Deep learning-based accurate and efficient identification and classification method for shale bedding structure

Abstract

The invention relates to a shale bedding structure accurate and efficient identification and classification method based on deep learning, and the method comprises the steps: calculating shale bedding structure pre-classification parameters according to a fast shear wave time difference curve, a slow shear wave time difference curve and a well logging curve having correlation with a shale bedding structure in target block well logging data; according to shale bedding structure pre-classification parameters, a segmented electric imaging logging diagram and a preliminary classification result are obtained; according to the preliminary classification result, the segmented electrical imaging logging graph is input into different stratification recognition network models, a shale stratification structure recognition result is obtained, the recognition result is analyzed through image morphology, a final classification result is obtained, and the stratification recognition network models are constructed based on a Swin Transform network model and obtained based on training of a training set. The invention provides a new division method for refined shale bedding structure types.

Description

Technical Field

[0001] The present invention relates to the technical field of shale classification, and in particular to a method for accurately and efficiently identifying and classifying shale bedding structures based on deep learning. Background Art

[0002] In recent years, significant breakthroughs have been made in the exploration and development of continental shale oil, and industrial production has been achieved. Continental shale reservoirs are characterized by well-developed bedding, rich types, and diverse combinations. Research shows that frequently developed laminae and thin interlayers will enhance the heterogeneity of shale reservoirs, which are favorable conditions for the stable seepage of shale oil and also favorable fracturing transformation channels. Therefore, accurate identification of laminae is the key to selecting geological sweet spots.

[0003] At present, conventional logging curve feature analysis and reconstruction methods are usually used to identify the development of bedding and interlayers in shale sedimentary structures. However, due to the low accuracy of logging curves and many influencing factors, large prediction errors occur in the types of shale bedding structures. Micro-resistivity imaging logging (electric imaging logging) has high resolution and a relatively complete coverage range, and can visually and visually present geological feature information such as the shape and spatial position of the formation. Therefore, electric imaging logging can play an important role in identifying fractures and bedding in oil and gas reservoirs. When using electric imaging logging to identify the types of shale bedding structures, traditional methods mainly rely on manual evaluation methods to divide the bedding and thin interlayers in the reservoir, and calibrate the categories of shale bedding structures by the shape and quantity of the bedding. The method of manual identification has the problems of low efficiency and being affected by human experience and subjective factors.

[0004] And the existing technology has the following disadvantages:

[0005] 1. The variation characteristics of conventional logging curves are affected by various factors. Through the analysis of conventional logging curves, the development characteristics of shale bedding can be characterized to a certain extent, but the types of shale bedding structures cannot be accurately distinguished. Especially for shale reservoirs with coexisting laminated, layered, and massive structures, the distinguishing effect is very unsatisfactory. 2. At present, when processing electric imaging logging pictures, the pictures are cut at equal-height intervals, which may cut a set of bedding into two pictures respectively, resulting in errors when using machine learning methods to judge the types of shale bedding structures. In addition, when evaluating the pictures used, the same set of artificial intelligence network models are used, resulting in low accuracy and efficiency. 3. When using machine learning methods to identify shale bedding based on electric imaging logging results, the more efficient and accurate Swin Transformer artificial intelligence technology has not been used. Summary of the Invention

[0006] The object of the present invention is to provide a precise and efficient recognition and classification method for shale bedding structures based on deep learning. On the basis of traditional logging curves, a new method for calculating characteristic parameters representing shale bedding structure types is proposed, and criteria for segmenting resistivity image logging maps are constructed based on the characteristic parameters. For different types of shale bedding structures, a new method for identifying and obtaining shale bedding geometric parameters by using different Swin Transformer artificial intelligence network models is proposed, and the types of shale bedding structures are judged by methods of bedding density and proportion. This method further improves the accuracy and speed of classifying shale bedding structure types, providing theoretical and technical support for identifying "sweet spots" in shale reservoirs.

