Lithology identification method based on logging curve morphological characteristics
Through the lithology recognition method based on the morphological characteristics of the logging curve, small-scale and multi-scale models are constructed using ResNet and Transformer networks, the problems of high computing resources, strong image dependence and incomplete data sets in the existing technology are solved, and efficient and accurate lithology recognition is achieved.
Patent Information
- Application Number
- CN202510454039.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-11
- Publication Date
- 2025-07-18
AI Technical Summary
The prior art has problems such as relying on image data in lithology recognition, high computing resource requirements, high sensitivity to image quality, incomplete data sets and imbalance of sample categories, resulting in low recognition accuracy and lack of generalization.
A lithology recognition method based on the morphological characteristics of the well logging curve was constructed. By constructing a centering section logging curve data set, pre-training was used using ResNet18 and ResNet50 network structures, shallow network parameters were frozen, and small-scale and multi-scale lithology recognition was performed in combination with Transformer network model, similar image samples were generated, transfer learning and global modeling were performed.
The efficiency and accuracy of lithology recognition are improved, real-time or near-real-time lithology recognition is achieved, computing resource requirements are reduced, the robustness and automation of the model are enhanced, lithology recognition errors are overcome, and identification accuracy and stability are improved.
Smart Images
Figure CN120339795A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the technical field of lithology interpretation of logging data. Specifically, it relates to a lithology identification method based on the morphological characteristics of logging curves. Background Technique
[0002] Lithology identification is an important part of reservoir evaluation and the basis for reservoir description and obtaining reservoir parameters. An accurate lithology identification model is crucial in the process of oil and gas development. Deep learning uses multi-layer networks to process data to extract features. With the gradual development of deep learning theory and technology, researchers have combined intelligent methods with logging. Starting from the morphological combination characteristics of logging curve images corresponding to different lithologies, the lithology identification problem is transformed into an image feature extraction and pattern recognition problem, and a lithology identification model based on image features is constructed.
[0003] The non-patent literature (Zheng Zunkai. Research on cuttings image recognition under deep learning model [D]. Yangtze University, 2019.) proposed a cuttings identification method based on Triplets and Resnet18 convolutional neural networks. It uses a triple input method and uses three identical convolutional networks to extract features and map them to a metric space for classification. The non-patent literature (Li Na, Gu Qing, Jiang Feng, etc. A feature representation method of microscopic sandstone images based on convolutional neural network [J]. Journal of Software, 2020, 31(11):
[0004] 3621 - 3639. LI N, GU Q, JIAN F, et al. Feature representation method of microscopic sandstone images based on convolutional neural network [J]. Journal of Software, 2020, 31(11): 3621 - 36 39.) proposed a feature representation network FeRNet based on convolutional neural network for small-scale data sets to effectively capture the semantic information of microscopic sandstone images and improve the feature representation ability of microscopic sandstone images. The non-patent literature (Li Teng. Research on rock thin section image recognition method based on VGG [D]. Xi'an Shiyou University, 2023. DOI: 10.27400 / d.cnki.
[0005] gxasc.2023.000396.) is based on a deep learning method that uses the VGG neural network combined with rock image features and network structure features to improve feature extraction and network accuracy.
[0006] Compared with the lithology identification model based on logging numerical characteristics, the model based on image data has made great progress. It can fully consider the continuity between data during the data processing, the model construction work has stronger logic, and the lithology identification result of the model is more accurate. However, some images have the disadvantages of low identification accuracy, high requirements for sample homogeneity, single algorithm, and lack of practicality of research results.
[0007] Non-patent literature (Zhang Ye, Li Mingchao, Han Shuai. Automatic Lithology Identification and Classification Method Based on Deep Learning of Rock Images [J]. Acta Petrologica Sinica, 2018, 34(02): 333-342.) Based on the Inception-v3 model, through the way of transfer learning, the images of three kinds of rocks, granite, phyllite and breccia, were identified and classified. For the images in the training set, the model can correctly classify with a probability of more than 90%. The following is the core content of this technology:
[0008] Dataset: The researchers collected 173 images of granite, 152 images of phyllite and 246 images of breccia, and divided them into training set and test set.
[0009] Model construction: Use Google's Inception-v3 pre-trained model, which has been trained on a large number of images and contains about 25 million parameters. The researchers used this model to extract the features of rock images.
