A lithology intelligent classification method based on a multi-well graph neural network

Abstract

This invention discloses an intelligent lithology classification method based on a multi-well graph neural network. First, engineering parameters and lithology data from multiple wells are collected and preprocessed to construct graph nodes at depth points. Then, a graph structure containing intra-well edges and inter-well edges is constructed, where intra-well edges connect adjacent nodes within the same well, and inter-well edges connect nodes in different wells with similar logging responses. Next, a graph attention network is used to aggregate neighborhood information and update node feature representations. The updated features are then passed through a fully connected layer and a Softmax function to output lithology classification probabilities. Finally, a weighted cross-entropy loss function is used to train the model, and micro-batch sampling is employed to reduce computational overhead. This method introduces inter-well geological similarity through inter-well edges and adaptively fuses neighborhood features using an attention mechanism, effectively solving the problem of traditional methods neglecting inter-well correlations. It significantly improves the accuracy of cross-well lithology classification and the ability to identify rare lithologies, possessing significant engineering application value.

Description

Technical Field

[0001] This invention relates to the fields of geological exploration and oil drilling technology, and in particular to a lithology intelligent classification method based on a multi-well map neural network. Background Technology

[0002] In oil and gas exploration and development, lithology identification is fundamental to formation evaluation and reservoir characterization, playing a crucial supporting role in drilling parameter optimization, reservoir sweet spot prediction, and geological steering decisions. Traditional lithology identification methods mainly include coring analysis, cuttings logging, and well logging interpretation. While coring analysis offers high accuracy, it is expensive and cannot achieve continuous coverage; cuttings logging suffers from significant time lag, making it difficult to meet real-time decision-making needs; and well logging interpretation relies heavily on the experience of interpreters and is susceptible to ambiguity.

[0003] With the development of machine learning technology, researchers have begun to explore intelligent lithology identification using well logging data and drilling engineering parameters. For example, Liang et al. studied intelligent lithology identification technology based on multi-parameter fusion of logging while drilling in their paper "Research on Intelligent Recognition Technology in Lithology Based on Multi-parameter Fusion of Logging While Drilling"; Qingfeng et al. proposed a comprehensive machine learning model for lithology identification while drilling in their paper "A comprehensive machine learning model for lithology identification while drilling"; and Zhang et al. achieved real-time lithology identification by combining self-attention and cross-attention mechanisms with wavelet transform in their paper "Real-time lithology identification from drilling data with self & cross attention model and wavelet transform".

[0004] However, the methods described above are essentially purely data-driven modeling approaches, focusing only on feature extraction and mapping relationships from single-point or single-well data, without considering the continuity of strata within the well and the geological correlations between wells. Actual drilling data possesses inherent graph structure characteristics: lithology at adjacent depths within the same well exhibits geological continuity, and depths with similar logging responses across different wells may correspond to similar lithologies. Traditional machine learning models treat each depth point as an independent, identically distributed sample, failing to effectively model such complex spatial relationships. This results in insufficient generalization ability when applying the model across wells, especially in blocks with significant variations in geological conditions, where model performance deteriorates considerably.

[0005] Therefore, there is an urgent need for an intelligent lithology classification method that can simultaneously utilize the continuity of strata within wells and the geological similarity between wells, in order to improve the accuracy and generalization ability of lithology identification in cross-well scenarios. Summary of the Invention

[0006] To address the shortcomings of existing lithology identification methods in utilizing inter-well geological correlation information and exhibiting poor cross-well generalization ability, this invention proposes an intelligent lithology classification method based on a multi-well graph neural network. This method constructs a multi-well graph structure including intra-well and inter-well edges, and utilizes a graph attention network to aggregate neighborhood node information, achieving effective fusion of cross-well geological information and accurate lithology classification.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0008] A lithology intelligent classification method based on multi-well map neural networks includes the following steps:

[0009] 1) Collect drilling engineering parameters and lithological logging data from multiple wells, perform data cleaning and standardization, and construct graph nodes for each depth point of each well. The node features are composed of the drilling engineering parameters corresponding to that depth point, and the lithological logging data is used as node training labels.

