A method for filling in missing values of spatial domain logging data based on DWGAT

Abstract

The application relates to a filling method for missing values of spatial domain logging data based on DWGAT and belongs to the technical field of logging data missing filling. The filling method for missing values of spatial domain logging data based on DWGAT provided by the application firstly designs logging data graph representation construction according to the spatial characteristics of logging data, adopts an optimized and improved Delaunay triangulation algorithm to construct a more reasonable logging data graph, and provides a data set for subsequent model training; secondly, the spatiality of logging data is fully considered, distance weight information is integrated to improve the traditional graph attention mechanism, and the logging data missing value filling problem to be solved is more effective; finally, due to the nonlinear characteristics of logging data, an MLP neural network is added to the graph attention layer of the model, and the representation ability of the model for logging data is improved. The application can solve the logging data missing value problem from the spatial domain angle, provide a filling solution with higher precision, and provide guarantee for the completeness and stability of logging data.

Description

Technical Field

[0001] This invention relates to a method for filling missing values ​​in spatial domain logging data based on DWGAT, belonging to the field of logging data missing value filling technology. Background Technology

[0002] Well logging data is an important geological data collected using well logging technology in uranium exploration. It mainly includes density, resistivity, porosity, permeability, and natural gamma rays. These data reflect the lithology, mineralization, and other valuable information of the uranium exploration area, and help to determine the geological structure, stratigraphic properties, and distribution of uranium resources.

[0003] Complete well logging datasets are a key factor in ensuring the accuracy of geological interpretation and the reliability of resource assessment. However, during the well logging data acquisition process, due to issues with logging instruments, geological conditions, operational errors, data transmission, or recording problems, well logging data may be missing, abnormal, or erroneous. These problems can affect the quality and reliability of well logging data, thereby impacting the evaluation and development of mineral resources. Therefore, filling in the missing well logging data with appropriate methods to make it as close to the actual situation as possible and improve the completeness and accuracy of the data is an important aspect of well logging data quality control.

[0004] With the rapid development of computer science and artificial intelligence, more machine learning methods are being used for missing value imputation tasks. Compared to traditional statistical methods, machine learning models can accurately and comprehensively predict missing values ​​by capturing complex relationships and patterns in the data, especially when there are non-linear and complex relationships between data. In terms of dynamic data learning, machine learning models do not rely on prior assumptions and can adaptively capture the relationships between features based on the characteristics and changes of the data, thereby better imputing missing values ​​and providing more flexible and accurate solutions when facing different types of data and missing value patterns.

[0005] However, there is currently limited research on using machine learning for missing value imputation in well logging data. This is partly due to the relatively late start of intelligent applications in the geological industry, the complexity of geological data labeling and cleaning, and the low level of data quality and standardization, which prevents machine learning from being directly applied to the problem of missing value imputation in well logging data. On the other hand, the high latitude, multiple attributes, and complex spatial relationships of well logging data make it difficult for existing machine learning imputation models to fully capture the complex geological structures reflected in the data, resulting in less than ideal performance in imputing missing values. Summary of the Invention

[0006] To address the problems existing in the prior art, this invention provides a method for filling missing values ​​in spatial domain logging data based on DWGAT. This invention provides a more accurate and effective solution for filling missing values ​​in logging data, ensuring the integrity and stability of logging data.

[0007] To achieve the above objectives, the technical solution provided by this invention is a method for filling missing values ​​in spatial domain logging data based on DWGAT, comprising the following steps:

[0008] (1) Data collection to form a dataset: Collect logging data with the same attributes from different wells and latitude and longitude coordinate data of each logging borehole. Integrate and reorganize logging data with the same attributes from different wells to form a logging data dataset with the same attributes from different wells. Integrate latitude and longitude coordinate data of each logging borehole to form a logging borehole coordinate information dataset.

[0009] (2) Forming well logging data map data: The well logging data in the well logging data dataset of step (1) is used to construct a well logging data node feature matrix through node representation. The data in the well logging borehole coordinate information dataset is used to construct an adjacency matrix and a distance matrix through edge representation based on the improved Delaunay triangulation algorithm. These three matrices together form well logging data map data.

[0010] (3) Output node feature matrix: Input the logging data graph data constructed in step (2) into the DWGAT model. The message passing layer based on the graph attention mechanism with fused distance weights calculates and updates the logging data node feature vectors, and then inputs them into the MLP neural network layer for calculation, finally outputting the complete logging data node feature matrix.

