Multi-well lithofacies identification method and device and medium

Abstract

The invention discloses a multi-well lithofacies recognition method and device and a medium, and the method comprises the steps: carrying out the data preprocessing of obtained original logging data, obtaining a training data set, training an initial recognition model based on the training data set, obtaining a target recognition model, and carrying out the recognition of the target recognition model. And respectively inputting well data, natural potential data and logging item data of a target well needing to be identified into corresponding branch models in the target identification model, carrying out weighted fusion on each output branch feature to obtain a corresponding target fusion feature, and carrying out target identification on the target well based on an intermediate frequency feature of the target fusion feature. The recognition result is output through the global discriminator and the local discriminator after adversarial alignment training in the target recognition model, so that the geological features are recognized based on the branch models at different angles, the recognition reliability is improved, alignment is performed through the intermediate-frequency features of the target fusion features, and the recognition accuracy is improved. And high-frequency and low-frequency information reflecting real geological differences is reserved, so that the recognition accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the field of oil and gas exploration and development, and in particular to a multi-well lithofacies identification method, device and medium. Background Art

[0002] In the process of oil and gas exploration and development, lithofacies identification is the basis for reservoir evaluation, reserve calculation and development plan formulation.

[0003] Existing solutions for lithofacies identification are typically based on machine learning solutions and logging data from a single depth point, or deep learning of the deep characteristics of logging curves to achieve lithofacies identification. However, the lithofacies identification solution based on logging data from a single depth point through machine learning ignores the continuity and spatial correlation of stratigraphic deposition, resulting in low identification accuracy. The solution of using deep learning to capture the deep characteristics of logging curves to achieve lithofacies identification has serious cross-well generalization problems, resulting in low recognition reliability.

[0004] Therefore, the existing scheme has the defects of low accuracy and reliability of lithofacies identification results. Summary of the Invention

[0005] The technical problem to be solved by the present invention is that the existing lithofacies identification scheme has the defects of low recognition accuracy and reliability. The purpose is to provide a multi-well lithofacies identification method, device and medium to solve the problems of low recognition accuracy and reliability in the existing lithofacies identification scheme.

[0006] The present invention is achieved through the following technical solutions: In a first aspect, the present application provides a multi-well lithofacies identification method, comprising: Acquiring raw well logging data and preprocessing the raw well logging data to obtain a training data set, wherein the training data set includes multiple sets of wellbore data, the wellbore data including logging project data at different depths of each wellbore and facies labels corresponding to the depths, wherein the facies labels are used to indicate lithofacies categories annotated by experts; Based on the training data set, the initial recognition model is trained to obtain a target recognition model; Acquire target well data of a target well to be identified, and input the target well data into an adaptive wavelet transform branch model in the target recognition model to obtain multi-scale features corresponding to an output of the adaptive wavelet transform branch model; Inputting the spontaneous potential data in the well logging project data of the target well data into the spontaneous potential transformation branch model in the target identification model to obtain the sedimentary rhythm characteristics corresponding to the output of the spontaneous potential transformation branch model; Inputting a key item logging curve pair in the target well data into a logging curve analysis branch model in the target identification model to obtain a petrophysical relationship feature corresponding to an output of the logging curve analysis branch model, wherein the key item logging curve pair is used to indicate a corresponding relationship between two different item data in the logging item data; Performing weighted fusion on the multi-scale features, the sedimentary rhythm features, and the rock physical relationship features to obtain corresponding target fusion features; Based on the intermediate frequency features of the target fusion features, the global discriminator and the local discriminator after adversarial alignment training in the target recognition model output recognition results, wherein the recognition results include the lithologic classification results of each depth point in the target wellbore and the lithologic transition boundary position.

[0007] In one possible design, preprocessing the original well logging data to obtain a training data set includes: Performing data cleaning on the original well logging data; performing normalization processing on the well logging data after data cleaning, organizing the normalized well logging data into three-dimensional vectors and two-dimensional vectors, and obtaining the training data set according to the three-dimensional vectors and the two-dimensional vectors; Among them, the three-dimensional vector is used to indicate the logging project data of different depth positions of each well, and the two-dimensional vector is used to indicate the phase similarity label of the corresponding depth position. According to the same depth position, the two-dimensional vector is associated with the corresponding three-dimensional vector.

[0008] In one possible design, obtaining the training data set according to the three-dimensional vector and the two-dimensional vector includes: Construct a sliding window of preset width; Sliding the sliding window along the direction of the three-dimensional vector indicating the wellbore depth, and performing sliding sampling on the three-dimensional vector and the corresponding two-dimensional vector to obtain a plurality of window data; The plurality of window data are used as the training data set.

[0009] In one possible design, obtaining multi-scale features corresponding to the output of the adaptive wavelet transform branch model includes: Establishing input indexes corresponding to the well logging project data according to different depth positions of the well logging project data in the target well data; According to the input index of each well logging project data, performing odd-even separation on each well logging project data in the target well data to obtain odd samples and even samples; Based on the odd samples and the even samples, multiple detail coefficients and approximation coefficients are calculated and obtained through the learnable prediction operator weights and update operator weights in the adaptive wavelet transform branch model, and the multiple detail coefficients and the multiple approximation coefficients are used as the multi-scale features.

[0010] In one possible design, obtaining a sedimentation rhythm feature corresponding to an output of the spontaneous potential transformation branch model includes: establishing a spontaneous potential curve based on the spontaneous potential data; According to the spontaneous potential curve and the preset mother wavelet, a continuous wavelet transform is performed through the spontaneous potential transformation branch model to extract the corresponding deposition rhythm features.

[0011] In a possible design, the correspondence between two different project data in the well logging project data includes: a first corresponding relationship, wherein the first corresponding relationship is used to indicate a relationship between mud content and porosity; A second corresponding relationship, wherein the second corresponding relationship is used to indicate a relationship between skeletal density and propagation speed; a third corresponding relationship, wherein the third corresponding relationship is used to indicate a difference in resistivity between deep and shallow areas; A fourth corresponding relationship is used to indicate the relationship between permeability and mud content.