[0010] 2) Construct a graph structure and define two types of edges: the first type is intra-well edges, which connect adjacent depth nodes in the same well; the second type is inter-well edges, which connect nodes with similar logging response characteristics in different wells.

[0011] 3) Construct a graph neural network model, taking the node features from step 1) and the graph structure from step 2) as input, and using a graph attention layer to aggregate neighborhood node information and update the node feature representation;

[0012] 4) Input the updated node features into the classification layer. After fully connected mapping and Sofimax function calculation, take the category corresponding to the maximum probability as the lithology classification result of each node.

[0013] 5) Use the lithology logging data collected in step 1) as supervised label training model, optimize model parameters with cross-entropy loss function, and evaluate the lithology classification accuracy of the model in cross-well scenarios through validation set.

[0014] The drilling engineering parameters collected in step 1) include: natural gamma (GR), resistivity (RT), sonic transit time (AC), compensated density (DEN), compensated neutron (CNL), caliper diameter (CAL), drilling time (ROP), hook load (HKLD), torque (TORQUE), drilling fluid density (MDIA), drilling fluid displacement (MFIA), standpipe pressure (SPPA), rotary table speed (RPMB), and weight on bit (WOBB); the lithological logging data is lithology (LITH). The data preprocessing process includes:

[0015] 1.1) Perform outlier detection on the collected drilling engineering parameters and replace outliers with the median value within a sliding window;

[0016] 1.2) Standardize all parameters to eliminate scale differences between different units;

[0017] 1.3) Sampling at fixed depth intervals, the continuous depth data is discretized into a sequence of depth points, each depth point is a graph node, and the node feature vector is composed of all the parameters of that depth point.

[0018] In step 2), a multi-well map structure containing intra-well edges and inter-well edges is constructed, as follows:

[0019] 2.1) In-well edge construction method: For each well, connect two adjacent nodes in depth with an edge and set the weight of the edge to 1 to reflect the continuity of formation deposition;

[0020] 2.2) The method for constructing well edges is as follows: calculate the similarity between features of different well nodes, use cosine similarity as the similarity metric, set a similarity threshold, select all cross-well nodes with similarity exceeding the threshold for each node to establish well edges, and set the weight of the well edge to the cosine similarity value between the two nodes.

[0021] In step 3), a graph neural network model is constructed to update node features, specifically using a graph attention network (GAT), including:

[0022] 3.1) For each node, calculate its attention coefficient with all neighboring nodes, including intra-well neighbors and inter-well neighbors. The formula for calculating the attention coefficient is:

[0023]

[0024] Among them, h iLet w be the input feature vector of node i, w be the trainable weight matrix, a be the trainable attention vector, and || denote the vector concatenation operation. Let i be the set of neighboring nodes of node i;

[0025] 3.2) Use attention coefficients to perform weighted summation of the features of neighboring nodes and update the node feature representation:

[0026]

[0027] Where σ is the ELU activation function;

[0028] 3.3) Stack two graph attention layers to aggregate first-order and second-order neighborhood information respectively, and finally output a high-dimensional feature representation of each node.

[0029] In step 4), the specific process of the classification layer is as follows:

[0030] 4.1) Input the high-dimensional features of the nodes output by the graph neural network into a fully connected network layer. Set the output dimension of the fully connected network to the total number of lithology categories.

[0031] 4.2) The output of the fully connected network is calculated using the Soffmax function to obtain the probability distribution of each node belonging to a certain type of lithology;

[0032] 4.3) Take the index corresponding to the maximum probability and map the index to a specific lithology name as the classification result output.

[0033] In step 5), the model training framework is constructed, specifically including:

[0034] 5.1) A weighted cross-entropy loss function is used as the training objective to address the lithological class imbalance problem. The formula is:

[0035]

[0036] Where N is the number of samples in the batch, C is the total number of lithological categories, and y i，c The true label for sample i belonging to category c. w represents the probability value predicted by the model that sample i belongs to class c. c The weight coefficient for category c is calculated by normalizing the coefficients based on the reciprocal of the number of samples in each category in the training set.

