an important factor restricting the intelligentization of electrical imaging work, resulting in the field exploration interpretation work still relying on the expert knowledge and manual identification of field engineers. In order to reduce manual dependence and improve field work efficiency, how to realize the automation of reservoir natural fracture detection process is an important issue that needs to be studied urgently.

[0004] With the advent of the artificial intelligence era, a large number of algorithms based on machine learning or deep learning have been used to try to solve the problem of crack detection automation. Academia and industry have begun to try to use some intelligent models to identify natural cracks, but there are still some limitations.

[0005] 1) Traditional computer vision-driven methods are usually limited to processing well-imaged natural fractures, i.e. complete continuous sinusoidal curves. The key reason is that edge detection and segmentation algorithms cannot recognize the physical meaning of pixels and have difficulty processing broken fractures.

[0006] 2) Similarly, when holes and cracks intersect, it is difficult for a purely visually driven model to accurately separate the two, resulting in rough model recognition accuracy.

[0007] 3) The crack area in electrical imaging is considered as a whole, ignoring the prior knowledge contained in the correlation between pixels. This knowledge contains rich geographic and engineering information and can provide key characterization for crack detection.

[0008] Based on this, there is an urgent need for a high-precision, automated method and system for rapid identification of reservoir fractures based on electrical imaging logging, which can provide an intelligent solution for complex field logging exploration and interpretation tasks. Summary of the invention

[0009] The purpose of the present invention is to provide a method and system for online identification of electrical imaging well logging fractures driven by graph representation learning, which can realize intelligent online identification of incomplete fractures based on electrical imaging on site.

[0010] To achieve the above object, the present invention provides the following solution.

[0011] In a first aspect, the present invention provides a graph representation learning-driven electrical imaging logging fracture identification method, and the graph representation learning-driven electrical imaging logging fracture identification method includes:

[0012] The acquired real-time electrical imaging logging fracture data is segmented into a plurality of first sequence electrical imaging segment images.

[0013] Each of the first sequence of electrical imaging fragment images is binarized to obtain a first binarized image, wherein the first binarized image is a plurality of binarized images containing only cracks and holes.

[0014] The first binarized image is subjected to denoising to obtain a first binarized segment map; the first binarized segment map is a plurality of images containing only natural fractures and broken natural fracture segments as research objects.

[0015] According to the first binarized fragment graph, a first original path graph is determined; the first original path graph is an image obtained by connecting the first nodes according to the open interval path connectivity; the first node is a bright pixel in the first binarized fragment graph.

[0016] Based on engineering-geological prior knowledge, the spatial characteristics and morphological characteristics of each first node in the first original path diagram are determined; the spatial characteristics are the horizontal and vertical coordinates of all first nodes in the two-dimensional space coordinate system; the morphological characteristics are the length, horizontal width, vertical width, horizontal aspect ratio, vertical aspect ratio and inclination of each first node and each first node's neighboring nodes within the circular verification threshold.

[0017] The spatial features and morphological features of each first node are embedded into the corresponding first node to obtain a first reconstructed knowledge path graph.

[0018] The first reconstructed knowledge path graph is projected into a high-dimensional feature space, and the high-dimensional space distances of the knowledge path graphs corresponding to different fragments are calculated.

[0019] The segments of the knowledge path graph whose high-dimensional space distance is lower than a first preset threshold are connected to obtain a first reorganized meta-crack knowledge graph.

[0020] The first reorganized meta-fracture knowledge graph is input into a graph representation learning-driven fracture intelligent identification model to obtain fracture types; the graph representation learning-driven fracture intelligent identification model is a model established based on a graph convolutional neural network.

[0021] In a second aspect, the present invention provides a graph representation learning-driven electrical imaging well logging fracture identification system, and the graph representation learning-driven electrical imaging well logging fracture identification system includes.

[0022] The segmentation module is used to segment the acquired real-time electrical imaging logging fracture data into a plurality of first sequence electrical imaging segment images.

[0023] The binarization processing module is used to perform binarization processing on each of the first sequence of electrical imaging fragment images to obtain a first binarized image, wherein the first binarized image is a plurality of binarized images containing only cracks and holes.

[0024] The denoising processing module is used to perform denoising on the image after the first binarization processing to obtain a first binarization segment image; the first binarization segment image is a plurality of images containing only natural fractures and broken natural fracture segments as research objects.

[0025] The first original path map determination module is used to determine the first original path map according to the first binarized fragment map; the first original path map is an image obtained by connecting the first nodes according to the open interval path connectivity; the first node is a bright pixel in the first binarized fragment map.

[0026] The spatial feature and morphological feature determination module is used to determine the spatial features and morphological features of each first node in the first original path diagram based on engineering-geological prior knowledge; the spatial features are the horizontal and vertical coordinates of all first nodes in the two-dimensional spatial coordinate system; the morphological features are the length, horizontal width, vertical width, horizontal aspect ratio, vertical aspect ratio and inclination of each first node and each first node's neighboring nodes within the circular verification threshold.

[0027] The embedding module is used to embed the spatial features and morphological features of each first node into the corresponding first node to obtain a first reconstructed knowledge path graph.

[0028] A projection module is used to project the first reconstructed knowledge path graph into a high-dimensional feature space and calculate the high-dimensional space distance between the knowledge path graphs corresponding to different fragments.

[0029] The fragment connection module is used to connect the fragments of the knowledge path graph whose high-dimensional space distance is lower than a first preset threshold to obtain a first recombined meta-crack knowledge graph.

