A deep learning-based intelligent lithology identification method and system

Abstract

The present invention relates to the field of geological exploration technology, and in particular to a method and system for intelligent lithology identification based on deep learning. The method comprises the following steps: obtaining original core image data; extracting texture hierarchy features based on the original core image data to generate core significant texture structure feature data; performing fine rock layer image segmentation processing based on the core significant texture structure feature data to generate fine rock layer segmentation map data; constructing an intelligent lithology identification model; transmitting the fine rock layer segmentation map data to the intelligent lithology identification model to identify the target lithology attributes, and predicting the time series lithology evolution boundary to obtain predicted lithology evolution boundary data; performing lithology spatial distribution pattern recognition based on the predicted lithology evolution boundary data to generate lithology spatial distribution pattern data. The present invention realizes high-precision, automated and rapid lithology discrimination by finely segmenting the rock layers of core images, and significantly improves the efficiency of resource assessment in mining areas.

Description

Technical Field

[0001] The present invention relates to the field of geological exploration technology, and in particular to a deep learning-based intelligent lithology identification method and system. Background Art

[0002] Lithology identification is a vital part of geological exploration and resource development. As the demand for new energy and mineral resources continues to grow around the world, accurate understanding of underground rock structure and composition is particularly important. Accurate understanding of the structure and composition of underground rocks can not only help determine the distribution and reserves of mineral resources, but also help optimize resource exploration paths, reduce exploration risks, reduce exploration costs, and improve the efficiency and sustainability of resource development. However, traditional lithology identification methods mainly rely on geologists' visual observation and analysis of cores and thin sections. This method is not only time-consuming and labor-intensive, but also limited by the personal experience of experts, and has certain subjectivity and errors. The existing lithology identification method based on deep learning has limitations in the accuracy of lithology attribute discrimination, and fails to fully consider the spatial distribution law of lithology attributes, which can easily lead to misjudgment or omission. Summary of the invention

[0003] Based on this, the present invention provides a deep learning-based intelligent lithology identification method and system to solve at least one of the above technical problems.

[0004] To achieve the above purpose, a deep learning-based intelligent lithology identification method includes the following steps:

[0005] Step S1: acquiring original core image data; performing image equalization processing on the original core image data to generate enhanced core image data; performing texture image reconstruction processing on the enhanced core image data to generate multi-scale core texture image data;

[0006] Step S2: extracting texture hierarchical features based on the multi-scale core texture image data to generate core significant texture structure feature data;

[0007] Step S3: construct super-pixel lithology nodes according to the core significant texture structure feature data to generate super-pixel core node data; perform fine rock layer image segmentation processing on the enhanced core image data through the super-pixel core node data to generate fine rock layer segmentation map data; perform rock layer attribute aggregation analysis on the fine rock layer segmentation map data to generate rock layer super-pixel aggregation data;

[0008] Step S4: constructing an intelligent lithology recognition model based on a preset deep convolutional neural network model; transmitting the rock layer superpixel aggregation data to the intelligent lithology recognition model for target lithology attribute recognition, and generating target lithology recognition label data;

[0009] Step S5: using the target lithology identification label data to perform lithology profile deformation correction on the fine rock layer segmentation map data to generate a three-dimensional lithology curved surface model;

[0010] Step S6: Predict the time-series lithologic evolution boundary based on the three-dimensional lithologic surface model to obtain predicted lithologic evolution boundary data; perform boundary confidence correction processing on the predicted lithologic evolution boundary data to generate corrected lithologic evolution boundary data; perform lithologic spatial distribution pattern recognition based on the corrected lithologic evolution boundary data to generate lithologic spatial distribution pattern data.

[0011] The present invention uses advanced image acquisition technology to obtain real and accurate core image data, enhances the original core image, improves the overall brightness and contrast of the image, enhances the visual effect of the image, and highlights the texture details of the core. Multi-scale texture analysis is performed on the enhanced core image to highlight the core texture features at different scales. Feature extraction of multi-scale core texture images can obtain significant texture structure information of the core image, including the texture direction, texture roughness, texture contrast, etc. of the core. These features can effectively represent the texture characteristics of the core, eliminate the influence of noise and interference, and enhance the accuracy of lithology identification. By constructing super-pixel lithology nodes, the core can be divided into small areas, each of which has similar texture structure characteristics, which helps to improve the accuracy of lithology identification. Then, the enhanced core image data is finely segmented by super-pixel core node data, which can accurately identify the different rock layers of the core and obtain clear rock layer boundaries and structural information. Aggregation analysis of rock formation attributes on the fine rock formation segmentation map can obtain the overall characteristics of the rock formation, including the thickness, distribution range, lithology type, etc., so as to achieve more accurate intelligent identification of rock formations. An intelligent model specifically for lithology identification is constructed. Through the analysis and learning of superpixel aggregation data of rock formations, the depth characteristics required for lithology identification are extracted to improve the accuracy and intelligence level of lithology identification. Through automatic recognition of the model, the lithology attributes of each area in the fine rock formation segmentation map data can be obtained. Using the target lithology identification label data, the fine rock formation segmentation map is corrected for lithology profile deformation, which can eliminate the deformation and distortion of the rock formation during the sedimentation process, reconstruct the true shape and structure of the rock formation, and more accurately reflect the spatial distribution and morphology of the rock formation. By using time series analysis technology to predict the boundaries of lithologic evolution, we can analyze the changes and evolution processes of lithologic changes over time, predict the changing trends of lithologic boundaries in different time periods, and perform confidence analysis and correction on the predicted lithologic evolution boundaries, which can eliminate the uncertainty and noise effects in the prediction, improve the accuracy and reliability of lithologic evolution boundaries, reveal the spatial arrangement law, distribution characteristics and reservoir structure of lithologics, and provide important basis for oil and gas exploration and geological research. Therefore, a lithologic intelligent recognition method based on deep learning of the present invention takes image samples of different lithologies in core images of mining areas as research objects, extracts texture features of different lithologic images, performs refined rock layer segmentation according to the extracted image texture features, and establishes lithologic intelligent lithologic recognition models of different lithologic images, intelligently identifies the lithologic attributes of target strata, and realizes high-precision rapid discrimination of lithologic attributes; predicts the evolution of strata boundaries according to the identified lithologic attributes, analyzes the distribution pattern of lithologies in space, improves the efficiency of resource assessment, and optimizes the mining process of mineral resources.

[0012] Preferably, step S1 comprises the following steps:

[0013] Step S11: obtaining original core image data;

[0014] Step S12: performing image color space conversion on the original core image data and performing adaptive image equalization processing to generate enhanced core image data;

[0015] Step S13: calculating the neighborhood pixel texture intensity of the core texture response data using a preset core image sliding window rule to generate core texture distribution response data;

[0016] Step S14: performing texture energy space scale decomposition on the core texture distribution response data to generate multi-scale texture energy space data;

[0017] Step S15: Using the multi-scale texture energy space data, texture image reconstruction is performed on the enhanced core image data to generate multi-scale core texture image data.

[0018] The present invention obtains real core image data through advanced image acquisition technology, converts the original core image from RGB color space to other color spaces, so as to better separate and process the color and brightness information of the image. According to the local brightness distribution of the image, the contrast and brightness of the image are adaptively adjusted to improve the overall visual effect and detail performance of the image, which is conducive to enhancing the texture characteristics of the core image. The core image is scanned and processed by a preset sliding window rule to ensure that each part of the image is analyzed, thereby improving the integrity and accuracy of the core texture feature extraction. By calculating the texture intensity of the neighborhood pixels, the local texture characteristics of the core image can be analyzed, and the texture direction, texture roughness and other information of the core can be identified. The texture features are decomposed into different spatial scales, which can highlight the texture details and structural information of the core at different scales. By performing multi-scale analysis on the core texture distribution response data, the texture characteristics and energy distribution of the core at different spatial scales are revealed, and the local and global texture information of the core image is captured. The texture image is reconstructed for the enhanced core image, the texture characteristics of the core image are enhanced, and high-quality image data is provided for intelligent lithology identification.

[0019] Preferably, step S2 comprises the following steps:

[0020] Step S21: extracting texture level features based on the multi-scale core texture image data to generate texture level feature data;

[0021] Step S22: performing structural tensor analysis on the texture level feature data to obtain core texture structural tensor data;

[0022] Step S23: performing local phase calculation according to the core texture structure tensor data to generate texture local phase data;

[0023] Step S24: using the local phase data of the texture to perform multi-directional phase consistency analysis on the texture hierarchical feature data, and performing significant texture tensor voting processing to generate significant texture structural feature data of the core.

[0024] The present invention analyzes multi-scale core texture image data, extracts the texture level features of the core image, and can identify texture details of different scales in the core image, including the texture direction, texture roughness, texture contrast, etc. of the core. Structural tensor analysis is performed on the texture level feature data to calculate the structural tensor matrix of the core image, which includes the shape, direction and texture information of the core image. Local phase calculation can reveal the local changes and unevenness of the core texture. Multi-directional analysis is performed on the texture level feature data to identify significant texture features with strong phase consistency in the core image. This analysis can eliminate the influence of noise and interference and enhance the accuracy of lithology identification. Voting processing is performed on the results of the multi-directional phase consistency analysis to select the most significant and representative texture features, which can strengthen the key texture features of the core image.

[0025] Preferably, step S3 comprises the following steps:

[0026] Step S31: construct super-pixel lithology nodes according to the core significant texture structure feature data to generate super-pixel core node data;

[0027] Step S32: Calculate the surrounding pixel texture similarity of the enhanced core image data based on the superpixel core node data, and perform iterative merging of homogeneous regions to generate initial core superpixel boundary data;

[0028] Step S33: performing fuzziness evaluation of neighboring superpixel boundaries according to the initial core superpixel boundary data to generate superpixel boundary fuzzy evaluation data;

[0029] Step S34: Perform fuzzy threshold judgment on the superpixel boundary fuzzy evaluation data by using a preset boundary fuzzy threshold. When the superpixel boundary fuzzy evaluation data is higher than or equal to the preset boundary fuzzy threshold, perform color texture difference calculation on the superpixel core node data to generate adjacent superpixel texture difference data. When the adjacent superpixel texture difference data is lower than or equal to the preset texture difference threshold, perform boundary penetration fusion processing on the initial core superpixel boundary data to generate fused superpixel core boundary data. Otherwise, retain the original initial core superpixel boundary data.

