High-precision logging lithology intelligent identification method and system based on DenseNet-Transform depth fusion, and storage medium

Abstract

The invention discloses a high-precision logging lithology intelligent identification method based on DenseNet-Transformer depth fusion. The method comprises the following steps: step 1, carrying out data preprocessing; and step 2, constructing a deep learning model in which a DenseNet mechanism and a Transform mechanism are fused. According to the method disclosed by the invention, efficient feature multiplexing is realized by using the dense connection characteristic of the DenseNet through a deep learning technology, and the global feature modeling capability is enhanced in combination with the Transform network, so that the lithology identification accuracy and generalization capability are remarkably improved. The method is suitable for intelligent analysis of complex logging data, especially shows excellent performance in the aspects of high-dimensional data processing and fine feature extraction, greatly improves the recognition difficulty of a traditional method in a complex stratum structure, and provides important technical support and efficient decision basis for the fields of oil-gas exploration, geological survey and the like.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent well logging data analysis and lithology identification in geological exploration, oil and gas development and groundwater resource management, and in particular to a well logging lithology identification method, system and storage medium based on the fusion of DenseNet and Transformer in this field. Background Art

[0002] Well logging lithology identification plays a vital role in key areas such as geological exploration, oil and gas development, and groundwater resource management. Lithology identification not only directly affects the accuracy of subsequent geological interpretation, reservoir evaluation, and resource assessment, but is also an important basis for formulating mining plans and optimizing production strategies. With the continuous advancement of exploration technology and the increasing abundance of geological data, how to efficiently and accurately extract lithology information from well logging data has become a core issue that geologists and engineers need to solve. Well logging data is the most direct and richest source of information on underground geological conditions. The physical quantities it contains (such as resistivity, acoustic wave velocity, natural gamma, etc.) are closely related to rock type, porosity, fluid properties, etc. Therefore, through advanced lithology identification technology, it is possible to achieve a detailed description of the complex underground geological structure, provide a scientific basis for oil and gas exploration and development, and also contribute to the efficient management and sustainable utilization of groundwater resources. In addition, high-precision lithology identification can also promote the in-depth development of geological science research and promote the continuous improvement of earth science theory.

[0003] Although the importance of logging lithology identification is self-evident, it still faces many challenges in practical applications. First, the logging data itself is highly complex and diverse. The logging response characteristics of different formations and different lithologies often overlap and cross, which makes accurate identification difficult. Well logging data contains a variety of physical quantities, such as resistivity, acoustic velocity, natural gamma, etc. These physical quantities are related to a variety of geological factors such as rock type, porosity, fluid properties, and there are complex nonlinear relationships between them. How to effectively extract features that play a key role in lithology identification from high-dimensional and complex logging data is a major challenge facing current technology. Secondly, traditional manual analysis methods rely on the experience and subjective judgment of geological experts, which is not only time-consuming and labor-intensive, but also easily affected by human factors, resulting in uncertainty and limitations in the identification results. In addition, with the increase in exploration depth and the complexity of the exploration environment, the quality and integrity of logging data are also facing challenges, further increasing the difficulty of lithology identification. The generalization ability of existing models is insufficient. Traditional machine learning algorithms and simple neural network models often have insufficient generalization ability when processing complex logging data. These models have difficulty adapting to data changes under different geological conditions, resulting in decreased recognition accuracy in applications in unknown or new areas.

[0004] In the field of deep learning, although convolutional neural networks (CNN) and recurrent neural networks (RNN) have achieved certain success in lithology identification, they still face some challenges. First, as the depth of the network increases, traditional deep neural networks are prone to gradient vanishing or gradient exploding problems, which makes model training difficult and unable to effectively learn the deep features of the data. Secondly, the overfitting problem is also common in deep learning. Especially when the amount of data is limited, the model is prone to overfitting the training data, thereby losing the ability to generalize to new data. Although deep learning can automatically learn feature representations from data, when processing high-dimensional and complex logging data, how to ensure that the model can accurately identify and enhance key features and avoid being interfered by non-key features is still a key problem facing current technology.