[0030] The crack identification module is used to input the first reorganized meta-crack knowledge graph into a crack intelligent identification model driven by graph representation learning to obtain crack types; the crack intelligent identification model driven by graph representation learning is a model established based on a graph convolutional neural network.

[0031] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0032] The present invention provides a method and system for online identification of electrical imaging logging fractures driven by graph representation learning. First, the acquired real-time electrical imaging logging fracture data is divided into a plurality of first-sequence electrical imaging fragment graphs; each of the first-sequence electrical imaging fragment graphs is binarized to obtain a first binarized image, which is a plurality of binarized images containing only fractures-holes; the first binarized image is denoised to obtain a first binarized fragment graph; the first binarized fragment graph is a plurality of images containing only natural fractures and broken natural fracture segments; secondly, according to the first binarized fragment graph, a first original path graph is determined; the first original path graph is an image obtained by connecting the first nodes according to the path connectivity between open intervals; the first node is a bright pixel in the first binarized fragment graph; further, according to engineering-geological prior knowledge, the first original path graph is determined. The spatial characteristics and morphological characteristics of each first node; the spatial characteristics are the horizontal and vertical coordinates of all first nodes in the two-dimensional space coordinate system; the morphological characteristics are the length, horizontal width, vertical width, horizontal aspect ratio, vertical aspect ratio and inclination of each first node and each first node's neighboring node within the circular verification threshold; the spatial characteristics and morphological characteristics of each first node are embedded in the corresponding first node to obtain a first reconstructed knowledge path graph; then, the first reconstructed knowledge path graph is projected into a high-dimensional feature space, and the high-dimensional space distance of the knowledge path graph corresponding to different fragments is calculated; the fragments whose high-dimensional space distance of the knowledge path graph is lower than the first preset threshold are connected to obtain a first reorganized meta-fracture knowledge graph; finally, the first reorganized meta-fracture knowledge graph is input into a fracture intelligent identification model driven by graph representation learning to obtain fracture types; the fracture intelligent identification model driven by graph representation learning is a model established based on a graph convolutional neural network. The present invention uses a graph representation learning method to empower the geological reservoir exploration process, embeds a priori knowledge in the geological field to construct a graph, and can achieve efficient identification of oil and gas reservoir fractures without relying on on-site manual work. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0034] Figure 1 A schematic flow chart of a method for identifying electrical imaging well logging fractures driven by graph representation learning is provided in accordance with an embodiment of the present invention.

[0035] Figure 2A schematic flow chart of an online intelligent identification method for reservoir electrical imaging logging fractures embedded with a priori knowledge graph representation provided by one embodiment of the present invention.

[0036] Figure 3 A schematic diagram of a micro-electric scanning imaging logging system provided in one embodiment of the present invention.

[0037] Figure 4 A schematic diagram of a segmented on-site microresistance scanning imaging logging image provided by an embodiment of the present invention.

[0038] Figure 5 A schematic diagram of graph structure data provided by an embodiment of the present invention.

[0039] Figure 6 A schematic diagram of converting electrical imaging data into a crack path diagram according to an embodiment of the present invention.

[0040] Figure 7 A schematic diagram of the construction and training process of a crack recognition model based on a graph convolutional neural network provided in one embodiment of the present invention.

[0041] Figure 8 A schematic diagram of storage fracture identification results of electrical imaging logging provided by an embodiment of the present invention.

[0042] Fig. 9 A schematic diagram of comparative test results of multiple methods-models in a data set (a) provided in an embodiment of the present invention.

[0043] Fig.10 A schematic diagram of comparative test results of multiple methods-models in data set (b) provided in one embodiment of the present invention.

[0044] Fig.11 A schematic diagram of comparative test results of setting a hole length noise reduction threshold L provided in an embodiment of the present invention.

[0045] Fig.12 A schematic diagram of functional modules of a graph representation learning-driven electrical imaging logging fracture identification system provided in one embodiment of the present invention. DETAILED DESCRIPTION

[0046] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0047] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0048] The purpose of the present invention is to solve the problems in the prior art of poor accuracy of incomplete fracture identification caused by reliance on manual experience annotation for electrical imaging logging interpretation and lack of prior knowledge guidance in the model.

[0049] To achieve the above objectives, the present invention provides a method and system for online identification of electrical imaging logging fractures driven by graph representation learning, which uses graph representation learning methods to empower the geological reservoir exploration process, embeds prior knowledge in the geological field to construct a map, and realizes efficient identification of oil and gas reservoir fractures without relying on on-site manual work. At the same time, by calculating the spatial characteristics and morphological characteristics of the fractures, the natural fracture segments are accurately identified and matched, and the on-site intelligent online identification of incomplete fractures based on electrical imaging is realized. The method has a simple structure, is easy to operate, and has a low implementation cost.

[0050] In an exemplary embodiment, Figure 1 and Figure 2 As shown, a graph representation learning-driven electrical imaging logging fracture identification method is provided, and the graph representation learning-driven electrical imaging logging fracture identification method includes.

[0051] S1: Segment the acquired real-time electrical imaging logging fracture data into a plurality of first sequence electrical imaging segment images.

[0052] S2: performing binarization processing on each of the first sequence of electrical imaging fragment images to obtain a first binarized image, wherein the first binarized image is a plurality of binarized images containing only cracks and holes.

[0053] S3: performing denoising processing on the image after the first binarization processing to obtain a first binarization segment image; the first binarization segment image is a plurality of images containing only natural fractures and broken natural fracture segments as research objects.