[0010] Transfer learning: Apply the Inception-v3 model to rock images, use its convolutional layer and pooling layer to extract image features, and then use these features to train a new Softmax neural network for classification.
[0011] Training and testing: The model is trained on the training set and evaluated on the test set.
[0012] However, this technology has the following disadvantages: 1. Mainly rely on image data, there are limitations in the direct integration of logging data, and it may not be directly applied to logging interpretation work.
[0013] 2. Depend on the quality and diversity of training data, a large amount of image data collection and processing are required, and the cost of collecting high-quality image data is high.
[0014] 3. There are problems with low identification accuracy and lack of generalization for individual lithologies.
[0015] Non-patent literature (Xu Zhenhao, Ma Wen, Lin Peng, et al. Intelligent Lithology Identification Based on Rock Image Transfer Learning [J]. Journal of Basic Science and Engineering, 2021, 29(05): 1075-1092. DOI: 10.16058 / j.issn.1005-0930.2021.05.002.) proposed an intelligent lithology identification method combining object detection preprocessing and deep learning transfer models, which improved the training efficiency and identification accuracy of the model through transfer learning.
[0016] Object detection preprocessing: Use a deep supervised object detection network (DSOD) to detect rocks in the image, automatically crop out the rock area, and construct a high-quality rock image dataset.
[0017] Deep learning transfer model: Combine the ResNet network to construct a deep learning transfer model for rock images, and use the residual network to extract rock feature information.
[0018] Transfer learning training: Train the model by loading pre-trained weights to achieve intelligent lithology identification.
[0019] Results: Confusion matrix, accuracy (ACC), precision (P), recall (R), and F1 value were used as evaluation indicators for the model accuracy. The recognition accuracy was high and the stability of each type of rock recognition was good.
[0020] However, this technology has the following disadvantages: 1. The ResNet-101 model is relatively complex and requires a large amount of computing resources for training and inference, which limits the application of the model in resource-constrained environments.
[0021] 2. The model is highly sensitive to image quality and requires high-quality and diverse image data to improve the robustness of the model.
[0022] 3. The research uses an incomplete dataset, the number of sample categories is unbalanced, and there is a lack of universality. Summary of the Invention
[0023] In view of this, the present application provides a lithology identification method based on the morphological characteristics of logging curves to achieve refined lithology identification and improve the efficiency and accuracy of lithology identification.
[0024] To achieve the above object, the technical solution adopted by the present application is as follows: A lithology identification method based on the morphological characteristics of logging curves, including: S1: Construct a core section logging curve dataset; S2: Construct a straight line dataset according to the color arrangement law of logging curves; S3: Select ResNet18 and ResNet50 network architectures and build a pre-trained model based on the straight-line dataset; S4: Freeze the parameters of the shallow network of the pre-trained model and update the parameters of the deep network according to the core-taking section logging curve dataset to build a small-scale lithology identification model, where the small scale means that the sampling interval of the logging curve data is less than a certain threshold; S5: Replace the fully connected layer of the small-scale lithology identification model with a Transformer network model to build a multi-scale lithology identification model, where the multi-scale means that the sampling interval of the logging curve data can have multiple values; S6: Train the multi-scale lithology identification model on the training set to achieve the effect of global modeling.
[0025] Further, the specific content of S1 is as follows: S1.1: Crop the core-taking section logging curve images; S1.2: Label each cropped core-taking section image with a lithology label according to the core-taking lithology observation results corresponding to the depth of the cropped image in the core-taking section; S1.3: Perform 0-pixel filling on the cropped images with lithology labels; S1.4: Scale the filled images while keeping the aspect ratio unchanged to obtain the final core-taking section logging curve dataset.
[0026] Further, the specific content of S2 is as follows: S2.1: Establish five datasets: RBG-I, RBG-II, BRG-I, BRG-II, and RGB; S2.2: Based on the core-taking section logging curve dataset constructed in step S1, generate sample images according to the RBG color arrangement pattern. Induce the images in which the spatial distributions of the red and blue curves in the sample images are in the first 30% area of the track into the RBG-I dataset; induce the images in which the spatial distributions of the red and blue curves are in the 30% - 50% area of the track into the RBG-II dataset. Generate sample images according to the BRG color arrangement pattern. Induce the images in which the spatial distribution of the green curve in the sample images is in the 40% - 60% area of the track into the BRG-I dataset; induce the images in which the spatial distribution of the green curve is in the 60% - 100% area of the track into the BRG-II dataset. Induce all the images generated under the RGB color arrangement pattern into the RGB dataset. The track is the picture channel of the logging curve in the axial direction representing the physical quantity, and it is defined that the direction towards the increase of the physical quantity is the front of the track.