[0012] In a possible design, outputting a recognition result by the global discriminator and the local discriminator after adversarial alignment training in the object recognition model includes: Decomposing the target fusion feature to obtain low-frequency features, the intermediate-frequency features, and high-frequency features, wherein the low-frequency features are used to indicate large-scale geological background information, the intermediate-frequency features are used to indicate instrument response difference information, and the high-frequency features are used to indicate rock texture detail information; The adversarial training of the global discriminator and the local discriminator is achieved through the first loss function of the global discriminator and the second loss function of the local discriminator, wherein the first loss function is used to eliminate the instrument response differences between different wells, and the second loss function is used to retain the facies classification characteristics of the target wellbore.

[0013] In a second aspect, the present application provides a multi-well lithofacies identification device, comprising: an acquisition module for acquiring raw logging data and preprocessing the raw logging data to obtain a training data set, wherein the training data set includes multiple sets of wellbore data, the wellbore data including logging project data at different depths of each wellbore, and facies labels corresponding to the depths, the facies labels indicating lithofacies categories annotated by experts, and training an initial recognition model based on the training data set to obtain a target recognition model; A first processing module is configured to obtain target wellbore data of a target wellbore to be identified, input the target wellbore data into an adaptive wavelet transform branch model in the target identification model to obtain multi-scale features corresponding to an output of the adaptive wavelet transform branch model, and input self-potential data in well logging project data of the target wellbore data into a self-potential transform branch model in the target identification model to obtain sedimentary rhythm features corresponding to an output of the self-potential transform branch model; a second processing module, configured to input the well logging project data of the target well data into the well logging curve analysis branch model in the target identification model to obtain petrophysical relationship features corresponding to the output of the well logging curve analysis branch model, wherein the key well logging curve pair is used to indicate the corresponding relationship between two different project data in the well logging project data; A first output module is configured to perform weighted fusion on the multi-scale feature, the sedimentary rhythm feature, and the rock physical relationship feature to obtain a corresponding target fusion feature; The second output module is used to output recognition results based on the intermediate frequency features of the target fusion features through the global discriminator and the local discriminator after adversarial alignment training in the target recognition model, wherein the recognition results include the lithologic classification results of each depth point in the target wellbore and the lithologic transition boundary position.

[0014] In a third aspect, the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are executed by a processor, they are used to implement a multi-well lithofacies identification method as described above.

[0015] Compared with the prior art, the present invention has the following advantages and beneficial effects: The present application provides a multi-well lithofacies identification method, device and medium, which preprocesses the acquired original logging data to obtain a training data set, trains the initial recognition model based on the training data set, and then inputs the wellbore data, natural potential data and logging project data of the target wellbore to be identified into the corresponding branch model in the target recognition model, and performs weighted fusion on the multi-scale features, sedimentary rhythm features and rock physical relationship features output by each branch model to obtain the corresponding target fusion feature. Based on the intermediate frequency features of the target fusion feature, the global discriminator and the local discriminator after adversarial alignment training in the target recognition model output the lithofacies classification results of each depth point and the recognition results of the lithofacies conversion boundary position. Therefore, the present application identifies geological features based on the branch models at different angles, improves the recognition reliability, and aligns the intermediate frequency features of the target fusion feature to retain the high-frequency and low-frequency information reflecting the real geological differences, thereby improving the recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the following briefly introduces the drawings required for use in the examples. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be considered as limiting the scope. A person of ordinary skill in the art can also derive other relevant drawings based on these drawings without inventive effort. In the drawings: Figure 1 Schematic diagram of the process of multi-well lithofacies identification method provided in the embodiment of this application Figure 1 ; Figure 2 Schematic diagram of the process of multi-well lithofacies identification method provided in the embodiment of this application Figure 2 ; Figure 3 A schematic diagram of sliding window extraction provided in an embodiment of the present application; Figure 4 A schematic diagram of the sedimentary microphases provided in the embodiments of the present application; Figure 5 Schematic diagram of wavelet transform time-frequency domain feature extraction provided in an embodiment of the present application; Figure 6 A schematic diagram comparing the lithofacies predictions and actual results provided in the examples of this application; Figure 7 This is a schematic diagram of the structure of the multi-well lithofacies identification device. DETAILED DESCRIPTION

[0017] Exemplary embodiments are described in detail herein, with examples illustrated in the accompanying drawings. When the following description refers to the drawings, identical numerals in different drawings represent identical or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments are not intended to represent all embodiments consistent with this application.

[0018] Identifying lithofacies types is a crucial step in oil and gas exploration and development. This is typically done by analyzing various logging data, such as GR (Gamma Ray), RT (Resistivity Tool), DEN (Density), and NPHI (Neutron Porosity Log).

[0019] In order to improve the efficiency of lithofacies identification, existing solutions usually adopt intelligent identification solutions, such as machine learning solutions, deep learning solutions, and transfer learning solutions to achieve lithofacies identification.

[0020] Traditional machine learning solutions, such as support vector machines and random forests, perform classification by extracting statistical features from well logging curves. However, these solutions treat the logging response at each depth point as an independent sample, ignoring the continuity and spatial correlation of stratigraphic deposition. Since lithofacies characteristics vary at different scales, existing solutions often use fixed-scale feature extraction and are unable to fully capture the multi-scale characteristics of lithofacies.

[0021] Deep learning solutions, such as convolutional neural networks and long short-term memory networks, can automatically learn the deep features of well logging curves. However, while these solutions perform well within a single well or region, they suffer from a serious cross-well generalization problem. Specifically, when a trained model is applied to a new well, the recognition accuracy typically decreases. This is because the logging responses of different wells are affected by non-geological factors such as instrument calibration, mud properties, and wellbore conditions, resulting in different logging characteristics for the same lithofacies in different wells.