[0054] S4: Determine a first original path graph according to the first binarized segment graph; the first original path graph is an image obtained by connecting the first nodes according to the open interval path connectivity; the first node is a bright pixel in the first binarized segment graph.

[0055] S5: According to engineering-geological prior knowledge, determine the spatial characteristics and morphological characteristics of each first node in the first original path diagram; the spatial characteristics are the horizontal and vertical coordinates of all first nodes in the two-dimensional space coordinate system; the morphological characteristics are the length, horizontal width, vertical width, horizontal aspect ratio, vertical aspect ratio and inclination of each first node and each first node's neighboring nodes within the circular verification threshold.

[0056] S6: Embed the spatial features and morphological features of each first node into the corresponding first node to obtain a first reconstructed knowledge path graph.

[0057] S7: Project the first reconstructed knowledge path graph into a high-dimensional feature space, and calculate the high-dimensional space distance between the knowledge path graphs corresponding to different fragments.

[0058] S8: Connect the fragments of the knowledge path graph whose high-dimensional space distance is lower than a first preset threshold to obtain a first reorganized meta-crack knowledge graph.

[0059] S9: Inputting the first recombined meta-fracture knowledge graph into a graph representation learning-driven fracture intelligent identification model to obtain fracture types; the graph representation learning-driven fracture intelligent identification model is a model established based on a graph convolutional neural network.

[0060] By implementing the above-mentioned steps S1 to S9, the present invention uses a graph representation learning method to empower the geological reservoir exploration process, embeds prior knowledge composition in the geological field, and can achieve efficient identification of oil and gas reservoir fractures without relying on on-site manual work.

[0061] As an optional implementation, in step S1, according to the instrument model and the on-site geological characteristics, a rectangular sliding window of a certain specification is used to segment the acquired continuous long-segment real-time electrical imaging logging data into a plurality of first sequence electrical imaging segment images.

[0062] As an optional implementation, in step S2, the original image is binarized according to the difference in microresistivity responses of the on-site rock formations and holes, and the fracture-hole objects to be studied are extracted from the formation background.

[0063] Specifically, Figure 3 As shown, the electrical imaging logging fracture data used in this embodiment is collected by field logging of a micro-resistance scanning imaging system. The micro-resistance scanning imaging system includes two parts: a micro-electric scanning logging device and a logging interpretation system. The micro-electric scanning logging device is used to continuously advance in the formation and record the micro-current value flowing in the rock formation-fracture. The logging interpretation system converts the measured discrete values ​​into interpretable electrical imaging data according to the current-resistance formula and the wellbore response. The instrument used is the micro-resistance imager MCI6575, and the key parameters of the instrument are shown in Table 1 below.

[0064] Table 1 Key parameters of microresistivity imager MCI6575

[0065]

[0066] In this example, the electrical imaging logging data with a depth of 3066-3723 was selected as the main analysis object, and the data of two other wells with similar depths and geological conditions were supplemented to test the generalization ability of the proposed method. Furthermore, a sliding window with a size of 360*360 pixels was used to segment the original long electrical imaging logging image into multiple sequentially spliced ​​sub-images. Through interlaced cutting and image rotation, 2112 field images were obtained for experimental verification. The obtained electrical imaging logging segmentation sub-images are shown in Figure 1. Figure 4 shown.

[0067] Specifically, relying on the prior knowledge of formation microresistance logging, combined with laboratory experimental rules and field engineering experience, the binary threshold that can accurately distinguish the formation background and the fracture-hole research object is determined according to the different resistivity responses caused by the formation-pore water content difference. The area with resistivity lower than the threshold is converted into a white pixel block, and vice versa, into a black pixel block, to obtain a binary image containing only fractures-holes. This embodiment uses the binarization algorithm encapsulated in the Python-Opencv module, sets the image light sensitivity coefficient to 25, and strips the fractures from the background rock layer to obtain a binary logging image containing fractures-holes.

[0068] As an optional implementation, in step S3, it specifically includes.

[0069] S31: using an edge detection algorithm to calculate all open interval connected domains in the first binarized image to obtain first labeled candidate individual objects; the first labeled candidate individual objects are a number of labeled candidate crack-hole-silt open interval individual objects.

[0070] S32: Calculate the basic morphological features of the first marked candidate individual object; the basic morphological features include: length, longitudinal width, transverse width, longitudinal aspect ratio and transverse aspect ratio.

[0071] S33: According to geological prior knowledge, on-site logging conditions and the basic morphological characteristics, a first length threshold is set to remove hole-sediment noise to obtain a first binarized segment map.

[0072] In step S3, an edge detection algorithm is used to calculate the connected domains of all open intervals in the fragment graph to obtain marked candidate fracture-pore-sediment individual objects; further, the holes and sediment in the fragment graph are removed by screening and denoising through a length threshold L to obtain a binary fragment graph containing only research objects such as natural fractures and broken natural fracture segments.

[0073] Specifically, five parameters, namely, length, longitudinal width, transverse width, longitudinal aspect ratio and transverse aspect ratio, of all candidate fracture-hole-sediment open intervals are calculated to describe their basic morphological characteristics. Among them, the maximum path length of all nodes in the open interval is traversed, and the maximum value is taken as the length of the open interval; the transverse open length of all nodes is traversed and averaged as the transverse width of the open interval; the longitudinal open length of all nodes is traversed and averaged as the longitudinal width of the interval; further, the ratio of length to transverse width is recorded as the transverse aspect ratio, and the ratio of length to longitudinal width is recorded as the longitudinal aspect ratio. Finally, according to geological prior knowledge and field logging conditions, the threshold L=70 is set in combination with the parameters of each open interval to remove hole-sediment noise.