[0027] Further, S3 includes: S3.1: Select ResNet18 and ResNet50 network models; S3.2: Test both the ResNet18 and ResNet50 network models on the validation set and test set of the straight-line dataset; S3.3: Analyze the test results and use the winner of the two network models as the pre-trained model.
[0028] Furthermore, the said S4 includes: S4.1: Based on the pre-trained model, perform small-sample transfer learning according to the logging curve image dataset corresponding to different lithologies in the cored section; S4.2: During the transfer learning process, freeze the parameters of the Conv1 layer (the first convolutional layer) and Conv2 layer (the second convolutional layer) in the pre-trained model network and do not update them, retaining the parameters in the shallow network that can identify the common features of the image data; update the parameters of the deep network according to the logging curve image dataset corresponding to different lithologies in the cored section, so as to construct a small-scale lithology identification model.
[0029] Furthermore, the said S6 includes: S6.1: In the self-attention layer of the Transformer, project the input image data into vectors in three dimensions of Q, K, and V, and calculate the score S between different input vectors according to the vectors in the Q dimension and K dimension. The said S is obtained by multiplying the Q and K matrices; S6.2: Perform a normalization operation on S using different normalization functions to obtain S n : S n =FNormalize(S) S6.3: According to the normalization exponential function, turn S n into probability P, and describe the score between different vectors in percentage: P = FNormalized_Exponential(S n ) S6.4: Multiply Z and P to obtain the weighted numerical matrix Zout; S6.5: Describe the positional relationship in the data through the PE vector: Wherein logs pos represents the absolute position of the sample data in the input logging data (the pos th multi-dimensional logging data sample), i represents the dimension of the current position sample, d represents the dimension of the input data; S6.6: Embed the PE vector into the input data of the next layer network of the self-attention layer. During the training process, the Zout value passes the input vector and the position information between vectors to achieve the effect of global modeling.
[0030] Compared with the prior art, the beneficial effects of the present application are as follows: 1. Optimize the model and improve efficiency: The lithology identification model constructed based on image data has a simple data processing method and intuitive results, which is conducive to summarization and understanding. The input of the model is an image, enabling lithology identification even with only the archived map data, overcoming the problems of large lithology identification errors and low lithology interpretation accuracy caused by differences in logging values for the same lithology in different regions.
[0031] 2. Expand the dataset and improve the accuracy of data feature extraction: Generate a similar dataset by extracting the features of existing data to ensure that correct geological features can be extracted during the model training process, guarantee the convergence of model training, and complete high-precision identification using conventional logging data with lower cost.
[0032] 3. Construction of a multi-scale lithology identification model: Use the ResNet network to fully extract the core data features and geological understanding, and step by step construct a small-scale model that can complete the lithology identification task based on the image features of the sampling interval. Based on the Transformer network, serialize the lithology identification results of the sampling interval obtained by the small-scale lithology identification model, and correct and splice them according to the vertical depth to construct a multi-scale lithology identification model.
[0033] 4. Improve the degree of automation: By changing the input of the lithology identification task, the logic of the model construction process and the data processing process is made stronger, thereby improving the degree of automation of well logging interpretation work.
[0034] 5. The model of the present application has a fast identification speed in practical applications and can meet the real-time or near-real-time lithology identification requirements, which is a significant improvement in the computational efficiency compared to the prior art. Description of the Drawings
[0035] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings required in the embodiments. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as limiting the scope. For those of ordinary skill in the art, other relevant drawings can be obtained based on these drawings without creative efforts.
[0036] Figure 1 It is a flowchart of a lithology identification method based on the morphological features of well logging curves according to the present application; Figure 2 It is a schematic diagram of the construction of the well logging curve dataset for the cored section according to the present application; Figure 3 This is the construction flow chart of the small-scale lithology identification model of this application; Figure 4 This is the construction flow chart of the multi-scale lithology identification model of this application; Figure 5 This is the core sampling result and model identification result diagram in Well D4 in the specific implementation manner of this application. Specific implementation manner
[0037] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of this application. Obviously, the described embodiments are some, but not all, of the embodiments of this application.