[0011] (4) Fill in missing data: Use the complete node feature matrix output in step (3) to fill in the missing logging data at the missing locations.

[0012] The further improvement to the above technical solution is as follows:

[0013] The specific steps in step (1) of forming a well logging data dataset with the same attributes from different wells include: integrating and reorganizing well logging data with the same attributes from different wells according to the requirements of the DWGAT model input data, and dividing the data into 512 dimensions. When the remaining data is less than 512 dimensions, it is padded with 0 to reach 512 dimensions to ensure the consistency of the dataset.

[0014] Step (2) specifically includes:

[0015] 1) Construction of node representation for well logging data graphs:

[0016] Node representation construction converts well logging data stored in text form into a well logging data node feature matrix that can be trained by the DWGAT model. For a certain attribute of the i-th well logging data node in the well logging data map, its feature vector is represented as h. i And record them sequentially downwards according to the formation. When there are n attributes in the well logging data node, h i The definition is as follows:

[0017] h i ={x1,x2,…,x n n>2;

[0018] Let H represent the feature vector set of a certain attribute of all nodes in the entire exploration area. When there are m logging points in the exploration area, H is defined as follows:

[0019] H = {h1,h2,…,h} m m>2;

[0020] 2) Construction of edge representation for well logging data maps:

[0021] Building upon the node representation, the edge representation completes the entire graph construction. Based on the geological spatial characteristics of the well logging data, the edge representation model is constructed using spatial coordinate information. For the location of the k-th well logging borehole in the well logging data graph, its spatial coordinate information is represented as p. k p k The definition is as follows:

[0022] p k =(x k ,y k )

[0023] Let P represent the detailed set of spatial locations of all logging boreholes in the entire exploration area. When there are m logging boreholes in the exploration area, P is defined as follows:

[0024] P = (p1, p2, p3, ..., p m )

[0025] In the formula, x k The x-coordinate of the k-th logging borehole is represented by its relative spatial position, i.e., its latitude. k The ordinate of the relative spatial position of the kth logging borehole is represented by the longitude of the relative spatial position.

[0026] An improved Delaunay triangulation algorithm is used to locally optimize the edge representation set generated by Delaunay triangulation. The local optimization target is the subset of edge representations at the outermost boundary position of the entire graph. The optimization process is as follows:

[0027] First, determine the representation subset of the outermost edge of the graph and define it using the shared boundary method. If an edge is not shared by two triangles at the same time, then the edge must be on the outermost boundary of the graph. The edge is then marked as the edge to be optimized, and then it is further determined whether the edge needs to be removed.

[0028] Secondly, the triangular elimination rule is determined. If two logging data nodes are connected by an abnormal edge, then there must be a logging data node that is relatively closer to them in space to connect them.

[0029] Finally, according to the triangle culling rule, if an edge is defined as an abnormal edge, then this edge must be the longest edge in the triangle it is in. This rule is used to detect the edge to be optimized. If the edge to be optimized is the longest edge in the triangle it is in, then the edge is deleted; otherwise, the edge is kept.

[0030] 3) Global construction of well logging data maps:

[0031] The edge representation set constructed by the Delaunay triangulation algorithm is transformed into an n×n dimensional adjacency matrix A, where n is the number of logging data nodes. The definition of A is as follows:

[0032]

[0033] In the formula, Aij = 1 indicates that there is an edge between the i-th logging data node and the j-th logging data node, and Aij = 0 indicates that there is no edge between the i-th logging data node and the j-th logging data node.

[0034] The well logging data graph structure was further optimized and improved by adding an edge connecting each node to itself, and defining the adjacency matrix as follows. That The mathematical expression of the calculation process is as follows:

[0035]

[0036] In the formula, I is an n×n identity matrix;

[0037] The definition is as follows:

[0038]

[0039] Based on the output of the Delaunay triangulation algorithm, the spatial distance between edges is calculated, and a distance matrix Dis representing the distance between two connected logging data nodes is defined. The distance from the logging data node to the connecting edge is defined as infinite. The definition of Dis is as follows:

[0040]

[0041] In the formula, the values ​​corresponding to the i-th row and j-th column represent the spatial distance between the i-th logging data node and the j-th logging data node.