[0074] In an exemplary embodiment, step S4 specifically includes:

[0075] S41: Projecting the first binarized segment image into a two-dimensional space coordinate system of the same size to obtain the space coordinates of each first pixel point.

[0076] S42: Determine a first original path map according to the spatial coordinates of each of the first pixel points.

[0077] Specifically, the first binary fragment graph is projected into a two-dimensional space coordinate system of the same size to obtain the spatial coordinates of each pixel point; then, according to the graph representation learning theory, the bright pixels in the fragment graph are regarded as nodes, and the nodes are connected according to the open interval path connectivity to form a first original path graph, that is, a graph structure representation form of the natural fracture segment; and according to the spatial correlation degree calculated by Euclidean distance, edge connections are created between the nodes, and the open intervals of the alternative fracture segment path are converted into a path graph matching the original form.

[0078] The process of determining the neighboring nodes of a node is shown in the following formula.

[0079]

[0080] Among them, D(v i ,v j ) represents node v i and node v j The Euclidean distance in rectangular coordinates, ψ(V p ) represents node v p The set of k minimum distance neighbor nodes of Indicates v p A neighbor node of dist-k It means that it contains a set of k minimum distance values, (x i ,y i ) is the node v i The coordinates of (x j ,y j) is the node v j The coordinates of .

[0081] Furthermore, the spatial and morphological characteristics of the nodes are calculated based on engineering-geological prior knowledge. The graph structure data conversion form is as follows: Figure 5 As shown in the figure, the pixel graph structure conversion process of the connected domain open interval is as follows Figure 6 shown.

[0082] Specifically, the binary fragment graph is projected into a two-dimensional space coordinate system of the same size to obtain the spatial coordinates of each pixel. Furthermore, according to the graph representation learning theory, the pixel points are regarded as nodes, and edge connections are created between the nodes according to the spatial correlation degree calculated by the Euclidean distance, and the corresponding open intervals of the alternative crack fragment path are converted into a path graph that matches the original form. Among them, the process of determining the neighboring nodes of the node is shown in the following formula:

[0083] Furthermore, edge connections are created between adjacent nodes. The calculation and storage of edge connections between nodes can be expressed as follows:

[0084]

[0085] Among them, e(v i ,v j ) represents node v i and node v j There are edge connections between them, and the matrix A adj It is the adjacency matrix of graph structure data, which is used to record the edge connection relationship between nodes and participate in graph convolution calculation.

[0086] In step S5, the horizontal and vertical coordinates of all nodes in the graph in the two-dimensional space coordinate system are traversed and recorded as the graph space feature set; the lengths, horizontal widths, vertical widths, horizontal aspect ratios, vertical aspect ratios and inclinations of all nodes in the graph and their neighboring nodes within the circular verification threshold are traversed as node morphological features, and the calculation method is the same as that described in step S3 and will not be repeated here.

[0087] In another exemplary embodiment of the present invention, in step S6, the spatial features and morphological features calculated by prior knowledge are embedded into the nodes to obtain a first reconstructed knowledge path graph;

[0088] Furthermore, in step S7, the similarity of each knowledge path graph in the fragment graph in the high-dimensional feature space is calculated, and in step S8, a first preset threshold is set to match the similarity graph to combine the natural fracture segments to obtain a first recombined meta-fracture knowledge graph.

[0089] Specifically, each node and its neighboring nodes in the verification domain are calculated, and the morphological features and spatial features determined by prior knowledge and calculated together are embedded as node features to obtain a knowledge embedding path graph containing the spatial and morphological knowledge of neighboring nodes. Specifically, the knowledge path graph containing spatial features and morphological features is projected into a high-dimensional feature space, and the high-dimensional space distance of the knowledge path graph corresponding to different fragments is calculated. Among them, the high-dimensional space distance of the knowledge path graph is measured by the Mahalanobis distance function, and the formula is as follows:

[0090]

[0091] Among them, Dist M (G i ,G j ) is the high-dimensional space distance between the knowledge path graph of the readout vector of graph i and the readout vector of graph j, G i is the readout vector of graph i, G j is the readout vector of graph i, ∑ -1 is the covariance matrix of the overall distribution of data in the high-dimensional data space.

[0092] Furthermore, in step S9, the fragments with similar distances are connected to obtain a first reorganized meta-crack knowledge graph corresponding to the complete original crack.

[0093] In another exemplary embodiment of the present invention, in step S9, a graph representation learning-driven fracture intelligent identification model based on a graph convolutional neural network is built, and a meta-fracture knowledge graph training model converted from historical electrical imaging logging fracture data collected at the exploration site is used to obtain a trained graph representation learning-driven fracture intelligent identification model.

[0094] The determination process of the graph representation learning-driven crack intelligent identification model is as follows.

[0095] S91: Segment the acquired historical electrical imaging logging fracture data into a plurality of second sequence electrical imaging segment images.

[0096] S92: performing binarization processing on each of the second sequence of electrical imaging fragment images to obtain a second binarized image, wherein the second binarized image is a plurality of binarized images containing only cracks and holes.