[0022] Transfer learning schemes, such as domain adaptation neural networks and conditional domain adaptation networks, use adversarial training to reduce data distribution differences between different wells. However, these schemes typically align all features indiscriminately, resulting in the loss of important geological information. Existing methods for multi-channel well logging curves typically use simple concatenation or averaging, without considering the relative importance of different curves in different geological environments. Furthermore, purely data-driven methods fail to incorporate geological prior knowledge, such as sedimentary rhythms and lithofacies combinations. This can lead to identification results that violate basic geological laws, resulting in low recognition reliability.

[0023] Therefore, the present application provides a multi-well lithofacies identification method, which performs data preprocessing on the acquired original logging data to obtain a training data set, trains the initial recognition model based on the training data set, and obtains the target recognition model. The wellbore data, natural potential data and logging project data of the target wellbore to be identified are respectively input into the corresponding branch models in the target recognition model, and the multi-scale features, sedimentary rhythm features and rock physical relationship features output by each branch model are weightedly fused to obtain the corresponding target fusion features. Based on the intermediate frequency features of the target fusion features, the global discriminator and the local discriminator after adversarial alignment training in the target recognition model output the lithofacies classification results of each depth point and the recognition results of the lithofacies conversion boundary position. Therefore, the present application identifies geological features based on branch models at different angles, improves recognition reliability, and aligns the intermediate frequency features of the target fusion features to retain high-frequency and low-frequency information reflecting the real geological differences, thereby improving recognition accuracy.

[0024] The following uses specific embodiments to describe in detail the technical solution of the present application and how the technical solution of the present application solves the above-mentioned technical problems. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present application will be described below in conjunction with the accompanying drawings.

[0025] Example Figure 1 Schematic diagram of the process of multi-well lithofacies identification method provided in the embodiment of this application Figure 1 ,like Figure 1 As shown, the method includes: S101: Acquire original well logging data, and pre-process the original well logging data to obtain a training data set.

[0026] Specifically, the logging data of each wellbore is collected as raw logging data. The logging data for each wellbore includes a variety of different logging items such as neutron porosity, gamma ray, shallow resistivity, acoustic wave time difference, spontaneous potential, density, density correction and deep induction resistivity. The raw logging data also includes phase labels, which are used to indicate the lithofacies category marked by experts at each depth point.

[0027] Furthermore, when preprocessing the original well logging data, it is necessary to perform data cleaning and normalization processing on the original well logging data to obtain a preprocessed training data set.

[0028] S102: Based on the training data set, train the initial recognition model to obtain a target recognition model.

[0029] Specifically, after obtaining the preprocessed training data set, the initial recognition model is trained based on the training data set, and the trained initial recognition model is used as the target recognition model, where the target recognition model includes multiple branch models such as an adaptive wavelet transform branch model, a natural potential transform branch model, and a logging curve analysis branch model.

[0030] S103 , acquiring target well data of a target well to be identified, and inputting the target well data into an adaptive wavelet transform branch model in the target identification model to obtain multi-scale features corresponding to an output of the adaptive wavelet transform branch model.

[0031] Specifically, after obtaining the trained target recognition model, the target wellbore data of the target wellbore to be identified is input into the adaptive wavelet transform branch model in the target recognition model, where the target wellbore data includes logging project data at different depth positions of the target wellbore. After the logging project data of the target wellbore is separated, predicted and updated by the adaptive wavelet transform branch model, multiple detail coefficients and approximate coefficients are obtained, and the obtained multiple detail coefficients and multiple approximate coefficients are used as corresponding multi-scale features.

[0032] S104: Input the spontaneous potential data in the well logging project data of the target well data into the spontaneous potential transformation branch model in the target recognition model to obtain the sedimentary rhythm characteristics corresponding to the output of the spontaneous potential transformation branch model.

[0033] Specifically, based on the natural potential data, i.e., spontaneous potential, in the target wellbore data, a corresponding natural potential curve is generated, where the natural potential curve is used to indicate the natural potential change trend at each depth point in the continuous wellbore depth direction. Based on the preset target wavelet, such as the Morlet wavelet, and the natural potential of each depth point in the natural potential curve, the corresponding sedimentary rhythm characteristics are extracted.

[0034] S105 , inputting the logging project data of the target well data into the logging curve analysis branch model in the target recognition model to obtain the petrophysical relationship characteristics corresponding to the output of the logging curve analysis branch model.

[0035] Specifically, two selected project data are determined in the logging project data as the corresponding correlation coefficient pair, and the corresponding key logging curve pair is determined based on the correlation coefficient pair, wherein the key logging curve pair includes the selected first logging project curve and the second logging project curve. The corresponding delayed physical relationship characteristics are extracted based on the project data of each depth point indicated by the first logging project curve and the second logging project curve through the logging curve analysis branch model.

[0036] S106 , performing weighted fusion on the multi-scale features, the sedimentary rhythm features, and the rock physical relationship features to obtain corresponding target fusion features.

[0037] Specifically, after obtaining the multi-scale features, sedimentary rhythm features and rock physical relationship features output by the branch model in the target recognition model, the wavelet energy spectrum-driven attention mechanism is used to perform weighted fusion of the multi-scale features, sedimentary rhythm features and rock physical relationship features to obtain the corresponding target fusion features.

[0038] S107. Based on the intermediate frequency features of the target fusion features, the global discriminator and the local discriminator after adversarial alignment training in the target recognition model output a recognition result.

[0039] Specifically, after obtaining the fused target fusion features, the target fusion features are feature decomposed to obtain the intermediate frequency features indicating the instrument response differences. According to the intermediate frequency features obtained by the decomposition, a first loss function for the global discriminator and a second loss function for the local discriminator are established according to the intermediate frequency features to realize adversarial alignment training of the global discriminator and the local discriminator. The target fusion features output recognition results through the global discriminator and the local discriminator after adversarial alignment training, where the recognition results include the facies classification results of each depth point in the target wellbore and the facies conversion boundary position.