[0007] To achieve the above object, the present invention provides the following solutions:

[0008] A precise and efficient recognition and classification method for shale bedding structures based on deep learning, comprising:

[0009] Calculating pre-classification parameters for shale bedding structures according to the fast shear wave slowness curve, slow shear wave slowness curve and logging curves related to shale bedding structures in the logging data of the target block;

[0010] Obtaining the segmented resistivity image logging map and preliminary classification results according to the pre-classification parameters of the shale bedding structures;

[0011] According to the preliminary classification results, inputting the segmented resistivity image logging map into different bedding recognition network models to obtain recognition results of shale bedding structures, and analyzing the recognition results by using image morphology to obtain final classification results, wherein the bedding recognition network model is constructed based on the Swin Transformer network model and trained based on a training set, and the training set includes resistivity image logging maps labeled with shale beddings of different types of bedding structures.

[0012] Optionally, calculating the pre-classification parameters for shale bedding structures includes:

[0013] Calculating the anisotropy coefficient at different depths of the shale reservoir according to the fast shear wave slowness curve and slow shear wave slowness curve;

[0014] Calculating the box-counting dimension of the logging curve at different depths of the reservoir according to the logging curve related to the shale bedding structure, wherein the logging curve related to the shale bedding structure is an acoustic slowness curve, a natural gamma curve or a neutron curve;

[0015] Obtaining the pre-classification parameters for the shale bedding structures according to the anisotropy coefficient and the box-counting dimension.

[0016] Optionally, calculating the anisotropy coefficient at different depths of the shale reservoir is:

[0017]

[0018] The box-counting dimension of the logging curves at different depths of the reservoir is calculated as follows:

[0019]

[0020] The pre-classification parameters of the shale bedding structure are obtained as follows:

[0021] P p = D f ANI

[0022] In the formula, ANI is the shear wave travel time anisotropy coefficient; s 1 is the fast shear wave travel time; s 2 is the slow shear wave travel time; D f is the box-counting dimension; r is the length of the box; N(r) is the minimum number of boxes covering the logging curve, and P p is the pre-classification parameter of the shale bedding structure.

[0023] Optionally, obtaining the segmented electrical image logging map and the preliminary classification result includes:

[0024] Normalize the pre-classification parameters of the shale bedding structure to obtain the normalized pre-classification parameters;

[0025] Compare the normalized pre-classification parameters with the first preset classification standard to obtain the preliminary classification result;

[0026] Cut the electrical image logging map in a sliding window manner according to the pre-classification parameters of the shale bedding structure to obtain the segmented electrical image logging map.

[0027] Optionally, the bedding recognition network model includes: a complex bedding recognition network model and a simple bedding recognition network model;

[0028] The complex bedding recognition network model is constructed based on the Swin Transformer network model and is used for the recognition of laminated and layered shale bedding structures;

[0029] The simple bedding recognition network model is constructed based on the Swin Transformer network model and is used for the recognition of massive shale bedding structures;

[0030] Among them, the Swin Transformer network model uses Swin-L as the benchmark network. For the Swin-L network model that constructs the complex bedding recognition network model, high feature dimensions and the number of attention heads are selected, and the window size of the attention layer is set according to the hybrid strategy. For the Swin-L network model that constructs the simple bedding recognition network model, low feature dimensions and the number of attention heads are selected, and the window for the attention layer processing is set to a medium window.

[0031] Optionally, the bedding recognition network model adopts a total loss function including cross-entropy loss and IOU loss;

[0032] The total loss function is:

[0033] Total Loss = α·CE + β·IoU Loss

[0034]

[0035]

[0036] In the formula, y is the true label; is the predicted output of the model; y i is the true label of the i-th sample; is the predicted probability of the i-th sample; N is the number of samples; A is the true region; B is the predicted region; α is the weight of the cross-loss function; β is the weight of the IoU loss function.

[0037] Optionally, before obtaining the segmented electrical imaging logging map, it includes:

[0038] Complement the blank strips in the electrical imaging logging map and perform threshold segmentation to output a binary image;

[0039] Filter the binary image using the Sobel operator to obtain the preprocessed electrical imaging logging map.