[0038] As Figure 1 shown, a lithology identification method based on the morphological characteristics of logging curves includes: In this specific implementation manner, the lithology identification scheme is specifically described by taking breccia, dolomite, and limestone as examples.
[0039] S1: Construct a logging curve data set for the cored section; Furthermore, the specific content of S1 is as follows: S1.1: Crop the logging curve image of the cored section; The logging curve image is a data type sensitive to scale. Excessive magnification of the logging curve image will change image features such as the morphological changes of the curve and the interval size between different curves, resulting in the model being unable to learn the correct information in the original image. To obtain as many samples with clear image details as possible, set the page depth scale in the single-well diagram to 1:17, the page zoom to 211%, the thickness of the logging curve line to 4 pixels, and the cropping interval to 0.125 m. Crop the composite image of the three curves corresponding to different lithologies in the cored layer section to obtain sample image data with a single size of 500×110 pixels.
[0040] It should be noted that the number of logging curves can be set to be more or less than three as required, and the number of curves can be adjusted according to factors such as identification accuracy and calculation efficiency.
[0041] S1.2: Label each cropped image of the cored section with a lithology label according to the cored lithology observation result corresponding to the depth of the cropped image in the cored section; S1.3: Perform 0-pixel filling on the cropped images with lithology labels; Considering the size requirements of the network input data, perform 0-pixel filling on the sample image with a size of 500×110 pixels vertically until the image size reaches 500×500 pixels and then stop the 0 filling.
[0042] S1.4: Scale the filled image while keeping the aspect ratio unchanged to obtain the final core section log curve dataset.
[0043] Finally, scale the size of the filled sample image to 480×480 pixels while keeping the aspect ratio unchanged to obtain the final core section log curve dataset. The construction process of the core section log curve dataset is as Figure 2 shown.
[0044] S2: Construct a straight line dataset according to the color arrangement rule of the log curve; Furthermore, the specific content of S2 is as follows: S2.1: Establish five datasets of RBG-I, RBG-II, BRG-I, BRG-II, and RGB; By analyzing the color arrangement order and spatial distribution law of the three normalized log curves corresponding to different lithologies in the standard sample image data of breccia, dolomite, and limestone, it is found that there are 5 comprehensive color arrangement and spatial distribution patterns for different lithologies. Thus, five datasets of RBG-I, RBG-II, BRG-I, BRG-II, and RGB are established corresponding to 5 comprehensive classification mode distributions.
[0045] S2.2: Based on the core section log curve dataset constructed in step S1, generate sample images according to the RBG color arrangement mode. Induce the images in which the spatial distributions of the red and blue curves are in the first 30% area of the track in the sample images into the RBG-I dataset; induce the images in which the spatial distributions of the red and blue curves are in the 30% - 50% area of the track into the RBG-II dataset; generate sample images according to the BRG color arrangement mode. Induce the images in which the spatial distribution of the green curve is in the 40% - 60% area of the track in the sample images into the BRG-I dataset; induce the images in which the spatial distribution of the green curve is in the 60% - 100% area of the track into the BRG-II dataset; induce all the images generated under the RGB color arrangement mode into the RGB dataset; the track is the picture channel of the log curve in the axial direction representing the physical quantity (as shown within the red frame in Figure 2 ), and it is defined that the direction towards the increase of the physical quantity is the front of the track.
[0046] Specifically, conduct manual statistics on the core section log curve dataset constructed in step S1, and then generate corresponding pictures according to the features summarized manually by the computer, that is: the computer first generates a white sample size plate, and then prints the straight lines on the plate according to the distribution features. Here, a straight line sample dataset is generated.
[0047] By extracting and summarizing the logging response characteristics of different lithologies, similar image samples (linear datasets) are generated to construct a pre-training dataset, which is a novel data augmentation method that effectively solves the problem of insufficient core sample quantity.
[0048] S3: Select the ResNet18 and ResNet50 network architectures and construct a pre-training model based on the linear dataset; Further, the S3 includes: S3.1: Select the ResNet18 and ResNet50 network models; Network training parameter setting: Since the number of samples in the linear dataset is small and the sample image features are relatively simple, in this application, ResNet18 with a relatively shallow network depth and ResNet50 with a moderate network depth are selected to construct a pre-training model based on the linear dataset.