[0040] A multi-well lithofacies identification method provided in an embodiment of the present application performs data preprocessing on the acquired original logging data to obtain a training data set, trains an initial recognition model based on the training data set, and after obtaining a target recognition model, inputs the wellbore data, natural potential data, and logging project data of the target wellbore to be identified into the corresponding branch models in the target recognition model respectively, and performs weighted fusion on the multi-scale features, sedimentary rhythm features, and rock physical relationship features output by each branch model to obtain the corresponding target fusion features. After that, based on the intermediate frequency features of the target fusion features, the global discriminator and the local discriminator after adversarial alignment training in the target recognition model output the lithofacies classification results of each depth point and the recognition results of the lithofacies conversion boundary position. Therefore, the present application identifies geological features based on branch models at different angles, improves recognition reliability, and aligns the intermediate frequency features of the target fusion features to retain high-frequency and low-frequency information reflecting the real geological differences, thereby improving recognition accuracy.

[0041] Example Figure 2 Schematic diagram of the process of multi-well lithofacies identification method provided in the embodiment of this application Figure 2 , Figure 3 The sliding window extraction diagram provided in the embodiment of the present application is combined with Figure 2 and Figure 3 As shown, the method includes: S201: Acquire original well logging data, and perform data cleaning on the original well logging data.

[0042] Specifically, after collecting the original logging data, data cleaning is performed on the original logging data, where the data cleaning action includes removing null values ​​in the well section, deleting duplicate records, and eliminating abnormal values ​​such as jump values ​​and physically unreasonable intervals.

[0043] S202 : performing normalization processing on the well logging data after data cleaning, and organizing the normalized well logging data into three-dimensional vectors and two-dimensional vectors.

[0044] Specifically, after completing the data cleaning of the logging data, the data of each logging project in the logging data is normalized, and the logging curves corresponding to the data of each logging project are mapped to the standard numerical range. The normalized and standardized logging data are organized into a three-dimensional vector, as shown in the following formula (1): (1); in, is the number of wells logged, is the number of depth sampling points for each well, is the number of logging projects, and the labels of each phase in the logging data are organized into corresponding two-dimensional vectors, as shown in the following formula (2): (2); in, is the total number of lithofacies classifications, For the The first in the well The facies categories of each depth point are marked by experts. The three-dimensional vector is used to indicate the logging project data at different depth positions of each well. The two-dimensional vector is used to indicate the facies label of the corresponding depth position. According to the same depth position, the two-dimensional vector is associated with the corresponding three-dimensional vector.

[0045] S203: Obtain a training data set according to the three-dimensional vector and the two-dimensional vector.

[0046] Specifically, if Figure 3 As shown, a sliding window of preset width is constructed, and the sliding window is moved along the direction indicating the well depth in the three-dimensional vector to perform sliding sampling on the three-dimensional vector and the corresponding two-dimensional vector.

[0047] Furthermore, using a fixed width The sliding window is sampled in the depth direction with a step size of 1 to construct a training sample sequence. The sliding window size can be adjusted in the range of 20-200 sampling points according to the specific geological characteristics and sampling rate. For each depth position , construct window data, as shown in the following formula (3): (3); in, is the fixed width of the sliding window, is the depth position, and the corresponding phase labels for each window data are as shown in the following formula (4): (4); in, For the corresponding depth position, after obtaining multiple window data, the multiple window data are used as training data sets.

[0048] S204: Based on the training data set, train the initial recognition model to obtain a target recognition model.

[0049] S205: Acquire target well data of the target well to be identified.

[0050] Specifically, target wellbore data of a target wellbore currently requiring lithofacies identification is acquired, wherein the target wellbore data includes a variety of different well logging items.

[0051] S206 , performing odd-even separation on the well logging item data in the target well data by using the adaptive wavelet transform branch model in the target recognition model to obtain odd samples and even samples.

[0052] Specifically, the target well data is input into the adaptive wavelet transform branch model in the target recognition model. In the adaptive wavelet transform branch model, the input index of the corresponding logging project data is established according to the different depth positions of each logging project data in the target well data. According to the odd input index and the even input index, the logging project data is separated into odd samples and even samples, as shown in the following formula (5): (5); in, For even-numbered samples, is an odd number of samples.

[0053] S207. Based on the odd samples and the even samples, multiple detail coefficients and approximation coefficients are calculated and obtained through the learnable prediction operator weights and update operator weights in the adaptive wavelet transform branch model, and the multiple detail coefficients and the multiple approximation coefficients are used as the multi-scale features.

[0054] Specifically, after obtaining the odd and even samples, the predicted odd samples are calculated through the learnable prediction operator weights in the adaptive wavelet transform branch model, and the difference between the actual odd sample value and the predicted odd sample value is used as the detail coefficient, as shown in the following formula (6): (6); in, is the learnable prediction operator weight, is the detail coefficient, i.e., the difference between the actual odd value and the predicted odd sample value; Furthermore, the learnable update operator weights in the adaptive wavelet transform branch model and the detail coefficients at this time are used to update the corresponding predicted even samples, and the sum of the actual even sample value and the predicted even sample value is used as the corresponding approximate coefficient, as shown in the following formula (7): (7); in, is the learnable update operator weight, After obtaining the detail coefficients and approximation coefficients of the first layer at this time, the approximation coefficients of the first layer are used as the input of the parity separation of the second layer, and the above prediction and update steps are repeated based on the approximation coefficients of the first layer to obtain the detail coefficients and approximation coefficients corresponding to the second layer. After the preset number of layers, such as three to five layers, multiple detail coefficients and multiple approximation coefficients are obtained as multi-scale features. .

[0055] S208 , performing continuous wavelet transform on the spontaneous potential curve corresponding to the spontaneous potential data and the preset mother wavelet through the spontaneous potential transformation branch model in the target recognition model, so as to extract the corresponding deposition rhythm features.