[0037] 5.2) The Adam optimizer is used to iteratively update the model parameters. The optimized model parameters include the weight matrix W of the graph attention layer, the attention vector a, and the weights and biases of the fully connected network in the classification layer.

[0038] 5.3) A micro-batch sampling strategy is used for training. In each iteration, a batch of nodes and their corresponding neighborhood subgraphs are randomly selected to participate in the calculation, so as to reduce the computational overhead of large-scale multi-well maps.

[0039] 5.4) On the validation set, the overall accuracy, macro F1 score and recall rate for each lithology class are used as evaluation indicators. If the overall accuracy of the validation set does not improve for several consecutive training rounds, the training is terminated early and the optimal model parameters are saved.

[0040] The trained model, by constructing a multi-well map structure that incorporates intra-well continuity and inter-well similarity, and utilizing a graph attention network to aggregate neighborhood information, can achieve high-precision classification of lithology in single-well and multi-well scenarios. This method fully explores the inherent correlation of geological responses between different wells, effectively improving the model's generalization ability in cross-well applications.

[0041] The advantages of this invention compared to the prior art are mainly reflected in:

[0042] (1) Multi-well information fusion, breaking through the limitations of single-well modeling: This invention constructs well edges to associate nodes with similar logging responses in different wells, realizing the effective transmission of geological information across wells and overcoming the defect that traditional single-well modeling cannot utilize information from neighboring wells.

[0043] (2) Graph structure modeling conforms to the essential characteristics of geological data: Graph neural networks are used to model drilling data. The inner edge of the well depicts the vertical continuity of the formation, and the outer edge of the well captures the lateral similarity between different wells. This structure is highly consistent with the essential characteristics of geological data and improves the model's representation ability.

[0044] (3) Attention mechanism to achieve adaptive information aggregation: Introducing a graph attention mechanism to assign adaptive weights to different neighboring nodes can distinguish the degree of contribution of different neighboring nodes to the lithology judgment of the current node and enhance the model's ability to extract key information.

[0045] In summary, this invention, with "multi-well map construction - graph attention aggregation - cross-well information fusion" as its core, constructs a set of intelligent lithology classification methods applicable to multi-well scenarios, providing reliable technical support for regional geological research and oil and gas exploration and development. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of the construction of a multi-well diagram structure;

[0047] Figure 2 Create a flowchart for the well-to-well boundary;

[0048] Figure 3 The diagram shows the structure of the graph attention network model. Detailed Implementation

[0049] Example 1:

[0050] This embodiment uses logging and drilling engineering data from 10 wells for method validation. The dataset covers depths from 1000m to 4500m, with approximately 50,000 valid node samples. The lithology is divided into six categories: mudstone (0), siltstone (1), fine sandstone (2), medium sandstone (3), coarse sandstone (4), and gravelly sandstone (5). The lithological distribution varies among the wells, providing a good foundation for validating the model's generalization ability in multi-well scenarios. Figure 1 A schematic diagram illustrating the construction of a multi-well map structure is shown.

[0051] The present invention provides a lithological intelligent classification method based on a multi-well map neural network, comprising the following steps:

[0052] 1) Collect drilling engineering parameters and lithological logging data from multiple wells, perform data cleaning and standardization, and construct graph nodes for each depth point of each well. The node features are composed of the drilling engineering parameters corresponding to that depth point, and the lithological logging data is used as node training labels.

[0053] 2) Construct a graph structure and define two types of edges: the first type is intra-well edges, which connect adjacent depth nodes in the same well; the second type is inter-well edges, which connect nodes with similar logging response characteristics in different wells.

[0054] 3) Construct a graph neural network model, taking the node features from step 1) and the graph structure from step 2) as input, and using a graph attention layer to aggregate neighborhood node information and update the node feature representation;

[0055] 4) Input the updated node features into the classification layer. After fully connected mapping and softmax function calculation, take the category corresponding to the maximum probability as the lithology classification result of each node.