[0040] Optionally, using image morphology to analyze the recognition result to obtain the final classification result includes:

[0041] Adopt image morphology to perform edge detection on the recognized bedding, calculate the thickness of each bedding and count the quantity;

[0042] According to the thickness and quantity of each bedding, calculate the bedding proportion and bedding density. Among them, the bedding proportion is the ratio of the total thickness of the bedding in the lithological body region to the thickness of the lithological body, and the bedding density is the ratio of the number of beddings in the lithological body region to the thickness of the lithological body;

[0043] Compare the bedding proportion and bedding density with the second preset classification standard to obtain the final classification result.

[0044] The beneficial effects of the present invention are as follows: The present invention first proposes pre-classification parameters for shale bedding structures and uses them as the basis for segmenting electrical imaging maps, enabling more accurate segmentation of electrical imaging logging maps. Using the Swin Transformer deep learning network, two network models for bedding recognition are established for complex-structured bedding (laminated, layered) and simple-structured bedding (massive). For shale reservoirs with coexisting laminated, layered, and massive beddings, when using the dual network model to identify beddings, the efficiency of bedding recognition is further improved on the basis of ensuring the accuracy of bedding recognition. On this basis, a new method using dual parameters of bedding density and proportion is proposed to discriminate the types of shale bedding structures, providing an important basis for the economy, efficiency, and rationality of shale oil reservoir evaluation and "sweet spot" identification. Description of the Drawings

[0045] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0046] Figure 1 It is a diagram of the pre-classification parameter calculation of well logging curves and the classification result of shale bedding structures in electrical imaging logging for the embodiments of the present invention;

[0047] Figure 2 It is the segmentation result of different types of shale bedding structures in the electrical imaging logging map for the embodiments of the present invention;

[0048] Figure 3 It is the result of the Swin Transformer in the embodiments of the present invention for identifying laminated shale bedding;

[0049] Figure 4 It is the result of the Swin Transformer in the embodiments of the present invention for identifying layered shale bedding;

[0050] Figure 5 It is the result of the Swin Transformer in the embodiments of the present invention for identifying massive shale bedding;

[0051] Figure 6 It is a flowchart of a precise and efficient identification and classification method for shale bedding structures based on deep learning in the embodiments of the present invention. Detailed Embodiments

[0052] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0053] To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0054] The electrical imaging logging technology can accurately characterize the development characteristics of shale bedding, and the artificial intelligence system can accurately and efficiently identify the bedding in the electrical imaging logging diagram. Therefore, constructing an artificial intelligence system that can automatically and accurately identify the bedding in the electrical imaging logging diagram is of great significance, which can effectively solve the problem of time-consuming and laborious manual interaction for identifying bedding and avoid the problem of discrimination errors in shale bedding structures caused by subjective evaluations of human experience.

[0055] As Figure 6 shown, this embodiment provides a precise and efficient recognition and classification method for shale bedding structures based on deep learning, including:

[0056] According to the fast shear wave slowness curve, slow shear wave slowness curve, and logging curves related to shale bedding structures in the logging data of the target block, calculate the pre-classification parameters of shale bedding structures;

[0057] According to the pre-classification parameters of shale bedding structures, obtain the segmented electrical imaging logging diagram and the preliminary classification result;

[0058] According to the preliminary classification result, input the segmented electrical imaging logging diagram into different bedding recognition network models to obtain the recognition result of shale bedding structures, and use image morphology to analyze the recognition result to obtain the final classification result. Among them, the bedding recognition network model is constructed based on the Swin Transformer network model and trained based on a training set. The training set includes electrical imaging logging diagrams labeled with different types of bedding structures. The recognition result includes: dividing the regional features of the bedding in the segmented electrical imaging logging diagram into white and the formation matrix into black.

[0059] Further, calculating the pre-classification parameters of shale bedding structures includes:

[0060] According to the fast shear wave slowness curve and slow shear wave slowness curve, calculate the anisotropy coefficient at different depths of the shale reservoir;

[0061] Calculate the box-counting dimension of the logging curves at different depths of the reservoir according to the logging curves that are correlated with the shale bedding structure, where the logging curves that are correlated with the shale bedding structure are the acoustic travel-time curve, the natural gamma curve, or the neutron curve;

[0062] Obtain the pre-classification parameters of the shale bedding structure based on the anisotropy coefficient and the box-counting dimension.