[0056] Specifically, the logging items in the target well data are taken as spontaneous potential data, and the corresponding spontaneous potential curve is subjected to continuous wavelet transform to extract the sedimentary rhythm characteristics, as shown in the following formula (8): (8); in, is the mother wavelet function, is the complex conjugate of the mother wavelet, a is the scale parameter that controls the expansion and contraction of the wavelet, and b is the translation parameter that controls the position of the wavelet. For the mother wavelet function, it is as shown in the following formula (9): (9); Further, As the mother wavelet function, the scale parameter a is set within a preset range such as 2 0 to 2 8 By transforming between them, we can obtain sedimentary rhythm characteristics of different scales.

[0057] S209: Input the key item logging curve pair of the target well data into the logging curve analysis branch model in the target recognition model to obtain the rock physical relationship characteristics corresponding to the output of the logging curve analysis branch model.

[0058] Specifically, two project data are determined from multiple well logging project data of the target well data as key well logging projects, and the well logging curves corresponding to the two key well logging project data are used as a key well logging curve pair, where the key well logging curve pair includes the selected first well logging project curve and the second well logging project curve. According to the project data of each depth point indicated by the first well logging project curve and the second well logging project curve, the rock physical relationship characteristics are extracted as shown in the following formula (10): (10); in, is the correlation coefficient of the two logging curves at depth position i, that is, the rock physical relationship characteristics, W is the size of the sliding window, is the logging value of the first logging project curve X at depth position i+j, is the average value of the first logging project curve X in the sliding window, is the logging value of the second logging project curve Y at depth i+j, is the average value of the second logging project curve Y within the sliding window.

[0059] Furthermore, the key logging curve pairs are used to indicate the correspondence between two different project data in the logging project data. The correspondence includes a first correspondence for indicating the relationship between mud content and porosity, a second correspondence for indicating the relationship between skeleton density and propagation velocity, a third correspondence for indicating the difference in deep and shallow resistivity, and a fourth correspondence for indicating the relationship between permeability and mud content.

[0060] S210 , performing weighted fusion on the multi-scale features, the sedimentary rhythm features, and the rock physical relationship features to obtain corresponding target fusion features.

[0061] Specifically, after obtaining the multi-scale features, sedimentary rhythm features, and rock physical relationship features output by the branch model in the target recognition model, the multi-scale features, sedimentary rhythm features, and rock physical relationship features are mapped to a unified dimension through the fully connected layer, which is achieved by the following formula (11): (11); in is the original eigenvector of the i-th branch, is the learnable mapping matrix of the i-th branch, is the original feature dimension, It is a unified mapping dimension; is the bias vector of the i-th branch.

[0062] Furthermore, based on the multi-scale features, sedimentary rhythm features, and rock physical relationship features mapped to a unified dimension, the energy score weight corresponding to each feature is calculated using the following formula (12): (12); in, is a learnable energy evaluation vector used to measure the importance of each branch feature, is a scalar, is the energy score weight of the i-th branch, which is used to indicate the information richness of the corresponding branch feature.

[0063] Furthermore, based on the features mapped to the unified dimension and the corresponding energy score weights, weighted fusion is performed to obtain the target fusion feature, which is calculated by the following formula (13):

[0064] in, is the Hadamard product, broadcasting the attention weights to Element-wise multiplication is performed after the same dimension.

[0065] S211. Decompose the target fusion features to obtain low-frequency features, medium-frequency features, and high-frequency features.

[0066] Specifically, after obtaining the target fusion features, the target fusion features are decomposed into low-frequency features, medium-frequency features and high-frequency features, as shown in the following formula (14): (14); in, is a low-frequency feature, is the mid-frequency characteristic, The low-frequency feature is used to indicate large-scale geological background information, the medium-frequency feature is used to indicate instrument response difference information, and the high-frequency feature is used to indicate rock texture detail information.

[0067] S212: Implement adversarial training of the global discriminator and the local discriminator through the first loss function of the global discriminator and the second loss function of the local discriminator.

[0068] Specifically, in order to eliminate the systematic differences caused by instrument responses between different wells while retaining the true geological information, the intermediate frequency features representing the instrument response difference information are aligned in the adversarial domain. When the global discriminator is trained adversarially, the first loss function is expressed as follows (15): (15); Where N is the total number of wells logged, is the data set of the i-th well logging, For a multi-classification network, output N-dimensional probability vector ,in Represents the probability that the input feature comes from the i-th well log; by minimizing the first loss value , the discriminator learns to recognize the unique instrument response pattern of each well, and the feature extractor maximizes this first loss value through the gradient reversal layer, making the mid-frequency features indistinguishable between different wells, thereby eliminating the systematic differences between logs caused by instrument response.

[0069] Furthermore, during the adversarial training of the local discriminator, the lithologic classification characteristics of the target wellbore are retained through the second loss function. For the second loss function, it is shown in the following formula (16):

[0070] Where C is the total number of lithofacies categories, is the data subset belonging to the cth lithofacies in the i-th well, is the one-hot encoding vector representing the c-th class, is the element-wise Hadamard product, The function is that the universal mask selectively retains the c-th class features, and the rest of the features are set to zero. It is a local discriminator for the cth lithofacies. Different from the global discriminator, considering that different lithofacies may have different domain migration patterns between different wells, an independent local discriminator is set for each lithofacies category.

[0071] Furthermore, based on the first A distance of the global distribution corresponding to the global discriminator and the second A distance of the local distribution corresponding to the local discriminator, the balance coefficient between the first loss value and the second loss value is calculated. The balance coefficient is shown in the following formula (17): (17); Where C is the total number of lithofacies categories, is the first A distance, The first A-distance is used to indicate the average A-distance between all source wells and target wellbores, and the second A-distance is used to indicate the average A-distance between the source wells and target wellbores within each lithofacies category.

[0072] Furthermore, A-distance, which is a standard metric in domain adaptation, is used to measure the difference between global distribution and local distribution, as shown in (18): (18); in, It is the error rate of the newly trained binary classifier to distinguish the two domains. The larger the A-distance is, the greater the difference between the two domains is, and the stronger the domain adaptation is.

[0073] S213. Based on the intermediate frequency features of the target fusion features, the global discriminator and the local discriminator after adversarial alignment training in the target recognition model output recognition results.

