Method for identifying middle-level boundary of geological envelope based on artificial intelligence

Abstract

The invention provides a method for identifying a middle-level boundary of a geological envelope based on artificial intelligence. The method comprises the following steps: determining seismic exploration data and logging data of a target work area; extracting seismic reflection features from the seismic exploration data based on the trained first neural network model, and extracting logging sequence features from the logging data by using the trained second neural network model; according to the seismic reflection characteristics and the logging sequence characteristics, geological characteristic representation of the target work area is generated; based on formation contact relation constraint and lithologic sequence constraint which are constructed for a target work area, according to geological characteristic representation, determining implicit topological structure representation of a representation geological envelope body in the target work area; and processing the implicit topological structure representation based on the trained third neural network model to predict the horizon boundary of the geological envelope in the target work area. According to the invention, the prediction result of the horizon boundary of the geological envelope body better conforms to the actual geological law.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of mineral resource exploration and the field of data interpretation, and in particular to a method for identifying a layer boundary in a geological envelope based on artificial intelligence. BACKGROUND

[0002] In the fields of oil and gas resource exploration, mineral resource development, and engineering geological survey, accurately determining the layer boundary of the geological envelope in the target work area is one of the core technical requirements. The accurate positioning of the layer boundary of the geological envelope directly affects the scientificity of stratigraphic division, resource reserve estimation, and development plan design, especially in complex geological structure areas, the boundary prediction needs to be realized through comprehensive analysis of multi-source geological data to meet the fine demand of geological information in actual engineering.

[0003] The existing scheme for predicting the layer boundary of the geological envelope mainly uses seismic exploration data as the main input, supplemented by a small amount of well logging data for auxiliary verification. After denoising the seismic data by a fixed parameter filtering algorithm, the threshold segmentation method is used to extract the seismic reflection phase axis as the preliminary indication of the layer boundary. Finally, the traditional statistical regression model is relied on, and the well logging lithology parameters are artificially screened to fit the preliminary indication boundary, and the layer boundary prediction result is obtained.

[0004] The above-mentioned existing scheme has obvious technical limitations: since only a single type of seismic feature is processed by a fixed algorithm, and no adaptation mechanism is established with the actual geological regularity of the target work area (such as stratigraphic contact relationship, lithology sequence distribution), the complementary information of multi-source data (seismic, well logging) cannot be effectively integrated. In complex geological conditions (such as diverse stratigraphic contact relationship, complex vertical transition of lithology), the predicted layer boundary is prone to spatial continuity fracture or large deviation from the actual geological envelope boundary, which is difficult to meet the demand of high-precision geological survey. SUMMARY

[0005] (I) Invention purpose

[0006] The main purpose of the present application is to provide a method for identifying a layer boundary in a geological envelope based on artificial intelligence, to overcome or alleviate the above problems in the prior art.

[0007] (II) Technical solutions

[0008] The present application provides a method for identifying a layer boundary in a geological envelope based on artificial intelligence, comprising:

[0009] Step 1, determining the seismic exploration data and well logging data of the target work area;

[0010] Step 2, extracting seismic reflection features from the seismic exploration data based on the first trained neural network model, and extracting logging sequence features from the logging data based on the second trained neural network model;

[0011] Step 3, generating a geological property representation of the target work area according to the seismic reflection features and the logging sequence features;

[0012] Step 4, determining an implicit topological structure representation of the geological envelope in the target work area according to the geological property representation based on the stratum contact relationship constraint and the lithology sequence constraint constructed for the target work area;

[0013] Step 5, processing the implicit topological structure representation based on the third trained neural network model to predict the stratigraphic boundary of the geological envelope in the target work area.

[0014] (III) Beneficial effects

[0015] 1. In view of the limitations of "fixed algorithm processing single type seismic features and manual screening of logging lithology parameters", the first trained neural network model is used in step 2 to extract seismic reflection features from seismic exploration data, and the second trained neural network model is used to extract logging sequence features from logging data. Unlike the traditional fixed parameter filtering method which can only process single seismic features and is easily affected by subjective factors, the two neural network models can dynamically learn the extraction rules of seismic reflection features and logging sequence features through training, without relying on fixed algorithms or manual intervention, and can more comprehensively capture multi-dimensional reflection information in seismic data and vertical sequence information in logging data, providing more abundant basic features for multi-source data integration and avoiding the limitations of single feature extraction.

[0016] 2. In view of the problem that the complementary information of multi-source data (seismic and logging) cannot be effectively integrated, the geological property representation of the target work area is generated according to the seismic reflection features and the logging sequence features in step 3. Unlike the passive role of logging data in the traditional scheme, this step actively combines the two features to make the macroscopic stratum reflection information reflected by seismic data and the microscopic lithology sequence information reflected by logging data complementary, directly converting into a geological property representation containing multi-source information, effectively solving the technical bottleneck of difficult integration of complementary information of multi-source data.

[0017] 3. For the defect of "not establishing an adaptive mechanism with the actual geological regularity (such as stratigraphic contact relationship, lithological sequence distribution) of the target work area": the step 4 of the present application determines the implicit topological structure representation based on the constructed stratigraphic contact relationship constraints and lithological sequence constraints of the target work area, combined with the geological property representation, which is different from the general processing mode of the traditional scheme which is independent of the actual geological regularity of the target work area. Through the targeted construction of geological regularity constraints matched with the target work area, it is ensured that the generated implicit topological structure representation conforms to the actual stratigraphic contact relationship and lithological sequence distribution of the target work area, avoiding the boundary prediction deviation caused by deviating from the actual geological regularity, and laying a foundation for subsequent accurate prediction of horizon boundary which conforms to the geological reality.

[0018] 4. For the problem of "the horizon boundary predicted in a complex geological area is easy to have spatial continuity fracture or large deviation": the step 5 of the present application processes the implicit topological structure representation based on the trained third neural network model to predict the horizon boundary - unlike the traditional statistical regression model which can only simply fit the preliminary boundary, the third neural network model can learn the spatial distribution regularity of the horizon boundary more accurately based on the implicit topological structure representation which conforms to the geological regularity of the target work area, and in the complex geological condition area (such as various stratigraphic contact relationship, complex vertical transition of lithology), the spatial continuity of the predicted horizon boundary is higher, and the deviation from the actual geological envelope boundary is lower, effectively solving the boundary prediction defects of the traditional scheme in the complex area, and meeting the demand of high-precision geological exploration in the background technology. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 A flowchart of a method for identifying a horizon boundary in a geological envelope based on artificial intelligence according to an embodiment of the present application. DETAILED DESCRIPTION

[0020] Figure 1 A flowchart of a method for identifying a horizon boundary in a geological envelope based on artificial intelligence according to an embodiment of the present application. As shown in Figure 1 , it includes:

[0021] Step 1, determining the seismic exploration data and logging data of the target work area;

[0022] Step 2, extracting the seismic reflection features from the seismic exploration data based on the trained first neural network model, and extracting the logging sequence features from the logging data using the trained second neural network model;

[0023] Step 3, generating the geological property representation of the target work area according to the seismic reflection features and the logging sequence features;

[0024] Step 4, based on the construction of the stratum contact relationship constraint and the lithology sequence constraint for the target work area, according to the geological characteristic representation, determine the implicit topological structure representation of the geological envelope in the target work area.

[0025] Step 5, based on the third neural network model trained, process the implicit topological structure representation to predict the stratigraphic boundary of the geological envelope in the target work area.

[0026] Optionally, step 1, determine the seismic exploration data and logging data of the target work area, specifically comprising:

[0027] Step 1.1, perform adaptive noise suppression processing on the original seismic exploration data of the target work area to generate seismic exploration denoising data;

[0028] Step 1.2, perform depth correction and outlier rejection processing on the original logging data of the target work area to generate logging data correction results;

[0029] Step 1.3, match the data space coordinates of the seismic exploration denoising data and the logging data correction results to generate the seismic exploration data and the logging data of the target work area.

[0030] Preferably, the specific implementation process of step 1.2 is as follows: the processing object is the original logging data of the target work area (the data is stored in a vertical sequence, and each data point contains an "original logging depth" field and a "physical parameter value" field, wherein the "physical parameter value" field covers parameters such as resistivity, acoustic time difference, density, and other parameters reflecting differences in formation lithology, and there are two types of problems in the data: one is that the "original logging depth" deviates from the actual formation depth (due to wellbore inclination), and the other is that the "physical parameter value" abnormally jumps (due to instrument interference or measurement error)); first, perform depth correction: obtain the wellbore trajectory data of the target work area (the data records the inclination angle and azimuth angle of each depth point of the wellbore, reflecting the deviation degree of the actual wellbore direction from the vertical well), and construct a depth correction matrix - the rows of the matrix represent the "original logging depth point number" of the original logging data (each number corresponds to an original logging depth value), the columns represent the "actual depth point number" (each number corresponds to an actual formation depth value), and the elements at the intersection of the matrix rows and columns are "original logging depth to actual depth mapping coefficients" (obtained by wellbore trajectory geometric calculation, for example, in the inclined well scenario, the mapping coefficient = cos (wellbore inclination angle), and the actual depth = original logging depth x mapping coefficient); substitute each "original logging depth value" in the original logging data into the depth correction matrix, and calculate the corresponding "actual depth value" through matrix operation, replace the "original logging depth" field in the original logging data with the "actual depth value", and generate the logging data after depth correction (the depth field of the data is consistent with the actual formation depth, solving the depth deviation problem caused by the traditional reliance on instrument recording depth, and laying a foundation for subsequent spatial matching with seismic data).

[0031] Then, the outlier elimination is performed: taking the "physical parameter value sequence" (such as the resistivity sequence, the acoustic travel time sequence, each sequence is arranged in the order of the actual depth) of the depth-corrected logging data as the processing object, calculating the vertical statistical range (statistical range = mean value ± 3 times the standard deviation of the parameter sequence, the range covers the reasonable fluctuation interval of the parameter under normal formation conditions) for each physical parameter sequence; constructing an outlier judgment vector - the length of the vector is consistent with the "actual depth point number" of the depth-corrected logging data, if the "physical parameter value" of a certain actual depth point is outside the statistical range of the corresponding parameter sequence, the element at the corresponding position of the vector is identified as "abnormal", otherwise it is identified as "normal"; for the depth points identified as "abnormal" in the outlier judgment vector, taking the "normal" depth points adjacent to it as the basis (usually selecting 2-3 continuous normal points above and below), using the linear interpolation method to calculate the "interpolated physical parameter value" of the abnormal point (by fitting the parameter variation trend of the adjacent normal points, ensuring that the interpolation result conforms to the gradual change law of the formation lithology), replacing the original abnormal value with the "interpolated physical parameter value"; finally generating the logging data correction result (the "actual depth" of the data is accurate, the "physical parameter value" has no abnormal jump and is continuous vertically, unlike the traditional processing method of only deleting abnormal values, this step avoids the data missing caused by subsequent feature extraction fault by interpolation completion, ensuring the vertical integrity of the logging data).

[0032] Preferably, in the specific technical implementation of step 1.3, the processing object is the denoised seismic exploration data and the corrected logging data (wherein the denoised seismic exploration data is three-dimensional grid data with a coordinate system of "lateral line X-longitudinal line Y-depth Z", and each grid point records "seismic amplitude value"; the corrected logging data is vertical point data with a coordinate system of "wellhead lateral X-wellhead longitudinal Y-actual depth Z", and each depth point records "actual depth" and "physical parameter value"; due to the difference in acquisition methods, the two types of data have different coordinate systems, and spatial consistency needs to be achieved through matching); first, a coordinate mapping table is constructed: taking the unified spatial coordinate system (such as the Gauss plane coordinate system + elevation depth coordinate system) of the target work area as the reference, the "three-dimensional grid coordinate set" of the denoised seismic exploration data (each element is the "lateral line X value-longitudinal line Y value-depth Z value" of a single grid point) and the "wellbore coordinate set" of the corrected logging data (each element is the "wellhead lateral X value-wellhead longitudinal Y value-actual depth Z value" of a single depth point) are extracted; for each depth point in the wellbore coordinate set, the spatial Euclidean distance between it and all grid points in the three-dimensional grid coordinate set is calculated, and the grid points with a distance less than half of the seismic data sampling interval are determined as "matching grid points" (because the grid points within this distance range are closest to the logging depth point in space, ensuring the rationality of parameter assignment); a coordinate mapping table is constructed, which represents the "depth point number" of the corrected logging data in the rows and the "grid point number" of the denoised seismic exploration data in the columns, and each cell in the table records the "matching grid point number" and the "corresponding spatial Euclidean distance" (the smaller the distance, the stronger the spatial correlation between the grid point and the logging depth point).

[0033] Subsequently, parameter assignment and data integration are performed: based on the coordinate mapping table, the "physical parameter value" (such as the resistivity value, the acoustic travel time value) of each depth point in the well logging data correction result is assigned to the "physical parameter field" of the corresponding "matching grid point" according to the "spatial Euclidean distance weighting" rule (the weighting rule is: the weight of a certain matching grid point = 1 / the spatial Euclidean distance between the grid point and the well logging depth point, and the total weight is normalized to 1 to ensure that the closer the grid point, the higher the parameter assignment proportion), and the coordinate-matched well logging data (which takes the three-dimensional grid of the denoised seismic exploration data as the carrier to realize the accurate mapping of well logging physical parameters to seismic space) is generated; the "seismic amplitude value" of the denoised seismic exploration data and the "physical parameter value" of the coordinate-matched well logging data are integrated one by one according to the "three-dimensional grid coordinates" (that is, the same grid point retains both the "seismic amplitude value" and the "physical parameter value", if there is no corresponding "matching grid point" for a certain grid point, only the "seismic amplitude value" is retained, and the relevant information is supplemented through the feature fusion of step 3); and finally the seismic exploration data and well logging data of the target work area are generated (the data is a multi-source data set with completely consistent spatial coordinates, which solves the fusion error problem caused by the disconnection of traditional multi-source data, and the spatial consistency provides a necessary prerequisite for the correlation extraction of "seismic reflection features" and "well logging sequence features" in step 2, ensuring that the two types of features correspond to the same stratigraphic unit during subsequent feature fusion).