[0063] Furthermore, the calculation of the anisotropy coefficient at different depths of the shale reservoir is as follows:

[0064]

[0065] The calculation of the box-counting dimension of the logging curves at different depths of the reservoir is as follows:

[0066]

[0067] The obtaining of the pre-classification parameters of the shale bedding structure is as follows:

[0068] P p = D f ANI (3)

[0069] In the formula, ANI is the shear-wave travel-time anisotropy coefficient; s 1 is the fast shear-wave travel-time; s 2 is the slow shear-wave travel-time; D f is the box-counting dimension; r is the length of the box; N(r) is the minimum number of boxes covering the logging curve, and P p is the pre-classification parameter of the shale bedding structure.

[0070] Specifically, the calculation of the pre-classification parameters of the shale bedding structure includes: ① Based on the fast shear-wave travel-time curve and the slow shear-wave travel-time curve in the conventional logging curves, calculate the anisotropy at different depths of the shale reservoir according to formula (1). ② According to the box-counting dimension theory in fractal geometry, use the logging curves that are correlated with the shale bedding structure to calculate the box-counting dimension of the logging curves at different depths of the reservoir according to formula (2). The box-counting dimension of the logging curve is calculated by means of a sliding window, where the size of the sliding window is set to n data points, the sliding step is set to 1 logging data point, and the depth corresponding to the middle point of the window is the depth of this box-counting dimension. ③ The anisotropy reflects the quantity, shape, and type of bedding, and the box-counting dimension characterizes the complexity of the bedding. Therefore, based on the anisotropy and the box-counting dimension, establish a calculation method for the pre-classification parameters of the shale bedding structure, as shown in formula (3).

[0071] Further, the obtaining of the segmented electrical image logging map and the preliminary classification result includes:

[0072] Normalize the pre-classification parameters of the shale bedding structure to obtain the normalized pre-classification parameters;

[0073] Compare the normalized pre-classification parameters with the first preset classification criteria to obtain a preliminary classification result;

[0074] According to the shale bedding structure pre-classification parameters, cut the electrical image logging map in the form of a sliding window to obtain the segmented electrical image logging map.

[0075] Specifically, the segmentation and preliminary classification of the electrical image logging map include: ① Based on the conventional logging curves, calculate the pre-classification parameters (P p ) of the shale bedding structure and normalize it. ② For the normalized pre-classification parameters, propose the classification criteria for the shale bedding structure, that is, the first preset classification criteria: when P p <0.3, it is massive shale; when 0.3 ≤ P p ≤ 0.6, it is laminated shale; when 0.6 < P p , it is laminar shale. ③ Establish the cutting criteria for the electrical image logging map. According to the pre-classification parameters calculated from the conventional logging curves, cut the electrical image logging map dynamically in the form of a sliding window. ④ Since laminar and laminated shales have the characteristics of thin bedding, large quantity, and complex structure, the recognition difficulty is much higher than that of laminated shale. Therefore, to improve the recognition accuracy, it is necessary to set an upper limit for the cutting height of the electrical images of laminar and laminated shales.

[0076] Specifically, the bedding recognition network model includes: a complex bedding recognition network model and a simple bedding recognition network model;

[0077] The complex bedding recognition network model is constructed based on the Swin Transformer network model and is used for the recognition of the bedding structures of laminar and laminated shales;

[0078] The simple bedding recognition network model is constructed based on the Swin Transformer network model and is used for the recognition of the bedding structures of massive shales;

[0079] Among them, the Swin Transformer network model uses Swin-L as the benchmark network. The Swin-L network model for constructing the complex bedding recognition network model selects a high feature dimension and the number of attention heads, and sets the size of the attention layer window according to the hybrid strategy. The Swin-L network model for constructing the simple bedding recognition network model selects a low feature dimension and the number of attention heads, and the window processed by the attention layer is set as a medium window.