[0074] Specifically, the target fusion feature is input into the global discriminator and local discriminator after adversarial alignment training in the target recognition model. Based on the first loss value and the second loss value of the global discriminator and the local discriminator, the overall loss function is constructed. The overall loss function is shown in the following formula (19): (19); in, is the lithofacies classification loss at each depth point, is the boundary detection loss; based on the overall loss function, the trained target recognition model outputs the recognition results, which include the lithofacies classification results of each depth point in the target wellbore and the location of the lithofacies transition boundary.

[0075] A multi-well lithofacies identification method provided in an embodiment of the present application performs data preprocessing on the acquired original logging data to obtain a training data set, trains an initial recognition model based on the training data set, and after obtaining a target recognition model, inputs the wellbore data, natural potential data, and logging project data of the target wellbore to be identified into the corresponding branch models in the target recognition model respectively, and performs weighted fusion on the multi-scale features, sedimentary rhythm features, and rock physical relationship features output by each branch model to obtain the corresponding target fusion features. After that, based on the intermediate frequency features of the target fusion features, the global discriminator and the local discriminator after adversarial alignment training in the target recognition model output the lithofacies classification results of each depth point and the recognition results of the lithofacies conversion boundary position. Therefore, the present application identifies geological features based on branch models at different angles, improves recognition reliability, and aligns the intermediate frequency features of the target fusion features to retain high-frequency and low-frequency information reflecting the real geological differences, thereby improving recognition accuracy.

[0076] This application provides specific embodiments. Figure 4 A schematic diagram of the deposition microphase provided in the embodiment of this application, Figure 5 Schematic diagram of wavelet transform time-frequency domain feature extraction provided in the embodiment of the present application, Figure 6 A schematic diagram comparing the lithofacies prediction and actual results provided in the embodiments of the present application is provided to further illustrate the above embodiments.

[0077] In this specific example, 43 vertical wells (N = 43) were sampled, covering the main reservoir interval. The logging data sampling interval was 0.1 m, and the average number of depth points per well was approximately . Eleven well logs (C = 11) were collected at each depth point as raw logging data, including conventional logging items such as depth (DEPT), neutron porosity (NPHI), gamma ray (GR), micro shallow probe resistivity (MSFL), sonic transit time (DT), sonic transit time correction (DT_CALI), spontaneous potential (SP), wellbore caliper (CALI), density (RHOB), density correction (RHOB_CALI), and deep induction resistivity (ILD).

[0078] like Figure 4 As shown in the figure, three geological interpretation experts (E=3) conducted sedimentary microfacies interpretation for all well sections, and the label space was divided into R=6 lithofacies categories, namely Outer Ramp, Distal Mid-Ramp, Proximal Mid-Ramp, High Shoal, Lagoon, Basin, and Medium-low shal.

[0079] Data preprocessing and sample construction of raw logging data.

[0080] Specifically, null values ​​and abnormal jump values ​​in the logging curves were removed; data of abnormally enlarged well diameter sections (CALI>15 inches) were marked; and each curve was normalized using z-score to make its mean 0 and standard deviation 1.

[0081] Furthermore, the sliding window size W is set to 101, which corresponds to a depth range of 10.1 meters; sliding is performed with a step size of 1, and context features are constructed for each depth point to obtain a training dataset.

[0082] Furthermore, the training dataset was divided, 30 wells were selected as the source domain training set, 8 wells were selected as the validation set, and 5 wells were retained as the target domain test set.

[0083] Feature extraction is performed through the adaptive wavelet transform branch model, the spontaneous potential transform branch model and the logging curve analysis branch model in the target recognition model.

[0084] Specifically, a 4-layer lifting scheme is used to decompose the adaptive wavelet transform branch model; both the prediction operator and the update operator are implemented using 1D convolution with a kernel size of 7.

[0085] Further, if Figure 5 As shown in the figure, for the natural potential transformation branch model, the natural potential curve is subjected to Morlet wavelet transformation; the scale parameter a is selected : Extract the energy distribution of each scale and form a 128-dimensional feature vector.

[0086] Furthermore, for the well logging curve analysis branch model, the sliding correlation coefficients of eight pairs of key curve combinations are calculated, such as GR-NPHI (shale-porosity relationship), RHOB-DT (density-sonic wave relationship), oILD-MSFL (deep-shallow resistivity difference), SP-GR (permeability-shale relationship), RHOB-NPHI (density-neutron relationship), DT-NPHI (sonic wave-neutron relationship), CALI-GR (wellbore diameter-shale relationship), and ILD-SP (resistivity-spontaneous potential relationship).

[0087] The features extracted by each branch model are trained with attention fusion and selective domain adaptation.

[0088] Specifically, the features of each branch are fused, and the three branch features are mapped to a 256-dimensional unified space through a fully connected layer; the attention weight is calculated based on the wavelet energy spectrum; the typical weight distribution is: homogeneous layer segment (0.25, 0.35, 0.40), thin interlayer segment (0.45, 0.30, 0.25); a three-layer wavelet decomposition is used to separate the fused features into low, medium, and high frequency components, and only the medium frequency features such as the second layer detail coefficient d are used. 2 Conduct domain confrontation, Furthermore, dynamic weight adjustment is performed, and the initial ω=0.5 is set. During the training process, it is automatically adjusted according to the A-distance. The final convergence value is: ω≈0.68, indicating that local alignment is more important in this dataset.

[0089] When training an object recognition model, the weight of the boundary detection task in the overall loss function is =0.5, adversarial loss weight =0.1.

[0090] As shown in the figure, Figure 6 As shown in the figure, the recognition results on 5 test wells show that for cross-well performance comparison, the method of the present invention has an average accuracy of 87.3% and a performance degradation of only 12.7%; the version without domain adaptation has an average accuracy of 72.5% and a performance degradation of 27.5%; the classic machine learning has an average accuracy of 61.2% and a performance degradation of 38.8%; and the consistency of expert interpretation is about 82.6%.