[0034] Optionally, step 1.1, the original seismic exploration data of the target work area is subjected to adaptive noise suppression processing to generate denoised seismic exploration data, specifically including:

[0035] Step 1.1.1, the original seismic exploration data of the target work area is divided into spatially continuous windows to calculate the amplitude standard deviation, frequency distribution characteristics and amplitude change rate between adjacent windows for each divided window data, and determine the seismic data window statistical characteristics;

[0036] Step 1.1.2, according to the seismic data window statistical characteristics, combined with the pre-labeled effective reflection signal samples and noise samples, the original seismic exploration data of the target work area is subjected to window-by-window signal type determination, dynamic weight filtering is applied to the window identified as noise, and the original characteristics are retained for the window identified as effective signal, to generate denoised seismic exploration data.

[0037] Preferably, the specific implementation process of step 1.1.1 is as follows: the processing object is the original seismic exploration data of the target work area (the data is stored in the form of a three-dimensional data body, each data point records the seismic amplitude value corresponding to the "lateral line X-longitudinal line Y-depth Z" spatial position, and the data contains effective reflection signals of mixed strata (corresponding to the vibration response of continuous stratum interface, high spatial continuity), instrument electronic noise (randomly distributed, no spatial rule), and surface environmental interference noise (such as surface wave, energy concentration but spatial continuity is significantly different from effective signal)); first, according to the spatial sampling rule of the original seismic exploration data (preset sampling interval along the X-axis lateral line, Y-axis longitudinal line, and Z-axis depth direction), perform spatial continuous window division-the window is distributed in a continuous order in X, Y, and Z dimensions, and adjacent windows retain overlapping areas in each dimension (the overlapping ratio is determined based on the sampling interval, such as 50%, the design purpose is to avoid the feature loss of effective reflection signals at the window boundary due to fracture), to obtain the divided window data (each window data contains all seismic amplitude values within the corresponding spatial range, and the window numbers are arranged in order of X-Y-Z, which is convenient for subsequent feature tracing);

[0038] Then, perform multi-dimensional feature calculation on each divided window data: first, calculate the amplitude standard deviation (statistical standard deviation of all seismic amplitude values in the window, reflecting the signal fluctuation degree in the window-the effective reflection signal has low fluctuation degree due to stratum continuity, and the noise component has high fluctuation degree); second, calculate the frequency distribution feature by time-frequency conversion (perform Fourier transform on the window data to convert the time domain amplitude signal into frequency domain energy signal, and calculate the main frequency value with the highest energy in the frequency domain and the frequency band width value with energy greater than 50% of the main frequency energy); third, calculate the amplitude change rate between adjacent windows (for each window, select the upper, lower, left, right, front and back six adjacent windows, calculate the difference between the amplitude mean value of adjacent windows and the amplitude mean value of the current window, and the proportion is (the amplitude mean value of adjacent windows-the amplitude mean value of the current window) / the amplitude mean value of the current window); the amplitude change rate matrix between adjacent windows is obtained (the row of the matrix corresponds to the window number, the column corresponds to the six adjacent directions, and the matrix element is the amplitude change rate in the corresponding direction-the effective reflection signal has low change rate due to stratum continuity, and the noise component has high change rate);

[0039] Finally, the window amplitude standard deviation vector, the window frequency distribution feature matrix, and the amplitude change rate matrix between adjacent windows are associated and integrated according to the window number, to ensure that the three types of features of the same window correspond to each other, and to generate the seismic data window statistical features (the features completely record the signal fluctuation, frequency distribution, and spatial variation characteristics of each window, and provide quantitative basis for the signal type determination of step 1.1.2, and each feature component can directly reflect the difference between the effective signal and the noise in the window).

[0040] Preferably, in the specific technical implementation of step 1.1.2, the processing objects are the seismic data window statistical features and the pre-labeled effective reflection signal samples and noise samples (wherein the effective reflection signal samples and the noise samples are both selected from the seismic profile of the target work area that has been verified by drilling - the effective reflection signal samples correspond to the seismic event area at the stratum interface revealed by drilling, and the seismic data window data classification statistical features of the area are extracted as a reference; the noise samples correspond to the blank area without stratum reflection or the known interference source area revealed by drilling, and the seismic data window statistical features of the area are extracted as a reference); first, a signal type determination reference is constructed: for the seismic data window data classification statistical features of the effective reflection signal samples, the mean range of each feature component (such as the mean of the amplitude standard deviation ± 1 times the standard deviation, the mean of the main frequency ± 1 times the standard deviation) is calculated to obtain the effective signal feature reference range; for the seismic data window statistical features of the noise samples, the mean range of each feature component is also calculated to obtain the noise feature reference range;

[0041] Then, the window-by-window signal type determination is performed: taking each window feature in the seismic data window statistical features as a processing unit, the amplitude standard deviation, the main frequency value, and the amplitude change rate between adjacent windows of the window are matched with the effective signal feature reference range and the noise feature reference range respectively - the number of feature components that meet the effective signal feature reference range and the number of feature components that meet the noise feature reference range of the window are counted; if the number of components that meet the effective signal feature reference range is more, the window is identified as “effective signal”; if the number of components that meet the noise feature reference range is more, the window is identified as “noise”, and a window signal type identification vector is generated (the length of the vector is consistent with the number of windows, and the elements are “effective signal” or “noise”, and the element subscript corresponds to the window number one by one);

[0042] Subsequently, dynamic weight filtering is performed on the windows identified as "noise": based on the window signal type identification vector, all windows identified as "noise" are screened out, the noise proportion of each noise window is calculated (the proportion of the amplitude value exceeding the amplitude standard deviation range in the effective signal characteristic reference range in the window - the higher the proportion, the higher the noise content in the window), the filtering weight is dynamically adjusted according to the noise proportion (the noise proportion is positively correlated with the filtering weight, for example, when the noise proportion is 80%, the filtering weight is 0.9, when the noise proportion is 50%, the filtering weight is 0.5, the design reason is to avoid the excessive suppression of the effective signal in the low noise proportion window by the fixed weight filtering); for each noise window, a Gaussian filter kernel matching the window size is used, and filtering is applied according to the dynamically adjusted filtering weight (the filter kernel filtering weight is realized by Gaussian function distribution, the greater the weight, the stronger the smoothing suppression of the amplitude value in the window), to obtain the filtered noise window data;

[0043] The windows identified as "effective signal" in the window signal type identification vector are directly retained with the original amplitude characteristics to avoid weakening of the effective reflection signal by filtering; finally, the filtered noise window data and the effective signal window data with the original characteristics are re-integrated according to the spatial coordinates of the original three-dimensional data body to ensure that the amplitude value of each spatial position corresponds to the corrected window data, and the denoised seismic exploration data is generated (the spatial continuity of the effective reflection events in the data is high, and the energy of the noise component is significantly suppressed - unlike the limitation of traditional fixed parameter filtering which can only optimize a single noise type, this step dynamically adjusts the noise content of different windows through dynamic weight adaptation, suppressing noise while improving the retention of effective signal, and the denoised seismic exploration data will be used as the seismic end basic data matched in step 1.3).

[0044] Optionally, in step 2, the first neural network model is trained to extract seismic reflection features from the seismic exploration data, and the second neural network model is trained to extract logging sequence features from the logging data, specifically including:

[0045] Step 2.1, based on the trained first neural network model, multi-scale feature extraction is performed on the denoised seismic exploration data to obtain initial seismic reflection features;

[0046] Step 2.2, based on the trained second neural network model, vertical sequence feature extraction is performed on the logging data to generate initial logging sequence features;

[0047] Step 2.3, the redundant features irrelevant to the formation boundary in the initial seismic reflection features and the initial logging sequence features are removed to generate screened seismic reflection features and screened logging sequence features.

[0048] Optionally, step 2.1, based on the first neural network model trained, multi-scale feature extraction is performed on the denoised seismic exploration data to obtain initial seismic reflection features, specifically including:

[0049] Step 2.1.1, based on the multi-scale convolution feature extraction layer in the first neural network model trained, feature extraction at different spatial scales is performed on the denoised seismic exploration data to generate preliminary multi-scale seismic feature maps;

[0050] Step 2.1.2, based on the cross-scale feature attention fusion layer in the first neural network model trained, attention weight calculation and feature fusion are performed on the preliminary multi-scale seismic feature maps to generate fused seismic feature maps;

[0051] Step 2.1.3, based on the residual noise suppression layer in the first neural network model trained, residual operation is performed on the fused seismic feature maps to generate denoised seismic feature maps;

[0052] Step 2.1.4, based on the feature dimension compression and selection layer in the first neural network model trained, dimension compression is performed on the denoised seismic feature maps to eliminate redundant feature dimensions, effective feature selection is performed to retain features strongly related to the horizon boundary, to generate initial seismic reflection features.

[0053] Preferably, the specific implementation process of step 2.1.1 is as follows: the processing object is the denoised seismic exploration data (the data is three-dimensional grid data, each grid point records the denoised seismic amplitude value corresponding to the "horizontal line X-vertical line Y-depth Z" spatial position, the spatial continuity of the effective reflection signal in the data is high, the residual noise energy is low, and it provides a high-quality data basis for feature extraction); relying on the multi-scale convolution feature extraction layer in the first neural network model trained (three three-dimensional convolution kernels of different sizes are built in this layer, the design purpose is to capture seismic reflection features of different spatial scales respectively - small size convolution kernel is suitable for local reflection feature extraction of fine stratigraphic interface, medium size convolution kernel is suitable for medium scale reflection feature extraction of stratigraphic section, large size convolution kernel is suitable for macro reflection feature extraction of regional stratigraphic distribution), scale feature extraction is performed on the denoised seismic exploration data:

[0054] First, a small size three-dimensional convolution kernel is used to perform convolution operation on the denoised seismic exploration data (the convolution kernel slides along the X-Y-Z three-dimensional direction, the sliding step is matched with the seismic data sampling interval each time, the edge features are preserved during the operation process to avoid loss of boundary information), the reflection event detail features at the stratigraphic interface (such as reflection amplitude mutation point, continuity inflection point of reflection event) are extracted, and a small scale seismic feature map is obtained;

[0055] Then, the medium-size three-dimensional convolution kernel is adopted to perform convolution operation on the denoised seismic exploration data (the convolution kernel covers a larger range than the small-size convolution kernel, and the sliding step is consistent with the small-size), to extract the reflection structure features (such as the combined reflection mode of multiple adjacent stratigraphic interfaces, and the reflection energy distribution corresponding to the stratigraphic thickness variation) in the stratigraphic section, to obtain a medium-scale seismic feature map;

[0056] Finally, the large-size three-dimensional convolution kernel is adopted to perform convolution operation on the denoised seismic exploration data (the convolution kernel covers the largest range, and is suitable for regional scale feature extraction), to extract the macroscopic distribution features (such as the overall tilting direction of the strata, and the continuity trend of the reflection layers in the region) of the strata in the target work area, to obtain a large-scale seismic feature map;

[0057] The small-scale seismic feature map, the medium-scale seismic feature map, and the large-scale seismic feature map are associated and integrated according to the spatial coordinates (to ensure that the different scale features at the same spatial position are one-to-one corresponding, and to facilitate subsequent fusion), to generate a preliminary multi-scale seismic feature map (the feature map contains three scale seismic reflection features, each scale feature can reflect the stratigraphic properties from different dimensions, which is different from the traditional single-scale feature extraction which can only capture local or macroscopic single information).

[0058] Preferably, in the specific technical implementation of step 2.1.2, the processing object is the preliminary multi-scale seismic feature map (the feature map contains the small-scale seismic feature map, the medium-scale seismic feature map, and the large-scale seismic feature map, the spatial coordinate systems of the three types of sub-maps are consistent, but the feature dimensions and the reflected geological information are different - the small-scale features focus on interface details, the medium-scale features focus on intra-segment structure, and the large-scale features focus on regional distribution); relying on the cross-scale feature attention fusion layer in the first neural network model trained, the attention weights of the three types of sub-maps in the preliminary multi-scale seismic feature map are calculated first:

[0059] Referring to the labeled stratigraphic boundary samples in the target work area (the samples contain the scale features corresponding to the known stratigraphic boundary positions), the similarity of each feature point in the preliminary multi-scale seismic feature map and the stratigraphic boundary sample features is calculated (the higher the similarity, the stronger the relevance of the feature point to the stratigraphic boundary), and the attention weights are assigned according to the similarity (the stronger the relevance of the feature point, the higher the weight, and the weaker the relevance, the lower the weight), to generate a small-scale feature attention weight map, a medium-scale feature attention weight map, and a large-scale feature attention weight map;

[0060] Subsequently, the three types of sub-maps are respectively subjected to point-by-point weighted operation (i.e., sub-map feature value x corresponding weight value) with the corresponding attention weight maps, to strengthen the features strongly related to the stratigraphic boundary and weaken the irrelevant redundant features, to obtain a weighted small-scale seismic feature map, a weighted medium-scale seismic feature map, and a weighted large-scale seismic feature map;

[0061] Finally, the three types of weighted subgraphs are executed to perform feature fusion (the spatial coordinates are kept aligned during the fusion process, the three types of weighted feature values at the same spatial position are spliced in the channel dimension to form a multi-channel feature vector), to generate a fused seismic feature map (the feature map integrates key reflection features of different scales, and highlights the horizon boundary related features through the attention mechanism, solving the problem of dilution of effective features caused by traditional simple feature splicing).