[0030] Step S35: performing fine rock formation image segmentation processing on the enhanced core image data based on the fused super-pixel core boundary data and the initial core super-pixel boundary data to generate fine rock formation segmentation map data;

[0031] Step S36: Perform rock layer attribute aggregation analysis based on the fine rock layer segmentation map data to generate rock layer superpixel aggregation data.

[0032] The present invention clusters pixels with similar texture features together to form representative lithology nodes, which can reflect the texture features of the core and provide accurate lithology information for rock formation image segmentation. Based on superpixel core node data, the texture similarity of the surrounding pixels of the enhanced core image is calculated. Pixels with similar texture features are merged together to form homogeneous regions, which can eliminate the influence of noise and interference and enhance the overall consistency of the core image. The initial core superpixel boundary data is analyzed to evaluate the boundary blur between adjacent superpixels and quantify the blur of adjacent superpixel boundaries in the core image. When the superpixel boundary blur evaluation data reaches a threshold, the color texture difference between adjacent superpixels is calculated to eliminate the boundary difference between adjacent superpixels and generate more continuous and accurate fused superpixel core boundary data. Through fine rock formation image segmentation, a fine rock formation segmentation map of the core can be obtained, which clearly presents the different rock formations, boundaries and structural information of the core. Through the aggregation analysis of rock formation attributes, the overall characteristic data of the rock formation can be obtained, including the average thickness of the rock formation, the total distribution range, the main lithology type, etc., which can effectively segment the rock formation image of the enhanced core image, accurately identify the different rock formations in the core, and obtain the attribute information of the rock formation.

[0033] Preferably, step S4 comprises the following steps:

[0034] Step S41: performing self-organizing mapping processing on the rock layer super-pixel aggregation data to obtain rock layer super-pixel attribute feature data;

[0035] Step S42: acquiring a plurality of core image sample data; performing image sample data expansion processing according to the plurality of core image sample data to generate expanded image sample data;

[0036] Step S43: performing lithology label mapping processing according to the expanded image sample data to generate lithology image label sample data;

[0037] Step S44: performing image feature engineering processing on the lithology image label sample data to generate core image attribute feature data;

[0038] Step S45: randomly dividing the core image attribute feature data to obtain a lithology sample training set and a lithology sample verification set data;

[0039] Step S46: Obtaining a preset deep convolutional neural network model and pre-trained model weight parameters;

[0040] Step S47: initializing the preset deep convolutional neural network model through the pre-trained model weight parameters, and performing SE attention mechanism integration processing, so as to construct an initial intelligent lithology recognition model;

[0041] Step S48: using the lithology sample training set to perform model training operations on the initial intelligent lithology identification model, and performing model hyperparameter tuning processing to obtain a trained intelligent lithology identification model; using the lithology sample verification set data to perform optimal model verification processing on the trained intelligent lithology identification model to obtain an intelligent lithology identification model;

[0042] Step S49: transmitting the lithology superpixel attribute feature data to the intelligent lithology recognition model for target lithology attribute recognition to generate target lithology recognition label data.

[0043] The present invention can obtain rock formation superpixel attribute feature data through self-organizing mapping processing, which can retain the key features and topological structure of rock formation superpixel aggregation data and simplify data complexity. Through image sample data expansion processing, richer and more diverse core image sample data can be obtained, including core images of different types, different scales and different angles. Through lithology label mapping processing, lithology image label sample data can be obtained, which contains core image samples and their corresponding lithology category labels, providing supervision information required for training for deep convolutional neural network models. Through image feature engineering processing, core image attribute feature data can be obtained, which contains key features of core images, such as texture features, color features, shape features, etc. Through random partitioning processing, lithology sample training set and lithology sample verification set data can be obtained, providing the data set required for training and verification for the deep convolutional neural network model, which is conducive to the training and optimization of the lithology recognition model and improving the accuracy and generalization ability of the model. Using the pre-trained model weight parameters, the model weights pre-trained on other large data sets can be used to accelerate the model convergence speed and improve the effect of the lithology recognition model. The SE attention mechanism can adaptively adjust the importance of different feature channels and improve the recognition and classification effect of the lithology recognition model on lithology features. Hyperparameter tuning can adjust the learning rate, regularization strength, etc. of the model. Through the best model verification process, the best lithology recognition model can be selected. Inputting the lithology superpixel attribute feature data into the intelligent lithology recognition model can identify the target lithology attributes of the core image. This processing can automatically identify the lithology attributes of the core, including lithology category, lithology characteristics, etc.

[0044] Preferably, step S5 comprises the following steps:

[0045] Step S51: performing dynamic contour tracking according to the fine lithology segmentation map data to obtain dynamic lithology contour interface data;

[0046] Step S52: using a preset standard lithology attribute matching library to perform standard lithology attribute matching processing on the target lithology identification tag data to generate standard lithology attribute matching data;

[0047] Step S53: performing lithology texture area filling processing on the dynamic lithology contour interface data through standard lithology attribute matching data to generate texture consistency lithology map data;

[0048] Step S54: performing core depth correlation processing on the texture consistency lithology map data, and performing layered virtual splicing of the core images to obtain lithology layer texture structure data;

[0049] Step S55: performing lithology profile deformation correction on the lithology hierarchy texture structure data according to preset geological structure constraint rules to generate corrected lithology hierarchy structure map data;

[0050] Step S56: Perform lithology boundary surface interpolation processing according to the corrected lithology hierarchy structure map data to generate a three-dimensional lithology surface model.

[0051] The present invention can analyze the changes and movements of lithology contours by identifying and extracting lithology contour interfaces based on fine lithology segmentation map data. Standard lithology attribute matching data can be obtained through standard lithology attribute matching processing, which contains the matching results of the target lithology identification tag and the standard lithology attribute library, including lithology category, lithology code, lithology name and other information. Through standard lithology attribute matching data, dynamic lithology contour interface data is filled with lithology texture areas, and texture information in lithology contour interface data can be filled according to the lithology attribute matching results, eliminating blank or inconsistent areas in lithology contour interface data. Core depth association processing is performed on texture consistency lithology map data, which can analyze the depth information of core images, and superimpose lithology map data of different depths to form a lithology hierarchical texture structure. Through preset geological structure constraint rules, lithology hierarchical texture structure data is corrected for lithology profile deformation, which can eliminate deformation and distortion in lithology hierarchical texture structure and correct lithology hierarchical structure map data. Surface interpolation processing of lithology boundaries can reflect the spatial distribution and morphology of lithology boundaries based on lithology hierarchy structure map data.

[0052] Preferably, step S6 comprises the following steps:

[0053] Step S61: performing three-dimensional lithology voxel analysis according to the three-dimensional lithology surface model, and performing voxel connection processing to generate lithology connected area data;

[0054] Step S62: performing fractal dimension calculation on the lithology connected area data to obtain a lithology distribution complexity index; performing lithology contact network analysis on the lithology connected area data using the lithology distribution complexity index to generate lithology contact network relationship characteristic data;

[0055] Step S63: mining lithology evolution rules according to standard lithology attribute matching data to generate lithology evolution rule data;

[0056] Step S64: performing transfer learning on the lithology evolution rule data through a preset long short-term memory network model, and performing core evolution time series modeling to obtain a lithology evolution prediction model;

[0057] Step S65: using the lithology evolution prediction model to perform time series lithology evolution boundary prediction on the lithology contact network relationship characteristic data to obtain predicted lithology evolution boundary data;

[0058] Step S66: performing boundary confidence correction processing on the predicted lithology evolution boundary data to generate corrected lithology evolution boundary data;

[0059] Step S67: performing lithology spatial autocorrelation analysis based on the corrected lithology evolution boundary data to obtain lithology spatial distribution map data;

[0060] Step S68: Perform lithology spatial distribution pattern recognition based on the lithology spatial distribution map data to generate lithology spatial distribution pattern data.

[0061] The present invention is based on a three-dimensional lithology surface model, decomposes a complex lithology surface model into smaller voxel units, and establishes a topological relationship between lithology voxels. The complexity index of lithology distribution is calculated according to the lithology connection area data, which can quantify the complexity and irregularity of the lithology connection area, and can reflect the complexity and irregularity of the lithology connection area. Through lithology evolution rule mining, lithology evolution rule data can be obtained, which contains the potential laws, trends and constraints of lithology evolution. By using a preset long-short term memory network model to transfer learning of lithology evolution rule data, the knowledge and experience learned in other fields or tasks can be used to improve the effect and generalization ability of the lithology evolution prediction model. Using the lithology evolution prediction model, the lithology contact network relationship feature data is predicted for time series lithology evolution boundaries, and the lithology contact network relationship data can be analyzed to predict the boundaries and trends of lithology evolution. The predicted lithology evolution boundary data is analyzed to correct the uncertainty and noise influence of the lithology evolution boundary, which can improve the accuracy and reliability of the lithology evolution boundary. Through the lithology spatial autocorrelation analysis, the correlation and consistency of lithology spatial distribution can be reflected. Through the lithology spatial distribution pattern recognition, the lithology spatial distribution map data can be analyzed to extract the characteristics and trends of lithology spatial distribution.