[0005] In recent years, combining deep learning methods with other models, such as deep fusion models based on DenseNet and Transformer, has begun to become an effective way to solve this problem. DenseNet enhances feature reuse and gradient transfer through dense connections, and can effectively extract local features in logging data; while Transformer performs well in processing time series data, can capture long-range dependencies in depth series, and is particularly suitable for modeling logging depth data. By combining the advantages of these two models, not only can the accuracy of lithology identification be improved, but also the complex nonlinear relationships and multiple feature dependencies in logging data can be better handled. At present, the model based on the combination of DenseNet and Transformer is still in the exploratory stage, but its application in lithology identification shows great potential. Through effective feature fusion and global modeling, this type of model can achieve efficient processing of complex logging data without the need for manual feature extraction, thereby overcoming the limitations of traditional methods in terms of accuracy and generalization ability. Summary of the invention

[0006] The technical problem to be solved by the present invention is to provide a high-precision well logging lithology intelligent identification method, system and storage medium based on DenseNet-Transformer deep fusion. This method significantly improves the accuracy of lithology identification by combining the local feature extraction capability of DenseNet and the global modeling capability of Transformer, especially when processing complex well logging data, and can effectively capture the long-range dependency characteristics of the data and the relationship between depth sequences.

[0007] The present invention adopts the following technical solution:

[0008] A high-precision well logging lithology intelligent identification method based on DenseNet-Transformer deep fusion, the improvement of which is that it includes the following steps:

[0009] Step 1, data preprocessing:

[0010] Normalize the parameter column values ​​and standardize the depth column data;

[0011] The normalization formula is:

[0012]

[0013] In the above formula, x is the original input data, x norm is the normalized output data, x min is the minimum value of the original input data, x max is the maximum value of the original input data;

[0014] The standardization formula is:

[0015]

[0016] In the above formula, x is the original input data, z is the normalized output data, μ is the mean of the original input data, and σ is the standard deviation of the original input data;

[0017] Step 2: Build a deep learning model that integrates DenseNet and Transformer mechanisms:

[0018] The dense block DenseNet consists of multiple convolutional layers, where the input of each convolutional layer is the concatenation of the outputs of all previous layers in the channel dimension, and the output of each layer is the cascade of the outputs of all previous layers; each convolutional layer in the dense block is followed by a batch normalization and a ReLU activation function;

[0019] In the dense block DenseNet, suppose there is a network consisting of L layers, and the input x of the Lth layer is l is concatenated from the outputs of all previous layers:

[0020] x l =H l ([x0,x1,…,x l-1 ])

[0021] In the above formula, H l is the nonlinear transformation function of the lth layer, [x0,x1,…,x l-1 ] means concatenating the outputs of all previous layers together;

[0022] The transition layer is located between two dense blocks and consists of a convolutional layer and a pooling layer;

[0023] The attention Zout is expressed as:

[0024] Zout=X·sigmoid(W2·ReLU(W1·X))

[0025] In the above formula, X is the input feature, W1 and W2 are the weight matrices of linear transformation, and Sigmoid() is the Sigmoid activation function;

[0026] The calculation process of Transformer is described by the following formula:

[0027] Q=XW Q ,K=XW K ,V=XW V

[0028] In the above formula, X is the input feature, W Q is the weight matrix of query Q, W K is the weight matrix of key K; W V is the weight matrix of value V;

[0029] Calculate the attention weight through the self-attention mechanism:

[0030]

[0031] In the above formula, d k is the dimension of the key vector, and the softmax function is used to normalize the weights;

[0032] Transformer-Encoder structure: By stacking multiple Transformer-Encoder layers, the Transformer model is able to extract global information of input features. Each Encoder layer consists of a self-attention layer and a feed-forward neural network FFN:

[0033] Encoder=FFN(Attention(Q,K,V))

[0034] In the above formula, FFN consists of two fully connected layers and ReLU activation function, and Attention() is the attention mechanism function;

[0035] In classification tasks, the cross entropy loss function Loss is used to measure the difference between the predicted output and the true label:

[0036]

[0037] In the above formula, C is the number of categories, y i is the unique hot encoding of the true label, is the model's predicted probability for the ith category.

[0038] A high-precision well logging lithology intelligent identification system based on DenseNet-Transformer deep fusion, the improvement of which lies in comprising: a processor; a memory storing executable instructions of the processor; the processor being configured to execute the steps of the above method by executing the executable instructions.

[0039] A computer-readable storage medium is used to store a program, wherein the improvement lies in that the steps of the above method are implemented when the program is executed.