[0042] Step (3) specifically includes:

[0043] 1) Graph attention mechanism with distance weights:

[0044] A Gaussian kernel function-based mapping distance weighting method is designed, which conforms to the representation of geological stratigraphic structure. Based on the distance matrix Dis of the constructed well logging data map, the Gaussian kernel function-based mapping distance weighting method is defined, and its mathematical expression is as follows:

[0045]

[0046] In the formula, Dis ij σ represents the distance between logging data node i and logging data node j in a two-dimensional coordinate system, where σ is a parameter that controls the mapping of the distance to appropriate weighting coefficients; ij The distance weighting coefficient between well logging data node i and well logging data node j is calculated in the final calculation. Their underground geological structures are more similar, and the characteristics of their well logging data nodes are also more similar. Therefore, the distance weighting coefficient between them is higher than that between well logging data nodes that are farther away from the target well logging point.

[0047] Distance weighting coefficient ω ij This indirectly reflects the similarity of features between nodes. Integrating this into the graph attention mechanism suppresses computational bias caused by missing node features, ensuring the accuracy of attention coefficient calculation. The attention coefficient z of the graph attention mechanism based on fused distance weights... ij The mathematical expression is as follows:

[0048]

[0049] The attention coefficients are normalized and converted into transmission weights required for message passing in the DWGAT model. The transmission weights between nodes are defined as d. ij d ij The mathematical expression is as follows:

[0050]

[0051] z ij Substituting the mathematical expression into d ij In the mathematical expression, the transmitted weight d is expressed more coherently. ij The calculation process, substituting z ij After the mathematical expression, d ij The mathematical expression is as follows:

[0052]

[0053] In the formula, and These are the feature representations of well logging data nodes i and j in the well logging data map, respectively, i.e., the feature vectors of well logging data nodes i and j in the well logging data map. W is a weight matrix used to linearly transform the feature representations of the well logging data nodes, a is a parameter vector used to calculate the attention weights, || denotes the concatenation operation of feature vectors, and LeakyReLU is an activation function that fuses the distance weight coefficients ω between well logging data nodes i and j after Gaussian kernel mapping. ij By introducing a trainable parameter β, the model can adjust the value of this parameter through training to set the degree of influence of distance weight on the final calculation result, and suppress the problem of attention coefficient calculation deviation caused by missing node features.

[0054] 2) Message Passing Layer

[0055] The message passing layer first learns and calculates the transmission weight d between logging data nodes in the logging data graph based on a graph attention mechanism with fused distance weights. ij , through d ij Weighted aggregation of feature vectors of neighboring logging data nodes To update the embedded feature vector of each logging data node The message passing between well logging data nodes is completed, and the data representation of the message passing layer is as follows:

[0056]

[0057] In the formula, σ is the Tanh activation function, W is a weight matrix used to linearly transform the feature representation of well logging data nodes, consistent with the description in 1) the graph attention mechanism with fused distance weights, d ij It is the transfer weight between logging data node i and its neighboring logging data node j. The feature vector of neighboring logging data node j, l represents the number of the current graph attention layer, N i Represents the set of nodes connected to node i;

[0058] 3) MLP Neural Network Layer

[0059] The MLP neural network consists of an input layer, hidden layers, and an output layer. The MLP neural network is designed based on the node features in the well logging data map. The well logging data node feature vectors are updated after message passing through the message passing layer. The input is fed into the MLP neural network layer for learning and training; the complete mathematical expression for the graph attention layer of the DWGAT model is as follows:

[0060]

[0061] In the formula, σ is the Tanh activation function, and N i Let d represent the set of nodes connected to node i. ij is the transfer weight between node i and its neighbor node j, W is a weight matrix used to linearly transform the feature representation of logging data nodes, β is the weight matrix of the MLP neural network layer, b is the bias term of the MLP layer, and l represents the number of attention layers in the current graph.