[0062] Preferably, in one scenario, when step 2.1.3 is specifically implemented, the processing object is the fused seismic feature map (the feature map is a multi-channel three-dimensional feature map, each channel corresponds to a type of weighted scale feature, and there may still be a small amount of residual noise superimposed on the effective reflection feature in the feature map, such as abnormal fluctuations in the feature value of the local region, which needs to be further suppressed to improve the feature purity); relying on the residual noise suppression layer in the first neural network model trained (the layer includes a residual connection and a noise suppression branch, and the design logic is to separate the effective feature and the residual noise through residual operation to avoid excessive suppression of the effective feature by traditional denoising methods), residual denoising processing is performed on the fused seismic feature map:

[0063] First, the fused seismic feature map is input into the noise suppression branch (the branch learns the residual noise distribution pattern in the fused seismic feature map through convolution and activation function operation, and outputs a noise feature map - the spatial dimensions of the noise feature map are the same as those of the fused seismic feature map, and each feature point value corresponds to the predicted residual noise intensity); Figure 1

[0064] Next, a residual operation relationship is established: based on the fused seismic feature map, the corresponding feature values in the noise feature map are subtracted (point-by-point subtraction to ensure accurate separation of effective features and noise at the same spatial position), to obtain a preliminary denoised feature map;

[0065] Finally, the effective feature components (such as feature points strongly related to the horizon boundary) in the fused seismic feature map are directly transmitted to the preliminary denoised feature map through the residual connection, to compensate for the loss of weak effective features in the residual operation, to generate a denoised seismic feature map (the residual noise energy in the feature map is significantly reduced, and the integrity and clarity of the effective reflection feature are higher, providing high-purity feature data for subsequent dimension compression and screening).

[0066] Preferably, in the specific technical implementation of step 2.1.4

[0067] The processing object is the denoised seismic feature map (the feature map is a multi-channel three-dimensional feature map, the number of channels is the same as that of the fused seismic feature Figure 1 ​The feature dimensions of part of the channels are redundant, such as different channel features reflecting the same geological information, and the feature effectiveness needs to be improved through compression and screening; relying on the feature dimension compression and screening layer in the first neural network model completed by training (the layer integrates a dimension compression module and a feature screening module to realize integrated "compression-screening" processing and avoid feature loss caused by traditional staged processing), the feature optimization processing is performed on the denoised seismic feature map:

[0068] Firstly, the dimension compression module is started: dimension reduction operation is performed on the multi-channel features of the denoised seismic feature map (by analyzing the variance contribution degree of each channel feature, the channel with high variance contribution degree is retained, and the redundant channel with very low variance contribution degree is removed-the variance contribution degree reflects the information amount of the channel feature, and the higher the contribution degree, the richer the geological information contained), to obtain the compressed seismic feature map (the channel number of the feature map is lower than that of the denoised seismic feature map, the meaningless redundant dimension is removed, and the complexity of subsequent processing is reduced);

[0069] Then, the feature screening module is started: taking the prior geological law of the target work area stratum boundary (such as the variance range of the reflection characteristics at the stratum boundary and the difference range with the adjacent stratum characteristics) as a reference, the correlation of each feature point in the compressed seismic feature map with the stratum boundary is calculated (the correlation is quantified by the matching degree of the feature value and the prior law, the higher the matching degree, the stronger the correlation);

[0070] Finally, the feature points with strong correlation in the compressed seismic feature map are retained, and the feature points with weak correlation are removed (the removal threshold is set based on the prior geological law, to ensure that the retained features all have strong association with the boundary), to generate the initial seismic reflection feature (the feature is a three-dimensional spatial distribution feature set, each feature point corresponds to clear "spatial coordinate-reflection feature value" information, directly provides processing objects for the redundant feature removal of step 2.3, and the feature purity is high, which can accurately reflect the seismic reflection properties related to the stratum boundary).

[0071] The first neural network model is trained with the core objective of "extracting seismic reflection features strongly related to the boundary of the layer". The training data uses seismic exploration data samples (including denoised seismic exploration data) and corresponding labeled stratum boundary samples (the labeled samples include known stratum boundary positions and corresponding seismic reflection feature labels) of the same type of target work area. During training, the denoised seismic exploration data samples are first input into the multi-scale convolution feature extraction layer of the model. The model learns the local, intra-segment, and regional scale reflection features through different size convolution kernels to generate preliminary multi-scale seismic feature map samples. Then, the cross-scale feature attention fusion layer is input. The model learns the correlation weight of different scale features and the stratum boundary based on the labeled samples, optimizes the feature fusion strategy, and highlights the boundary-related features in the fused seismic feature map. Subsequently, the residual noise suppression layer separates the residual noise in the fused feature samples by learning the residual mapping to generate denoised seismic feature map samples. Finally, the feature dimension compression and selection layer learns the redundancy judgment rule and effective feature selection standard of the feature dimension to retain the features strongly related to the boundary of the layer. During the training process, the mean square error loss function is used to compare the initial seismic reflection feature samples output by the model with the reflection feature labels in the labeled samples. The convolution kernel parameters, attention weights, and residual mapping parameters of each layer are iteratively adjusted through back propagation until the loss function value converges to the preset threshold, and the model training is completed. The model has the ability to extract high-quality seismic reflection features from seismic data.

[0072] Optionally, step 2.2, based on the trained second neural network model, vertical sequence feature extraction is performed on the logging data to generate initial logging sequence features, specifically including:

[0073] Step 2.2.1, based on the vertical window sequencing layer in the trained second neural network model, vertical window division and sequence conversion are performed on the logging data to generate a set of vertical logging sequence segments;

[0074] Step 2.2.2, based on the sequence residual denoising layer in the trained second neural network model, residual operation is performed on the set of vertical logging sequence segments to generate a set of denoised vertical logging sequence segments;

[0075] Step 2.2.3, based on the multi-feature attention reinforcement layer in the trained second neural network model, attention weight calculation is performed on the set of denoised vertical logging sequence segments to generate a set of attention-reinforced logging sequence features;

[0076] Step 2.2.4, based on the sequence feature selection layer in the trained second neural network model, feature correlation analysis and selection are performed on the set of attention-reinforced logging sequence features to generate initial logging sequence features.

[0077] Preferably, the specific implementation process of step 2.2.1 is as follows: the processing object is the well logging data (the data is the target work area well logging data generated in step 1.3, which contains the "lateral line X-longitudinal line Y-actual depth Z" spatial coordinates and the corresponding "resistivity-acoustic travel time-density" physical parameter sequence, the physical parameters are continuously distributed along the actual depth direction, and the parameter value of each depth point reflects the lithology attribute of the corresponding formation, which needs to be captured through serialization processing to capture the vertical lithology change rule); relying on the vertical window serialization layer in the second neural network model trained, first perform vertical window division:

[0078] According to the actual depth sampling interval of the well logging data (such as one basic unit per preset depth interval along the Z axis), continuous vertical windows are divided - the windows are only distributed along the actual depth direction, adjacent windows retain overlapping areas (the overlapping ratio is determined based on the sampling interval, such as 50%, the design purpose is to avoid the loss of sequence characteristics of the lithology transition section due to window cracking), and the window numbers are arranged in order from shallow to deep according to the actual depth, to obtain the divided vertical windows (each window contains all physical parameter values in the corresponding depth range, such as resistivity, acoustic travel time, and density data of "depth Z1-Z2");

[0079] Then perform sequence conversion: arrange the physical parameter values in each divided vertical window in the order of "actual depth increasing" to form a one-dimensional parameter sequence (the sequence length is consistent with the number of depth points in the window, and the sequence elements are combined vectors of "resistivity value-acoustic travel time value-density value" in turn), and associate the window number with the corresponding one-dimensional parameter sequence to ensure that each sequence can be traced back to the original depth range;

[0080] Finally, integrate the one-dimensional parameter sequences of all divided vertical windows to generate a vertical well logging sequence segment set (the segment set is a three-dimensional structure with dimensions of "window number-sequence length-parameter type", each segment can reflect the vertical lithology parameter change in the local depth range, which is different from the problem that local lithology details are diluted by traditional full-depth sequence processing).

[0081] Preferably, in the specific technical implementation of step 2.2.2, the processing object is the vertical well logging sequence segment set (the segment set contains multiple vertical well logging sequence segments, and there may still be a small amount of measurement noise in the segment - such as abnormal jump of resistivity value at a certain depth point, which is inconsistent with the lithology change rule of adjacent depth points, and needs to be removed to retain the true vertical lithology sequence characteristics); relying on the sequence residual denoising layer in the second neural network model trained, first construct a residual processing branch:

[0082] The vertical logging sequence fragment set is input into a noise extraction branch, which learns the distribution pattern of noise in the fragment through convolution and cyclic activation function operation, such as random jump parameter value characteristics, and outputs a noise sequence fragment set, the dimension of which is consistent with that of the vertical logging sequence fragment set, and each element corresponds to a predicted noise value;

[0083] Then, a residual operation is performed: the noise values corresponding to the noise sequence fragment set are subtracted from the vertical logging sequence fragment set element by element (i.e., the parameter values of the base sequence - the noise values), to separate a preliminary denoising sequence fragment set (which has removed most of the random noise, but may have lost weak lithology transition characteristics);

[0084] Finally, the effective features related to lithology transition in the vertical logging sequence fragment set are directly transmitted to the preliminary denoising sequence fragment set to correct the feature deviation caused by the residual operation, generating a denoised vertical logging sequence fragment set (the physical parameter sequence of which has high vertical continuity, and the lithology change rule is clear, solving the problem of easy misdeletion of weak lithology characteristics in traditional fixed threshold denoising).

[0085] Preferably, in a scenario, when step 2.2.3 is implemented, the processing object is the denoised vertical logging sequence fragment set (which contains multiple denoised vertical logging sequence fragments, and different physical parameters in the fragments have different indications of lithology boundaries - for example, resistance mutations often correspond to sand-shale boundaries, and acoustic travel time gradients often correspond to changes in the density of the same rock, and the parameter characteristics strongly related to lithology boundaries need to be strengthened through attention mechanisms); relying on the multi-feature attention strengthening layer in the second neural network model trained, the attention weight is first calculated:

[0086] Referring to the labeled lithology boundary samples of the target work area (which contain the sequence characteristics of each physical parameter corresponding to the known lithology boundary depth point), the correlation between each sequence element in the denoised vertical logging sequence fragment set and the lithology boundary sample characteristics is calculated (the correlation is quantified by the matching degree of the parameter value change rate and the sample change rate, and the higher the matching degree, the stronger the correlation between the element and the lithology boundary);

[0087] According to the correlation, the attention weight is assigned (the stronger the correlation, the higher the weight of the sequence element, and the weaker the correlation, the lower the weight), and a parameter attention weight matrix is generated (the dimension of which is consistent with that of the denoised vertical logging sequence fragment set, the row corresponds to the window number, the column corresponds to the sequence length, the channel corresponds to the parameter type, and the element is the attention weight at the corresponding position);

[0088] Then the weighted reinforcement is performed: the denoised vertical logging sequence segment set is multiplied element by element with the parameter attention weight matrix (i.e. sequence parameter value x corresponding weight value), to reinforce the characteristics related to lithology boundary such as resistivity mutation, density sudden change, and to weaken the irrelevant characteristics such as acoustic travel time slight fluctuation, to obtain the weighted vertical logging sequence segment set;

[0089] Finally, the weighted vertical logging sequence segment set is integrated to generate the attention-reinforced logging sequence feature set (the feature set highlights the key parameter features of vertical lithology change, laying a foundation for subsequent screening of sequence features related to stratigraphic boundary).

[0090] Preferably, in the specific technical implementation of step 2.2.4, the processing object is the attention-reinforced logging sequence feature set (the feature set contains multiple attention-reinforced logging sequence features, some of which are irrelevant to the stratigraphic boundary, such as density slight fluctuation features within the same lithology section, which need to be screened to improve feature effectiveness); relying on the sequence feature screening layer in the second neural network model trained, first perform feature relevance analysis:

[0091] Taking the labeled stratigraphic boundary samples of the target work area as a reference (the samples contain sequence features corresponding to known stratigraphic boundary depth points), the cosine similarity between each feature in the attention-reinforced logging sequence feature set and the stratigraphic boundary sample features is calculated (the higher the similarity, the stronger the relevance of the feature to the stratigraphic boundary), and a similarity threshold is set (the threshold is determined based on sample statistics, such as similarity greater than a predetermined threshold is determined as "relevant feature", otherwise as "redundant feature");

[0092] Then perform feature screening: retain the relevant features in the attention-reinforced logging sequence feature set with similarity greater than the threshold, and eliminate redundant features, such as eliminating resistivity smooth sequence features within the same sandstone layer, and retaining resistivity mutation sequence features at sand-shale interfaces;

[0093] Finally, the screened relevant features are integrated according to the "window number-actual depth" association to ensure that each feature can be traced back to the original spatial location, to generate the initial logging sequence feature (the feature is a mapping set of "spatial coordinates-vertical lithology change sequence", only contains vertical sequence features strongly related to stratigraphic boundary, directly provides processing object for step 2.3 redundant feature elimination, and the feature is highly targeted, can accurately reflect the lithology sequence change at the stratigraphic boundary).