[0062] Preferably, step S66 includes the following steps:

[0063] Step S661: extracting the lithology boundary skeleton according to the predicted lithology evolution boundary data to generate lithology prediction boundary skeleton data;

[0064] Step S662: dividing the lithology prediction boundary skeleton data into boundary points and performing continuous boundary interpolation processing to generate optimized prediction lithology boundary data;

[0065] Step S663: Calculate boundary displacement according to the optimized predicted lithology boundary data, and perform boundary stability assessment to generate boundary evolution stability assessment data;

[0066] Step S664: performing natural breakpoint activity level division according to the boundary evolution stability evaluation data to generate boundary activity partition data, wherein the boundary activity partition data includes high activity boundary data, medium activity boundary data and low activity boundary data;

[0067] Step S665: randomly taking a partition disturbance intensity value according to the boundary activity partition data, and determining a partition disturbance intensity range by using a preset partition boundary disturbance intensity index to generate partition disturbance intensity range data;

[0068] Step S666: Based on the partition disturbance intensity range data, using the boundary activity partition data, perform boundary morphological disturbance simulation on the predicted lithology evolution boundary data to obtain boundary morphological disturbance simulation data;

[0069] Step S667: performing boundary morphology probability statistics according to the boundary morphology disturbance simulation data to obtain boundary morphology probability distribution data;

[0070] Step S668: Based on the boundary morphology probability distribution data, the predicted lithology evolution boundary data is corrected in terms of boundary confidence to obtain corrected lithology evolution boundary data.

[0071] The present invention extracts the skeleton from the predicted lithologic evolution boundary data, which is essentially a simplification and refinement of the complex boundary, retaining only the basic framework that constitutes the boundary shape, which helps to focus more on the main direction and structure of the boundary during processing, reduce unnecessary detail interference, and facilitate more efficient analysis and simulation. The boundary skeleton is divided into multiple key points, and then a smooth and continuous boundary line is generated between these points through interpolation technology, which helps to restore and optimize the true form of the predicted boundary and improve the geometric accuracy of the boundary. Calculating the displacement of the boundary over time can intuitively reflect the dynamic changes of the lithologic boundary. Combining the displacement to conduct stability assessment can identify which boundary areas are more prone to change. By dividing the boundary data into high, medium and low activity levels, the classification management of the dynamic behavior of the lithologic boundary is realized. According to the boundary activity level, a reasonable range of disturbance intensity indicators is set for different regions. This strategy not only takes into account the natural differences in geological activities, but also sets a scientific parameter framework for subsequent simulation experiments to ensure the reality and effectiveness of the simulation results. By applying different degrees of disturbance to the predicted boundary, the boundary changes occurring in nature are simulated, and the response of the lithologic boundary under different disturbance conditions can be predicted. The confidence correction of the predicted lithologic evolution boundary data is performed to generate the corrected lithologic evolution boundary data. The uncertainty and noise influence of the lithologic evolution boundary can be adjusted by combining the results of the boundary morphology disturbance simulation and the probability statistical analysis. The probability-based correction strategy can effectively integrate a variety of boundary morphology changes, improve the robustness and reliability of the final prediction results, and ensure that the scientificity and practicality of the lithologic evolution analysis reach the highest level.

[0072] The present invention also provides a deep learning-based intelligent lithology identification system, which executes the deep learning-based intelligent lithology identification method as described above. The deep learning-based intelligent lithology identification system includes:

[0073] The core image preprocessing module is used to obtain the original core image data; perform image equalization processing on the original core image data to generate enhanced core image data; perform texture image reconstruction processing on the enhanced core image data to generate multi-scale core texture image data;

[0074] The texture feature extraction module is used to extract texture hierarchical features based on multi-scale core texture image data to generate significant core texture structure feature data;

[0075] The fine rock formation segmentation module is used to construct super-pixel lithology nodes based on the core significant texture structure feature data to generate super-pixel core node data; perform fine rock formation image segmentation processing on the enhanced core image data through the super-pixel core node data to generate fine rock formation segmentation map data; perform rock formation attribute aggregation analysis on the fine rock formation segmentation map data to generate rock formation super-pixel aggregation data;

[0076] The intelligent lithology recognition module is used to build an intelligent lithology recognition model based on a preset deep convolutional neural network model; transmit the rock layer superpixel aggregation data to the intelligent lithology recognition model to identify the target lithology attributes and generate target lithology recognition label data;

[0077] The 3D core model building module is used to use the target lithology identification label data to perform lithology profile deformation correction on the fine rock layer segmentation map data to generate a 3D lithology surface model;

[0078] The lithology evolution prediction module is used to predict the time-series lithology evolution boundary based on the three-dimensional lithology surface model to obtain the predicted lithology evolution boundary data; perform boundary confidence correction processing on the predicted lithology evolution boundary data to generate corrected lithology evolution boundary data; perform lithology spatial distribution pattern recognition based on the corrected lithology evolution boundary data to generate lithology spatial distribution pattern data. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 It is a schematic diagram of the steps of the intelligent lithology identification method based on deep learning of the present invention;

[0080] Figure 2 for Figure 1 Detailed implementation steps of step S2 in the flowchart;

[0081] Figure 3 for Figure 1 Detailed implementation steps of step 53 are shown in the flowchart;

[0082] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0083] The technical method of the present invention is described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by technicians in this field without creative work are within the scope of protection of the present invention.

[0084] In addition, the accompanying drawings are only schematic illustrations of the present invention and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and their repeated description will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities. The functional entities can be implemented in software form, or implemented in one or more hardware modules or integrated circuits, or implemented in different network and / or processor methods and / or microcontroller methods.

[0085] It should be understood that, although the terms "first", "second", etc. may be used herein to describe various units, these units should not be limited by these terms. These terms are used only to distinguish one unit from another unit. For example, without departing from the scope of the exemplary embodiments, the first unit may be referred to as the second unit, and similarly the second unit may be referred to as the first unit. The term "and / or" used herein includes any and all combinations of one or more of the listed associated items.

[0086] To achieve this, please refer to Figures 1 to 3 The present invention provides a method for intelligent lithology identification based on deep learning, comprising the following steps:

[0087] Step S1: acquiring original core image data; performing image equalization processing on the original core image data to generate enhanced core image data; performing texture image reconstruction processing on the enhanced core image data to generate multi-scale core texture image data;

[0088] Step S2: extracting texture hierarchical features based on the multi-scale core texture image data to generate core significant texture structure feature data;

[0089] Step S3: construct super-pixel lithology nodes according to the core significant texture structure feature data to generate super-pixel core node data; perform fine rock layer image segmentation processing on the enhanced core image data through the super-pixel core node data to generate fine rock layer segmentation map data; perform rock layer attribute aggregation analysis on the fine rock layer segmentation map data to generate rock layer super-pixel aggregation data;

[0090] Step S4: constructing an intelligent lithology recognition model based on a preset deep convolutional neural network model; transmitting the rock layer superpixel aggregation data to the intelligent lithology recognition model for target lithology attribute recognition, and generating target lithology recognition label data;

[0091] Step S5: using the target lithology identification label data to perform lithology profile deformation correction on the fine rock layer segmentation map data to generate a three-dimensional lithology curved surface model;

[0092] Step S6: Predict the time-series lithologic evolution boundary based on the three-dimensional lithologic surface model to obtain predicted lithologic evolution boundary data; perform boundary confidence correction processing on the predicted lithologic evolution boundary data to generate corrected lithologic evolution boundary data; perform lithologic spatial distribution pattern recognition based on the corrected lithologic evolution boundary data to generate lithologic spatial distribution pattern data.

[0093] In the embodiment of the present invention, reference Figure 1The above is a schematic diagram of the steps of a deep learning-based intelligent lithology identification method of the present invention. In this embodiment, the deep learning-based intelligent lithology identification method includes the following steps:

[0094] Step S1: acquiring original core image data; performing image equalization processing on the original core image data to generate enhanced core image data; performing texture image reconstruction processing on the enhanced core image data to generate multi-scale core texture image data;

[0095] In an embodiment of the present invention, a core scanner is used to obtain original core image data with a resolution of 4096x4096 pixels. The original image is processed by an image enhancement algorithm based on histogram equalization. Specifically, after the image is converted into a grayscale image, the distribution histogram of the pixel grayscale value is statistically calculated, and the histogram is nonlinearly transformed using the cumulative distribution function to remap the pixel value to a wider range, thereby improving the contrast and clarity of the image and generating enhanced core image data. The image is decomposed into sub-bands of different scales. Then, texture features are extracted from each sub-band, such as using grayscale co-occurrence matrix, Gabor filter and other methods to extract texture features. Finally, these features are used to reconstruct a multi-scale texture image to highlight the texture information of the core at different scales. The generated multi-scale core texture image data may include texture images of multiple scales such as the original scale, 1 / 2 scale and 1 / 4 scale.

[0096] Step S2: extracting texture hierarchical features based on the multi-scale core texture image data to generate core significant texture structure feature data;

[0097] In an embodiment of the present invention, for example, a Local Binary Pattern (LBP) operator is used to extract texture hierarchical features from multi-scale core texture image data, and a binary code is generated by comparing the grayscale values ​​of the central pixel and its neighborhood pixels to reflect local texture features. Specifically, a circular neighborhood with a radius of 3 is set with each pixel as the center, and the grayscale value of the pixel in the neighborhood is compared with the grayscale value of the central pixel. The grayscale value greater than the central pixel value is encoded as 1, otherwise it is encoded as 0, and an 8-bit binary code is generated. LBP encoding is performed on each pixel in the multi-scale texture image data, and the frequency of occurrence of different encodings at each scale is counted to construct an LBP histogram as a texture feature descriptor at that scale. Finally, the LBP histograms of all scales are connected in series to generate significant texture structure feature data of the core.

[0098] Step S3: construct super-pixel lithology nodes according to the core significant texture structure feature data to generate super-pixel core node data; perform fine rock layer image segmentation processing on the enhanced core image data through the super-pixel core node data to generate fine rock layer segmentation map data; perform rock layer attribute aggregation analysis on the fine rock layer segmentation map data to generate rock layer super-pixel aggregation data;

[0099] In the embodiment of the present invention, the SLIC algorithm is used to construct superpixel lithology nodes based on significant texture structure feature data, and the texture similarity of the node neighborhood pixels is calculated for initial region division. Then, the texture feature differences of the initial boundary pixels are analyzed, and the adjacent regions with high similarity are fused. Finally, the fused boundary and the initial boundary are used as hard constraints and soft constraints respectively, and applied to the graph cut algorithm to obtain fine rock formation segmentation map data by minimizing the energy function. Finally, attribute aggregation analysis is performed on each rock formation area after segmentation to generate rock formation superpixel aggregation data.