[0097] S93: performing denoising processing on the second binarized image to obtain a second binarized segment image; the second binarized segment image is a plurality of images containing only natural fractures and broken natural fracture segments as research objects.

[0098] S94: Determine a second original path graph according to the second binarized segment graph; the second original path graph is an image obtained by connecting the second nodes according to the open interval path connectivity; the second nodes are bright pixels in the second binarized segment graph.

[0099] S95: Determine the spatial characteristics and morphological characteristics of each second node in the second original path diagram based on engineering-geological prior knowledge; the spatial characteristics are the horizontal and vertical coordinates of all second nodes in the two-dimensional space coordinate system; the morphological characteristics are the length, horizontal width, vertical width, horizontal aspect ratio, vertical aspect ratio and inclination of each second node and each second node's neighboring nodes within the circular verification threshold.

[0100] S96: Embed the spatial features and morphological features of each second node into the corresponding second node to obtain a second reconstructed knowledge path graph.

[0101] S97: Project the second reconstructed knowledge path graph into a high-dimensional feature space, and calculate the high-dimensional space distance between the knowledge path graphs corresponding to different fragments.

[0102] S98: Connect the fragments of the knowledge path graph whose high-dimensional space distance is lower than a second preset threshold to obtain a second reorganized meta-crack knowledge graph.

[0103] S99: Taking the second reorganized meta-fracture knowledge graph as input and the fracture types corresponding to the second reorganized meta-fracture knowledge graph as output, the graph convolutional neural network is trained to obtain a fracture intelligent recognition model driven by graph representation learning.

[0104] Specifically, the graph convolutional neural network model is composed of three layers of Chebyshev graph convolutional neural network (GCN) and two layers of fully connected layers (FC_Layer) in sequence; the model uses leaky rectifier unit (LeakyRelu) as the activation function between the networks, and uses cross entropy loss (CE_Loss) loss function as the training function of the model. The model receives graph structure data and activates the dropout mechanism to suppress possible overfitting trends. Among them, the Chebyshev graph convolution operation is expressed as:

[0105]

[0106] L=In-D -1 / 2 WD -1 / 2 (7).

[0107]

[0108] Among them, D i,i is a degree matrix containing only diagonal elements, A i,jis the adjacency matrix of the graph, L is the Laplacian matrix, In is the identity matrix, W is the computable weight matrix, F∈R n×m is a graph signal, representing the input of the Chebyshev convolution operation, represents the output of the Chebyshev convolution operation, U is the eigenvector matrix of the feature quantity in L, K represents the highest order of the Chebyshev polynomial, k represents the order of the Chebyshev polynomial, represents the feature matrix, represents the K-order Chebyshev polynomial, Represents the Chebyshev coefficient vector. At this time, the output Y∈R of the ChebGCN convolutional layer n×s It is expressed as:

[0109]

[0110] Where W∈R m×s Represents the weight matrix learned by the ChebGCN convolutional layer.

[0111] On this basis, the constructed graph convolution model is as follows:

[0112]

[0113] Among them, W( 1 )∈R m×l is the trainable weight matrix of the first graph convolutional layer, W (2) ∈R l×v is the trainable weight matrix of the second graph convolutional layer, W( 3 )∈R v×g is the trainable weight matrix of the third graph convolution layer, Cheb(·) is the Chebyshev graph convolution function, m is the scale of the input vector, l is the first scale of the output vector, v is the second scale of the output vector, g is the third scale of the output vector, Y( 1 )∈R n×l represents the output of the first convolutional layer, Y (2) ∈R n×v represents the output of the second convolutional layer, Y( 3 )∈R n×g represents the output of the third convolutional layer, σ represents the leaky rectifier unit (LeakyRelu) activation function; the model structure and parameter details are shown in Table 2.

[0114] Table 2 Structure and parameter details of the crack recognition model based on graph convolutional neural network

[0115]

[0116] Specifically, the mathematical expression of the activation function LeakyReLU described in this embodiment is shown in formula (11), and the mathematical expression of the CE_Loss loss function is shown in formula (12).

[0117]

[0118] Wherein, LeakyReLU(x) is the activation function, x is the input, α is a constant, CE_Loss(p,q) is the loss function, C is the total number of categories, in this embodiment, C=4, p i is the true value of the sample, q i The predicted value output by the model.

[0119] As an optional implementation, in step S93, it specifically includes.

[0120] S931: Calculate all connected domains of open intervals in the second binarized image using an edge detection algorithm to obtain second labeled candidate individual objects; the second labeled candidate individual objects are a number of labeled candidate crack-hole-silt open interval individual objects.

[0121] S932: Calculate the basic morphological features of the second marked candidate individual object; the basic morphological features include: length, longitudinal width, transverse width, longitudinal aspect ratio and transverse aspect ratio.

[0122] S933: According to geological prior knowledge, on-site logging conditions and the basic morphological characteristics, a second length threshold is set to remove hole-sediment noise to obtain a second binarized segment map.

[0123] As an optional implementation, in step S94, it specifically includes.

[0124] S941: Project the second binarized segment image into a two-dimensional space coordinate system of the same size to obtain the space coordinates of each second pixel point.

[0125] S942: Determine a second original path map according to the spatial coordinates of each of the second pixel points.

[0126] In an exemplary embodiment, the real-time electrical imaging data collected by the microresistance logging instrument is converted into the corresponding meta-fracture knowledge graph according to the above steps, and these graphs are sequentially sent to the trained graph representation learning-driven fracture intelligent identification model for fracture classification and identification, so as to realize online fracture intelligent identification as the instrument advances. The model training process driven by graph representation learning is as follows: Figure 7 shown.