[0040] The beneficial effects of the present invention are:

[0041] The method disclosed in the present invention uses deep learning technology, utilizes the dense connection characteristics of DenseNet to achieve efficient feature reuse, and combines the Transformer network to enhance the global feature modeling capability, significantly improving the accuracy and generalization ability of lithology identification. This method is suitable for intelligent analysis of complex well logging data, especially in high-dimensional data processing and fine feature extraction. It has shown excellent performance, greatly improving the recognition difficulties of traditional methods in complex stratigraphic structures, and providing important technical support and efficient decision-making basis for oil and gas exploration, geological surveys and other fields.

[0042] The method disclosed in the present invention aims to solve the problems of low accuracy and insufficient generalization ability in traditional methods and existing deep learning models in well logging lithology identification, and provide more efficient and accurate technical support for geological exploration, oil and gas development, and groundwater resource management. DenseNet achieves efficient reuse of features and effective transmission of gradients through dense connections, effectively alleviating the gradient vanishing problem in traditional neural networks; while the Transformer network can capture long-range dependencies in the data, automatically learn and strengthen the weights of key features through the self-attention mechanism (Self-Attention), so that the model pays more attention to feature information that has an important impact on lithology identification. The introduction of Transformer enhances the model's ability to model global features, and is particularly suitable for processing well logging data with long time series dependencies. This combination not only improves the expressiveness and accuracy of the model, but also enhances the model's adaptability to complex data, and is particularly suitable for processing high-dimensional and complex well logging data.

[0043] The method disclosed in the present invention has shown significant beneficial effects in many aspects. First, the recognition accuracy is significantly improved. By integrating the dense connection characteristics of DenseNet and the global modeling capabilities of the Transformer network, it can automatically learn and strengthen key features from complex logging data, effectively overcoming the shortcomings of traditional methods in feature extraction and key feature identification, thereby significantly improving the accuracy and precision of lithology identification. This is of great significance for decision-making in the fields of geological exploration, oil and gas development, etc., and can reduce the risk of misjudgment and improve the reliability of resource assessment. Secondly, the generalization ability is enhanced. The dense connection structure of DenseNet effectively alleviates the gradient disappearance problem in deep learning, so that the model can more effectively learn the deep features of the data during the training process. At the same time, the introduction of the Transformer network enables the model to automatically adjust the degree of attention to key features, and enhances the model's adaptability to data under different geological conditions. This combination not only improves the recognition accuracy of the model in known areas, but also significantly enhances its generalization ability in unknown or new areas, providing more robust technical support for practical applications. Moreover, it can improve data processing efficiency and automatically extract and learn feature representation from raw logging data without manual intervention, greatly improving the efficiency and automation level of data processing. In addition, the real-time performance of the system has also been optimized, and it can complete the analysis and identification of a large amount of logging data in a relatively short time, providing a strong guarantee for rapid decision-making.

[0044] Here, DenseNet is responsible for extracting local features, while the Transformer network captures global features and long-range dependencies through the self-attention mechanism. The combination of the two enables this method to take into account both local and global information when processing complex logging data, improving the accuracy and generalization ability of lithology identification, especially showing superior performance when facing high-dimensional and complex geological environments.

[0045] In summary, the method disclosed in the present invention has shown significant beneficial effects in improving recognition accuracy, enhancing generalization ability and improving data processing efficiency, and has provided more efficient and accurate technical support for the fields of geological exploration, oil and gas development and groundwater resource management. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is the overall architecture diagram of the deep learning model;

[0047] Figure 2 It is a curve chart of the changes in training loss and test loss;

[0048] Figure 3 It is a curve chart of the changes in training accuracy and test accuracy;

[0049] Figure 4This is a comparison chart of the recognition accuracy of various recognition methods;

[0050] Figure 5 It is a recognition result diagram of the method disclosed in the present invention. DETAILED DESCRIPTION

[0051] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0052] Example 1. This example discloses a high-precision well logging lithology intelligent identification method based on DenseNet--Transformer deep fusion, which integrates the dense connection structure of DenseNet and the global modeling capability of the Transformer network. Efficient reuse of features and effective transfer of gradients are achieved through DenseNet, while the self-attention mechanism (Self-Attention) in the Transformer network is used to automatically learn and strengthen the weights of key features, so that the model can focus more on feature information that is critical to lithology identification. This combination not only overcomes the limitations of traditional methods in processing complex well logging data, but also significantly improves the expression ability and recognition accuracy of the model, providing more reliable technical support for geological exploration, oil and gas development and other fields. It includes the following steps:

[0053] Step 1: Preprocessing of deep well data:

[0054] The ranges of the exploration well depth and various parameters vary greatly, so the parameter column values ​​are normalized and the depth column data is standardized to prevent a certain feature from having too much influence on the model, to ensure that each feature has a similar influence on the model, and to improve the stability and convergence speed of the later model;

[0055] Normalization: Each parameter column in the logging data (such as resistivity, acoustic velocity, etc.) is normalized to ensure that the data is within the same numerical range and to avoid training instability due to dimensional differences. The normalization formula is:

[0056]

[0057] In the above formula, x is the original input data, x norm is the normalized output data, x min is the minimum value of the original input data, x max is the maximum value of the original input data;

[0058] Standardization: Standardize the depth column data and convert it into a normal distribution with a mean of 0 and a standard deviation of 1. The standardization formula is:

[0059]

[0060] In the above formula, x is the original input data, z is the normalized output data, μ is the mean of the original input data, and σ is the standard deviation of the original input data;

[0061] The logging data of this embodiment comes from the Hugoton and Panoma oil and gas fields in North America, including 8 wells with clear lithology, a total of 3232 samples, and a sampling interval of 15.24 cm. This embodiment selects five logging characteristic variables, namely natural gamma (GR), deep induction (ILDlog10), lithology density (PE), neutron density porosity (PHIND) and average neutron density porosity (DeltaPHI). Five logging characteristic variables and different logging methods have different response characteristics for the same lithology. As shown in Table 1 below. The data are divided into a training set (85%), a test set (10%) and a validation set (5%) in order of depth.

[0062] Table 1: Well logging data table

[0063]

[0064] Table 2: Rock categories

[0065]

[0066] Step 2: Build a deep learning model that combines DenseNet and Transformer:

[0067] like Figure 1 As shown in the figure, a deep fusion network based on DenseNet and Transformer is designed. DenseNet achieves efficient reuse of features and effective transmission of gradients through its densely connected structure, which helps to alleviate the problem of gradient disappearance. The self-attention mechanism in the Transformer network is used to automatically learn and strengthen the weights of key features, so that the model can more accurately focus on the feature information that is critical to lithology identification.

[0068] The dense block DenseNet consists of multiple convolutional layers, where the input of each convolutional layer is the concatenation of the outputs of all previous layers in the channel dimension, and the output of each layer is the cascade of the outputs of all previous layers. Each convolutional layer in the dense block is followed by a batch normalization and a ReLU activation function;

[0069] In the dense block DenseNet, suppose there is a network consisting of L layers, and the input x of the Lth layer is l is concatenated from the outputs of all previous layers:

[0070] x l =H l ([x0,x1,…,x l-1 ])

[0071] In the above formula, H l is the nonlinear transformation function of the lth layer, [x0,x1,…,x l-1 ] means concatenating the outputs of all previous layers together;

[0072] The transition layer is located between two dense blocks and consists of a convolutional layer and a pooling layer.

[0073] The attention Zout is expressed as:

[0074] Zout=X·sigmoid(W2·ReLU(W1·X))

[0075] In the above formula, X is the input feature, W1 and W2 are the weight matrices of linear transformation, and Sigmoid() is the Sigmoid activation function;

[0076] In order to effectively model the deep sequence relationship in the logging data, the Transformer network is integrated into the DenseNet structure. The self-attention mechanism of the Transformer can capture long-range dependencies and fully explore the complex relationships between deep sequences.

[0077] The input of Transformer is the feature map extracted from DenseNet, which is processed by the Embedding Layer and then sent to the Transformer model. The self-attention mechanism of Transformer generates weighted feature representation by calculating the correlation between input features.

[0078] The calculation process of Transformer is described by the following formula:

[0079] Q=XW Q ,K=XW K ,V=XW V

[0080] In the above formula, X is the input feature, W Q is the weight matrix of query Q, W K is the weight matrix of key K; W V is the weight matrix of value V;

[0081] Calculate the attention weight through the self-attention mechanism:

[0082]

[0083] In the above formula, d k is the dimension of the key vector, and the softmax function is used to normalize the weights so that the final output can focus on important features;

[0084] Transformer-Encoder structure: By stacking multiple Transformer-Encoder layers, the Transformer model can extract global information of input features. Each Encoder layer consists of a self-attention layer and a feed-forward neural network (FFN):

[0085] Encoder=FFN(Attention(Q,K,V))

[0086] In the above formula, FFN consists of two fully connected layers and ReLU activation function, and Attention() is the attention mechanism function;

[0087] In classification tasks, the cross entropy loss function Loss is used to measure the difference between the predicted output and the true label:

[0088]

[0089] In the above formula, C is the number of categories, y i is the unique hot encoding of the true label, is the model's predicted probability for the ith category.