[0062] As can be seen from the above technical solution, the present invention provides a method for filling missing values ​​in spatial domain logging data based on DWGAT. From the perspective of the spatial domain, and based on the spatial characteristics of logging data, it first designs a logging data graph representation construction, employing the Delaunay triangulation algorithm and optimizing its shortcomings to make the logging data graph constructed by this algorithm more reasonable, providing data integration for subsequent model training. Secondly, it fully considers the spatial nature of logging data, incorporating distance weight information to improve the traditional graph attention mechanism, making it more effective in solving the problem of filling missing values ​​in logging data. Finally, due to the nonlinear characteristics of logging data, an MLP neural network is added to the graph attention layer of the model, improving the model's ability to represent logging data. Compared with existing technologies, the present invention has the following advantages:

[0063] (1) Because the technical solution adopted in this invention fully considers the complex spatial characteristics of well logging data, this invention solves the problem that the missing value filling effect is not ideal in the current spatial missing value filling methods. For example, the spatial interpolation algorithm in geostatistics only uses specific distance rules to gather the feature information of the surrounding neighbors to fill in the missing data locations, and the missing value filling effect is not ideal.

[0064] (2) Because the technical solution adopted in this invention improves the spatial correlation characterization capability of well logging data, this invention improves the accuracy of filling missing values ​​compared with the prior art, especially in well logging data with a large proportion of missing values.

[0065] (3) Because the technical solution adopted in this invention uses the improved Delaunay triangulation algorithm to solve the problem of edge representation construction of well logging data map, this invention can solve the abnormal phenomenon that the edge representation generated by the Delaunay triangulation algorithm on the outermost boundary of the entire map will have an excessively long edge. Attached Figure Description

[0066] Figure 1Overall architecture diagram of the DWGAT spatial domain filling model;

[0067] Figure 2 Flowchart of the improved Delaunay triangulation algorithm for handling abnormal edges;

[0068] Figure 3 A complete flowchart for constructing well logging data graphical representations;

[0069] Figure 4 A flowchart illustrating the overall process of a graph attention mechanism that incorporates distance weights;

[0070] Figure 5 Attention layer structure diagram;

[0071] Figure 6 A schematic diagram illustrating the complete calculation process of the DWGAT model;

[0072] Figure 7 Flowchart of the complete training process for the DWGAT model. Detailed Implementation

[0073] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments, but the scope of protection of the present invention is not limited to the following embodiments.

[0074] The present invention provides a method for filling missing values ​​in spatial domain logging data based on DWGAT. The overall architecture diagram of the DWGAT spatial domain filling model is shown below. Figure 1 As shown, it includes the following steps:

[0075] (1) Data collection to form a dataset: Collect logging data with the same attributes from different wells and latitude and longitude coordinate data of each logging borehole. Integrate and reorganize logging data with the same attributes from different wells to form a logging data dataset with the same attributes from different wells. Integrate latitude and longitude coordinate data of each logging borehole to form a logging borehole coordinate information dataset.

[0076] The specific steps to form a dataset of logging data with the same attributes from different wells include: integrating and reorganizing logging data with the same attributes from different wells according to the requirements of the DWGAT model input data, and dividing the data into 512 dimensions. When the remaining data is less than 512 dimensions, it is padded with 0 to reach 512 dimensions to ensure the consistency of the dataset.

[0077] (2) Forming the logging data graph: The logging data in the logging data dataset from step (1) is used to construct a logging data node feature matrix through node representation. The data in the logging borehole coordinate information dataset is used to construct an adjacency matrix and a distance matrix through edge representation based on the improved Delaunay triangulation algorithm. These three matrices together form the logging data graph. The complete flowchart for constructing the logging data graph representation is shown below. Figure 3As shown;

[0078] 1) Construction of node representation for well logging data graphs:

[0079] Node representation construction converts well logging data stored in text form into a well logging data node feature matrix that can be trained by the DWGAT model. For a certain attribute of the i-th well logging data node in the well logging data map, its feature vector is represented as h. i And record them sequentially downwards according to the formation. When there are n attributes in the well logging data node, h i The definition is as follows:

[0080] h i ={x1,x2,…,x n n>2;

[0081] Let H represent the feature vector set of a certain attribute of all nodes in the entire exploration area. When there are m logging points in the exploration area, H is defined as follows:

[0082] H = {h1,h2,…,h} m m>2;

[0083] 2) Construction of edge representation for well logging data maps:

[0084] Based on the node representation construction, the edge representation is constructed to complete the entire graph construction. According to the geological spatial characteristics of well logging data, the edge representation model is constructed using spatial coordinate information. For the position of the k-th well logging point in the well logging data graph, k>2, its spatial coordinate information is represented as p. k p k The definition is as follows:

[0085] p k =(x k ,y k )

[0086] Let P represent the detailed set of spatial locations of all logging boreholes in the entire exploration area. When there are m logging boreholes in the exploration area, P is defined as follows:

[0087] P = (p1, p2, p3, ..., p m )

[0088] In the formula, x k The x-coordinate of the k-th logging borehole is represented by its relative spatial position, i.e., its latitude. k The ordinate of the relative spatial position of the kth logging borehole is represented by the longitude of the relative spatial position.