[0094] Preferably, the specific implementation process of step 2.3 is as follows, and the processing objects are the initial seismic reflection features and the initial logging sequence features (wherein the initial seismic reflection features are the three-dimensional feature set generated in step 2.1.4, each feature point contains the “lateral line X-longitudinal line Y-depth Z” coordinates and the corresponding reflection feature value, reflecting the spatial reflection attribute of the formation boundary; the initial logging sequence features are the vertical sequence feature set generated in step 2.2.4, each feature segment contains the “wellhead X-wellhead Y-actual depth Z” coordinates and the corresponding physical parameter sequence value, reflecting the vertical lithology change attribute of the formation, and there are redundant features in both types of features that are irrelevant to the formation boundary, such as background noise residual features in the seismic reflection features and smooth parameter features in the same lithology section in the logging sequence features);

[0095] First, the redundant features in the initial seismic reflection features are removed: taking the labeled formation boundary samples (the samples contain the known formation boundary position and the corresponding reflection features) of the target work area as the reference, the correlation degree of each feature point in the initial seismic reflection features and the formation boundary sample features is calculated (the correlation degree is quantified by the matching of the feature value change trend and the sample change trend, the higher the matching degree, the higher the correlation degree), and a seismic feature correlation matrix is generated (the rows of the matrix correspond to the feature point numbers of the initial seismic reflection features, the columns correspond to the formation boundary sample numbers, and the elements at the intersection of the matrix are the correlation degree values of the corresponding feature points and samples); a correlation threshold is set (the threshold is determined based on sample statistics, such as determining that it is a “correlation feature with the formation boundary” if the correlation degree is greater than a preset threshold, otherwise it is a “redundant feature”), the feature points in the seismic feature correlation matrix with a correlation degree greater than the threshold are retained, the redundant feature points are removed, and the preliminary screened seismic reflection features are generated (this feature set has removed most of the reflection features irrelevant to the formation boundary, but still needs to be cross-verified with the logging sequence features to avoid cross-feature redundancy);

[0096] Then, the redundant features in the initial logging sequence features are removed: also taking the formation boundary samples as the reference (the samples contain the known formation boundary depth points and the corresponding logging parameter sequences), the similarity of each feature segment in the initial logging sequence features and the formation boundary sample sequence is calculated (the similarity is quantified by the fitting degree of the vertical parameter change rate and the sample change rate, the higher the fitting degree, the higher the similarity), and a logging feature similarity vector is generated (the length of the vector is consistent with the number of feature segments of the initial logging sequence features, and the vector elements are the similarity values of the corresponding feature segments and samples, and the element subscript is one-to-one corresponding to the feature segment number); a similarity threshold is set (the threshold is determined based on sample statistics, such as determining that it is a “correlation feature with the formation boundary” if the similarity is greater than a preset threshold, otherwise it is a “redundant feature”), the feature segments in the logging feature similarity vector with a similarity greater than the threshold are retained, the redundant feature segments are removed, and the preliminary screened logging sequence features are generated;

[0097] Finally, cross-feature redundancy checking is performed: the mutual information value between the preliminary screened seismic reflection features and the preliminary screened well sequence features is calculated (the mutual information value reflects the degree of information overlap between the two types of features, the higher the value, the more similar the geological information reflected by the two types of features, and there is redundancy); set the mutual information threshold (such as if the mutual information value is greater than the preset threshold, it is determined as "cross-feature redundancy"); for the features determined as cross-feature redundancy, such as a certain seismic reflection feature point and a certain well sequence feature segment both reflect the same sand-shale boundary, the feature with more information is retained (preferably the well sequence feature, because the well data directly reflects the lithology, and the indication of the stratigraphic boundary is more direct), and the other type of redundant feature is removed;

[0098] Integrate the features after cross-feature redundancy checking to generate screened seismic reflection features and screened well sequence features (both types of features only retain features strongly related to the stratigraphic boundary, and there is no cross-feature information redundancy, unlike the information duplication problem caused by traditional single-feature screening, this step significantly improves the effectiveness of the features through "single-class screening + cross-class checking" double processing, and provides high-quality input for the subsequent step 3 feature fusion).

[0099] The second neural network model focuses on "letting the model efficiently extract well sequence features reflecting the vertical change of lithology", and the training data is the well data samples (including well data correction results) and the corresponding labeled lithology sequence labels of the same type target work area. The labeled samples include the known lithology boundary depth and the corresponding well sequence feature label. During training, first input the well data sample into the vertical window sequence layer, the model learns the optimal division size of the vertical window and the sequence conversion rule, and converts the well data sample into a vertical well sequence fragment set sample; then input the sequence residual denoising layer, remove the measurement noise in the fragment set sample by learning the sequence residual mapping, and generate a denoised vertical well sequence fragment set sample; the multi-feature attention reinforcement layer learns the correlation weight of different well parameter samples (resistivity, acoustic travel time, etc.) and the lithology boundary based on the labeled sample, strengthens the sequence features of key parameters, and generates an attention-reinforced well sequence feature set sample; finally, the sequence feature screening layer learns the correlation judgment standard of the features and the lithology boundary, and screens out the effective features. The training uses the cross-entropy loss function, compares the initial well sequence feature sample output by the model with the lithology sequence label in the labeled sample, and iteratively optimizes the window division parameters, residual mapping coefficients and attention weights through back propagation. When the loss function value is stable and the model's extraction accuracy of the lithology sequence meets the standard, the training is completed, so that the model can extract accurate vertical well sequence features from the well data.

[0100] Optionally, step 3, generating a geological property representation of the target work area according to the seismic reflection features and the well sequence features, specifically includes:

[0101] Step 3.1, based on the spatial coordinate system of the target work area, the spatial coordinate of the seismic reflection feature and the logging sequence feature is anchored, and then the correlation weight is calculated according to the correlation degree of the spatial coordinate anchored seismic reflection feature and the logging sequence feature and the stratum boundary, to generate a feature correlation mapping table;

[0102] Step 3.2, based on the feature correlation mapping table, the seismic reflection feature and the logging sequence feature are fused by weighting, and the fused contradictory features are removed by checking the stratum topological relationship, to generate a fused geological feature set;

[0103] Step 3.3, based on the regional geological background of the target work area, the fused geological feature set is processed by spatial continuity, and the corresponding geological significance of each feature is labeled, to generate the geological characteristic representation of the target work area.

[0104] Preferably, the specific implementation process of step 3.1 is as follows, the processing object is the spatial coordinate system of the target work area, the screened seismic reflection feature and the screened logging sequence feature (wherein the spatial coordinate system of the target work area is defined as a "horizontal survey line X-vertical survey line Y-depth Z" three-dimensional rectangular coordinate system, the X axis corresponds to the work area horizontal exploration line number, the Y axis corresponds to the vertical exploration line number, and the Z axis corresponds to the stratum vertical depth, the value increases downward from the surface; the screened seismic reflection feature is the three-dimensional feature set generated in step 2.3, containing "X-Y-Z" coordinates and reflection feature values, reflecting the spatial reflection properties of the stratum interface; the screened logging sequence feature is the vertical feature set generated in step 2.3, containing "wellhead X-wellhead Y-actual depth Z" coordinates and sequence feature values, reflecting the vertical lithology change properties of the stratum, and the two types of features need to be unified in space through coordinate anchoring);

[0105] First, the spatial coordinate anchoring is performed: for the screened seismic reflection feature, the original "X-Y-Z" coordinates are directly matched to the spatial coordinate system of the target work area, to ensure that each reflection feature point accurately corresponds to the specific spatial position of the work area, to generate the coordinate anchored seismic reflection feature; for the screened logging sequence feature, the "wellhead X-wellhead Y" coordinates are matched to the X-Y plane of the coordinate system, and the "actual depth Z" coordinates are matched to the Z axis, to realize the coordinate alignment of the vertical sequence feature and the three-dimensional space of the work area, to generate the coordinate anchored logging sequence feature (this process solves the coordinate disconnection problem of the two types of features caused by different acquisition methods, and lays a spatial foundation for subsequent fusion);

[0106] Then the correlation weight is calculated: taking the stratum boundary sample verified by drilling in the target work area as the reference (the sample contains "X-Y-Z" coordinates and corresponding "seismic reflection standard feature-log sequence standard feature", such as the seismic reflection standard feature at a certain stratum boundary is "continuous high-amplitude event" and the log sequence standard feature is "sudden increase in resistivity"), for each feature point in the coordinate-anchored seismic reflection feature, the similarity between the reflection feature value and the seismic reflection standard feature at the corresponding spatial position in the sample is calculated (the similarity is quantified by the matching degree of the feature value change trend, such as the closer the change slope of the reflection amplitude to the standard feature, the higher the matching degree, and the higher the correlation degree), and the seismic feature correlation weight is assigned according to the matching degree (the higher the correlation degree, the higher the weight value) ; for each feature segment in the coordinate-anchored log sequence feature, the similarity between the sequence feature value and the log sequence standard feature at the corresponding depth in the sample is calculated, and the log feature correlation weight is assigned.

[0107] Finally, the "feature number (including seismic / log identification) - spatial coordinates (X-Y-Z) - seismic feature correlation weight - log feature correlation weight" is arranged and integrated by row to generate a feature correlation mapping table (the rows of the table correspond to the feature numbers, the columns correspond to the spatial coordinates, the seismic feature correlation weight, and the log feature correlation weight, and the elements at the intersection correspond to the specific parameters of the corresponding features, ensuring that the weight of each feature can be traced back to the sample reference process, avoiding subjectivity in weight assignment).

[0108] Preferably, in the specific technical implementation of step 3.2, the processing objects are the feature correlation mapping table, the coordinate-anchored seismic reflection feature, and the coordinate-anchored log sequence feature (all three types of objects need to be implemented simultaneously to ensure that the fused features conform to the natural laws of the stratum) ;

[0109] First, perform weighted fusion: traverse each spatial coordinate (X-Y-Z) in the feature correlation mapping table, if there are both coordinate-anchored seismic reflection feature points and coordinate-anchored log sequence feature segments at this coordinate, calculate the fusion feature value of this coordinate according to the corresponding weights in the table by "fusion feature value = seismic reflection feature value x seismic feature correlation weight + log sequence feature value x log feature correlation weight"; if there is only one type of feature at a certain coordinate (such as only coordinate-anchored seismic reflection feature), then the feature value is directly taken as the fusion feature value to ensure that there is no feature missing; integrate the fusion feature values of all spatial coordinates according to "X-Y-Z-fusion feature value" to generate a preliminary fused geological feature set (this set covers most of the spatial positions in the work area, but there may be a small number of features that contradict the laws of the stratum due to feature errors) ;

[0110] Then, the stratum topology relationship verification is performed: taking the known topology relationship of the stratum in the target work area as the verification basis (the known topology relationship includes the vertical sequence rule, such as the fixed vertical sequence of “upper mudstone-middle sandstone-lower limestone”, and the lateral continuity rule, such as the distribution range and continuity of the same stratum in the X-Y plane), and traversing the preliminary fused geological feature set: if the stratum attribute (such as “limestone”) indicated by the fused feature value of a spatial coordinate violates the vertical sequence rule with the attribute (such as “sandstone”) indicated by the fused feature value of the spatial coordinate directly above it, or if a stratum feature is suddenly interrupted in lateral extension without a tectonic fault explanation, the feature is marked as a contradictory feature; all the contradictory features are removed, and the remaining features after the removal are reordered according to the “X-Y-Z” coordinates to ensure that the spatial distribution of the features meets the stratum topology relationship, and a fused geological feature set is generated (the set has no topology contradiction, which is different from the traditional processing method of fusing by only considering the weight without considering the stratum rule, and the feature authenticity is improved).

[0111] Preferably, in a scenario, when step 3.3 is specifically implemented, the processing object is the target work area regional geological background and the fused geological feature set (wherein the target work area regional geological background includes a sedimentary environment type, such as “delta front sedimentation”, a tectonic movement influence, such as “gentle fold and small stratum dip angle”, and a stratum development rule, such as “lateral continuous distribution of sandstone and interbedded mudstone”, which provides a basis for feature continuity and geological significance annotation; the fused geological feature set has no topology contradiction, but may have a spatial vacancy area due to insufficient original data acquisition density).

[0112] Firstly, the spatial continuity processing is performed: identifying the spatial vacancy area in the fused geological feature set (i.e. a coordinate point without “X-Y-Z-fused feature value” record, such as the edge of the work area or a well logging sparse area), selecting an interpolation method based on the rule in the target work area regional geological background— if the vacancy area is located in a sedimentary facies continuous distribution area (such as a delta front sand body), trend interpolation is adopted (based on the variation trend of the fused feature values of the peripheral effective feature points, the feature value of the vacancy point is calculated); if the vacancy area is affected by tectonics (such as a gentle fold), structure-guided interpolation is adopted (combining with the stratum dip angle parameter, the interpolated feature is ensured to be consistent with the structure shape); the fused feature value obtained by interpolation is supplemented to the spatial vacancy area, integrated with the original fused geological feature set, and a continuous fused geological feature set is generated (the set covers the complete spatial range of the work area without feature vacancy, solving the problem of the subsequent modeling fault caused by the spatial discontinuity of the traditional fused feature).

[0113] Then mark the geological significance: according to the corresponding relationship of "fusion feature value-geological attribute" in the regional geological background of the target work area (such as "high amplitude fusion feature value" corresponds to "sandstone formation interface", based on the wave impedance difference between sandstone and mudstone in delta deposition; "resistivity sudden rise fusion feature value" corresponds to "sand-mudstone boundary", based on the lithology indication rule of logging sequence), mark the corresponding geological attribute (such as "stratum interface", "lithological boundary", "porosity anomaly zone") and interpretation basis (such as "based on the sand body distribution rule of delta front deposition, determine that the high amplitude feature is a sandstone interface") for each fusion feature value in the continuous fusion geological feature set;

[0114] Finally, the "X-Y-Z coordinates-continuous fusion feature value-geological attribute-interpretation basis" are associated and integrated to generate the geological property representation of the target work area (the representation realizes the accurate mapping of the spatial position and geological property of the work area, each feature is accompanied by clear geological significance, which provides clear geological reference for the construction of stratum contact relationship constraint and lithological sequence constraint in step 4, and ensures that the subsequent implicit topological structure representation can accurately reflect the properties of the geological envelope).

[0115] Optionally, step 3, generating the geological property representation of the target work area according to the seismic reflection features and logging sequence features, further comprises:

[0116] Step 3.0, generating the regional geological background of the target work area according to the existing geological outcrop data and drilling core description data of the target work area and its surrounding area.