[0127] Specifically, the electrical imaging data measured by multiple wells on site are collected for model training to obtain a trained fracture recognition model based on graph convolutional neural network. Furthermore, the online measured electrical imaging data is collected and cut using a sliding time window, and converted into a knowledge graph structure according to the above steps. These graphs are sequentially sent to the trained fracture intelligent recognition model for fracture classification and recognition, thus realizing online fracture intelligent recognition as the instrument advances. The fracture recognition results are shown in Figure 2. Figure 8 shown.

[0128] Specifically, in order to verify the performance of the fracture identification method provided in this embodiment, we selected some advanced similar methods for comparative experiments in the electrical imaging data collected from two wells, including edge detection method (EDM), path morphology method (PMM), artificial neural network (ANN) and mask regional convolutional neural network (M-RCNN). The average of ten results was taken as the final effect of each method for comparison, and the model performance was evaluated using indicators including detection rate, accuracy, F1 score and time consumption, and their formulas are shown below.

[0129]

[0130] Among them, TP is a positive sample predicted to be positive, TN is a negative sample predicted to be positive, FP is a positive sample predicted to be negative, FN is a negative sample predicted to be negative, P is precision (Precision, P), R is recall (Recall, R), and the F1 score is a comprehensive performance indicator.

[0131] Fig. 9 and Fig.10The results show that in the field logging data set (a), all models have achieved good recognition results. At the same time, affected by the low quality of the image in the data set (b), the effects of all models involved in the evaluation have declined to varying degrees. This phenomenon shows that although the traditional machine learning-deep learning driven fracture recognition model can generalize and process heterogeneous logging data, it will be directly affected by the quality of the original data. Among them, the EDM based on computer vision is the most affected, because the feature extraction without prior knowledge makes it unable to distinguish fractures from noise. To address this problem, the present invention minimizes the impact of data migration by capturing more detailed spatial features, because the node feature embedding based on prior knowledge is suitable for representing various natural fractures at any location. However, in addition to the morphological-based method that requires the use of multiple operators to calculate all open interval connected domains for each node in the fracture path, the model proposed by this method consumes the most time. It can be said that graph representation learning is to mine additional features inside the object through additional computing resource consumption, thereby improving the task accuracy without relying on more input information.

[0132] The hyperparameters are verified below.

[0133] In an exemplary embodiment, the hole length noise reduction threshold L is set as: Fig.11 As shown in the figure, the performance curve of the proposed method shows a trend of first rising and then falling, and the best performance is shown at L = 70, that is, 96.37% accuracy and all 2451 edges. This shows that a smaller length L is difficult to remove larger noise holes, while a larger length L will mistakenly remove smaller crack fragments. It is worth noting that Fig.11 The slope of the edge number curve is large in the early stage of medium threshold growth (less than 70), and the slope gradually slows down in the later stage. This shows that the small L value in the early stage quickly identifies and deletes a large number of small holes, while the large L value in the later stage deletes a smaller number of large holes and small holes.

[0134] Setting the order of Chebyshev graph convolution kernel K: For the Chebyshev graph convolutional neural network model, the selection of Chebyshev polynomial K is particularly critical. It represents a convolutional network that aggregates information from a node's K-hop neighbor nodes. Too small a K value will limit the network's ability to mine information from graph structure data, while too large a K value will multiply the consumption of computing resources and introduce unnecessary graph computing noise. In order to fully tap the potential of the model, a comparative experiment was designed to determine the optimal K value. The experimental results are shown in Table 3.

[0135] Table 3 Comparative experimental results under different graph convolution parameter K orders

[0136]

[0137] As shown in Table 3, the fitting accuracy of the model constructed by the present invention at different orders of K values ​​is ideal, and when K=2, the fitting accuracy is the lowest. At the same time, when K=3 and 4, the model is close to the best experimental results in both indicators (accuracy is 96.37%~95.04%, F1 score is 94.25%~95.04). Subsequently, further increase in the K value will lead to loss of accuracy and increased consumption of computing resources. This is because aggregating related nodes for learning can enhance the model effect. On the contrary, too much node aggregation will bring a lot of noise information, reduce the accuracy, and increase the consumption of computing resources. Furthermore, since most of the valid node information has been aggregated, continuing to increase the K value will cause irrelevant nodes to be incorrectly aggregated, thereby reducing the model accuracy. Therefore, the present invention selects K=4 as the optimal hyperparameter of the network layer.

[0138] Based on the same inventive concept, the embodiment of the present invention also provides a graph representation learning driven electrical imaging logging fracture identification system for implementing the graph representation learning driven electrical imaging logging fracture identification method involved above. The implementation scheme for solving the problem provided by the system is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more graph representation learning driven electrical imaging logging fracture identification system embodiments provided below can refer to the limitations of the graph representation learning driven electrical imaging logging fracture identification method above, and will not be repeated here.

[0139] In an exemplary embodiment, Fig.12 As shown, a graph representation learning-driven electrical imaging logging fracture identification system is provided, and the graph representation learning-driven electrical imaging logging fracture identification system includes.

[0140] The segmentation module M1 is used to segment the acquired real-time electrical imaging logging fracture data into a plurality of first sequence electrical imaging segment images.

[0141] The binarization processing module M2 is used to perform binarization processing on each of the first sequence of electrical imaging fragment images to obtain a first binarized image, wherein the first binarized image is a plurality of binarized images containing only cracks and holes.