[0100] Step S4: constructing an intelligent lithology recognition model based on a preset deep convolutional neural network model; transmitting the rock layer superpixel aggregation data to the intelligent lithology recognition model for target lithology attribute recognition, and generating target lithology recognition label data;

[0101] In an embodiment of the present invention, based on the pre-trained weight parameters, the deep convolutional neural network model combined with the SE attention mechanism is initialized, and its classification number is modified to the number of categories of the lithology image (6 categories). The samples are expanded by operations such as image flipping, adding Gaussian noise, and rotation, and the training set and the test set are divided into 8:2 ratios. Using the Adam optimizer, the learning rate reduction strategy is set, and the cross entropy loss function is used for training until the model accuracy is stable and meets the initial discrimination conditions. The rock formation superpixel aggregation data is transmitted to the trained intelligent lithology recognition model. The model predicts the lithology category to which each rock formation superpixel area belongs based on the input feature data, and outputs the probability value of each category. The category with the highest probability value is selected as the prediction result for the area, and the target lithology recognition label data is generated.

[0102] Step S5: using the target lithology identification label data to perform lithology profile deformation correction on the fine rock layer segmentation map data to generate a three-dimensional lithology curved surface model;

[0103] In the embodiment of the present invention, the active contour model is used to generate dynamic lithology contour interface data, and the standard texture information of each lithology area is obtained by comparing with the standard lithology attribute matching library. Then, the standard texture is filled into the corresponding area, and layered and spliced ​​according to the depth information to generate texture consistent lithology map data. On this basis, the geological structure constraint rules are applied to perform geometric transformation to simulate stratum deformation. Finally, a surface interpolation algorithm is used to generate a smooth surface connecting all boundary points to construct a three-dimensional lithology surface model.

[0104] Step S6: Predict the time-series lithologic evolution boundary based on the three-dimensional lithologic surface model to obtain predicted lithologic evolution boundary data; perform boundary confidence correction processing on the predicted lithologic evolution boundary data to generate corrected lithologic evolution boundary data; perform lithologic spatial distribution pattern recognition based on the corrected lithologic evolution boundary data to generate lithologic spatial distribution pattern data.

[0105] In an embodiment of the present invention, the three-dimensional lithology surface model is discretized into voxels, and connectivity analysis is performed to construct a lithology contact network. Then, the lithology evolution rules are extracted from the historical data using a data mining algorithm, and a suitable time series model (such as LSTM) is selected for training to learn the law of lithology spatial distribution changing over time. Finally, the current lithology contact network is input into the model to predict the change trend of the future lithology connected area, and the lithology boundary information is extracted to generate predicted lithology evolution boundary data. The predicted boundaries are classified according to the confidence of the model prediction results, and high confidence boundaries and low confidence boundaries are distinguished. The characteristic indicators of different lithologies in space (such as fractal dimension, Moran index, etc.) are calculated and visualized for analysis. On this basis, pattern recognition methods such as cluster analysis are applied to classify areas with similar spatial distribution and similar lithology, identify typical patterns of lithology spatial distribution, and generate lithology spatial distribution pattern data.

[0106] Preferably, step S1 comprises the following steps:

[0107] Step S11: obtaining original core image data;

[0108] Step S12: performing image color space conversion on the original core image data and performing adaptive image equalization processing to generate enhanced core image data;

[0109] Step S13: calculating the neighborhood pixel texture intensity of the core texture response data using a preset core image sliding window rule to generate core texture distribution response data;

[0110] Step S14: performing texture energy space scale decomposition on the core texture distribution response data to generate multi-scale texture energy space data;

[0111] Step S15: Using the multi-scale texture energy space data, texture image reconstruction is performed on the enhanced core image data to generate multi-scale core texture image data.

[0112] In an embodiment of the present invention, a high-resolution core scanner is used to obtain original core image data. For example, the resolution of the scanner is set to 4096x 4096 pixels, and the RGB three-channel color mode is used for acquisition. Each pixel records the color information of the corresponding position on the core surface. In order to better extract the core texture information, it is necessary to perform color space conversion on the original core image data. Common color space conversion methods include RGB to HSV, RGB to LAB, etc., which are completed by the color space conversion formula to convert the RGB value of each pixel into a Lab value. The distribution of pixel values ​​is automatically adjusted according to the local contrast of the image to improve the clarity and detail expression of the image. For example, the image is divided into several sub-blocks, each sub-block is histogram-equalized, and the block effect is eliminated by a bilinear interpolation algorithm, and the enhanced core image data is finally generated. For example, the sliding window size is set to 5x5 pixels, and each pixel is centered to calculate the statistical characteristics of the pixel values ​​within its neighborhood, such as variance, contrast, entropy, etc. These statistical characteristics can reflect the texture changes in the local area. For example, a higher variance represents a rougher texture, and a lower variance represents a smoother texture. The calculated texture feature value is assigned to the central pixel point to generate the core texture distribution response data. The core texture distribution response data is decomposed into texture energy space scales using two-dimensional discrete wavelet transform. For example, Haar wavelet or Daubechies wavelet is used to perform multi-level wavelet decomposition on the core texture distribution response data. Each level of decomposition will produce four sub-band images, representing the low-frequency information and high-frequency information of the image in the horizontal, vertical and diagonal directions. By adjusting the number of decomposition layers, the scale range of texture feature extraction can be controlled. For example, more decomposition layers can extract more subtle texture information. The energy values ​​of sub-band images of each scale are combined to generate multi-scale texture energy space data. The multi-scale texture energy space data is used as a weight coefficient to perform weighted superposition on the enhanced core image data. Texture energy information of different scales contributes to different degrees to the final reconstructed image. For example, high-frequency information mainly reflects the edges and details of the image, and low-frequency information mainly reflects the overall structure of the image. By adjusting the weights of texture energy information of different scales, the texture detail of the final reconstructed image can be controlled to generate multi-scale core texture image data.

[0113] Preferably, step S1 comprises the following steps:

[0114] Step S131: Perform Gabor texture filtering on the enhanced core image data to generate core texture response data;

[0115] Step S132: performing local binary pattern encoding processing according to the core texture response data, and calculating the neighborhood pixel texture intensity to generate local core texture feature value data;

[0116] Step S133: performing morphological gradient processing according to the local core texture feature value data to generate core texture gradient data;

[0117] Step S134: constructing a texture scale space based on the core texture gradient data, and performing multi-scale edge flow analysis to generate texture flow field feature data;

[0118] Step S135: performing curvature scale space transformation processing on the texture flow field feature data to obtain multi-scale curvature feature map data;

[0119] Step S136: performing texture topology continuity analysis based on the multi-scale curvature feature map data to obtain texture topology feature data;

[0120] Step S137: Match the texture energy distribution of the enhanced core image data with the texture topological feature data to generate core texture distribution response data.

[0121] In an embodiment of the present invention, a Gabor filter group is used to perform multi-directional and multi-scale texture filtering on the enhanced core image data. In a specific operation, a plurality of Gabor filters of different directions (e.g., 0°, 45°, 90°, 135°) and a plurality of Gabor filters of different scales (e.g., 3x3, 5x5, 7x7) can be set to perform convolution operation on the enhanced core image data. Each Gabor filter generates a response map reflecting the texture information of the image in the direction and scale. The core texture response data is subjected to local binary pattern (LBP) encoding. For example, a 3x3 neighborhood window is set with each pixel as the center. Each pixel value in the neighborhood window is compared with the center pixel value. If it is greater than or equal to the center pixel value, it is encoded as 1, otherwise it is encoded as 0. The encoding results are arranged in a clockwise direction to form an 8-bit binary number, which is the LBP value of the pixel. Morphological gradient processing is performed on the local core texture feature value data. In a specific operation, morphological operations such as expansion and corrosion can be used to calculate the gradient of the image. For example, a 3x3 structure element can be used to dilate and erode an image, and then the dilation result and the erosion result can be subtracted to obtain the morphological gradient of the image. Texture scale space refers to representing an image at different scales. There are many ways to construct texture scale space, such as Gaussian scale space, wavelet scale space, etc. Taking Gaussian scale space as an example, the image representation at different scales can be obtained by performing Gaussian blurring on the core texture gradient data to different degrees. For example, the edge information of the image at different scales is extracted using methods such as the Canny edge detection algorithm, and then the direction and intensity of the edge are calculated using methods such as the optical flow method to generate texture flow field feature data. The texture flow field feature data is processed, for example, using a preset curvature calculation method, such as the second-order derivative method, to calculate the curvature of texture flow field data at different scales. The curvature data at different scales are integrated to generate multi-scale curvature feature map data. By analyzing the changes in the texture structure at different scales, its topological persistence can be determined. In specific operations, topological data analysis methods such as PersistentHomology can be used to calculate the topological characteristics of the texture structure at different scales, such as Betti number, persistence map, etc., to obtain texture topological feature data.

[0122] As an example of the present invention, refer to Figure 2 As shown, Figure 1 Detailed implementation steps of step S2 in the embodiment are shown in the flowchart. In this embodiment, step S2 includes:

[0123] Step S21: extracting texture level features based on the multi-scale core texture image data to generate texture level feature data;

[0124] In an embodiment of the present invention, for example, multi-scale core texture image data is grouped according to preset scale levels, for example, divided into three levels: fine scale, medium scale, and coarse scale. Then, for each scale level, LBP operators with different radii are used to extract texture information of different scales from the image. For example, for fine scale, LBP operators with radii of 1 and 2 can be used; for medium scale, LBP operators with radii of 3 and 4 can be used; for coarse scale, LBP operators with radii of 5 and 6 can be used. Finally, the LBP feature maps extracted at each scale are connected in series to form texture hierarchical feature data containing multi-scale texture information.