[0117] Optionally, step 3.0, generating the regional geological background of the target work area according to the existing geological outcrop data and drilling core description data of the target work area and its surrounding area, specifically comprises:

[0118] Step 3.0.1, classifying and labeling the lithology type, sedimentary structure and fossil combination of the existing geological outcrop data and drilling core description data of the target work area and its surrounding area to generate a classified geological basic data set;

[0119] Step 3.0.2, vertically sequence correlation analysis of the lithological combination and sedimentary structure in the classified geological basic data set, extracting key features reflecting the sedimentary environment to generate a sedimentary environment feature set;

[0120] Step 3.0.3, based on the sedimentary environment feature set, combining the existing tectonic movement record of the target work area and its surrounding area, establishing the time sequence correlation between the sedimentary environment feature and the tectonic evolution stage to generate a tectonic evolution time sequence table;

[0121] Step 3.0.4, based on the sedimentary environment feature set and the tectonic evolution time sequence table, comprehensively verifying the sedimentary environment feature and the tectonic evolution stage, and supplementing the stratum development rule description to generate the regional geological background of the target work area.

[0122] Preferably, the specific implementation process of step 3.0.1 is as follows: the processing object is the existing geological outcrop data and drilling core description data of the target work area and its surroundings (the geological outcrop data contains the "spatial coordinates (X-Y-Z)-lithology appearance description-sedimentary structure photo-fossil species record" of the outcrop point, reflecting the direct geological properties of the surface exposed stratum; the drilling core description data contains the "hole number-depth section-lithology name-core color-sedimentary structure type-fossil assemblage list" of the drilling, reflecting the continuous geological information of the underground vertical stratum, and both types of data need to be structured through classification labeling);

[0123] First, sort the data dimensions: for geological outcrop data, extract the four core dimensions of "spatial coordinates, lithology type, sedimentary structure, and fossil assemblage"; for drilling core description data, extract the four core dimensions of "hole number-depth section, lithology type, sedimentary structure, and fossil assemblage", ensuring that the labeling dimensions of both types of data are unified for subsequent integration;

[0124] Then perform classification labeling: for the "lithology type" dimension, set classification standards according to common lithology in the target work area (such as sandstone, mudstone, limestone, and shale), and match the lithology appearance description of the geological outcrop data and the lithology name of the drilling core description data to the corresponding lithology type (such as labeling "gray fine-grained, well-sorted particles" as "sandstone"); for the "sedimentary structure" dimension, set classification standards according to structural forms (such as horizontal bedding, cross-bedding, ripple marks, and mud cracks), and label the sedimentary structure photos of the geological outcrop data and the sedimentary structure types of the drilling core description data as the corresponding structure types; for the "fossil assemblage" dimension, set classification standards according to fossil classes (such as foraminifera, ostracods, and plant spores), and label the fossil species records of the geological outcrop data and the fossil assemblage lists of the drilling core description data as the corresponding fossil assemblage types;

[0125] Finally, integrate the labeling results: integrate the "spatial coordinates-lithology type-sedimentary structure-fossil assemblage" of the geological outcrop data and the "hole number-depth section-lithology type-sedimentary structure-fossil assemblage" of the drilling core description data by row to generate the classified geological basic data set (the rows of this data set correspond to individual data samples, the columns correspond to the above core dimensions, and the elements at the intersection are specific labeling results, such as a row record of "outcrop point A (X1, Y1, Z1)-sandstone-cross-bedding-foraminifera assemblage", ensuring that the labeling of each sample can be traced back to the original data description, avoiding labeling ambiguity).

[0126] Preferably, in the specific technical implementation of step 3.0.2, the processing object is the classified geological basic data set (which contains a large number of labeled geological samples, and needs to be associated and analyzed vertically to mine the internal relationship between samples and extract key features that can reflect the sedimentary environment).

[0127] Firstly, vertical continuous samples are screened: from the classified geological foundation data set, drill core description data are screened out (as drill core data are continuously distributed according to depth sections, they are more suitable for vertical sequence analysis), arranged according to the "hole number-depth section increasing" order, and a single-hole vertical geological sequence is formed (for example, the sequence of hole 1 is "depth 0-5m: mudstone-horizontal bedding-no fossils→ depth 5-~12m: sandstone-cross bedding-foraminifera→ depth 12-~20m: limestone-horizontal bedding-ostracoda"), and each drill hole corresponds to a single-hole vertical geological sequence;

[0128] Then, vertical sequence correlation analysis is performed: for each single-hole vertical geological sequence, the "lithology type combination rule" and "sedimentary structure evolution trend" are analyzed according to the depth section: if the lithology type in the vertical sequence of a certain drill hole presents a repeated combination of "mudstone→ sandstone→ mudstone", and the corresponding sedimentary structure transitions from "horizontal bedding" to "cross bedding" and then back to "horizontal bedding", it is determined as a typical vertical combination mode of "delta front underwater distributary channel-interdistributary bay"; if the lithology type presents a gradual change of "limestone→ shale", and the sedimentary structure is always "horizontal bedding", it is determined as a vertical evolution mode of "shallow sea-semi deep sea";

[0129] Finally, key features are extracted: the representative "lithology type combination (such as mudstone-sandstone-mudstone), sedimentary structure sequence (such as horizontal bedding-cross bedding-horizontal bedding), and fossil combination (such as foraminifera combination)" in the above vertical combination mode and evolution mode are taken as key features reflecting the sedimentary environment, integrated according to "sedimentary environment type-lithology combination feature-sedimentary structure feature-fossil combination feature", and a sedimentary environment feature set is generated (each sedimentary environment type in the set corresponds to clear multi-dimensional features, which is different from the traditional method of relying on single lithology to determine the sedimentary environment, and improves the reliability of environment determination).

[0130] Preferably, in a scenario, when step 3.0.3 is implemented, the processing object is the sedimentary environment feature set and the existing tectonic movement record around the target work area (the tectonic movement record contains "tectonic movement name-occurrence time-movement property (such as fold, fault, uplift and down) -influence range", such as "Yanshan late tectonic movement-100-80 Ma from now- regional uplift-covering the target work area and adjacent area", and the correspondence between the sedimentary environment and the tectonic movement needs to be established through time sequence correlation);

[0131] Firstly, the tectonic movement time sequence is sorted: the existing tectonic movement record around the target work area is sorted according to the "occurrence time from early to late" to form a tectonic movement time sequence (such as "Indo-Chinese tectonic movement→ Yanshan early tectonic movement→ Yanshan late tectonic movement→ Himalayan tectonic movement"), and the time range and movement property of each tectonic movement stage are determined;

[0132] Then the correlation relationship is established: traverse each sedimentary environment type in the sedimentary environment feature set, combine its corresponding geological sample time information (such as the geological age corresponding to the drilling core depth section, determined by fossil assemblage dating), and match to the corresponding stage in the tectonic movement time sequence: if the geological sample dating result of a certain sedimentary environment type (such as "delta front") is "90-85 Ma from now on", the corresponding tectonic movement time sequence is "Yanshanian late tectonic movement" (100-80 Ma from now on), and the tectonic movement property is "regional uplift" - uplift leads to sea level drop, which is beneficial to the development of delta front sedimentation, it is determined that there is a time sequence correlation between the sedimentary environment type and the Yanshanian late tectonic movement stage;

[0133] Finally, the correlation results are integrated: according to "tectonic movement stage-occurrence time-movement property-associated sedimentary environment feature", the tectonic evolution time sequence table is generated (the rows of the table correspond to the tectonic movement stages, the columns correspond to the occurrence time, the movement property and the associated sedimentary environment feature, and the elements at the intersection are the specific correlation contents, such as "Yanshanian late tectonic movement-100-80 Ma from now on-regional uplift-delta front, fluvial sedimentary environment", which ensures that the correlation of sedimentary environment features and tectonic movement has clear time and geological mechanism support).

[0134] Preferably, in the specific technical implementation of step 3.0.4, the processing object is the sedimentary environment feature set and the tectonic evolution time sequence table (the two types of objects need to be verified to ensure consistency, and then the stratum development rule is generated to generate the regional geological background);

[0135] First, perform comprehensive verification: traverse each tectonic movement stage in the tectonic evolution time sequence table, check if the associated sedimentary environment features are consistent with the description in the sedimentary environment feature set: if the associated sedimentary environment features of a certain tectonic movement stage are "fluvial facies", but the lithology combination characteristics (such as "gravel-sandstone-mudstone") corresponding to "fluvial facies" in the sedimentary environment feature set are inconsistent with the lithology combination of the geological sample of the tectonic movement stage (the actual sample is "limestone-shale"), then modify the correlation in the tectonic evolution time sequence table and remove the inconsistent sedimentary environment features; if there is no corresponding tectonic movement stage associated with a certain environment type in the sedimentary environment feature set, supplement the dating analysis and match to the appropriate tectonic movement stage to ensure that there is no logical contradiction between the sedimentary environment feature set and the tectonic evolution time sequence table;

[0136] Then complement the stratum development law description: based on the verified sedimentary environment feature set and the tectonic evolution time sequence table, extract the vertical development law of the stratum in the target work area (such as "from early to late, the vertical sequence of the stratum is'shallow sea limestone→delta sandstone→river gravel', corresponding to the tectonic movement from'regional subsidence' to'regional uplift'), the lateral development law (such as "the delta sandstone is thicker in the middle of the work area, gradually transitions to mudstone towards the edge, controlled by the ancient topography"), and the lithology distribution law (such as "limestone is mainly distributed in the deep part of the work area, controlled by the early range of marine transgression");

[0137] Finally, integrate all information: integrate "verified sedimentary environment feature set-verified tectonic evolution time sequence table-vertical development law-lateral development law-lithology distribution law" to generate the regional geological background of the target work area (this background covers multi-dimensional information of sedimentation, structure, and stratum distribution, providing a comprehensive basis for the feature space continuous and geological significance labeling in step 3.3, which is different from the traditional regional background which only contains simple sedimentary environment description, and the information is more complete).

[0138] Optionally, step 4, based on the stratum contact relationship constraint and lithology sequence constraint for the target work area, according to the geological characteristic representation, determine the implicit topological structure representation of the geological envelope in the target work area, specifically:

[0139] Based on the seismic exploration interpretation profile data of the target work area, the stratum contact relationship constraint and lithology sequence constraint for the target work area are constructed, and the implicit topological structure representation of the geological envelope in the target work area is determined according to the geological characteristic representation.

[0140] Optionally, step 4, based on the seismic exploration interpretation profile data of the target work area, the stratum contact relationship constraint and lithology sequence constraint for the target work area are constructed, and the implicit topological structure representation of the geological envelope in the target work area is determined according to the geological characteristic representation, specifically including:

[0141] Step 4.1, based on the seismic exploration interpretation profile data of the target work area, generate the target work area contact relationship analogy rule;

[0142] Step 4.2, based on the target work area contact relationship analogy rule, generate the target work area stratum contact relationship constraint and the target work area lithology sequence constraint;

[0143] Step 4.3, based on the target work area stratum contact relationship constraint and the target work area lithology sequence constraint, according to the geological characteristic representation, determine the implicit topological structure representation of the geological envelope in the target work area.

[0144] Optionally, step 4.1, based on the seismic exploration interpretation profile data of the target work area, generate the target work area contact relationship analogy rule, including the following steps:

[0145] Step 4.1.1: Based on the seismic exploration interpretation profile data of the target work area, perform parametric extraction of the spatial location, dip angle, and interface continuity of the stratigraphic contact interface in the seismic exploration interpretation profile to generate the contact feature set of the target work area profile.

[0146] Step 4.1.2: Based on the contact feature set of the target work area profile, and combined with the verified analog geological model data of the neighboring areas of the target work area, establish an analog mapping between the profile contact features and the contact relationship of the neighboring areas to generate the contact relationship analogy rules of the target work area.

[0147] Preferably, the specific implementation process of step 4.1.1 is as follows: the processing object is the seismic exploration interpretation profile data of the target work area (this data includes two-dimensional or three-dimensional visualization profiles. The coordinate system of the two-dimensional profile is defined as "X-axis - transverse survey line direction, Z-axis - vertical depth direction". The three-dimensional profile additionally adds "Y-axis - longitudinal survey line direction". Each profile is marked with "survey line number - depth range" information. The profile clearly presents the reflection phase axis morphology of the stratigraphic contact interface. The interface is represented by the interruption of the phase axis, the change of attitude or the sudden change of amplitude, which provides a visual basis for parameter extraction).

[0148] First, the spatial location of the stratigraphic contact interface is extracted: using the coordinate system of the seismic exploration interpretation profile data as a reference, the stratigraphic contact interface is identified point by point along the XZ (or XYZ) trajectory of the profile. For two-dimensional profiles, key points with obvious reflection characteristics on the interface (such as amplitude peak points, phase axis inflection points) are selected, and the coordinates of each key point are recorded (X value - lateral survey line position, Z value - vertical depth). For three-dimensional profiles, the Y value - longitudinal survey line position is additionally recorded to form an interface coordinate sequence (e.g., the coordinate sequence of interface 1 in a two-dimensional profile is "(X1,Z1), (X2,Z2)...(X...Z1)"). n Z n )”, where X1-X n Representing different positions of the transverse survey line, Z1-Z n (Representing the interface depth at the corresponding location), this interface coordinate sequence directly reflects the spatial orientation of the interface;

[0149] Next, tilt angle extraction is performed: for the interface coordinate sequence of each interface, the positions of adjacent key points (such as (X)) are calculated. i Z i ) and (X i+1 Z i+1 The slope of the tangent line (slope k = (Z)) i+1 -Z i ) / (X i+1 -X i), the slope of the X-Z plane in the three-dimensional profile, and then the angle between the tangent and the horizontal direction (X-axis positive direction) of the profile is calculated by the inverse tangent function to obtain the interface inclination angle (the angle value is positive when the interface inclines downward along the X-axis positive direction, and is negative when it inclines upward);

[0150] Then the interface continuity extraction is performed: for the interface coordinate sequence of each interface, two parameters are counted: the total length of continuous segments (referring to the length of coordinate point series without interruption of the same phase axis, such as (X1, Z1) to (X5, Z5) without interruption, the length of continuous segments is X5-X1), and the number of interruptions (referring to the number of times that the same phase axis is obviously blank or the occurrence changes abruptly in the coordinate sequence); the continuity ratio is calculated according to "continuity ratio = total length of continuous segments / total length of profile transverse (two-dimensional) or total range of profile X-Y plane (three-dimensional)", and the interruption frequency is calculated according to "interruption frequency = interruption number / total length of interface coordinate sequence", which together constitute the interface continuity parameter (such as the continuity parameter of interface 1 is "continuity ratio 82%, interruption frequency 0.15 times per unit length", where "unit length" is the basic measurement unit of the profile transverse, such as meter);

[0151] Finally, the "interface number-interface coordinate sequence-interface inclination angle-interface continuity parameter" is integrated by row to generate the target work area profile contact feature set (the rows of the feature set correspond to the interface numbers, and the columns correspond to the above four dimensions respectively, and the elements at the intersection are specific parameter values, such as a row record is "interface 1-[(X1, Z1)...(X n ,Z n )]-α1-continuity ratio 82%, interruption frequency 0.15 times per unit length"), which ensures that the parameters of each interface can be traced back to the original visualization information of the seismic exploration interpretation profile data, avoiding the subjectivity of parameter extraction.