[0091] In terms of phase category performance, for High Shoal: recognition accuracy is 92.1%, boundary detection accuracy is 89.3%; Proximal Mid-Ramp: recognition accuracy is 88.7%, boundary detection accuracy is 86.2%; Distal Mid-Ramp: recognition accuracy is 85.4%, boundary detection accuracy is 83.1%; Lagoon: recognition accuracy is 86.9%, boundary detection accuracy is 84.7%; Outer Ramp: recognition accuracy is 84.2%, boundary detection accuracy is 81.5%; Basin: recognition accuracy is 85.5%, boundary detection accuracy is 82.8%.

[0092] In an embodiment of the present invention, the electronic device or main control device can be divided into functional modules according to the above method example. For example, each functional module can be divided according to each function, or two or more functions can be integrated into one processing unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional modules. It should be noted that the division of modules in the embodiment of the present invention is schematic and is only a logical functional division. In actual implementation, there may be other division methods.

[0093] Figure 7 A schematic diagram of the structure of a multi-well lithofacies identification device provided in an embodiment of the present application is shown in FIG. Figure 7 As shown, the apparatus 700 includes: An acquisition module 701 is configured to acquire raw logging data and preprocess the raw logging data to obtain a training data set, wherein the training data set includes multiple sets of wellbore data, the wellbore data including logging project data at different depths of each wellbore, and facies labels corresponding to the depths, the facies labels indicating lithofacies categories annotated by experts. An initial recognition model is trained based on the training data set to obtain a target recognition model. The first processing module 702 is configured to obtain target wellbore data of a target wellbore to be identified, input the target wellbore data into an adaptive wavelet transform branch model in the target identification model to obtain multi-scale features corresponding to the output of the adaptive wavelet transform branch model, and input self-potential data in well logging project data of the target wellbore data into a self-potential transform branch model in the target identification model to obtain sedimentary rhythm features corresponding to the output of the self-potential transform branch model; The second processing module 703 is configured to input the well logging project data of the target well data into the well logging curve analysis branch model in the target identification model to obtain the petrophysical relationship characteristics corresponding to the output of the well logging curve analysis branch model, wherein the key well logging curve pair is used to indicate the corresponding relationship between two different project data in the well logging project data; A first output module 704 is configured to perform weighted fusion on the multi-scale feature, the sedimentary rhythm feature, and the rock physical relationship feature to obtain a corresponding target fusion feature; The second output module 705 is used to output recognition results based on the intermediate frequency features of the target fusion features through the global discriminator and the local discriminator after adversarial alignment training in the target recognition model, wherein the recognition results include the facies classification results of each depth point in the target wellbore and the facies transition boundary position.

[0094] Furthermore, the acquisition module 701 is specifically configured to clean the original logging data, perform normalization processing based on the cleaned logging data, organize the normalized logging data into three-dimensional vectors and two-dimensional vectors, and obtain the training data set based on the three-dimensional vectors and the two-dimensional vectors; Among them, the three-dimensional vector is used to indicate the logging project data of different depth positions of each well, and the two-dimensional vector is used to indicate the phase similarity label of the corresponding depth position. According to the same depth position, the two-dimensional vector is associated with the corresponding three-dimensional vector.

[0095] Furthermore, the acquisition module 701 is specifically used to construct a sliding window of a preset width, and slide the sliding window along the direction indicating the shaft depth in the three-dimensional vector to sample the three-dimensional vector and the corresponding two-dimensional vector to obtain multiple window data, and use the multiple window data as the training data set.

[0096] Furthermore, the first processing module 702 is specifically configured to establish an input index corresponding to the well logging project data according to different depth positions of the well logging project data in the target well data; According to the input index of each well logging project data, performing odd-even separation on each well logging project data in the target well data to obtain odd samples and even samples; Based on the odd samples and the even samples, multiple detail coefficients and approximation coefficients are calculated and obtained through the learnable prediction operator weights and update operator weights in the adaptive wavelet transform branch model, and the multiple detail coefficients and the multiple approximation coefficients are used as the multi-scale features.

[0097] Furthermore, the first processing module 702 is specifically configured to establish a self-potential curve based on the self-potential data; and perform a continuous wavelet transform based on the self-potential curve and a preset mother wavelet through the self-potential transformation branch model to extract the corresponding deposition rhythm features.

[0098] Furthermore, the second processing module 703, the correspondence between two different project data in the logging project data, includes: a first correspondence, the first correspondence is used to indicate the relationship between mud content and porosity; a second correspondence, the second correspondence is used to indicate the relationship between skeleton density and propagation velocity; a third correspondence, the third correspondence is used to indicate the difference in deep and shallow resistivity; a fourth correspondence, the fourth correspondence is used to indicate the relationship between permeability and mud content.

[0099] Furthermore, the second output module 705 is specifically configured to decompose the target fusion feature to obtain low-frequency features, the intermediate-frequency features, and high-frequency features, wherein the low-frequency features are used to indicate large-scale geological background information, the intermediate-frequency features are used to indicate instrument response difference information, and the high-frequency features are used to indicate rock texture detail information; The adversarial training of the global discriminator and the local discriminator is achieved through the first loss function of the global discriminator and the second loss function of the local discriminator, wherein the first loss function is used to eliminate the instrument response differences between different wells, and the second loss function is used to retain the facies classification characteristics of the target wellbore.

[0100] The present application also provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer-executable instructions. When a processor executes the computer-executable instructions, the multi-well lithofacies identification method as described above is implemented.