[0089] An improved Delaunay triangulation algorithm is used to locally optimize the edge representation set generated by Delaunay triangulation. The local optimization target is the subset of edge representations at the outermost boundary position of the entire graph. The optimization process is as follows: Figure 2 As shown, the specific steps are as follows:

[0090] The first step is to traverse the already constructed set of edge representations;

[0091] The second step is to determine whether the edge representation set has been traversed completely. If the traversal is complete, the optimization ends; otherwise, proceed to the third step.

[0092] The third step is to select an edge that has not been traversed in the edge representation set and define it using the shared boundary method. If an edge is not shared by two triangles at the same time, then the edge must be on the outermost boundary of the graph. This edge is then identified as the edge to be optimized and proceeds to the fourth step. If the edge is not the edge to be optimized, return to the second step.

[0093] The fourth step is to determine whether the edge to be optimized needs to be removed and to determine the triangular removal rule. If two logging data nodes are connected by an abnormal edge, then there must be a logging data node that is relatively closer to both of them in terms of spatial distance. Connect the two logging data nodes to the logging data node that is relatively closer to both of them in terms of spatial distance, and then delete the abnormal edge. If the edge to be optimized satisfies the triangular removal rule, remove the edge; otherwise, return to the second step.

[0094] 3) Global construction of well logging data maps:

[0095] The edge representation set constructed by the Delaunay triangulation algorithm is transformed into an n×n dimensional adjacency matrix A, where n is the number of logging data nodes. The definition of A is as follows:

[0096]

[0097] In the formula, A ij =1 indicates that there is an edge between the i-th logging data node and the j-th logging data node, A ij =0 indicates that there is no edge between the i-th logging data node and the j-th logging data node;

[0098] The well logging data graph structure was further optimized and improved by adding an edge connecting each node to itself, and defining the adjacency matrix as follows. That The mathematical expression of the calculation process is as follows:

[0099]

[0100] In the formula, I is an n×n identity matrix;

[0101] The definition is as follows:

[0102]

[0103] Based on the output of the Delaunay triangulation algorithm, the spatial distance between edges is calculated, and a distance matrix Dis representing the distance between two connected logging data nodes is defined. The distance from the logging data node to the connecting edge is defined as infinite. The definition of Dis is as follows:

[0104]

[0105] In the formula, the values ​​corresponding to the i-th row and j-th column represent the spatial distance between the i-th logging data node and the j-th logging data node.

[0106] (3) Output node feature matrix: Input the well logging data map constructed in step (2) into the DWGAT model. A schematic diagram of the complete calculation process of the DWGAT model is shown below. Figure 6 As shown, the message passing layer based on the graph attention mechanism with fused distance weights calculates and updates the feature vectors of logging data nodes, and then inputs them into the MLP neural network layer for calculation, finally outputting the complete feature matrix of logging data nodes.

[0107] 1) Graph attention mechanism with distance weights:

[0108] The overall flowchart of the graph attention mechanism that integrates distance weights is as follows: Figure 4 As shown in the diagram, the attention layer structure is as follows: Figure 5 As shown;

[0109] A Gaussian kernel function-based mapping distance weighting method is designed, which conforms to the representation of geological stratigraphic structure. Based on the distance matrix Dis of the constructed well logging data map, the Gaussian kernel function-based mapping distance weighting method is defined, and its mathematical expression is as follows:

[0110]

[0111] In the formula, Dis ij σ represents the distance between logging data node i and logging data node j in a two-dimensional coordinate system, where σ is a parameter that controls the mapping of the distance to appropriate weighting coefficients; ij The distance weighting coefficient between well logging data node i and well logging data node j is calculated in the final calculation. Their underground geological structures are more similar, and the characteristics of their well logging data nodes are also more similar. Therefore, the influence weighting coefficient between them is higher than that between well logging data nodes that are farther away from the target well logging point.