[0152] Preferably, in the specific technical implementation of step 4.1.2, the processing object is the target work area profile contact feature set and the verified analog geological model data of the target work area adjacent area (wherein the analog geological model data contains the "stratum contact relationship type-corresponding profile contact feature parameter" mapping relationship verified by drilling in the adjacent area, such as the parameters corresponding to the "integrated contact" type in the adjacent area are "interface inclination angle deviation <12°, continuity ratio >78%, interface coordinate sequence presents a gentle curve", and the parameters corresponding to the "unconformable contact" are "interface inclination angle deviation >28°, continuity ratio <65%, interface coordinate sequence has obvious discontinuity", which has high reliability because it is verified by drilling);

[0153] First, the feature similarity calculation is performed: traverse each interface in the target work area profile contact feature set (process in order according to interface number 1 to n), compare the three core features of each interface, i.e. interface dip angle, interface continuity parameter (including continuity ratio and interruption frequency), and interface coordinate sequence form, with the corresponding parameters of all stratigraphic contact relationship types in the analog geological model data:

[0154] 1. Calculate the dip angle similarity: take the standard dip angle range of a certain contact relationship type in the adjacent area (such as the standard range of conformable contact is -10° to 10°) as the reference, if the dip angle of the target interface falls within the range, the similarity is 1; if it exceeds the range, calculate the similarity according to "similarity = 1 - |target angle - standard range midpoint| / standard range half-width" (for example, the target angle is 12°, the standard range midpoint is 0°, and the half-width is 10°, then the similarity = 1 - 12 / 10 = 0, to avoid negative similarity);

[0155] 2. Calculate the continuity similarity: continuity ratio similarity = target interface continuity ratio / adjacent area standard continuity ratio (the ratio needs to be controlled between 0-1, such as target ratio 82%, standard ratio 78%, then similarity = 82 / 78 ≈ 1, take 1); interruption frequency similarity = 1 - target interface interruption frequency / adjacent area standard interruption frequency (such as target frequency 0.15, standard frequency 0.3, similarity = 1 - 0.15 / 0.3 = 0.5), continuity similarity = (continuity ratio similarity × 0.6 + interruption frequency similarity × 0.4);

[0156] 3. Calculate the form matching degree: compare the form of the target interface coordinate sequence with the standard sequence in the adjacent area (such as the standard sequence of conformable contact is a smooth quadratic curve) through curve fitting, if the goodness of fit (such as R 2 ) ≥ 0.85, the matching degree is 1; otherwise, calculate the matching degree according to "matching degree = goodness of fit";

[0157] Calculate the total similarity of the target interface and each contact relationship type in the adjacent area by weighted calculation according to "total similarity = dip angle similarity × 0.4 + continuity similarity × 0.3 + form matching degree × 0.3";

[0158] Then establish the analogy mapping: for each target interface, select the contact relationship type in the adjacent area with the highest total similarity - if the highest similarity is ≥ 60%, directly establish the corresponding mapping of "target interface feature - adjacent contact relationship type" (such as the total similarity of target interface 1 and adjacent "conformable contact" is 76%, then the mapping is "conformable contact"); if the highest similarity is < 60%, modify it in combination with the regional geological background of the target work area (such as it is known that the tectonic activity in the work area during a certain period is weak, and there is no evidence of unconformity), preferentially map it to the contact relationship type under the scenario of weak tectonic activity (such as conformable contact), to avoid misjudgment due to data errors;

[0159] Finally, all analogy mapping relationships are integrated, sorted according to "target interface feature matching conditions - corresponding contact relationship types", and target work area contact relationship analogy rules are generated (rule example: "if the interface inclination angle falls within the range of -15° to 15°, the continuity proportion is ≥75%, the interruption frequency is ≤0.2 times / unit length, and the coordinate sequence fitting goodness is ≥0.8, then the corresponding contact relationship type is integrated contact"). Each matching condition in the rule corresponds to a specific parameter threshold, ensuring that in the subsequent step 4.2.1, the rule can be directly applied to determine the contact relationship type by parameter comparison. Unlike the traditional method of directly applying the adjacent area model, this step realizes dynamic mapping through multi-dimensional weighted similarity calculation, significantly improving the adaptability of the rule to the target work area.

[0160] Optionally, in step 4.2, based on the target work area contact relationship analogy rule, the target work area stratigraphic contact relationship constraint and the target work area lithology sequence constraint are generated.

[0161] Step 4.2.1, based on the target work area contact relationship analogy rule, the target work area stratigraphic contact relationship constraint is generated.

[0162] Step 4.2.2, based on the target work area stratigraphic contact relationship constraint, the corresponding association between the stratigraphic contact interface and the upper and lower lithology is established by combining the lithology vertical transition record in the target work area logging data, to generate the target work area lithology sequence constraint.

[0163] Preferably, the specific implementation process of step 4.2.1 is as follows, the processing object is the target work area contact relationship analogy rule (generated by step 4.1.2, containing the mapping relationship of "feature matching conditions - stratigraphic contact relationship type", such as "interface inclination angle deviation <15°, continuity proportion ≥75%→ integrated contact" "interface inclination angle deviation >28°, continuity proportion <65%→ unconformable contact", each mapping relationship corresponds to a specific parameter threshold).

[0164] First, the constraint clause extraction is performed: each mapping relationship in the target work area contact relationship analogy rule is traversed, and the "feature matching conditions" are converted into specific constraint clauses - for the "integrated contact" type, the constraint clause is extracted as "the interface inclination angle needs to fall within the range of [-15°, 15°], the continuity proportion is ≥75%, the interruption frequency is ≤0.2 times / unit length ('unit length' refers to the horizontal basic measurement unit of the target work area seismic exploration interpretation profile, such as meters)"; for the "unconformable contact" type, the constraint clause is extracted as "the interface inclination angle needs to exceed the range of [-28°, 28°], the continuity proportion is ≤65%, and the difference in occurrence of the upper and lower strata needs to be ≥15°"; all constraint clauses are integrated according to the contact relationship type to form a preliminary constraint clause set;

[0165] Then the constraint parameter matrix is constructed: taking "contact relationship type" as row and "constraint parameter item" as column (constraint parameter item includes tilt angle range, continuity proportion threshold, interruption frequency threshold, occurrence difference threshold, etc.), filling the parameter threshold in the preliminary constraint clause set into the matrix to generate the contact relationship constraint parameter matrix (the rows of the matrix correspond to integrated contact, unconformable contact, etc., the columns correspond to each constraint parameter, and the intersection elements are specific threshold ranges, such as "integrated contact-tilt angle range" corresponds to [-15°, 15°]);

[0166] Then the spatial consistency check is performed: based on the structural rules in the target work area regional geological background (such as the overall structure of the work area is gentle, and there is no large-scale fault), check whether there are contradictory clauses in the contact relationship constraint parameter matrix - if the constraint clause of a certain "unconformable contact" requires occurrence difference ≥ 15°, but the regional geological background shows that the tectonic activity of the work area is weak, and the occurrence difference is generally < 10°, then modify the clause threshold to "occurrence difference ≥ 10°"; At the same time, check the constraint compatibility of adjacent interfaces, such as interface 1 is integrated contact (tilt angle 5°), and if the interface 2 below it is labeled as unconformable contact, it needs to ensure that the tilt angle of interface 2 is different from interface 1 ≥ the modified occurrence difference threshold, to avoid spatial logical contradiction;

[0167] Finally, integrate the contact relationship constraint parameter matrix and the constraint clause after the check to generate the target work area stratigraphic contact relationship constraint (this constraint clearly specifies the quantitative parameter requirements and spatial compatibility rules for each contact relationship type, which is different from the traditional qualitative description of contact relationship constraints, and realizes the quantifiable and checkable constraints).

[0168] Preferably, in the specific technical implementation of step 4.2.2, the processing object is the target work area stratigraphic contact relationship constraint and the lithology vertical transition record in the target work area logging data (wherein the lithology vertical transition record is extracted from the target work area logging data generated in step 1.3, and contains "depth position-physical parameter mutation type-corresponding lithology change", such as "depth Z1-resistivity sudden rise 50%→mudstone→sandstone" and "depth Z2-sonic time difference sudden drop 30%→sandstone→limestone", the physical parameter mutation point corresponds to the lithology boundary, and there is a spatial correlation with the stratigraphic contact interface);

[0169] First, the lithology vertical transition record extraction is performed: traversing the well logging data of the target work area, identifying the vertical mutation points of the physical parameters (resistivity, acoustic time, density) (the mutation point determination standard is that the parameter value difference of adjacent depth points exceeds 20% of the average value of the parameter sequence), recording the depth position (Z value) of each mutation point, parameter mutation type (sudden rise / sudden drop), combining the existing lithology sample library of the target work area (such as the "resistivity range-lithology" corresponding relationship calibrated by drilling core), determining the upper and lower lithology corresponding to the mutation point (such as resistivity 80-150 Ω·m corresponding to sandstone, 2-20 Ω·m corresponding to mudstone), and generating a lithology vertical transition detail table (the rows of the table correspond to the mutation point number, and the columns correspond to the depth position, parameter mutation type, overlying lithology, underlying lithology);

[0170] Then the space correlation of the contact interface and the lithology transition is established: based on the interface coordinate sequence (including depth Z value) in the contact relationship constraint of the target work area, the depth range of the contact interface is matched with the depth position in the lithology vertical transition detail table - if the depth range of the interface coordinate sequence of a contact interface contains the depth position Z1 of a mutation point, and the contact relationship type of the contact interface is integrated contact (the constraint requires higher continuity, and the lithology transition is gentle), it is determined that the lithology change (such as mudstone→sandstone) corresponding to the mutation point is the upper and lower lithology of the contact interface; if the contact interface is unconformable contact (the constraint requires large occurrence difference, and there may be lithology loss), the lithology transition record with overlapping depth range and large parameter mutation amplitude (such as resistivity sudden rise 80%, corresponding to conglomerate→sandstone) is preferentially matched;

[0171] Then the interface-lithology corresponding relationship table is constructed: integrated according to "contact interface number-interface depth range-overlying lithology-underlying lithology-matching basis (parameter mutation type+contact relationship constraint clause)", the interface-lithology corresponding relationship table is generated (such as "interface 1-depth Z0-Z2-mudstone-sandstone-resistivity sudden rise 50%+integrated contact continuity constraint"), ensuring that the upper and lower lithology of each contact interface is supported by well logging data;

[0172] Finally, the lithology sequence constraint is generated: based on the interface-lithology corresponding relationship table, the upper and lower lithologies of all contact interfaces of the target work area are extracted, sorted by vertical depth from shallow to deep, forming a vertical sequence chain of "overlying lithology→underlying lithology" (such as "mudstone→sandstone→limestone→shale"), and combining the spatial compatibility rules in the contact relationship constraint of the target work area, the sequence constraint clauses are supplemented - such as "the lithology sequence corresponding to integrated contact needs to maintain continuous transition, without skipping lithology change" "the lithology sequence corresponding to unconformable contact allows 1-2 lithology losses, but the missing lithology needs to meet the sedimentary interval rules in the regional geological background", and finally the lithology sequence constraint of the target work area is generated (the constraint clearly defines the vertical lithology transition order and the matching rules with the contact relationship, solving the problem of disconnection between the traditional lithology sequence constraint and the contact relationship).

[0173] Preferably, the step 4.3, the target work area stratum contact relationship constraint and the target work area lithology sequence constraint, according to the geological property characterization, determine the implicit topological structure characterization of the geological envelope in the target work area, specifically including the following steps:

[0174] Step 4.3.1, based on the target work area stratum contact relationship constraint and the target work area lithology sequence constraint, the geological property characterization of the target work area is constrained and adapted to generate the constrained and adapted geological property characterization;

[0175] Step 4.3.2, based on the constrained and adapted geological property characterization, generate the implicit topological structure characterization of the geological envelope in the target work area.