[0049] S3.2: Test both the ResNet18 and ResNet50 network models on the validation set and the test set; Network structure adjustment: Since the logging curve images corresponding to different lithologies in the linear dataset are divided into 5 categories, it is necessary to adjust the structure of the fully connected layer in the network. The original ResNet50 network has 2048 input features and 1000 output features in the fully connected layer. The number of output features of the original fully connected layer is adjusted to 256, and an additional fully connected layer is added, with 256 input features and 5 output features. The number of input features of the fully connected layer in the ResNet18 network is 512, and the number of output features is directly changed to 5.
[0050] The software experimental platform in this specific embodiment is built based on the PyTorch framework, and the hardware platform is a Windows platform equipped with an NVIDIA RTX 3090 graphics card. Two ResNet models are trained, the upper and lower limits of the hyperparameters are set, and the dichotomy method is used for recursive search to determine the optimal training parameters.
[0051] S3.3: Analyze the test results and take the winner among the two network models as the pre-training model.
[0052] The construction time consumption of two network models, ResNet18 and ResNet50, and the recognition accuracy on the test set are shown in Table 1. The results in the table show that when the same network hyperparameters are set, training a deeper network takes more time; on the straight-line dataset, there is no obvious positive correlation between the recognition accuracy of the pre-trained model and the network depth. Compared with ResNet50 with a shallower network structure, the pre-trained model generated under the ResNet18 network structure has a higher recognition accuracy on the three-straight-line test set. Therefore, the model generated under the ResNet18 network structure is selected as the pre-trained model for subsequent establishment of the lithology recognition model.
[0053] Table 1 Construction Time Consumption of the Pre-trained Model and Recognition Accuracy on the Test Set S4: Freeze the parameters of the shallow network of the pre-trained model, and update the parameters of the deep network according to the logging curve dataset of the cored section to construct a small-scale lithology recognition model, where the small scale refers to the sampling interval of the logging curve data (i.e., the cropping size of the logging curve image) being less than a certain threshold. In this specific embodiment, the sampling interval of the logging curve data is 0.125 m, and the threshold can be set to 0.13 m.
[0054] Furthermore, the S4 includes: S4.1: Based on the pre-trained model, perform small-sample transfer learning according to the logging curve image dataset corresponding to different lithologies in the cored section. In order to enable the small-scale lithology recognition model to comprehensively utilize features in various images such as color differences of different logging curves, arrangement differences of multiple curves in the same channel, and morphological changes of different curves to complete accurate lithology recognition. Based on the pre-trained model, small-sample transfer learning is completed according to the logging curve image dataset corresponding to 300 different lithologies in the cored section.
[0055] S4.2: During the transfer learning process, freeze the parameters of the Conv1 layer (the first convolutional layer) and the Conv2 layer (the second convolutional layer) in the pre-trained model network without updating, and retain the parameters in the shallow network that can identify the common features of the image data; update the parameters of the deep network according to the logging curve image dataset corresponding to different lithologies in the cored section to construct a small-scale lithology recognition model.
[0056] The shallow network structure of the pre-trained model extracts general image information such as pixel colors and image contours in the picture. During the transfer learning process, the parameters of the Conv1 and Conv2 layers in the pre-trained model network are frozen and not updated, and the parameters in the shallow network that can identify the common features of these image data are retained. According to the core sampling section dataset, the parameters of the deep network close to the output layer are updated so that the model can extract more image features from the well logging curve image dataset corresponding to different lithologies in the core sampling section and accurately complete lithology identification. The construction flow chart of the small-scale lithology identification model is as Figure 3 shown.
[0057] Adjustment of the training parameters of the small-scale lithology identification network model: The research section is mainly divided into three lithologies: breccia, dolomite, and limestone; correspondingly, the number of output features of the last fully connected layer in the pre-trained model network structure is adjusted from 5 to 3. When using the core images for few-shot learning on the pre-trained model after freezing the shallow parameters and adjusting the network structure, set the learning rate in the training loop to 0.002; the batch size to 4; select SGD (stochastic gradient descent) as the optimizer; and select the cross-entropy loss function as the loss function. In the experiment, two trainings are set. The number of epochs for the first training is set to be relatively large, 500 epochs. During the first model training, observe that the loss function of the model converges at the 300th epoch. The number of epochs for the second training is set to 300. After the training is completed, a small-scale lithology identification model is obtained.