[0112] Distance weighting coefficient ω ij This indirectly reflects the similarity of features between nodes. Integrating this into the graph attention mechanism suppresses computational bias caused by missing node features, ensuring the accuracy of attention coefficient calculation. The attention coefficient z of the graph attention mechanism based on fused distance weights... ij The mathematical expression is as follows:

[0113]

[0114] The attention coefficients are normalized and converted into transmission weights required for message passing in the DWGAT model. The transmission weights between nodes are defined as d. ij d ij The mathematical expression is as follows:

[0115]

[0116] z ij Substituting the mathematical expression into d ij In the mathematical expression, the transmitted weight d is expressed more coherently. ij The calculation process, substituting z ij After the mathematical expression, d ij The mathematical expression is as follows:

[0117]

[0118] In the formula, and These are the feature representations of well logging data nodes i and j in the well logging data map, respectively, i.e., the feature vectors of well logging data nodes i and j in the well logging data map. W is a weight matrix used to linearly transform the feature representations of the well logging data nodes, a is a parameter vector used to calculate the attention weights, || denotes the concatenation operation of feature vectors, and LeakyReLU is an activation function that fuses the distance weight coefficients ω between well logging data nodes i and j after Gaussian kernel mapping. ij By introducing a trainable parameter β, the model can adjust the value of this parameter through training to set the degree of influence of distance weight on the final calculation result, and suppress the problem of attention coefficient calculation deviation caused by missing node features.

[0119] 2) Message Passing Layer

[0120] The message passing layer first learns and calculates the transmission weight d between logging data nodes in the logging data graph based on a graph attention mechanism with fused distance weights. ij , through d ij Weighted aggregation of feature vectors of neighboring logging data nodes To update the embedded feature vector of each logging data node The message passing between well logging data nodes is completed, and the data representation of the message passing layer is as follows:

[0121]

[0122] In the formula, σ is the Tanh activation function, W is a weight matrix used to linearly transform the feature representation of well logging data nodes, and d ij h is the attention coefficient between logging data node i and its neighboring logging data node j. j The feature vector of neighboring logging data node j, l represents the number of the current graph attention layer, N i Represents the set of nodes connected to node i;

[0123] 3) MLP Neural Network Layer

[0124] The MLP neural network consists of an input layer, hidden layers, and an output layer. The specific neural network structure is designed based on the node features in the well logging data map (can this be written as an MLP neural network? If not, the inventor needs to provide the specific name of the neural network structure or the specific design steps). The well logging data node feature vectors are updated after message passing through the message passing layer. The input is fed into the MLP neural network layer for learning and training; the complete mathematical expression for the graph attention layer of the DWGAT model is as follows:

[0125]

[0126] In the formula, σ is the Tanh activation function, and N i Let d represent the set of nodes connected to node i. ij is the transfer weight between node i and its neighbor node j, W is a weight matrix used to linearly transform the feature representation of logging data nodes, β is the weight matrix of the MLP neural network layer, b is the bias term of the MLP layer, and l represents the number of attention layers in the current graph.

[0127] (4) Fill in missing data: Use the complete node feature matrix output in step (3) to fill in the missing logging data at the missing locations.

[0128] In this embodiment, the complete training process flowchart of the DWGAT model is as follows: Figure 7 As shown, the implementation steps are as follows:

[0129] (1) Missing data is input into the model, and complete data is used as labels;

[0130] (2) Construct a logging data map based on the Delaunay triangulation algorithm and optimize the map structure according to the abnormal edge removal rules;

[0131] (3) Initialize the parameters of the DWGAT model and train it in batches;

[0132] (4) Forward propagation: In the well logging data graph, each node is calculated through two graph attention layers. The feature vector of node i calculated by the first graph attention layer is as follows:

[0133]

[0134] The feature vector of node i calculated after the second graph attention layer is as follows:

[0135]

[0136] (5) Calculate the loss by using the RMSE loss function to calculate the model outputs. The total error between the dataset and the corresponding complete labeled dataset;

[0137] (6) Calculate the gradient of the loss with respect to the model parameters using the backpropagation algorithm, and update the model parameters to minimize the loss function;

[0138] (7) Repeat the previous steps for multiple iterations until the specified stopping criteria or the specified number of training rounds are reached.

[0139] The above method provides a missing value filling method for spatial domain logging data based on DWGAT, offering a more accurate and effective solution for filling missing values ​​in logging data, and ensuring the integrity and stability of logging data.