[0176] Preferably, the specific implementation process of step 4.3.1 is as follows, the processing object is the target work area stratum contact relationship constraint, the target work area lithology sequence constraint, and the geological property characterization of the target work area (wherein the geological property characterization of the target work area contains "spatial coordinate (X-Y-Z)-geological attribute (lithology type, stratum interface type)-geological significance annotation" three-dimensional association information, which needs to be verified and adapted by two types of constraints to ensure that it conforms to the actual geological law of the target work area);

[0177] First, disassemble the geological property characterization elements: split the geological property characterization of the target work area by "spatial coordinate unit", each spatial coordinate unit (corresponding to a single X-Y-Z coordinate point) contains "lithology attribute value, interface attribute value (if it is an interface, mark the type, such as integrated contact / non-integrated contact, if it is not an interface, mark 'none')", generate a geological property element table (the rows of the table correspond to the spatial coordinate unit number, the columns correspond to the X value, Y value, Z value, lithology attribute value, interface attribute value, and the elements at the intersection are specific parameters, such as "unit 1-X1-Y1-Z1-sandstone-integrated contact");

[0178] Then perform stratum contact relationship constraint adaptation verification: based on the "contact relationship type-parameter threshold" (such as integrated contact requires inclination angle [-15°, 15°], continuity ratio ≥75%) in the target work area stratum contact relationship constraint, traverse all spatial coordinate units marked with "interface attribute value" in the geological property element table: if a unit is marked with "integrated contact" but the corresponding inclination angle (extracted from the interface parameter of the geological property characterization) is 18°, which exceeds the constraint threshold, mark the unit as a "constraint conflict feature unit"; if a unit is marked with "non-integrated contact" but the continuity ratio is 80%, which is higher than the constraint threshold 65%, it is also marked as a conflict unit, and a preliminary constraint conflict list is generated;

[0179] Then the lithology sequence constraint fitting check is performed: according to the "vertical lithology transition sequence" (such as "mudstone→sandstone→limestone") in the lithology sequence constraint of the target work area, the geological property element table is sorted in ascending order of Z value (vertical depth) to form the vertical lithology sequence of each X-Y plane: if the vertical lithology sequence of a certain X-Y position is "mudstone→limestone→sandstone", which is contradictory to the transition sequence in the constraint, then mark the spatial coordinate unit corresponding to the sequence as a "constraint conflict feature unit" and supplement it to the preliminary constraint conflict list to generate a complete constraint conflict list;

[0180] Then the constraint conflict feature units are processed: the conflicts are corrected in combination with the regional geological background of the target work area (such as the existence of local small faults and sedimentary discontinuities in the known work area) - if a certain "conformable contact" unit is marked as conflicting due to the inclination angle exceeding the threshold value, but the regional geological background shows that there is a small normal fault at this position, then the interface attribute value is corrected to "fault contact" and supplemented to the geological property element table; if a certain vertical lithology sequence conflict is caused by local sedimentary facies change, then the lithology attribute value is corrected by referring to the similar facies change rules of the adjacent area to generate a corrected geological property element table.

[0181] Finally, the corrected geological property element table is integrated, and the "lithology attribute value, interface attribute value, and geological significance label" are re-associated according to the "X-Y-Z spatial coordinates" to ensure that the geological properties of each spatial coordinate unit meet the stratigraphic contact relationship constraints and the lithology sequence constraints of the target work area, and a constraint-fitted geological property representation is generated (this representation solves the possible geological rule conflicts in the original geological property representation and provides compliant geological foundation data for the generation of subsequent implicit topological structure representations).

[0182] Preferably, in the specific technical implementation of step 4.3.2, the processing object is the constraint-fitted geological property representation (this representation contains compliant "spatial coordinate-geological attribute" association information, which needs to be converted into a structure representation that can reflect the topological relationship of the geological envelope body through implicit expression. The core of the implicit topological structure representation is to implicitly represent the spatial distribution and topological connection relationship of the geological body through continuous mathematical fields (such as scalar fields)).

[0183] First, the implicit scalar mapping rule of the geological envelope body is constructed: according to the geological attribute types in the constraint-fitted geological property representation, a unique mapping relationship between "geological attribute-scalar value" is set - such as "sandstone" corresponding to scalar value 1.0, "mudstone" corresponding to scalar value 2.0, "limestone" corresponding to scalar value 3.0, and stratigraphic interfaces (such as conformable contact interfaces and fault contact interfaces) corresponding to scalar gradient mutation zones (gradient value ≥ preset threshold value, such as 0.5 / unit length, "unit length" is the basic measurement unit of the spatial coordinates of the target work area, such as meters). The scalar values of the same geological envelope body (such as a set of continuous sandstone strata) remain consistent, and the scalar values of different geological envelope bodies have clear differences, and a geological envelope body scalar mapping table is generated.

[0184] Then generate the initial geological envelope scalar field: based on the spatial coordinate system (X-Y-Z) of the target work area, construct a three-dimensional space scalar field matrix (the rows of the matrix correspond to the X value, the columns correspond to the Y value, and the depth dimension corresponds to the Z value. The element at the intersection of the matrix is the scalar value to be assigned); traverse each spatial coordinate unit in the constraint-adapted geological property representation, and according to its lithology attribute value or interface attribute, extract the corresponding scalar value or scalar gradient rule from the geological envelope scalar mapping table, and assign it to the corresponding position of the empty scalar field matrix - for example, if the lithology attribute value of a certain spatial coordinate unit is "sandstone", fill 1.0 in the corresponding position of the scalar field matrix; if a unit is "integrated contact interface", a transition scalar value is assigned between its adjacent coordinate units at a gradient of 0.5 per unit length, generating an initial geological envelope scalar field;

[0185] Then optimize the topological continuity of the scalar field: based on the spatial compatibility rules in the stratigraphic contact relationship constraints of the target work area (such as the scalar gradient direction of the integrated contact interface needs to be consistent with the stratigraphic inclination direction), check the scalar value trend of the initial geological envelope scalar field - if the scalar value of a certain region suddenly changes from 1.0 (sandstone) to 3.0 (limestone) and there is no interface attribute annotation, it does not conform to the transition rule of "sandstone→mudstone→limestone", then insert a transition coordinate unit with a scalar value of 2.0 (mudstone) to correct the scalar field trend; for the scalar gradient at the fault contact interface, ensure that its direction is consistent with the fault trend, and generate an optimized geological envelope scalar field (the continuous change of this scalar field can implicitly reflect the spatial boundary and internal continuity of the geological envelope, i.e. the topological relationship);

[0186] Finally, label the topological significance of the scalar field: clarify the correspondence between "scalar value-geological envelope attribution" and "scalar gradient-stratigraphic boundary" in the optimized geological envelope scalar field - for example, the connected region with a scalar value of 1.0 is the "lower sandstone geological envelope in the target work area", and the region with a scalar gradient of ≥0.5 per unit length is the "stratigraphic boundary of the geological envelope". Integrate these topological significance annotations with the optimized geological envelope scalar field to generate an implicit topological structure representation that characterizes the geological envelope in the target work area (this representation implicitly represents the spatial distribution, stratigraphic boundary, and mutual topological connection relationship of the geological envelope through a continuous scalar field, which is different from the traditional explicit geometric modeling method that relies on discrete surfaces, and is more suitable for subsequent neural network model processing of continuous features).

[0187] Optionally, step 5, based on the third neural network model trained, processes the implicit topological structure representation to predict the stratigraphic boundary of the geological envelope in the target work area, specifically including:

[0188] Step 5.1, based on the implicit structure feature reinforcement layer in the third neural network model completed by training, boundary-sensitive feature enhancement is performed on the implicit topological structure representation to generate reinforced implicit structure features;

[0189] Step 5.2, based on the multi-scale boundary feature capture layer in the third neural network model completed by training, multi-scale boundary feature extraction is performed on the reinforced implicit structure features to generate a multi-scale boundary feature set;

[0190] Step 5.3, based on the preliminary horizon boundary prediction layer in the third neural network model completed by training, preliminary horizon boundary prediction is performed on the multi-scale boundary feature set to generate initial horizon boundary data;

[0191] Step 5.4, based on the boundary topology verification optimization layer in the third neural network model completed by training, topological verification and optimization are performed on the initial horizon boundary data to generate final horizon boundary data of the geological envelope in the target work area.

[0192] Preferably, the specific implementation process of step 5.1 is as follows, and the processing object is the implicit topological structure representation (which is a three-dimensional scalar field matrix, the rows of the matrix correspond to the X coordinates of the horizontal lines of the target work area, the columns correspond to the Y coordinates of the vertical lines, and the depth dimension corresponds to the Z coordinates. The elements at the intersection of the matrix are scalar values - regions with the same or similar scalar values correspond to the same geological envelope, and regions with significant scalar gradient changes correspond to the horizon boundaries of the geological envelope, such as a sandstone geological envelope corresponding to a scalar value of 1.0 and a mudstone corresponding to a scalar value of 2.0. The scalar gradient at the interface is ≥0.5 / unit length, and "unit length" is the basic measurement unit of the spatial coordinates of the target work area, such as meters);

[0193] Relying on the implicit structure feature reinforcement layer in the third neural network model completed by training (which has a built-in boundary-sensitive sensitivity enhancement algorithm, the design core of which is to focus on areas with significant scalar gradient changes and avoid boundary confusion caused by excessive enhancement of features in non-boundary areas), first calculate the scalar field gradient of the implicit topological structure representation: for each element (X, Y, Z) in the three-dimensional scalar field matrix, calculate the scalar value difference of the element with its adjacent 6 elements (up, down, left, right, front, back), and calculate the scalar field gradient of the element according to "gradient value = √[(ΔX scalar difference) 2 +(ΔY scalar difference) 2 +(ΔZ scalar difference) 2 ]”, generate a scalar field gradient matrix (with the same dimensions as the implicit topological structure representation, and the elements are gradient values at the corresponding positions);

[0194] Then the boundary candidate region is delineated: based on the gradient matrix of the scalar field, a gradient threshold is set (the threshold is the mean value of the gradient matrix of the scalar field + 1 times the standard deviation, for example, if the mean value is 0.3 per unit length and the standard deviation is 0.1, the threshold is 0.4 per unit length), the region with gradient value ≥ threshold is determined as the boundary candidate region, and the region with gradient value < threshold is determined as the non-boundary region, and the boundary candidate region mask is generated (the mask matrix has the same dimension as the scalar field, the boundary candidate region element is 1, and the non-boundary region is 0);

[0195] Then the boundary-sensitive feature enhancement is performed: the implicit topological structure representation is multiplied element by element with the boundary candidate region mask, the scalar values of the boundary candidate region are retained, and the scalar values of the non-boundary region are weakened (the scalar values of the non-boundary region are multiplied by a weakening coefficient of 0.5 to avoid interference with subsequent boundary extraction); at the same time, the scalar values in the boundary candidate region are gradient-weighted enhanced - the higher the gradient value of the scalar field, the greater the enhancement weight (for example, a gradient value of 0.6 per unit length corresponds to a weight of 1.2, and a gradient value of 0.4 per unit length corresponds to a weight of 1.0), calculated according to “enhanced scalar value = original scalar value × enhancement weight”, to generate the enhanced scalar field matrix;

[0196] Finally, the enhanced scalar field matrix is consistent with the spatial coordinate system of the original implicit topological structure representation, and the enhanced implicit structure feature is generated (in this feature, the scalar gradient of the boundary region of the layer is highlighted, and the redundant information of the non-boundary region is suppressed, which is different from the traditional uniform enhancement of the whole region which is easy to lead to confusion between boundary and non-boundary features).

[0197] Preferably, in the specific technical implementation of step 5.2, the processing object is the enhanced implicit structure feature (this feature is the enhanced three-dimensional scalar field matrix, and the scalar gradient feature of the boundary region of the layer has been highlighted, but the boundary details of different scales, such as micro-lithology thin interlayer boundary and macro-stratigraphic interface boundary, still need to be fully captured through multi-scale extraction);

[0198] Relying on the multi-scale boundary feature capture layer in the third neural network model trained (this layer is built-in with three different sizes of three-dimensional convolution kernels, small size kernel for micro boundary extraction, medium size kernel for medium scale boundary extraction, and large size kernel for macro boundary extraction, to avoid missing key boundary information in single scale extraction), first, the small scale boundary feature extraction is performed: the small size three-dimensional convolution kernel (such as 3×3×3 pixels, corresponding to the actual spatial range of the target work area which needs to match the micro boundary scale, such as 0.5m×0.5m×0.5m) is used to perform convolution operation on the enhanced implicit structure feature, and the sliding step is consistent with the sampling interval of the scalar field matrix, to extract the boundary feature of the micro thin interlayer (such as mudstone interlayer with thickness <2m), and generate a small scale boundary feature map (with the same dimension as the enhanced implicit structure feature, and the element is the micro boundary feature value);

[0199] Then the medium-scale boundary feature extraction is performed: the above-mentioned convolution operation is repeated by using a medium-size three-dimensional convolution kernel (such as 5x5x5 pixels, corresponding to an actual space of 1m x 1m x 1m) to extract the boundary features of a medium-scale stratum section (such as a sandstone section with a thickness of 2-10m), and a medium-scale boundary feature map is generated;

[0200] Then the large-scale boundary feature extraction is performed: a large-size three-dimensional convolution kernel (such as 7x7x7 pixels, corresponding to an actual space of 2m x 2m x 2m) is used to extract the boundary features of a macro-area stratum interface (such as a sand-mudstone interface with a thickness of >10m), and a large-scale boundary feature map is generated;

[0201] Finally, the small-scale boundary feature map, the medium-scale boundary feature map and the large-scale boundary feature map are integrated according to the "feature point number-scale type-feature value" to generate a multi-scale boundary feature set (the set is a three-dimensional matrix, the rows correspond to the target work area space feature point numbers, the columns correspond to the scale types (small / medium / large), and the depth dimension corresponds to the boundary feature values, such as "feature point 1-small scale-0.8" and "feature point 1-medium scale-0.6", and it is ensured that different scale boundary features at the same spatial position can be cooperatively involved in subsequent prediction).