[0125] Step S22: performing structural tensor analysis on the texture level feature data to obtain core texture structural tensor data;

[0126] In an embodiment of the present invention, the gradient information of each pixel in different directions is first calculated, such as horizontal direction, vertical direction, diagonal direction, etc. Then, a structure tensor matrix is ​​constructed based on the gradient information, and its eigenvalues ​​and eigenvectors are calculated. The size of the eigenvalue reflects the degree of change of the pixel in the direction of the corresponding eigenvector. For example, a larger eigenvalue indicates that the texture in this direction changes dramatically. The structure tensor eigenvalues ​​of each pixel are combined to obtain the core texture structure tensor data.

[0127] Step S23: performing local phase calculation according to the core texture structure tensor data to generate texture local phase data;

[0128] In the embodiment of the present invention, for example, the eigenvector direction of the structure tensor can be used as the direction of the local phase, and the eigenvalue size of the structure tensor can be used to calculate the frequency of the local phase. For example, the direction with the largest eigenvalue can be used as the direction of the local phase, and the size of the eigenvalue can be mapped to the range of 0 to 2π as the frequency of the local phase. The local phase information of each pixel is combined to generate texture local phase data.

[0129] Step S24: using the local phase data of the texture to perform multi-directional phase consistency analysis on the texture hierarchical feature data, and performing significant texture tensor voting processing to generate significant texture structural feature data of the core.

[0130] In an embodiment of the present invention, phase consistency refers to the consistency of phase information between adjacent pixels in an image, which can reflect the continuity and directionality of the texture. In specific operations, the variance of the phase difference between each pixel and its neighboring pixel can be calculated. The smaller the variance, the higher the phase consistency. In order to analyze the phase consistency in different directions, multiple neighborhood windows in different directions can be set, such as horizontal direction, vertical direction, diagonal direction, etc., and the variance of the phase difference can be calculated respectively. Each pixel votes for its neighboring pixel points based on its multi-directional phase consistency. The direction of the vote is consistent with the direction of the local phase, and the weight of the vote is proportional to the degree of phase consistency. For example, if a pixel has a high phase consistency in the horizontal direction, it will cast a higher weight to the pixels on its left and right sides. Finally, the voting results received by each pixel are counted to generate significant texture structure feature data of the core.

[0131] As an example of the present invention, refer to Figure 3 As shown, Figure 1 Detailed implementation steps of step S3 in the flowchart, in this example, step S3 includes:

[0132] Step S31: construct super-pixel lithology nodes according to the core significant texture structure feature data to generate super-pixel core node data;

[0133] In an embodiment of the present invention, a simple linear iterative clustering (SLIC) algorithm is used to construct superpixel lithology nodes based on the core significant texture structure feature data. The core significant texture structure feature data is used as input, and the preset number of superpixels and compactness parameters are set. The SLIC algorithm divides the image into multiple superpixel regions with uniform size and compact boundaries based on feature similarity and spatial distance. Each superpixel region represents a lithology region with similar texture features, and its center point is the superpixel lithology node. Finally, superpixel core node data is generated, and each node data contains two parts of information: one is the spatial position coordinates (x, y) of the node in the image, and the other is the average texture feature vector of the superpixel region to which the node belongs, which can be obtained by taking the average of the texture feature vectors of all pixels in the region.

[0134] Step S32: Calculate the surrounding pixel texture similarity of the enhanced core image data based on the superpixel core node data, and perform iterative merging of homogeneous regions to generate initial core superpixel boundary data;

[0135] In an embodiment of the present invention, based on the superpixel core node data, the enhanced core image data is subjected to the calculation of the texture similarity of the surrounding pixels. Each superpixel lithology node is traversed, and the similarity between the texture feature vector of each pixel in its neighborhood (for example, a 5x5 pixel area centered on the node) and the texture feature vector of the node is calculated. The cosine similarity can be used to measure the similarity between two vectors. The higher the similarity, the closer the texture features of the two pixels are. For example, a superpixel lithology node represents an area with lighter color and more uniform texture. There are some pixels with similar colors but slightly different textures in its neighborhood. By calculating the similarity of the texture features of these pixels with the node, it can be determined whether these pixels belong to the same lithology area. A similarity threshold (for example, 0.9) is set, and pixels with a similarity higher than the threshold are marked as belonging to the same area as the node. After performing the above operations on all superpixel lithology nodes, a preliminary regional division result is obtained.

[0136] Step S33: performing fuzziness evaluation of neighboring superpixel boundaries according to the initial core superpixel boundary data to generate superpixel boundary fuzzy evaluation data;

[0137] In an embodiment of the present invention, the boundary pixels between all adjacent superpixel regions are determined based on the initial core superpixel boundary data. For example, assuming that two adjacent superpixel regions are represented by the numbers 1 and 2, respectively, the boundary pixels can be represented as adjacent positions of pixels with values ​​1 and 2 in the image. For each boundary pixel, the texture feature difference within its neighborhood (for example, a 5x5 pixel area centered on the pixel) is calculated. The difference value of the gradient directional histogram (HOG) is used for calculation. The larger the difference value, the greater the difference in texture features on both sides of the boundary pixel, and the clearer the boundary; conversely, the smaller the difference value, the more blurred the boundary.

[0138] Step S34: Perform fuzzy threshold judgment on the superpixel boundary fuzzy evaluation data by using a preset boundary fuzzy threshold. When the superpixel boundary fuzzy evaluation data is higher than or equal to the preset boundary fuzzy threshold, perform color texture difference calculation on the superpixel core node data to generate adjacent superpixel texture difference data. When the adjacent superpixel texture difference data is lower than or equal to the preset texture difference threshold, perform boundary penetration fusion processing on the initial core superpixel boundary data to generate fused superpixel core boundary data. Otherwise, retain the original initial core superpixel boundary data.

[0139] In the embodiment of the present invention, a boundary blur threshold and a texture difference threshold are set. The boundary blur threshold is used to determine whether a boundary pixel is sufficiently blurred and needs further processing. The texture difference threshold is used to determine whether two adjacent super-pixel regions belong to the same lithology region, and these two thresholds can be adjusted according to the specific application scenario and data characteristics. For example, suppose there are two adjacent super-pixel regions, representing a sandstone region with a lighter color and uniform texture and a conglomerate region with a darker color and rough texture. The boundary pixel is located at the junction of the two regions, and both the color and texture features are in a transitional state. By calculating the color and texture feature differences, the color feature difference can be calculated using indicators such as color histogram difference or color moment difference. The texture feature difference can be calculated using indicators such as gray level co-occurrence matrix (GLCM) difference and LBP difference, and it can be determined whether there are obvious differences between the two regions. If the calculated color and texture feature differences are lower than the preset texture difference threshold, it is considered that the two super-pixel regions belong to the same lithology region and their boundaries need to be fused. For example, in the above example of sandstone and conglomerate areas, if the color and texture features of the boundary pixel are closer to the sandstone area, the pixel is merged into the sandstone area, thereby achieving boundary movement and fusion. Conversely, if the difference in color and texture features is higher than the preset texture difference threshold, the two superpixel areas are considered to belong to different lithology areas, and the boundaries do not need to be fused, and the original initial core superpixel boundary data is retained.

[0140] Step S35: performing fine rock formation image segmentation processing on the enhanced core image data based on the fused super-pixel core boundary data and the initial core super-pixel boundary data to generate fine rock formation segmentation map data;

[0141] In an embodiment of the present invention, the fused superpixel core boundary data is used as a hard constraint, and the initial core superpixel boundary data is used as a soft constraint, and applied to the graph cut algorithm (GraphCut). The hard constraint means that the fused boundary must be retained in the final segmentation result, while the soft constraint means that the initial boundary can be adjusted according to the optimization of the energy function. Next, an energy function is constructed. The design goal of the energy function is to ensure that the segmentation result not only satisfies the boundary constraints, but also maintains the consistency of texture and color within the region as much as possible. The energy function can be expressed by the following formula: E = E d +E s , where E d Represents a data item used to measure the matching degree between the segmentation result and the image data. For example, it can be calculated using the color histogram difference or texture feature difference. sRepresents a smoothness term, which is used to constrain adjacent pixels to belong to the same category. For example, it can be calculated using the boundary length or curvature. The energy function is minimized using the graph cut algorithm. The graph cut algorithm transforms the image segmentation problem into the minimum cut problem in graph theory. The optimal segmentation result is obtained by solving the minimum cut. It can effectively handle hard constraints and soft constraints, find the global optimal solution, and generate fine rock formation segmentation map data.

[0142] Step S36: Perform rock layer attribute aggregation analysis based on the fine rock layer segmentation map data to generate rock layer superpixel aggregation data.

[0143] In an embodiment of the present invention, attribute aggregation analysis is performed on each rock layer region according to the fine rock layer segmentation map data to generate rock layer superpixel aggregation data. For example, the average color, texture features, shape features, etc. of each rock layer region are counted. For example, it is assumed that the fine rock layer segmentation map data contains three regions, representing sandstone, mudstone, and conglomerate, respectively. By extracting and aggregating the color, texture, and shape features of each region, three feature vectors can be obtained, which respectively describe the typical features of the three rock layers.