[0056] 5) Use the lithology logging data collected in step 1) as supervised label training model, optimize model parameters with cross-entropy loss function, and evaluate the lithology classification accuracy of the model in cross-well scenarios through validation set.

[0057] In step 1, the following 14 drilling engineering parameters are collected in real time using logging instruments and the drilling monitoring system: natural gamma (GR), resistivity (RT), acoustic transit time (AC), compensated density (DEN), compensated neutron (CNL), caliper diameter (CAL), drilling time (ROP), hook load (HKLD), torque (TORQUE), drilling fluid density (MDIA), drilling fluid displacement (MFIA), standpipe pressure (SPPA), rotary table speed (RPMB), and weight on bit (WOBB); the target parameter is lithology (LITH). The collected data is then preprocessed, specifically including:

[0058] 1.1) Outlier detection was performed on the collected drilling parameters using a sliding window of length 5. The median within the window was calculated, and outliers exceeding ±3 standard deviations of the window mean were replaced. Table 1 shows a partial sample of the processed data:

[0059] Table 1 Sample Data Example

[0060]

[0061] 1.2) Subsequently, all parameters are standardized using the Z-score formula:

[0062]

[0063] Where μ and σ are the mean and standard deviation of each feature in the training set.

[0064] 1.3) Sampling is performed at a fixed depth interval of 0.125m. The continuous depth data is discretized into a sequence of depth points. Each depth point is treated as a graph node. The node feature vector consists of all 14 parameters of that depth point, resulting in the node feature matrix.

[0065] In this embodiment, step 2) involves constructing a multi-well map structure that includes both intra-well edges and inter-well edges:

[0066] 2.1) Construction of In-Well Edges: For each well, connect two adjacent nodes in depth with an edge, and set the weight of the edge to 1. For example, for three nodes in a well with depths of 1000.00m, 1000.25m and 1000.50m respectively, construct two in-well edges (1000.00, 1000.25) and (1000.25, 1000.50) respectively.

[0067] 2.2) Well-to-well edge construction: Calculate the cosine similarity between features of different well nodes, using the following formula:

[0068]

[0069] Set a similarity threshold θ = 0.85, select all cross-well nodes with similarity greater than 0.85 for each node to establish an inter-well edge, and set the weight of the inter-well edge to the cosine similarity value between the two nodes. Figure 2 The construction process of the well-side is demonstrated.

[0070] In this embodiment, step 3) involves constructing a graph attention network (GAT) model. The node features from step 1) and the graph structure from step 2) are used as input. The graph attention layer aggregates neighborhood node information and updates the node feature representation. Specifically, this includes:

[0071] 3.1) For each node, calculate its attention coefficient with all neighboring nodes, including intra-well neighbors and inter-well neighbors. The formula for calculating the attention coefficient is:

[0072]

[0073] Among them, h i Let be the input feature vector of node i, W be the trainable weight matrix, a be the trainable attention vector, and || denote the vector concatenation operation. Let f be the set of neighboring nodes, and let LeakyReLU have a negative slope of 0.2.

[0074] 3.2) Use attention coefficients to perform weighted summation of the features of neighboring nodes and update the node feature representation:

[0075]

[0076] Where σ is the ELU activation function;

[0077] 3.3) Stack two graph attention layers to aggregate first-order and second-order neighborhood information respectively, ultimately outputting a high-dimensional feature representation for each node. The first graph attention layer has an input feature dimension of 14, an output feature dimension of 64, and 4 attention heads; the second graph attention layer has an input feature dimension of 64, an output feature dimension of 128, and 4 attention heads. The output is a high-dimensional feature matrix of the nodes.