[0202] Preferably, in a scenario, when step 5.3 is implemented, the processing object is a multi-scale boundary feature set (the set contains boundary features of different scales, which need to be mapped to stratum boundary existence probability through classification operation to realize preliminary boundary positioning);

[0203] Relying on the stratum boundary preliminary prediction layer in the third neural network model (the layer is a full-connection classifier, the input is the feature vector of the multi-scale boundary feature set, and the output is the "stratum boundary existence probability" of the corresponding spatial position), first, the multi-scale boundary feature set is vector reshaped: the "small-scale feature value-medium-scale feature value-large-scale feature value" of each spatial feature point is spliced into a three-dimensional feature vector (such as the vector of feature point 1 is [0.8, 0.6, 0.5]), and a feature vector matrix is generated (the rows correspond to the feature point numbers, and the columns correspond to the three scale feature values);

[0204] Then the feature vector matrix is input into the classifier of the stratum boundary preliminary prediction layer, and the stratum boundary existence probability corresponding to each feature vector is output through the activation function operation (the probability value ranges from 0 to 1, and the value closer to 1 indicates that the feature point is more likely to belong to the stratum boundary), and a boundary probability vector is generated (the length is consistent with the number of feature points, and the elements are the probability values of the corresponding feature points);

[0205] Then set the probability threshold (threshold based on the target work area has been labeled stratigraphic boundary sample statistics, such as the minimum probability of boundary points in the sample is 0.6, the threshold is set to 0.6), traverse the boundary probability vector: the probability value ≥ threshold feature points marked as "initial boundary points", record its spatial coordinates (X, Y, Z); The probability value < threshold feature points marked as "non-boundary points", excluded from the boundary extraction range;

[0206] Finally, according to the adjacent relationship of the spatial coordinates, the "initial boundary points" are connected in the horizontal, vertical and depth directions to form continuous boundary line segments, and then the adjacent boundary line segments are integrated into complete boundary curves or surfaces to generate initial stratigraphic boundary data (the data contains "boundary number-boundary coordinate sequence-corresponding geological envelope pair", such as "boundary 1-[(X1, Y1, Z1)…(Xn, Yn, Zn)]-sandstone envelope-shale envelope", which clearly indicates the corresponding adjacent geological envelope of each initial boundary).

[0207] Preferably, the specific implementation process of step 5.4 is as follows, and the processing object is the initial stratigraphic boundary data (which contains the preliminary extracted stratigraphic boundary, but there may be two problems: one is that the boundary segment and the stratigraphic contact relationship constraint are contradictory, such as the inclination angle of a certain boundary segment is 20°, which exceeds the [-15°, 15°] range of the integrated contact constraint; the second is that the boundary is locally interrupted, such as the boundary line segment appears blank due to insufficient original data sampling density);

[0208] Relying on the boundary topological verification and optimization layer in the third neural network model trained (which integrates topological rule verification and interpolation completion algorithm to ensure that the optimized boundary conforms to the geological natural law), first perform topological verification: call the key parameters (inclination angle range, continuity proportion threshold) in the stratigraphic contact relationship constraint of the target work area, and traverse each boundary segment in the initial stratigraphic boundary data: if a certain boundary segment is labeled as "integrated contact boundary", but its inclination angle (calculated through the boundary coordinate sequence, such as the angle between adjacent two points (Xi, Zi) and (Xi+1, Zi+1)) is 18°, which exceeds the constraint range, then mark this segment as "topological contradiction segment"; if the continuity proportion (continuous segment length / segment total length) of a certain boundary segment is 70%, which is lower than the constraint threshold 75%, it is also marked as a contradiction segment, and a topological contradiction list is generated;

[0209] Then process the topological contradiction segment: for the marked contradiction segment, modify it according to the regional geological background of the target work area - if the contradiction is caused by local small fault interference, modify the contact relationship type of the segment from "integrated contact boundary" to "fault contact boundary", and update its inclination angle constraint (the constraint angle range of fault contact boundary is [-30°, 30°]); if the contradiction is caused by data error, delete the contradiction segment and keep the continuous effective boundary line segment at both ends;

[0210] Then the boundary continuity optimization is performed: for the boundaries that still have interruptions after the correction (such as there is a blank area between two effective boundary line segments), the boundary coordinate points of the blank area are calculated using linear interpolation or trend interpolation (based on the trend of the two end boundary line segments) to supplement into the boundary coordinate sequence, ensuring the spatial continuity of the boundary;

[0211] Finally, all the corrected boundary segments are integrated, reorganized according to "boundary number-complete boundary coordinate sequence-contact relationship type-corresponding geological envelope pair", and the final stratigraphic boundary data of the geological envelope in the target work area are generated (the spatial form of the boundary data conforms to the stratigraphic contact relationship constraint and has no obvious interruption, and can be directly used for subsequent boundary positioning and analysis of the geological envelope, which is different from the traditional boundary data that only relies on neural network prediction and has not been checked, and the reliability of boundary prediction is improved).

[0212] The third neural network model is trained to accurately predict the stratigraphic boundary of the geological envelope from the implicit topological structure representation. The training data is the implicit topological structure representation sample and the corresponding labeled stratigraphic boundary label (the labeled sample includes the spatial coordinates of the known stratigraphic boundary and the corresponding implicit representation feature label). During training, the implicit topological structure representation sample is first input into the implicit structure feature enhancement layer, the model learns the scalar gradient judgment rule and feature enhancement weight of the stratigraphic boundary area, and generates the enhanced implicit structure feature sample; the multi-scale boundary feature capture layer learns the parameters of different size convolution kernels to master the extraction strategy of micro, medium and macro scale boundary features, and generates a multi-scale boundary feature set sample; the stratigraphic boundary preliminary prediction layer learns the mapping relationship between multi-scale features and boundary existence probability, and outputs the initial stratigraphic boundary data sample; the boundary topology verification and optimization layer learns the adaptation rules of stratigraphic contact relationship constraint and lithology sequence constraint, and corrects the topological contradictions in the initial boundary sample. The Dice loss function (adapted to the class imbalance problem of boundary prediction) is used for training, the final stratigraphic boundary data sample output by the model is compared with the boundary coordinate label in the labeled sample, and the feature enhancement weight, convolution kernel parameter, prediction layer mapping coefficient and verification rule parameter are adjusted through back propagation iteration. When the loss function value converges and the spatial coincidence degree of the boundary predicted by the model and the labeled boundary meets the standard, the training is completed, and the model has the ability to predict reliable stratigraphic boundaries from implicit representation.

[0213] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any skilled person in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for predicting the stratigraphic boundaries of geological envelopes based on artificial intelligence, characterized in that, The method comprises the following steps: Step 1, determining seismic exploration data and logging data of a target work area; Step 2, extracting seismic reflection features from the seismic exploration data based on a trained first neural network model, and extracting logging sequence features from the logging data based on a trained second neural network model; Step 3, generating a geological property representation of the target work area according to the seismic reflection features and the logging sequence features; Step 4, determining an implicit topological structure representation of a geological envelope in the target work area according to the geological property representation based on constraints of stratum contact relationship and lithology sequence for the target work area; Step 5, processing the implicit topological structure representation based on a trained third neural network model to predict a horizon boundary of the geological envelope in the target work area.

2. The method of claim 1, wherein, Step 1, determining seismic exploration data and logging data of a target work area, specifically comprising: Step 1.1, performing adaptive noise suppression processing on original seismic exploration data of the target work area to generate seismic exploration denoised data; Step 1.2, performing depth correction and outlier rejection processing on original logging data of the target work area to generate logging data correction results; Step 1.3, performing data space coordinate matching on the seismic exploration denoised data and the logging data correction results to generate seismic exploration data and logging data of the target work area.

3. The method of claim 2, wherein, Step 1.1, performing adaptive noise suppression processing on original seismic exploration data of the target work area to generate seismic exploration denoised data, specifically comprising: Step 1.1.1, dividing the original seismic exploration data of the target work area according to spatially continuous windows to calculate amplitude standard deviation, frequency distribution features and amplitude change rate between adjacent windows for each divided window data and determine seismic data window statistics features accordingly; Step 1.1.2, performing window-by-window signal type determination on the original seismic exploration data of the target work area according to the seismic data window statistics features, combining pre-labeled effective reflection signal samples and noise samples, applying dynamic weight filtering to the windows identified as noise, and retaining the original features of the windows identified as effective signals to generate seismic exploration denoised data.

4. The method of claim 1, wherein, Step 2: based on the trained first neural network model, extract seismic reflection features from the seismic exploration data, and based on the trained second neural network model, extract logging sequence features from the logging data, specifically comprising: Step 2.1, based on the trained first neural network model, performing multi-scale feature extraction on the seismic exploration denoised data to obtain initial seismic reflection features; Step 2.2, based on the trained second neural network model, performing vertical sequence feature extraction on the logging data to generate initial logging sequence features; Step 2.3, eliminating redundant features in the initial seismic reflection features and the initial logging sequence features that are irrelevant to the stratum boundary to generate screened seismic reflection features and screened logging sequence features.

5. The method of claim 4, wherein, Step 2.1, based on the trained first neural network model, performing multi-scale feature extraction on the seismic exploration denoised data to obtain initial seismic reflection features, specifically comprising: Step 2.1.1, based on the multi-scale convolution feature extraction layer in the first neural network model trained, different spatial scale features are extracted from the denoised seismic exploration data to generate preliminary multi-scale seismic feature maps; Step 2.1.2, based on the cross-scale feature attention fusion layer in the first neural network model trained, attention weight calculation and feature fusion are performed on the preliminary multi-scale seismic feature maps to generate fused seismic feature maps; Step 2.1.3, based on the residual noise suppression layer in the first neural network model trained, residual operation is performed on the fused seismic feature maps to generate denoised seismic feature maps; Step 2.1.4, based on the feature dimension compression and selection layer in the first neural network model trained, dimension compression is performed on the denoised seismic feature maps to eliminate redundant feature dimensions, effective feature selection is performed to retain features strongly related to the stratigraphic boundary, and initial seismic reflection features are generated.

6. The method of claim 4, wherein, Step 2.2, based on the second neural network model trained, vertical sequence feature extraction is performed on the logging data to generate initial logging sequence features, specifically including: Step 2.2.1, based on the vertical window sequencing layer in the second neural network model trained, vertical window division and sequence conversion are performed on the logging data to generate a set of vertical logging sequence segments; Step 2.2.2, based on the sequence residual denoising layer in the second neural network model trained, residual operation is performed on the set of vertical logging sequence segments to generate a set of denoised vertical logging sequence segments; Step 2.2.3, based on the multi-feature attention enhancement layer in the second neural network model trained, attention weight calculation is performed on the set of denoised vertical logging sequence segments to generate a set of attention-enhanced logging sequence features; Step 2.2.4, based on the sequence feature selection layer in the second neural network model trained, feature correlation analysis and selection are performed on the set of attention-enhanced logging sequence features to generate initial logging sequence features.

7. The method of claim 1, wherein, Step 3, according to the seismic reflection features and logging sequence features, the geological characteristics representation of the target work area is generated, specifically including: Step 3.1, based on the spatial coordinate system of the target work area, the spatial coordinate anchoring of the seismic reflection features and logging sequence features is performed, and then the correlation weight is calculated according to the correlation degree of the spatial coordinate anchored seismic reflection features and logging sequence features with the stratigraphic boundary to generate a feature correlation mapping table; Step 3.2, based on the feature correlation mapping table, the weighted fusion of the seismic reflection features and logging sequence features is performed, and the fused contradictory features are removed through stratigraphic topological relationship verification to generate a set of fused geological features; Step 3.3, based on the regional geological background of the target work area, the spatial continuous processing of the set of fused geological features is performed, and the geological significance corresponding to each feature is labeled to generate the geological characteristics representation of the target work area.

8. The method of claim 7, wherein, Step 3, according to the seismic reflection features and logging sequence features, the geological characteristics representation of the target work area is generated, which also includes: Step 3.0, according to the target work area and the existing geological outcrop data and drilling core description data in the surrounding area, the regional geological background of the target work area is generated.

9. The method of claim 8, wherein, Step 3.0, according to the existing geological outcrop data, drilling core description data of the target work area and the surrounding area, generate the regional geological background of the target work area, specifically including: Step 3.0.1, the existing geological outcrop data, drilling core description data of the target work area and the surrounding area are classified and labeled according to the lithology type, sedimentary structure and fossil combination, to generate the classified geological basis data set; Step 3.0.2, the lithology combination and sedimentary structure in the classified geological basis data set are analyzed, and the key features reflecting the sedimentary environment are extracted to generate the sedimentary environment feature set; Step 3.0.3, based on the sedimentary environment feature set, combined with the existing tectonic movement record of the target work area and the surrounding area, the time sequence correlation between the sedimentary environment feature and the tectonic evolution stage is established to generate the tectonic evolution time sequence table; Step 3.0.4, based on the sedimentary environment feature set and the tectonic evolution time sequence table, the sedimentary environment feature and the tectonic evolution stage are comprehensively verified, and the stratum development rule description is supplemented to generate the regional geological background of the target work area.

10. The method of claim 1, wherein, Step 4, based on the stratum contact relationship constraint and lithology sequence constraint constructed for the target work area, according to the geological characteristic representation, determine the implicit topological structure representation of the geological envelope in the target work area, specifically: based on the seismic exploration interpretation profile data of the target work area, construct the stratum contact relationship constraint and lithology sequence constraint for the target work area, according to the geological characteristic representation, determine the implicit topological structure representation of the geological envelope in the target work area.

11. The method of claim 1, wherein, Step 5, based on the third neural network model trained, the implicit topological structure representation is processed to predict the stratigraphic boundary of the geological envelope in the target work area, specifically including: Step 5.1, based on the implicit structure feature enhancement layer in the third neural network model trained, the boundary sensitive feature of the implicit topological structure representation is enhanced to generate the enhanced implicit structure feature; Step 5.2, based on the multi-scale boundary feature capture layer in the third neural network model trained, the multi-scale boundary feature set is extracted from the enhanced implicit structure feature to generate the multi-scale boundary feature set; Step 5.3, based on the stratigraphic boundary preliminary prediction layer in the third neural network model trained, the multi-scale boundary feature set is preliminarily predicted to generate the initial stratigraphic boundary data; Step 5.4, based on the boundary topology verification and optimization layer in the third neural network model trained, the initial stratigraphic boundary data is topologically verified and optimized to generate the final stratigraphic boundary data of the geological envelope in the target work area.

Images