[0058] S5: Replace the fully connected layer of the small-scale lithology identification model with a Transformer network model to construct a multi-scale lithology identification model. The multi-scale means that the sampling interval of the well logging curve data (i.e., the cropping size of the well logging curve image) can have multiple values; Replace the fully connected layer of the small-scale lithology identification model with a Transformer, serialize the image features extracted by the small-scale lithology identification model as the input of the Transformer network, and still use the core lithology as the label to complete the construction of the multi-scale lithology identification model (as Figure 4 shown).
[0059] In this specific embodiment, the cropping interval when cropping the well logging curve image of the core sampling section is 0.125m, which is less than the set threshold of 0.13m, and is a small-scale cropping. The multi-scale means that the cropping scale (or the sampling interval of the well logging curve data) is not necessarily 0.125m and can have multiple scales, including cases greater than the set threshold.
[0060] S6: Train the multi-scale lithology identification model on the training set to achieve the effect of global modeling.
[0061] The training set includes a straight line dataset and a well logging curve dataset.
[0062] Further, the S6 includes: S6.1: In the self-attention layer of the Transformer, project the input image data into vectors of three dimensions Q, K, and V, calculate the score S between different input vectors according to the Q and K vectors, and the S is obtained by multiplying the Q and K matrices; S6.2: Perform a normalization operation on S using different normalization functions to obtain S n : S n = FNormalize(S) S6.3: According to the normalized exponential function, change S n into probability P, and describe the score between different vectors in percentage: P = FNormalized_Exponential(S n ) S6.4: Multiply Z and P to obtain the weighted numerical matrix Zout; S6.5: Describe the positional relationship in the data through the PE vector: where logs pos represents the absolute position of the sample data in the input logging data (the pos th multi-dimensional logging data sample); i represents the dimension of the current position sample; d represents the dimension of the input data. In this embodiment, it can be understood as three different data corresponding to three different logging curves, so d is 3; S6.6: The PE vector is embedded in the input data of the next layer network of the self-attention layer. During the training process, the Zout value transmits the input vector and the positional information between vectors, achieving the effect of global modeling.
[0063] In another embodiment of the present application, the ResNet50 network model can also be used as the pre-training model, and the final multi-scale lithology recognition model is constructed based on this. The core of this alternative solution is to use deep learning algorithms, especially the ResNet50 residual network, to identify and predict the lithology of rock images. This method constructs a lithology recognition model for rock images and tunes and validates the parameters of the network model according to the defined loss function. This model can achieve the prediction of the lithology of rock images, and in some cases, the recognition accuracy can reach 75% - 90%.
[0064] Example: In the Daniudi research area, the above method is used to construct a lithology identification model. 183 image data that were not used when constructing the lithology identification model are used as the test set for constructing the model. Among them, there are a total of 66 well logging curve image samples corresponding to dolomite, 61 well logging curve image samples corresponding to limestone, and 55 well logging curve image samples corresponding to breccia.
[0065] In contrast, when dealing with similar data by traditional methods, the overall identification accuracy is usually low, generally around 70%. As Figure 5 shown, the overall identification accuracy of the lithology identification model constructed this time reaches 89% on the test set. Among them, the identification accuracy of dolomite test samples is 95%, the identification accuracy of limestone test samples is 88%, and the identification accuracy of breccia test samples is 83%. The overall identification accuracy in Well D4 is 85%. It significantly reduces the interference of human factors, improves the accuracy and stability of lithology identification, and better adapts to lithology identification work under complex geological conditions.
[0066] Therefore, this study starts from the response characteristics of well logging curves, extracts the well logging curve image characteristics corresponding to different lithologies, generates similar image samples to construct a pre-training data set, combines the core-taking results with well logging data, and constructs a well logging response image data set corresponding to lithology. On this basis, the ResNet network model is used to train the two data sets step by step. Taking the output of the pre-training model as the input and combining with the core-taking lithology labels, a refined lithology identification model is trained to achieve high-efficiency and high-precision lithology identification.
[0067] The above is only the specific implementation manner of this application, but the protection scope of this application is not limited thereto. Any person skilled in the art within the technical scope disclosed by this application can easily think of changes or substitutions, which should all be covered within the protection scope of this application. Therefore, the protection scope of this application should be subject to the protection scope of the claims.