[0078] In this embodiment, in step 4), the updated node features are input into the classification layer. After fully connected mapping and Softmax function calculation, the category corresponding to the maximum probability is taken as the lithology classification result of each node. The specific process is as follows:

[0079] 4.1) High-dimensional features of nodes output by the graph neural network Input a fully connected layer network, and set the output dimension of the fully connected network to the total number of lithology categories C = 6;

[0080] 4.2) The output of the fully connected network is calculated using the Softmax function to obtain the probability distribution of each node belonging to different lithologies, as shown in the formula:

[0081]

[0082] Among them, Z i，c This is the fully connected output value of category c corresponding to node i;

[0083] 4.3) Take the index corresponding to the maximum probability and map the index to a specific lithology name as the classification result output.

[0084] In this embodiment, in step 5), the lithology logging data collected in step 1) is used as supervised label training for the model. The model parameters are optimized using the cross-entropy loss function, and the lithology classification accuracy of the model in the cross-well scenario is evaluated through a validation set. Specifically, this includes:

[0085] 5.1) A weighted cross-entropy loss function is used as the training objective to address the lithological class imbalance problem. The formula is:

[0086]

[0087] Where N is the number of samples in the batch, C=6 is the total number of lithological categories, and y i，c The true label for sample i belonging to category c. w represents the probability value predicted by the model that sample i belongs to class c. c The weight coefficient for category c is calculated by normalizing the coefficients based on the reciprocal of the number of samples in each category in the training set.

[0088] 5.2) The Adam optimizer was used to iteratively update the model parameters, with a learning rate of 0.001 and a weight decay of 5e-4. The optimized model parameters included the weight matrix W and attention vector a of the graph attention layer, as well as the weights and biases of the fully connected network in the classification layer.

[0089] 5.3) A micro-batch sampling strategy is used for training. In each iteration, 512 nodes and their corresponding second-order neighborhood subgraphs are randomly selected to participate in the calculation, so as to reduce the computational cost of large-scale multi-well maps.

[0090] 5.4) On the validation set, the overall accuracy, macro F1 score and recall rate for each lithology class are used as evaluation indicators. If the overall accuracy of the validation set does not improve for 15 consecutive rounds, the training is terminated early and the optimal model parameters are saved.

[0091] 5.5) To verify the effectiveness of this invention, data from 8 wells were used as the training set and data from 2 wells were used as the test set. Comparison methods included: the GAT model using only intra-well edges (without inter-well edges), the traditional random forest model, and the multilayer perceptron model.

[0092] Table 2 shows a comparison of the classification results of different methods on test wells.

[0093] Table 2 Comparison of Lithological Classification Results by Different Methods

[0094]

[0095] As shown in Table 2, the overall accuracy of the method of this invention reaches 87.3%, and the macroscopic F1 score reaches 82.6%, both significantly better than the comparative methods. The method of this invention achieves the best performance in recall for all types of lithology. Particularly noteworthy is that for siltstone, fine sandstone, and rare lithologies such as coarse sandstone and gravelly sandstone with smaller sample sizes, the recall rates of the method of this invention reach 78.4%, 80.5%, 67.8%, and 62.5%, respectively, representing improvements of 7.9%, 6.9%, 9.4%, and 10.8% compared to the GAT model using only well edges. This result indicates that the introduction of well edges effectively transmits geological similarity information between different wells, significantly improves the identification ability of rare lithologies, and alleviates model bias caused by lithology class imbalance.

Claims

1. A lithological intelligent classification method based on a multi-well map neural network, characterized in that, Includes the following steps: 1) Collect drilling engineering parameters and lithological logging data from multiple wells, perform data cleaning and standardization, and construct graph nodes for each depth point of each well. The node features are composed of the drilling engineering parameters corresponding to that depth point, and the lithological logging data is used as node training labels. 2) Construct a graph structure and define two types of edges: the first type is intra-well edges, which connect adjacent depth nodes in the same well; the second type is inter-well edges, which connect nodes with similar logging response characteristics in different wells. 3) Construct a graph neural network model, taking the node features from step 1) and the graph structure from step 2) as input, and using a graph attention layer to aggregate neighborhood node information and update the node feature representation; 4) Input the updated node features into the classification layer. After fully connected mapping and softmax function calculation, take the category corresponding to the maximum probability as the lithology classification result of each node. 5) Use the lithology logging data collected in step 1) as supervised label training model, optimize model parameters with cross-entropy loss function, and evaluate the lithology classification accuracy of the model in cross-well scenarios through validation set.