[0142] The denoising processing module M3 is used to perform denoising on the image after the first binarization processing to obtain a first binarization segment image; the first binarization segment image is a plurality of images containing only natural fractures and broken natural fracture segments as research objects.

[0143] The first original path map determining module M4 is used to determine the first original path map according to the first binarized fragment map; the first original path map is an image obtained by connecting the first nodes according to the open interval path connectivity; the first node is a bright pixel in the first binarized fragment map.

[0144] The spatial feature and morphological feature determination module M5 is used to determine the spatial features and morphological features of each first node in the first original path diagram based on engineering-geological prior knowledge; the spatial features are the horizontal and vertical coordinates of all first nodes in the two-dimensional spatial coordinate system; the morphological features are the length, horizontal width, vertical width, horizontal aspect ratio, vertical aspect ratio and inclination of each first node and each first node's neighboring nodes within the circular verification threshold.

[0145] The embedding module M6 is used to embed the spatial features and morphological features of each first node into the corresponding first node to obtain a first reconstructed knowledge path graph.

[0146] The projection module M7 is used to project the first reconstructed knowledge path graph into a high-dimensional feature space and calculate the high-dimensional space distance between the knowledge path graphs corresponding to different fragments.

[0147] The fragment connection module M8 is used to connect the fragments of the knowledge path graph whose high-dimensional space distance is lower than a first preset threshold to obtain a first recombined meta-crack knowledge graph.

[0148] The crack identification module M9 is used to input the first reorganized meta-crack knowledge graph into a crack intelligent identification model driven by graph representation learning to obtain crack types; the crack intelligent identification model driven by graph representation learning is a model established based on a graph convolutional neural network.

[0149] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0150] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A method for electrical imaging logging fracture identification driven by graph representation learning, characterized in that: The electrical imaging logging fracture identification method driven by graph representation learning includes: Segmenting the acquired real-time electrical imaging logging fracture data into a plurality of first sequence electrical imaging segment images; Binarization is performed on each of the first sequence of electrical imaging fragment images to obtain a first binarized image, wherein the first binarized image is a plurality of binarized images containing only cracks and holes; De-noising the image after the first binarization process to obtain a first binarization segment map; the first binarization segment map is a plurality of images containing only natural fractures and broken natural fracture segments as research objects; Determine a first original path graph according to the first binarized segment graph; the first original path graph is an image obtained by connecting the first nodes according to the open interval path connectivity; the first node is a bright pixel in the first binarized segment graph; According to engineering-geological prior knowledge, the spatial characteristics and morphological characteristics of each first node in the first original path diagram are determined; the spatial characteristics are the horizontal and vertical coordinates of all first nodes in the two-dimensional space coordinate system; the morphological characteristics are the length, horizontal width, vertical width, horizontal aspect ratio, vertical aspect ratio and inclination of each first node and each first node's neighboring nodes within the circular verification threshold; Embed the spatial features and morphological features of each first node into the corresponding first node to obtain a first reconstructed knowledge path graph; Projecting the first reconstructed knowledge path graph into a high-dimensional feature space, and calculating the high-dimensional space distances of the knowledge path graphs corresponding to different fragments; Connecting the fragments of the knowledge path graph whose high-dimensional space distance is lower than a first preset threshold to obtain a first reorganized meta-crack knowledge graph; The first reorganized meta-fracture knowledge graph is input into a graph representation learning-driven fracture intelligent identification model to obtain fracture types; the graph representation learning-driven fracture intelligent identification model is a model established based on a graph convolutional neural network.

2. The electrical imaging logging fracture identification method driven by graph representation learning according to claim 1 is characterized in that: The first binarized image is subjected to denoising to obtain a first binarized segment image, which specifically includes: An edge detection algorithm is used to calculate all open interval connected domains in the first binarized image to obtain first labeled candidate individual objects; the first labeled candidate individual objects are a number of labeled candidate crack-hole-silt open interval individual objects; Calculating the basic morphological features of the first marked candidate individual object; the basic morphological features include: length, longitudinal width, transverse width, longitudinal aspect ratio and transverse aspect ratio; According to geological prior knowledge, on-site logging conditions and the basic morphological characteristics, a first length threshold is set to remove hole-sediment noise to obtain a first binarized fragment map.

3. The electrical imaging logging fracture identification method driven by graph representation learning according to claim 1 is characterized in that: Determining a first original path graph according to the first binarized segment graph specifically includes: Projecting the first binarized segment image into a two-dimensional space coordinate system of the same size to obtain the space coordinates of each first pixel point; A first original path map is determined according to the spatial coordinates of each of the first pixel points.

4. The electrical imaging logging fracture identification method driven by graph representation learning according to claim 1 is characterized in that: The calculation formula of the high-dimensional space distance of the knowledge path graph is: Among them, Dist M (G i ,G j ) is the high-dimensional space distance between the knowledge path graph of the readout vector of graph i and the readout vector of graph j, G i is the readout vector of graph i, G j is the readout vector of graph i, ∑ -1 is the covariance matrix of the overall distribution of data in the high-dimensional data space.