[0144] Preferably, step S4 comprises the following steps:

[0145] Step S41: performing self-organizing mapping processing on the rock layer super-pixel aggregation data to obtain rock layer super-pixel attribute feature data;

[0146] Step S42: acquiring a plurality of core image sample data; performing image sample data expansion processing according to the plurality of core image sample data to generate expanded image sample data;

[0147] Step S43: performing lithology label mapping processing according to the expanded image sample data to generate lithology image label sample data;

[0148] Step S44: performing image feature engineering processing on the lithology image label sample data to generate core image attribute feature data;

[0149] Step S45: randomly dividing the core image attribute feature data to obtain a lithology sample training set and a lithology sample verification set data;

[0150] Step S46: Obtaining a preset deep convolutional neural network model and pre-trained model weight parameters;

[0151] Step S47: initializing the preset deep convolutional neural network model through the pre-trained model weight parameters, and performing SE attention mechanism integration processing, so as to construct an initial intelligent lithology recognition model;

[0152] Step S48: using the lithology sample training set to perform model training operations on the initial intelligent lithology identification model, and performing model hyperparameter tuning processing to obtain a trained intelligent lithology identification model; using the lithology sample verification set data to perform optimal model verification processing on the trained intelligent lithology identification model to obtain an intelligent lithology identification model;

[0153] Step S49: transmitting the lithology superpixel attribute feature data to the intelligent lithology recognition model for target lithology attribute recognition to generate target lithology recognition label data.

[0154] In an embodiment of the present invention, the high-dimensional feature vectors in the rock formation superpixel aggregation data are mapped to a low-dimensional feature space, and the topological relationship between the original data is retained. For example, a two-dimensional SOM network is used for iterative mapping, and each rock formation superpixel aggregation data sample is mapped to a specific node on the SOM network. The weight vector of the node is the rock formation superpixel attribute feature data of the sample. For example, the YXCJ-VII high-resolution core image acquisition instrument and its supporting system were used to collect image samples of the corresponding core samples. According to the professional description of the lithological attributes in the core images by geological experts, image data samples of different core lithological attributes were constructed, including 1000 core images of aeolian sand, coal, sandstone and conglomerate respectively. In order to increase the diversity of samples and improve the robustness of the model in data feature recognition, the samples were expanded by image flipping, adding Gaussian noise to the images, and rotating the images by 90 degrees and 180 degrees. The data set was randomly divided according to the ratio of training set to test set of 8:2. The SqueezeNet1_1 model combined with the SE attention mechanism was initialized using the pre-trained weight parameters, and its classification number was modified to 6 categories of lithological images. The data was normalized, the training rounds were set to 100 times, the Adam optimizer was used, the learning rate reduction strategy was set, the parameters were updated after every 5 training rounds, the multiplication factor of the updated learning rate was 0.5, and the cross entropy loss function was used for loss discrimination between the predicted value and the true value. Finally, the model is iteratively trained using the training set data, and the test set is used to test the data until the loss between the predicted value and the true value gradually decreases and stabilizes with the increase in the number of training times. After the accuracy of the model stabilizes and meets the initial accuracy discrimination conditions, an intelligent classification model for different lithology samples of core images is obtained. The components of the SqueezeNet model combined with the SE attention mechanism are as follows: SqueezeNet1_1 is used as the basic framework, in which there are 8 Fire layers, each layer contains a Squeeze module and an Extract module. The Squeeze module includes a 1×1 convolution layer. After extracting the compressed features, it enters the Extract module, which includes a 3×3 convolution layer, including convolution, BatchNorm layer and ReLU activation function layer. A new feature layer is obtained by parallelizing the 1×1 convolution layer features and the 3×3 convolution layer features. Adaptive average pooling is performed in the spatial dimension, channel attention is learned through two fully connected layers, Sigmoid is used for normalization to obtain the channel attention feature map, and finally the channel attention feature map is multiplied by the original Fire layer output feature to obtain the weighted feature. The intelligent lithology recognition model predicts the lithology category of each rock layer superpixel area based on the input feature data and outputs the probability value of each category. The category with the highest probability value can be selected as the prediction result of the area to generate the target lithology recognition label data.

[0155] Preferably, step S5 comprises the following steps:

[0156] Step S51: performing dynamic contour tracking according to the fine lithology segmentation map data to obtain dynamic lithology contour interface data;

[0157] Step S52: using a preset standard lithology attribute matching library to perform standard lithology attribute matching processing on the target lithology identification tag data to generate standard lithology attribute matching data;

[0158] Step S53: performing lithology texture area filling processing on the dynamic lithology contour interface data through standard lithology attribute matching data to generate texture consistency lithology map data;

[0159] Step S54: performing core depth correlation processing on the texture consistency lithology map data, and performing layered virtual splicing of the core images to obtain lithology layer texture structure data;

[0160] Step S55: performing lithology profile deformation correction on the lithology hierarchy texture structure data according to preset geological structure constraint rules to generate corrected lithology hierarchy structure map data;

[0161] Step S56: Perform lithology boundary surface interpolation processing according to the corrected lithology hierarchy structure map data to generate a three-dimensional lithology surface model.

[0162] In an embodiment of the present invention, an active contour model, such as a serpentine model, is used to track the boundary of the rock layer, and the position and shape of the contour curve are continuously updated. The contour curve will gradually approach the target lithology boundary and finally converge to a stable state to generate dynamic lithology contour interface data. A standard lithology attribute matching library is constructed. The library contains attribute information such as the name, color, and texture of various common lithologies. The target lithology identification label data is compared with the standard lithology attribute matching library. For example, if the lithology label of a certain area is "sandstone", the color, texture, and other attribute information corresponding to "sandstone" are searched in the standard lithology attribute matching library. The matched standard lithology attribute information is associated with the target lithology identification label data to generate standard lithology attribute matching data. According to the standard lithology attribute matching data generated in step S52, the standard texture information corresponding to each lithology area is obtained. For example, the sandstone area can be filled with a preset sandstone texture image; the mudstone area can be filled with a preset mudstone texture image. The standard texture information is filled into the corresponding lithology area using a texture synthesis algorithm. According to the core acquisition record, the depth information corresponding to each core image is obtained. According to the depth information, the texture consistency lithology map data is layered to obtain lithology images of different depths. The lithology images of different depths are virtually spliced. In the splicing process, it is necessary to align the overlapping parts of the core image and perform smoothing to eliminate the splicing traces. According to geological exploration data or manual experience, the geological structure constraint rules are preset. For example, a mathematical model of the stratigraphic layer can be established according to the strike, dip and other information of the stratum. Using the preset geological structure constraint rules, the lithology layer texture structure data is geometrically transformed, such as translation, rotation, scaling, etc., to simulate the deformation of the stratum. For example, according to the faults, folds and other features on the core image, the parameters of the stratigraphic layer can be adjusted so that the corrected lithology layer structure map data can more accurately reflect the actual geological structure. Using the surface interpolation algorithm, such as the Kriging interpolation method and the thin plate spline interpolation method, a smooth surface connecting all boundary points is generated. In the interpolation process, the parameters can be adjusted according to the actual situation to control the smoothness and approximation of the surface. The model can intuitively display the spatial distribution and morphological characteristics of different underground rock layers.

[0163] Preferably, step S6 comprises the following steps:

[0164] Step S61: performing three-dimensional lithology voxel analysis according to the three-dimensional lithology surface model, and performing voxel connection processing to generate lithology connected area data;

[0165] Step S62: performing fractal dimension calculation on the lithology connected area data to obtain a lithology distribution complexity index; performing lithology contact network analysis on the lithology connected area data using the lithology distribution complexity index to generate lithology contact network relationship characteristic data;

[0166] Step S63: mining lithology evolution rules according to standard lithology attribute matching data to generate lithology evolution rule data;

[0167] Step S64: performing transfer learning on the lithology evolution rule data through a preset long short-term memory network model, and performing core evolution time series modeling to obtain a lithology evolution prediction model;

[0168] Step S65: using the lithology evolution prediction model to perform time series lithology evolution boundary prediction on the lithology contact network relationship characteristic data to obtain predicted lithology evolution boundary data;

[0169] Step S66: performing boundary confidence correction processing on the predicted lithology evolution boundary data to generate corrected lithology evolution boundary data;

[0170] Step S67: performing lithology spatial autocorrelation analysis based on the corrected lithology evolution boundary data to obtain lithology spatial distribution map data;

[0171] Step S68: Perform lithology spatial distribution pattern recognition based on the lithology spatial distribution map data to generate lithology spatial distribution pattern data.

[0172] In an embodiment of the present invention, according to a preset voxel resolution, the three-dimensional lithologic surface model is discretized into a uniform three-dimensional grid, and each grid unit is a voxel. Each voxel inherits the lithologic properties of the lithologic surface model at its location, such as sandstone, mudstone, etc. Connectivity analysis is performed on all voxels. Adjacent voxels with the same lithologic properties can be connected to form connected regions using algorithms such as depth-first search or breadth-first search. For example, suppose there are two lithologic connected regions, one with a regular shape and the other with an irregular shape. By calculating the fractal dimension, the box dimension method, the grid dimension method, and the like can be used to calculate the fractal dimension, and it can be found that the irregularly shaped region has a higher fractal dimension, indicating that its spatial distribution is more complex. A lithologic contact network is constructed based on the lithologic connected region data. Each connected region is regarded as a node in the network. If two connected regions are adjacent in space, an edge is established between the corresponding nodes. The weight of the edge can represent the contact area or contact length between the two regions. Using data mining algorithms, such as association rule mining and decision trees, lithology evolution rules are extracted from lithology attributes and geological history data. For example, rules such as "sandstone is more likely to be deposited in a high-energy environment" and "sandstone gradually transitions to mudstone upward as the sea level rises" can be mined. Select a suitable time series model, such as a long short-term memory network (LSTM). LSTM networks can effectively process time series data and capture long-term dependencies in the data. During model training, historical core data can be used as training samples, such as core images from different periods, geological exploration data, etc. The goal of the model is to learn the law of lithology spatial distribution over time. The current lithology contact network relationship feature data is input into the lithology evolution prediction model to predict the change trend of each lithology connection area in the future, such as volume expansion, contraction, morphological change, etc. For example, if a sandstone connection area is predicted to expand to the surrounding area, it is necessary to update the contact relationship between the area and its adjacent areas, and adjust the nodes and edges in the network accordingly. According to the updated lithology contact network, the lithology boundary information is extracted to generate the predicted lithology evolution boundary data. According to the confidence of the model prediction results, such as probability value, variance, etc., the prediction boundaries are classified. For example, the boundaries with high confidence can be marked as "high confidence boundaries", and the boundaries with low confidence can be marked as "low confidence boundaries". The low confidence boundaries are corrected according to the knowledge of geological experts or other auxiliary information. For example, the Moran index of different lithologies in space can be calculated, and the results of spatial autocorrelation analysis can be visualized, such as drawing a spatial autocorrelation map to intuitively show the distribution of different lithologies in space. In order to identify the typical patterns of lithology spatial distribution, such as layered distribution, lens distribution, etc., it is necessary to perform pattern recognition on the lithology spatial distribution map data. Cluster analysis methods can be used to cluster areas with similar spatial distribution and similar lithology into one category, so as to identify different lithology combination patterns.