2. The method according to claim 1, characterized in that, The drilling engineering parameters in step 1) include: natural gamma ray (GR), resistivity (RT), sonic transit time (AC), compensated density (DEN), compensated neutron (CNL), borehole diameter (CAL), drilling time (ROP), hook load (HKLD), torque (TORQUE), drilling fluid density (MDIA), drilling fluid displacement (MFIA), standpipe pressure (SPPA), rotary table speed (RPMB), and drilling pressure (WOBB); the lithology logging data is lithology LITH.

3. The method according to claim 1, characterized in that, The data preprocessing in step 1) includes: 1.1) Detect outliers in the collected drilling parameters and replace outliers with the median value within a sliding window; 1.2) Standardize all parameters to eliminate scale differences between different units; 1.3) Sampling at fixed depth intervals, the continuous depth data is discretized into a sequence of depth points, each depth point is a graph node, and the node feature vector is composed of all the parameters of that depth point.

4. The method according to claim 1, characterized in that, The method for constructing the graph structure in step 2) includes: 2.1) Well edge construction method: For each well, connect two adjacent nodes in depth with an edge and set the weight of the edge; 2.2) Inter-well edge construction method: Calculate the similarity between features of different well nodes, use cosine similarity as the similarity metric, set a similarity threshold, select all cross-well nodes with similarity exceeding the threshold for each node to establish an inter-well edge, and set the weight of the inter-well edge to the cosine similarity value between the two nodes.

5. The method according to claim 1, characterized in that, The graph neural network in step 3) uses a graph attention network (GAT), specifically including: 3.1) For each node, calculate its attention coefficient with all neighboring nodes, including intra-well neighbors and inter-well neighbors. The formula for calculating the attention coefficient is: Among them, h i Let be the input feature vector of node i, W be the trainable weight matrix, a be the trainable attention vector, and || denote the vector concatenation operation. Let i be the set of neighboring nodes of node i; 3.2) Use attention coefficients to perform weighted summation of the features of neighboring nodes and update the node feature representation: Where σ is the ELU activation function; 3.3) Stack two graph attention layers to aggregate first-order and second-order neighborhood information respectively, and finally output a high-dimensional feature representation of each node.

6. The method according to claim 1, characterized in that, The specific process of the classification layer in step 4) is as follows: 4.1) Input the high-dimensional node features output by the graph neural network into a fully connected network layer, and set the output dimension of the fully connected network to the total number of lithology categories; 4.2) The output of the fully connected network is calculated using the Softmax function to obtain the probability distribution of each node belonging to various lithologies; 4.3) Take the index corresponding to the maximum probability and map the index to a specific lithology name as the classification result output.

7. The method according to claim 1, characterized in that, The model training in step 5) includes: 5.1) A weighted cross-entropy loss function is used as the training objective to address the lithological class imbalance problem. The formula is: Where N is the number of samples in the batch, C is the total number of lithological categories, and y i，c The true label for sample i belonging to category c. w represents the probability value predicted by the model that sample i belongs to class c. c The weight coefficient for category c is calculated by normalizing the coefficients based on the reciprocal of the number of samples in each category in the training set. 5.2) The Adam optimizer is used to iteratively update the model parameters. The optimized model parameters include the weight matrix W of the graph attention layer, the attention vector a, and the weights and biases of the fully connected network in the classification layer. 5.3) A micro-batch sampling strategy is used for training. In each iteration, a batch of nodes and their corresponding neighborhood subgraphs are randomly selected to participate in the calculation, so as to reduce the computational overhead of large-scale multi-well maps. 5.4) On the validation set, the overall accuracy, macro F1 score and recall rate for each lithology class are used as evaluation indicators. If the overall accuracy of the validation set does not improve for several consecutive training rounds, the training is terminated early and the optimal model parameters are saved.

Images