Claims
1. A lithology identification method based on the morphological characteristics of logging curves, characterized in that, Including: S1: Construct a core section logging curve data set; S2: Construct a straight line data set according to the color arrangement rule of logging curves; S3: Select the ResNet18 and ResNet50 network structures, and construct a pre-trained model based on the straight line data set; S4: Freeze the shallow network parameters of the pre-trained model, and update the deep network parameters according to the core section logging curve data set to construct a small-scale lithology identification model, where the small scale refers to the sampling interval of the logging curve data being less than a certain threshold; S5: Replace the fully connected layer of the small-scale lithology identification model with a Transformer network model to construct a multi-scale lithology identification model, where the multi-scale means that the sampling interval of the logging curve data can have multiple values; S6: Train the multi-scale lithology identification model on the training set to achieve the effect of global modeling.
2. The lithology identification method based on the morphological characteristics of logging curves according to claim 1, characterized in that The specific content of S1 is as follows: S1.1: Crop the core section logging curve image; S1.2: Based on the core lithology observation results corresponding to the depth of the cropped image in the core section, assign a lithology label to each cropped core section image; S1.3: Perform 0-pixel filling on the cropped image with a lithology label; 3. A lithology identification method based on the morphological characteristics of logging curves according to claim 1, characterized in that S1.4: Scale the filled image while keeping the aspect ratio unchanged to obtain the final core section logging curve data set. The specific content of S2 is as follows: S2.1: Establish five data sets: RBG-I, RBG-II, BRG-I, BRG-II, and RGB; 4. The lithology identification method based on the morphological characteristics of logging curves according to claim 1, wherein, S2.2: Based on the core section logging curve data set constructed in step S1, generate sample images according to the RBG color arrangement pattern. Induce the images in which the spatial distribution of the red and blue curves in the sample image is in the first 30% area of the track into the RBG-I data set; Induce the images in which the spatial distribution of the red and blue curves is in the 30% - 50% area of the track into the RBG-II data set; Generate sample images according to the BRG color arrangement pattern. Induce the images in which the spatial distribution of the green curve in the sample image is in the 40% - 60% area of the track into the BRG-I data set; Induce the images in which the spatial distribution of the green curve is in the 60% - 100% area of the track into the BRG-II data set; Induce all the images generated under the RGB color arrangement pattern into the RGB data set; The track is the image channel of the logging curve in the axial direction representing the physical quantity, and it is defined that the direction towards the increase of the physical quantity is the front of the track. The S3 includes: S3.1: Select the ResNet18 and ResNet50 network models; S3.2: Test both the ResNet18 and ResNet50 network models on the validation set and test set of the straight line data set; 5. A lithology identification method based on the morphological characteristics of logging curves according to claim 1, characterized in that, S3.3: Analyze the test results, and take the winner of the two network models as the pre-trained model. The S4 includes: S4.1: Based on the pre-trained model, perform small-sample transfer learning according to the logging curve image data set corresponding to different lithologies in the core section; S4.2: During the transfer learning process, freeze the parameters of the first convolutional layer and the second convolutional layer in the pre-trained model network without updating, and retain the parameters in the shallow network that can identify the common features of the image data; update the parameters of the deep network according to the logging curve image dataset corresponding to different lithologies in the coring section, so as to construct a small-scale lithology identification model.
6. The lithology identification method based on the morphological characteristics of logging curves according to claim 1, wherein, The said S6 includes: S6.1: In the self-attention layer of the Transformer, project the input image data into vectors in three dimensions of Q, K, and V, and calculate the score S between different input vectors according to the vectors in the Q dimension and the K dimension. The said S is obtained by multiplying the Q and K matrices; S6.2: Normalize S using different normalization functions to obtain S n : S n = FNormalize(S) S6.3: Convert S into probability P according to the normalized exponential function, and describe the scores between different vectors in percentage: n P = FNormalized_Exponential(S n ) S6.4: Use Z and P to do multiplication to obtain the weighted numerical matrix Zout; S6.5: Describe the positional relationship in the data through the PE vector: Among them logs pos represents the absolute position of the sample data in the input logging data, i represents the dimension of the sample at the current position, d indicates the dimension of the input data; S6.6: Embed the said PE vector in the input data of the next layer network of the self-attention layer. During the training process, the Zout value transmits the input vector and the positional information between vectors, achieving the effect of global modeling.