[0101] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A multi-well lithofacies identification method, characterized in that: The method comprises: Acquiring raw well logging data and preprocessing the raw well logging data to obtain a training data set, wherein the training data set includes multiple sets of wellbore data, the wellbore data including logging project data at different depths of each wellbore and facies labels corresponding to the depths, wherein the facies labels are used to indicate lithofacies categories annotated by experts; Based on the training data set, the initial recognition model is trained to obtain a target recognition model; Acquire target well data of a target well to be identified, and input the target well data into an adaptive wavelet transform branch model in the target recognition model to obtain multi-scale features corresponding to an output of the adaptive wavelet transform branch model; Inputting the spontaneous potential data in the well logging project data of the target well data into the spontaneous potential transformation branch model in the target identification model to obtain the sedimentary rhythm characteristics corresponding to the output of the spontaneous potential transformation branch model; Inputting a key item logging curve pair in the target well data into a logging curve analysis branch model in the target identification model to obtain a petrophysical relationship feature corresponding to an output of the logging curve analysis branch model, wherein the key item logging curve pair is used to indicate a corresponding relationship between two different item data in the logging item data; Performing weighted fusion on the multi-scale features, the sedimentary rhythm features, and the rock physical relationship features to obtain corresponding target fusion features; Based on the intermediate frequency features of the target fusion features, the global discriminator and the local discriminator after adversarial alignment training in the target recognition model output recognition results, wherein the recognition results include the lithologic classification results of each depth point in the target wellbore and the lithologic transition boundary position.

2. The multi-well lithofacies identification method according to claim 1, characterized in that: The preprocessing of the original well logging data to obtain a training data set includes: Performing data cleaning on the original well logging data; performing normalization processing on the well logging data after data cleaning, organizing the normalized well logging data into three-dimensional vectors and two-dimensional vectors, and obtaining the training data set according to the three-dimensional vectors and the two-dimensional vectors; Among them, the three-dimensional vector is used to indicate the logging project data of different depth positions of each well, and the two-dimensional vector is used to indicate the phase similarity label of the corresponding depth position. According to the same depth position, the two-dimensional vector is associated with the corresponding three-dimensional vector.

3. The multi-well lithofacies identification method according to claim 2, characterized in that: The acquiring the training data set according to the three-dimensional vector and the two-dimensional vector includes: Construct a sliding window of preset width; Sliding the sliding window along the direction of the three-dimensional vector indicating the wellbore depth, and performing sliding sampling on the three-dimensional vector and the corresponding two-dimensional vector to obtain a plurality of window data; The plurality of window data are used as the training data set.

4. The multi-well lithofacies identification method according to claim 1, characterized in that: The obtaining of multi-scale features corresponding to the output of the adaptive wavelet transform branch model includes: Establishing input indexes corresponding to the well logging project data according to different depth positions of the well logging project data in the target well data; According to the input index of each well logging project data, performing odd-even separation on each well logging project data in the target well data to obtain odd samples and even samples; Based on the odd samples and the even samples, multiple detail coefficients and approximation coefficients are calculated and obtained through the learnable prediction operator weights and update operator weights in the adaptive wavelet transform branch model, and the multiple detail coefficients and the multiple approximation coefficients are used as the multi-scale features.

5. The multi-well lithofacies identification method according to claim 1, characterized in that: The obtaining of the sedimentation rhythmic features corresponding to the output of the spontaneous potential transformation branch model includes: establishing a spontaneous potential curve based on the spontaneous potential data; According to the spontaneous potential curve and the preset mother wavelet, a continuous wavelet transform is performed through the spontaneous potential transformation branch model to extract the corresponding deposition rhythm features.

6. The multi-well lithofacies identification method according to claim 1, characterized in that: The corresponding relationship between two different project data in the well logging project data includes: a first corresponding relationship, wherein the first corresponding relationship is used to indicate a relationship between mud content and porosity; A second corresponding relationship, wherein the second corresponding relationship is used to indicate a relationship between skeletal density and propagation speed; a third corresponding relationship, wherein the third corresponding relationship is used to indicate a difference in resistivity between deep and shallow areas; A fourth corresponding relationship is used to indicate the relationship between permeability and mud content.

7. The multi-well lithofacies identification method according to claim 1, characterized in that: Outputting the recognition result by the global discriminator and the local discriminator after adversarial alignment training in the target recognition model includes: Decomposing the target fusion feature to obtain low-frequency features, the intermediate-frequency features, and high-frequency features, wherein the low-frequency features are used to indicate large-scale geological background information, the intermediate-frequency features are used to indicate instrument response difference information, and the high-frequency features are used to indicate rock texture detail information; The adversarial training of the global discriminator and the local discriminator is achieved through the first loss function of the global discriminator and the second loss function of the local discriminator, wherein the first loss function is used to eliminate the instrument response differences between different wells, and the second loss function is used to retain the facies classification characteristics of the target wellbore.

8. A multi-well lithofacies identification device, characterized in that: include: an acquisition module for acquiring raw logging data and preprocessing the raw logging data to obtain a training data set, wherein the training data set includes multiple sets of wellbore data, the wellbore data including logging project data at different depths of each wellbore, and facies labels corresponding to the depths, the facies labels indicating lithofacies categories annotated by experts, and training an initial recognition model based on the training data set to obtain a target recognition model; A first processing module is configured to obtain target wellbore data of a target wellbore to be identified, input the target wellbore data into an adaptive wavelet transform branch model in the target identification model to obtain multi-scale features corresponding to an output of the adaptive wavelet transform branch model, and input self-potential data in well logging project data of the target wellbore data into a self-potential transform branch model in the target identification model to obtain sedimentary rhythm features corresponding to an output of the self-potential transform branch model; a second processing module, configured to input the well logging project data of the target well data into the well logging curve analysis branch model in the target identification model to obtain petrophysical relationship features corresponding to the output of the well logging curve analysis branch model, wherein the key well logging curve pair is used to indicate the corresponding relationship between two different project data in the well logging project data; A first output module is configured to perform weighted fusion on the multi-scale feature, the sedimentary rhythm feature, and the rock physical relationship feature to obtain a corresponding target fusion feature; The second output module is used to output recognition results based on the intermediate frequency features of the target fusion features through the global discriminator and the local discriminator after adversarial alignment training in the target recognition model, wherein the recognition results include the lithologic classification results of each depth point in the target wellbore and the lithologic transition boundary position.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the multi-well lithofacies identification method according to any one of claims 1 to 7.

Images