Claims

1. A method for filling missing values ​​in spatial domain logging data based on DWGAT, characterized in that, Includes the following steps: (1) Data collection to form a dataset: Collect logging data with the same attributes from different wells and latitude and longitude coordinate data of each logging borehole. Integrate and reorganize logging data with the same attributes from different wells to form a logging data dataset with the same attributes from different wells. Integrate latitude and longitude coordinate data of each logging borehole to form a logging borehole coordinate information dataset. (2) Forming well logging data map data: The well logging data in the well logging data dataset of step (1) is used to construct a well logging data node feature matrix through node representation. The data in the well logging borehole coordinate information dataset is used to construct an adjacency matrix and a distance matrix through edge representation based on the improved Delaunay triangulation algorithm. These three matrices together form well logging data map data. (3) Output node feature matrix: Input the logging data graph data constructed in step (2) into the DWGAT model. The message passing layer based on the graph attention mechanism with fused distance weights calculates and updates the logging data node feature vectors, and then inputs them into the MLP neural network layer for calculation, finally outputting the complete logging data node feature matrix. (4) Fill in missing data: Use the complete node feature matrix output in step (3) to fill in the missing logging data at the missing locations.

2. The method for filling missing values ​​in spatial domain logging data based on DWGAT according to claim 1, characterized in that, The specific steps in step (1) of forming a well logging data dataset with the same attributes from different wells include: integrating and reorganizing well logging data with the same attributes from different wells according to the requirements of the DWGAT model input data, and dividing the data into 512 dimensions. When the remaining data is less than 512 dimensions, it is padded with 0 to reach 512 dimensions to ensure the consistency of the dataset.

3. The method for filling missing values ​​in spatial domain logging data based on DWGAT according to claim 1, characterized in that, Step (2) specifically includes: 1) Construction of node representation for well logging data graphs: Node representation construction converts well logging data stored in text form into a well logging data node feature matrix that can be trained by the DWGAT model. For a certain attribute of the i-th well logging data node in the well logging data map, its feature vector is represented as h. i And record them sequentially downwards according to the formation. When there are n attributes in the well logging data node, h i The definition is as follows: h i {x1,x2,…,x n }n>2\ Let H represent the feature vector set of a certain attribute of all nodes in the entire exploration area. When there are m logging points in the exploration area, H is defined as follows: H={h1,h2,…,h m }m>2; 2) Construction of edge representation for well logging data graphs: Based on the node representation construction, the edge representation is constructed to complete the entire graph construction. According to the geological spatial characteristics of well logging data, the edge representation model is constructed using spatial coordinate information. For the position of the k-th well logging borehole in the well logging data graph, k>2, its spatial coordinate information is represented as p. k p k The definition is as follows: p k =(x k ,y k ) Let P represent the detailed set of spatial locations of all logging boreholes in the entire exploration area. When there are m logging boreholes in the exploration area, P is defined as follows: P=(p1,p2,p3,…,p m ) In the formula, x k The x-coordinate of the k-th logging borehole is represented by its relative spatial position, i.e., its latitude. k The ordinate of the relative spatial position of the kth logging borehole is represented by the longitude of the relative spatial position. An improved Delaunay triangulation algorithm is used to locally optimize the edge representation set generated by Delaunay triangulation. The local optimization target is the subset of edge representations at the outermost boundary position of the entire graph. The optimization process is as follows: First, determine the representation subset of the outermost edge of the graph and define it using the shared boundary method. If an edge is not shared by two triangles at the same time, then the edge must be on the outermost boundary of the graph. The edge is then marked as the edge to be optimized, and then it is further determined whether the edge needs to be removed. Secondly, the triangular elimination rule is determined. If two logging data nodes are connected by an abnormal edge, then there must be a logging data node that is relatively closer to them in space to connect them. Finally, according to the triangle culling rule, if an edge is defined as an abnormal edge, then this edge must be the longest edge in the triangle it is in. This rule is used to detect the edge to be optimized. If the edge to be optimized is the longest edge in the triangle it is in, then the edge is deleted; otherwise, the edge is kept. 3) Global construction of well logging data maps: The edge representation set constructed by the Delaunay triangulation algorithm is transformed into an n×n dimensional adjacency matrix A, where n is the number of logging data nodes. The definition of A is as follows: In the formula, A ij =1 indicates that there is an edge between the i-th logging data node and the j-th logging data node, A ij =0 indicates that there is no edge between the i-th logging data node and the j-th logging data node; The well logging data graph structure was further optimized and improved by adding an edge connecting each node to itself, and defining the adjacency matrix as follows. That The mathematical expression of the calculation process is as follows: In the formula, I is an n×n identity matrix; The definition is as follows: Based on the output of the Delaunay triangulation algorithm, the spatial distance between edges is calculated, and a distance matrix Dis representing the distance between two connected logging data nodes is defined. The distance from the logging data node to the connecting edge is defined as infinite. The definition of Dis is as follows: In the formula, the values ​​corresponding to the i-th row and j-th column represent the spatial distance between the i-th logging data node and the j-th logging data node.