5. The electrical imaging logging fracture identification method driven by graph representation learning according to claim 1 is characterized in that: The determination process of the graph representation learning-driven crack intelligent identification model is as follows: Segmenting the acquired historical electrical imaging logging fracture data into a plurality of second sequence electrical imaging segment images; Binarization is performed on each of the second sequence of electrical imaging fragment images to obtain a second binarized image, wherein the second binarized image is a plurality of binarized images containing only cracks and holes; De-noising the image after the second binarization process to obtain a second binarization segment map; the second binarization segment map is a plurality of images containing only natural fractures and broken natural fracture segments as research objects; According to the second binarized fragment graph, a second original path graph is determined; the second original path graph is an image obtained by connecting the second nodes according to the open interval path connectivity; the second nodes are bright pixels in the second binarized fragment graph; According to engineering-geological prior knowledge, the spatial characteristics and morphological characteristics of each second node in the second original path diagram are determined; the spatial characteristics are the horizontal and vertical coordinates of all second nodes in the two-dimensional space coordinate system; the morphological characteristics are the length, horizontal width, vertical width, horizontal aspect ratio, vertical aspect ratio and inclination of each second node and each second node's neighboring nodes within the circular verification threshold; Embed the spatial features and morphological features of each second node into the corresponding second node to obtain a second reconstructed knowledge path graph; Projecting the second reconstructed knowledge path graph into a high-dimensional feature space, and calculating the high-dimensional space distances of the knowledge path graphs corresponding to different fragments; Connecting the fragments of the knowledge path graph whose high-dimensional space distance is lower than a second preset threshold to obtain a second reorganized meta-crack knowledge graph; Taking the second reorganized meta-fracture knowledge graph as input and the fracture types corresponding to the second reorganized meta-fracture knowledge graph as output, the graph convolutional neural network is trained to obtain a fracture intelligent recognition model driven by graph representation learning.

6. The electrical imaging logging fracture identification method driven by graph representation learning according to claim 5 is characterized in that: The second binarized image is subjected to denoising to obtain a second binarized segment image, specifically comprising: Using an edge detection algorithm to calculate all open interval connected domains in the second binarized image, to obtain second labeled candidate individual objects; the second labeled candidate individual objects are a number of labeled candidate crack-hole-silt open interval individual objects; Calculating the basic morphological features of the second marked candidate individual object; the basic morphological features include: length, longitudinal width, transverse width, longitudinal aspect ratio and transverse aspect ratio; According to geological prior knowledge, on-site logging conditions and the basic morphological characteristics, a second length threshold is set to remove hole-sediment noise to obtain a second binarized fragment map.

7. The electrical imaging logging fracture identification method driven by graph representation learning according to claim 5 is characterized in that: Determining a second original path graph according to the second binarized segment graph specifically includes: Projecting the second binarized segment image into a two-dimensional space coordinate system of the same size to obtain the space coordinates of each second pixel point; A second original path map is determined according to the spatial coordinates of each of the second pixel points.

8. The electrical imaging logging fracture identification method driven by graph representation learning according to claim 5 is characterized in that: The graph convolutional neural network is composed of three layers of Chebyshev graph convolutional neural network and two layers of fully connected layers, and a cross entropy loss function is used as a training function of the graph convolutional neural network; wherein the kernel K of the Chebyshev graph convolutional network is set to 4; and a leaky rectifier unit is used as an activation function between the Chebyshev graph convolutional neural networks.

9. The electrical imaging logging fracture identification method driven by graph representation learning according to claim 8 is characterized in that: The expression of the activation function is: The expression of the loss function is: Among them, LeakyReLU(x) is the activation function, x is the input, α is a constant, CE_Loss(p,q) is the loss function, C is the total number of categories, p i is the true value of the sample, q i The predicted value output by the model.

10. A graph representation learning-driven electrical imaging logging fracture identification system, characterized in that: The graph representation learning-driven electrical imaging logging fracture identification system includes: A segmentation module, used for segmenting the acquired real-time electrical imaging logging fracture data into a plurality of first sequence electrical imaging segment images; A binarization processing module, used for performing binarization processing on each of the first sequence of electrical imaging fragment images to obtain a first binarized image, wherein the first binarized image is a plurality of binarized images containing only cracks and holes; A denoising processing module is used to perform denoising on the image after the first binarization processing to obtain a first binarization segment image; the first binarization segment image is a plurality of images containing only natural fractures and broken natural fracture segments as research objects; A first original path graph determination module is used to determine a first original path graph according to the first binarized segment graph; the first original path graph is an image obtained by connecting the first nodes according to the open interval path connectivity; the first node is a bright pixel in the first binarized segment graph; A spatial feature and morphological feature determination module, for determining the spatial features and morphological features of each first node in the first original path diagram according to engineering-geological prior knowledge; the spatial features are the horizontal and vertical coordinates of all first nodes in a two-dimensional spatial coordinate system; the morphological features are the length, horizontal width, vertical width, horizontal aspect ratio, vertical aspect ratio and inclination of each first node and each first node's neighboring nodes within a circular verification threshold; An embedding module, used for embedding the spatial features and morphological features of each first node into the corresponding first node to obtain a first reconstructed knowledge path graph; A projection module, used to project the first reconstructed knowledge path graph into a high-dimensional feature space, and calculate the high-dimensional space distance between knowledge path graphs corresponding to different fragments; A fragment connection module, used to connect the fragments of the knowledge path graph whose high-dimensional space distance is lower than a first preset threshold to obtain a first recombined meta-crack knowledge graph; The crack identification module is used to input the first reorganized meta-crack knowledge graph into a crack intelligent identification model driven by graph representation learning to obtain crack types; the crack intelligent identification model driven by graph representation learning is a model established based on a graph convolutional neural network.

Images