[0173] Preferably, step S66 includes the following steps:

[0174] Step S661: extracting the lithology boundary skeleton according to the predicted lithology evolution boundary data to generate lithology prediction boundary skeleton data;

[0175] Step S662: dividing the lithology prediction boundary skeleton data into boundary points and performing continuous boundary interpolation processing to generate optimized prediction lithology boundary data;

[0176] Step S663: Calculate boundary displacement according to the optimized predicted lithology boundary data, and perform boundary stability assessment to generate boundary evolution stability assessment data;

[0177] Step S664: performing natural breakpoint activity level division according to the boundary evolution stability evaluation data to generate boundary activity partition data, wherein the boundary activity partition data includes high activity boundary data, medium activity boundary data and low activity boundary data;

[0178] Step S665: randomly taking a partition disturbance intensity value according to the boundary activity partition data, and determining a partition disturbance intensity range by using a preset partition boundary disturbance intensity index to generate partition disturbance intensity range data;

[0179] Step S666: Based on the partition disturbance intensity range data, using the boundary activity partition data, perform boundary morphological disturbance simulation on the predicted lithology evolution boundary data to obtain boundary morphological disturbance simulation data;

[0180] Step S667: performing boundary morphology probability statistics according to the boundary morphology disturbance simulation data to obtain boundary morphology probability distribution data;

[0181] Step S668: Based on the boundary morphology probability distribution data, the predicted lithology evolution boundary data is corrected in terms of boundary confidence to obtain corrected lithology evolution boundary data.

[0182] In an embodiment of the present invention, a morphological thinning algorithm is used to process the predicted lithologic evolution boundary data. The algorithm can gradually erode the boundary area until its center line is retained to form a boundary skeleton with a single pixel width. According to a preset point spacing, the lithologic prediction boundary skeleton data is divided into a series of discrete boundary points, and the selection of the point spacing needs to be adjusted according to the complexity of the boundary and the required accuracy. The current predicted lithologic evolution boundary data is compared with the boundary data at the previous moment, and the displacement vector of each boundary point is calculated. The size of the displacement vector represents the moving distance of the boundary point, and the direction represents the moving direction of the boundary point. According to the displacement vector of the boundary point, the displacement of the entire boundary is calculated, such as the average displacement, the maximum displacement, etc. According to the statistical analysis results of the boundary displacement, the evolution stability of the boundary is evaluated, and the boundary evolution stability evaluation data is generated. According to the statistical characteristics of the boundary displacement, such as the mean value, the standard deviation, etc., the boundary activity division threshold is set. For example, the boundary area with a displacement greater than the mean value plus one standard deviation can be divided into a high activity boundary, the boundary area with a displacement less than the mean value minus one standard deviation can be divided into a low activity boundary, and the remaining area can be divided into a medium activity boundary. According to the preset disturbance intensity indicators of the partition boundary, such as displacement, deformation degree, etc., as well as the knowledge of geological experts or empirical formulas, the disturbance intensity range of each activity level area is determined. For example, the disturbance intensity range of the high activity boundary can be set larger, while the disturbance intensity range of the low activity boundary can be set smaller. For each activity level area, randomly select values ​​within the preset disturbance intensity range to generate partition disturbance intensity range data. Apply the disturbance vector to the corresponding boundary point, update the position of the boundary point, and realize the disturbance of the boundary morphology. In the process of disturbance simulation, it is necessary to ensure that the topological structure of the boundary remains unchanged, such as avoiding the occurrence of boundary self-intersection, fracture, etc. Repeat the above steps, perform disturbance simulation on all boundary points, and generate multiple groups of different boundary morphology disturbance simulation data, reflecting the uncertainty of the boundary morphology. Convert the occurrence frequency of each pixel point into a probability value to generate boundary morphology probability distribution data. According to the boundary morphology probability distribution data, calculate the confidence of each boundary point. For example, the probability value of the boundary point can be used as its confidence, and the higher the probability value, the higher the confidence. According to the preset confidence threshold, classify the predicted boundary. For example, boundaries with confidence levels higher than a threshold can be marked as "high confidence boundaries", and boundaries with confidence levels lower than a threshold can be marked as "low confidence boundaries". For low confidence boundaries, they can be modified based on the information of surrounding high confidence boundaries or in combination with geological expert knowledge, such as deletion, movement, merging, etc., to generate modified lithology evolution boundary data.

[0183] The present invention also provides a deep learning-based intelligent lithology identification system, which executes the deep learning-based intelligent lithology identification method as described above. The deep learning-based intelligent lithology identification system includes:

[0184] The core image preprocessing module is used to obtain the original core image data; perform image equalization processing on the original core image data to generate enhanced core image data; perform texture image reconstruction processing on the enhanced core image data to generate multi-scale core texture image data;

[0185] The texture feature extraction module is used to extract texture hierarchical features based on multi-scale core texture image data to generate significant core texture structure feature data;

[0186] The fine rock formation segmentation module is used to construct super-pixel lithology nodes based on the core significant texture structure feature data to generate super-pixel core node data; perform fine rock formation image segmentation processing on the enhanced core image data through the super-pixel core node data to generate fine rock formation segmentation map data; perform rock formation attribute aggregation analysis on the fine rock formation segmentation map data to generate rock formation super-pixel aggregation data;

[0187] The intelligent lithology recognition module is used to build an intelligent lithology recognition model based on a preset deep convolutional neural network model; transmit the rock layer superpixel aggregation data to the intelligent lithology recognition model to identify the target lithology attributes and generate target lithology recognition label data;

[0188] The 3D core model building module is used to use the target lithology identification label data to perform lithology profile deformation correction on the fine rock layer segmentation map data to generate a 3D lithology surface model;

[0189] The lithology evolution prediction module is used to predict the time-series lithology evolution boundary based on the three-dimensional lithology surface model to obtain the predicted lithology evolution boundary data; perform boundary confidence correction processing on the predicted lithology evolution boundary data to generate corrected lithology evolution boundary data; perform lithology spatial distribution pattern recognition based on the corrected lithology evolution boundary data to generate lithology spatial distribution pattern data.

[0190] The beneficial effect of the present application is that through advanced image acquisition technology and image enhancement processing, clear, high-contrast core images can be obtained to highlight the texture details of the core. The multi-scale texture analysis method is used to extract the texture features of the core image, including texture direction, roughness and contrast, etc. These features can effectively eliminate noise and interference, overcome the subjectivity of manual observation, and provide richer and more objective information for lithology identification. By constructing superpixel lithology nodes, the core can be divided into small areas, each of which has similar texture structure characteristics, which helps to improve the accuracy of lithology identification. Then, the enhanced core image data is processed by fine rock layer image segmentation using superpixel core node data, which can accurately identify the different rock layers of the core, obtain clear rock layer boundaries and structural information, and thus realize refined rock layer segmentation and attribute identification. Build an intelligent model specifically for lithology identification. Through a deep learning algorithm, the model can automatically extract the key features of lithology identification from superpixel aggregation data, and learn the complex relationship between different lithologies, thereby achieving high-precision and automated lithology identification, overcoming the shortcomings of low efficiency and error-prone manual identification. The target lithology identification label data is used to correct the deformation of the segmentation results, eliminate the impact of rock deformation, and reconstruct the true rock morphology. On this basis, combined with time series analysis technology, the boundaries of lithology evolution can be predicted, and the laws of lithology change over time can be analyzed, providing more comprehensive and accurate information for oil and gas exploration and geological research.

[0191] Therefore, the embodiments should be regarded as illustrative and non-restrictive from all points, and the scope of the present invention is limited by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and range of equivalent elements of the application documents are included in the present invention.

[0192] The above description is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but should conform to the widest scope consistent with the principles and novel features invented herein.