4. The method for filling missing values ​​in spatial domain logging data based on DWGAT according to claim 1, characterized in that, Step (3) specifically includes: 1) Graph attention mechanism with distance weights: A Gaussian kernel function-based mapping distance weighting method is designed, which conforms to the representation of geological stratigraphic structure. Based on the distance matrix Dis of the constructed well logging data map, the Gaussian kernel function-based mapping distance weighting method is defined, and its mathematical expression is as follows: In the formula, Dis ij σ represents the distance between logging data node i and logging data node j in a two-dimensional coordinate system, where σ is a parameter that controls the mapping of the distance to appropriate weighting coefficients; ij The distance weighting coefficient between well logging data node i and well logging data node j is calculated in the final calculation. Their underground geological structures are more similar, and the characteristics of their well logging data nodes are also more similar. Therefore, the distance weighting coefficient between them is higher than that between well logging data nodes that are farther away from the target well logging point. Distance weight coefficient ω ij This indirectly reflects the similarity of features between nodes. Integrating this into the graph attention mechanism suppresses computational bias caused by missing node features, ensuring the accuracy of attention coefficient calculation. The attention coefficient z of the graph attention mechanism based on fused distance weights... ij The mathematical expression is as follows: The attention coefficients are normalized and converted into transmission weights required for message passing in the DWGAT model. The transmission weights between nodes are defined as d. ij d ij The mathematical expression is as follows: z ij Substituting the mathematical expression into d ij In the mathematical expression, the transmitted weight d is expressed more coherently. ij The calculation process, substituting z ij After the mathematical expression, d ij The mathematical expression is as follows: In the formula, and These are the feature representations of well logging data nodes i and j in the well logging data map, respectively, i.e., the feature vectors of well logging data nodes i and j in the well logging data map. W is a weight matrix used to linearly transform the feature representations of the well logging data nodes, a is a parameter vector used to calculate the attention weights, || denotes the concatenation operation of feature vectors, and LeakyReLU is an activation function that fuses the distance weight coefficients ω between well logging data nodes i and j after Gaussian kernel mapping. ij By introducing a trainable parameter β, the model can adjust the value of this parameter through training to set the degree of influence of distance weight on the final calculation result, and suppress the problem of attention coefficient calculation deviation caused by missing node features. 2) Message Passing Layer The message passing layer first learns and calculates the transmission weight d between logging data nodes in the logging data graph based on a graph attention mechanism with fused distance weights. ij , through d ij Weighted aggregation of feature vectors of neighboring logging data nodes To update the embedded feature vector of each logging data node The message passing between well logging data nodes is completed, and the data representation of the message passing layer is as follows: In the formula, σ is the Tanh activation function, W is a weight matrix used to linearly transform the feature representation of well logging data nodes, and d ij It is the transfer weight between logging data node i and its neighboring logging data node j. The feature vector of neighboring logging data node j, l represents the number of the current graph attention layer, N i Represents the set of nodes connected to node i; 3) MLP Neural Network Layer The MLP neural network consists of an input layer, hidden layers, and an output layer. The MLP neural network is designed based on the node features in the well logging data map. The well logging data node feature vectors are updated after message passing through the message passing layer. The input is fed into the MLP neural network layer for learning and training; the complete mathematical expression for the graph attention layer of the DWGAT model is as follows: In the formula, σ is the Tanh activation function, and N i Let d represent the set of nodes connected to node i. ij is the transfer weight between node i and its neighbor node j, W is a weight matrix used to linearly transform the feature representation of logging data nodes, β is the weight matrix of the MLP neural network layer, b is the bias term of the MLP layer, and l represents the number of the current graph attention layer.

Images