[0090] The model is trained and verified, and the forecast performance is evaluated, including indicators such as accuracy, stability, and real-time performance. Based on the evaluation results, the model is iteratively optimized, and the model structure and parameter settings are adjusted to meet the needs of different application scenarios.

[0091] During the model training process, the cross-validation strategy is used to divide the data set into training set, test set and validation set to ensure the generalization ability of the model. During the training process, the Adam optimizer and cross entropy are used as the loss function to optimize the model parameters. The training results are shown in Figure 2. Figure 2 and Figure 3 shown.

[0092] Table 3: Comparison of prediction results

[0093]

[0094] Figure 4 This is a comparison chart of the recognition accuracy of various recognition methods; Figure 5It is a recognition result diagram of the method disclosed in the present invention.

[0095] This embodiment also discloses a high-precision well logging lithology intelligent identification system based on DenseNet-Transformer deep fusion, comprising: a processor; a memory storing executable instructions of the processor; the processor is configured to execute the steps of the above method by executing the executable instructions. And a computer-readable storage medium for storing a program, which implements the steps of the above method when executed.

Claims

1. A high-precision well logging lithology intelligent identification method based on DenseNet-Transformer deep fusion, characterized in that: The steps include: Step 1, data preprocessing: Normalize the parameter column values ​​and standardize the depth column data; The normalization formula is: In the above formula, x is the original input data, x norm is the normalized output data, x min is the minimum value of the original input data, x max is the maximum value of the original input data; The standardization formula is: In the above formula, x is the original input data, z is the normalized output data, μ is the mean of the original input data, and σ is the standard deviation of the original input data; Step 2: Build a deep learning model that integrates DenseNet and Transformer mechanisms: The dense block DenseNet consists of multiple convolutional layers, where the input of each convolutional layer is the concatenation of the outputs of all previous layers in the channel dimension, and the output of each layer is the cascade of the outputs of all previous layers; each convolutional layer in the dense block is followed by a batch normalization and a ReLU activation function; In the dense block DenseNet, suppose there is a network consisting of L layers, and the input x of the Lth layer is l is concatenated from the outputs of all previous layers: x l =H l ([x0,x1,…,x l-1 ]) In the above formula, H l is the nonlinear transformation function of the lth layer, [x0,x1,…,x l-1 ] means concatenating the outputs of all previous layers together; The transition layer is located between two dense blocks and consists of a convolutional layer and a pooling layer; The attention Zout is expressed as: Zout=X·sigmoid(W2·ReLU(W1·X)) In the above formula, X is the input feature, W1 and W2 are the weight matrices of linear transformation, and Sigmoid() is the Sigmoid activation function; The calculation process of Transformer is described by the following formula: Q=XW Q ,K=XW K ,V=XW V In the above formula, X is the input feature, W Q is the weight matrix of query Q, W K is the weight matrix of key K; W V is the weight matrix of value V; Calculate the attention weight through the self-attention mechanism: In the above formula, d k is the dimension of the key vector, and the softmax function is used to normalize the weights; Transformer-Encoder structure: By stacking multiple Transformer-Encoder layers, the Transformer model can extract global information of input features. Each Encoder layer consists of a self-attention layer and a feed-forward neural network FFN. composition: Encoder=FFN(Attention(Q,K,V)) In the above formula, FFN consists of two fully connected layers and ReLU activation function, and Attention() is the attention mechanism function; In classification tasks, the cross entropy loss function Loss is used to measure the difference between the predicted output and the true label: In the above formula, C is the number of categories, y i is the unique hot encoding of the true label, is the model's predicted probability for the ith category.

2. A high-precision well logging lithology intelligent identification system based on DenseNet-Transformer deep fusion, characterized in that: include: processor; A memory in which executable instructions of the processor are stored; the processor is configured to perform the steps of the method of claim 1 by executing the executable instructions.

3. A computer-readable storage medium for storing a program, characterized in that: When the program is executed, the steps of the method according to claim 1 are implemented.

Images