Claims

1. A deep learning-based intelligent lithology identification method, characterized in that: The following steps are involved: Step S1: obtaining original core image data; Perform image equalization processing on the original core image data to generate enhanced core image data; Performing texture image reconstruction processing on the enhanced core image data to generate multi-scale core texture image data; Step S2: extracting texture hierarchical features based on the multi-scale core texture image data to generate core significant texture structure feature data; Step S3: construct superpixel lithology nodes according to the core significant texture structure feature data to generate superpixel core node data; perform fine rock layer image segmentation processing on the enhanced core image data through the superpixel core node data to generate fine rock layer segmentation map data; perform rock layer attribute aggregation analysis on the fine rock layer segmentation map data to generate rock layer superpixel aggregation data; wherein, step S3 is specifically as follows: Step S31: construct super-pixel lithology nodes according to the core significant texture structure feature data to generate super-pixel core node data; Step S32: Calculate the surrounding pixel texture similarity of the enhanced core image data based on the superpixel core node data, and perform iterative merging of homogeneous regions to generate initial core superpixel boundary data; Step S33: performing fuzziness evaluation of neighboring superpixel boundaries according to the initial core superpixel boundary data to generate superpixel boundary fuzzy evaluation data; Step S34: Perform fuzzy threshold judgment on the superpixel boundary fuzzy evaluation data by using a preset boundary fuzzy threshold. When the superpixel boundary fuzzy evaluation data is higher than or equal to the preset boundary fuzzy threshold, perform color texture difference calculation on the superpixel core node data to generate adjacent superpixel texture difference data. When the adjacent superpixel texture difference data is lower than or equal to the preset texture difference threshold, perform boundary penetration fusion processing on the initial core superpixel boundary data to generate fused superpixel core boundary data. Otherwise, retain the original initial core superpixel boundary data. Step S35: performing fine rock formation image segmentation processing on the enhanced core image data based on the fused super-pixel core boundary data and the initial core super-pixel boundary data to generate fine rock formation segmentation map data; Step S36: performing rock layer attribute aggregation analysis based on the fine rock layer segmentation map data to generate rock layer superpixel aggregation data; Step S4: construct an intelligent lithology recognition model based on a preset deep convolutional neural network model; transmit the rock layer superpixel aggregation data to the intelligent lithology recognition model to identify the target lithology attributes and generate target lithology recognition label data; wherein step S4 is specifically as follows: Step S41: performing self-organizing mapping processing on the rock layer super-pixel aggregation data to obtain rock layer super-pixel attribute feature data; Step S42: acquiring a plurality of core image sample data; performing image sample data expansion processing according to the plurality of core image sample data to generate expanded image sample data; Step S43: performing lithology label mapping processing according to the expanded image sample data to generate lithology image label sample data; Step S44: performing image feature engineering processing on the lithology image label sample data to generate core image attribute feature data; Step S45: randomly dividing the core image attribute feature data to obtain a lithology sample training set and a lithology sample verification set data; Step S46: Obtaining a preset deep convolutional neural network model and pre-trained model weight parameters; Step S47: initializing the preset deep convolutional neural network model through the pre-trained model weight parameters, and performing SE attention mechanism integration processing, so as to construct an initial intelligent lithology recognition model; Step S48: using the lithology sample training set to perform model training operations on the initial intelligent lithology identification model, and performing model hyperparameter tuning processing to obtain a trained intelligent lithology identification model; using the lithology sample verification set data to perform optimal model verification processing on the trained intelligent lithology identification model to obtain an intelligent lithology identification model; Step S49: transmitting the lithology superpixel attribute feature data to the intelligent lithology recognition model to perform target lithology attribute recognition and generate target lithology recognition label data; Step S5: using the target lithology identification label data to perform lithology profile deformation correction on the fine rock layer segmentation map data to generate a three-dimensional lithology curved surface model; Step S6: Predict the time-series lithologic evolution boundary based on the three-dimensional lithologic surface model to obtain predicted lithologic evolution boundary data; perform boundary confidence correction processing on the predicted lithologic evolution boundary data to generate corrected lithologic evolution boundary data; perform lithologic spatial distribution pattern recognition based on the corrected lithologic evolution boundary data to generate lithologic spatial distribution pattern data.

2. The deep learning-based intelligent lithology identification method according to claim 1 is characterized in that: Step S1 includes the following steps: Step S11: obtaining original core image data; Step S12: performing image color space conversion on the original core image data and performing adaptive image equalization processing to generate enhanced core image data; Step S13: calculating the neighborhood pixel texture intensity of the core texture response data using a preset core image sliding window rule to generate core texture distribution response data; Step S14: performing texture energy space scale decomposition on the core texture distribution response data to generate multi-scale texture energy space data; Step S15: Using the multi-scale texture energy space data, texture image reconstruction is performed on the enhanced core image data to generate multi-scale core texture image data.

3. The deep learning-based intelligent lithology identification method according to claim 2 is characterized in that: Step S13 includes the following steps: Step S131: Perform Gabor texture filtering on the enhanced core image data to generate core texture response data; Step S132: performing local binary pattern encoding processing according to the core texture response data, and calculating the neighborhood pixel texture intensity to generate local core texture feature value data; Step S133: performing morphological gradient processing according to the local core texture feature value data to generate core texture gradient data; Step S134: constructing a texture scale space based on the core texture gradient data, and performing multi-scale edge flow analysis to generate texture flow field feature data; Step S135: performing curvature scale space transformation processing on the texture flow field feature data to obtain multi-scale curvature feature map data; Step S136: performing texture topology continuity analysis based on the multi-scale curvature feature map data to obtain texture topology feature data; Step S137: Match the texture energy distribution of the enhanced core image data with the texture topological feature data to generate core texture distribution response data.

4. The deep learning-based intelligent lithology identification method according to claim 1 is characterized in that: Step S2 includes the following steps: Step S21: extracting texture level features based on the multi-scale core texture image data to generate texture level feature data; Step S22: performing structural tensor analysis on the texture level feature data to obtain core texture structural tensor data; Step S23: performing local phase calculation according to the core texture structure tensor data to generate texture local phase data; Step S24: using the local phase data of the texture to perform multi-directional phase consistency analysis on the texture hierarchical feature data, and performing significant texture tensor voting processing to generate significant texture structural feature data of the core.

5. The deep learning-based intelligent lithology identification method according to claim 1 is characterized in that: Step S5 includes the following steps: Step S51: performing dynamic contour tracking according to the fine lithology segmentation map data to obtain dynamic lithology contour interface data; Step S52: using a preset standard lithology attribute matching library to perform standard lithology attribute matching processing on the target lithology identification tag data to generate standard lithology attribute matching data; Step S53: performing lithology texture area filling processing on the dynamic lithology contour interface data through standard lithology attribute matching data to generate texture consistency lithology map data; Step S54: performing core depth correlation processing on the texture consistency lithology map data, and performing layered virtual splicing of the core images to obtain lithology layer texture structure data; Step S55: performing lithology profile deformation correction on the lithology hierarchy texture structure data according to preset geological structure constraint rules to generate corrected lithology hierarchy structure map data; Step S56: Perform lithology boundary surface interpolation processing according to the corrected lithology hierarchy structure map data to generate a three-dimensional lithology surface model.

6. The deep learning-based intelligent lithology identification method according to claim 1, characterized in that: Step S6 includes the following steps: Step S61: performing three-dimensional lithology voxel analysis according to the three-dimensional lithology surface model, and performing voxel connection processing to generate lithology connected area data; Step S62: performing fractal dimension calculation on the lithology connected area data to obtain a lithology distribution complexity index; performing lithology contact network analysis on the lithology connected area data using the lithology distribution complexity index to generate lithology contact network relationship characteristic data; Step S63: mining lithology evolution rules according to standard lithology attribute matching data to generate lithology evolution rule data; Step S64: performing transfer learning on the lithology evolution rule data through a preset long short-term memory network model, and performing core evolution time series modeling to obtain a lithology evolution prediction model; Step S65: using the lithology evolution prediction model to perform time series lithology evolution boundary prediction on the lithology contact network relationship characteristic data to obtain predicted lithology evolution boundary data; Step S66: performing boundary confidence correction processing on the predicted lithology evolution boundary data to generate corrected lithology evolution boundary data; Step S67: performing lithology spatial autocorrelation analysis based on the corrected lithology evolution boundary data to obtain lithology spatial distribution map data; Step S68: Perform lithology spatial distribution pattern recognition based on the lithology spatial distribution map data to generate lithology spatial distribution pattern data.

7. The deep learning-based intelligent lithology identification method according to claim 6 is characterized in that: Step S66 includes the following steps: Step S661: extracting the lithology boundary skeleton according to the predicted lithology evolution boundary data to generate lithology prediction boundary skeleton data; Step S662: dividing the lithology prediction boundary skeleton data into boundary points and performing continuous boundary interpolation processing to generate optimized prediction lithology boundary data; Step S663: Calculate boundary displacement according to the optimized predicted lithology boundary data, and perform boundary stability assessment to generate boundary evolution stability assessment data; Step S664: performing natural breakpoint activity level division according to the boundary evolution stability evaluation data to generate boundary activity partition data, wherein the boundary activity partition data includes high activity boundary data, medium activity boundary data and low activity boundary data; Step S665: randomly taking a partition disturbance intensity value according to the boundary activity partition data, and determining a partition disturbance intensity range by using a preset partition boundary disturbance intensity index to generate partition disturbance intensity range data; Step S666: Based on the partition disturbance intensity range data, using the boundary activity partition data, perform boundary morphological disturbance simulation on the predicted lithology evolution boundary data to obtain boundary morphological disturbance simulation data; Step S667: performing boundary morphology probability statistics according to the boundary morphology disturbance simulation data to obtain boundary morphology probability distribution data; Step S668: Based on the boundary morphology probability distribution data, the predicted lithology evolution boundary data is corrected in terms of boundary confidence to obtain corrected lithology evolution boundary data.

8. A deep learning-based intelligent lithology identification system, characterized in that: Used to execute the deep learning-based intelligent lithology identification method according to claim 1, the deep learning-based intelligent lithology identification system comprises: The core image preprocessing module is used to obtain the original core image data; perform image equalization processing on the original core image data to generate enhanced core image data; perform texture image reconstruction processing on the enhanced core image data to generate multi-scale core texture image data; The texture feature extraction module is used to extract texture hierarchical features based on multi-scale core texture image data to generate significant core texture structure feature data; The fine rock formation segmentation module is used to construct super-pixel lithology nodes based on the core significant texture structure feature data to generate super-pixel core node data; perform fine rock formation image segmentation processing on the enhanced core image data through the super-pixel core node data to generate fine rock formation segmentation map data; perform rock formation attribute aggregation analysis on the fine rock formation segmentation map data to generate rock formation super-pixel aggregation data; The intelligent lithology recognition module is used to build an intelligent lithology recognition model based on a preset deep convolutional neural network model; transmit the rock layer superpixel aggregation data to the intelligent lithology recognition model to identify the target lithology attributes and generate target lithology recognition label data; The 3D core model building module is used to use the target lithology identification label data to perform lithology profile deformation correction on the fine rock layer segmentation map data to generate a 3D lithology surface model; The lithology evolution prediction module is used to predict the time-series lithology evolution boundary based on the three-dimensional lithology surface model to obtain the predicted lithology evolution boundary data; perform boundary confidence correction processing on the predicted lithology evolution boundary data to generate corrected lithology evolution boundary data; perform lithology spatial distribution pattern recognition based on the corrected lithology evolution boundary data to generate lithology spatial distribution pattern data.

Images