[0080] Specifically, the network model for shale bedding recognition constructed based on Swin Transformer deep learning includes: ① Based on Swin Transformer deep learning technology, two sets of network models for shale bedding recognition are designed, which are respectively used to recognize laminated and layered bedding with complex bedding structures and massive bedding with simple bedding structures. ② For laminated and layered bedding with complex structures, alternating thicknesses, and high density, a Swin Transformer network model with high recognition accuracy but large computational complexity is constructed. Taking Swin-L as the benchmark network, with a smaller Patch Size, the window size of the attention layer is set according to the hybrid strategy, and high feature dimensions and the number of attention heads are selected in the four stages of image processing. ③ For massive bedding with simple structures and small quantities, a Swin Transformer network model with small computational complexity and high recognition efficiency is constructed while ensuring recognition accuracy. Taking Swin-T as the benchmark network, a larger Patch Size is selected, the window for attention layer processing is set as a medium window, and low feature dimensions and the number of attention heads are selected in the four stages of image processing. ④ The two Swin Transformer network models adopt the Adam algorithm structure to construct an adaptive learning rate. The loss function adopts a calculation method combining cross-entropy loss (Formula (4)) and IOU loss (Formula (5)), and its total loss function is shown in Formula (6). ⑤ n sample images are made by the method of manual annotation, the training set and the test set are set according to a ratio of 9:1, and the mean intersection over union (MIoU) index, recall rate (Recall), and precision rate (Precision) are used to evaluate the recognition accuracy of the model for bedding classification, and the network model is trained to complete the construction of the network model.

[0081] Furthermore, the bedding recognition network model adopts a total loss function including cross-entropy loss and IOU loss;

[0082] The total loss function is:

[0083] Total Loss=α·CE+β·IoU Loss (4)

[0084]

[0085] Where y is the true label; is the predicted output of the model; y i is the true label of the i-th sample; is the predicted probability of the i-th sample; N is the number of samples; A is the true region; B is the predicted region; α is the weight of the cross-loss function; β is the weight of the IoU loss function.

[0086] Further, before obtaining the segmented electrical imaging logging map, it includes:

[0087] Complete the blank strips in the electrical imaging logging map and perform threshold segmentation to output a binary image;

[0088] Filter the binary image using the Sobel operator to obtain the preprocessed electrical imaging logging map.

[0089] Furthermore, use image morphology to analyze the identified shale bedding to obtain the final classification results including:

[0090] Adopt image morphology to perform edge detection on the identified bedding, calculate the thickness of each bedding and count the quantity;

[0091] According to the thickness and quantity of each bedding, calculate the bedding proportion and bedding density. Among them, the bedding proportion is the ratio of the total thickness of the bedding in the lithological body area to the thickness of the lithological body, and the bedding density is the ratio of the number of beddings in the lithological body area to the thickness of the lithological body;

[0092] Compare the bedding proportion and bedding density with the second preset classification standard to obtain the final classification result.

[0093] Specifically, the image preprocessing and analysis include: ① First, complete the blank strips in the electrical imaging logging map. ② Perform threshold segmentation on the image to output a binary image. ③ Filter the binary image using the Sobel operator to enhance the extraction of horizontal direction features in the image. ④ Segment the electrical imaging logging map based on the pre-classification parameters of the shale bedding structure. ⑤ According to the result of segmenting the electrical imaging logging map by the pre-classification parameters, import the laminated and layered bedding pictures into the complex Swin Transformer network model for analysis, and import the massive bedding pictures into the simple Swin Transformer network model for analysis to identify the shale bedding structure. ⑥ Use image morphology to perform edge detection on the identified bedding, calculate the thickness of each bedding and count the quantity. ⑦ Define the bedding density as the ratio of the number of beddings in the lithological body area to the thickness of the lithological body, and the bedding proportion as the ratio of the total thickness of the beddings in the lithological body area to the thickness of the lithological body, as shown in formulas (6) and (7). According to the density and proportion of the beddings, propose the classification criteria for laminated, layered, and massive shale bedding structures. When the bedding proportion < 40%, and the bedding density > 0.4, it is determined as laminated shale; when the bedding proportion < 40%, and 0.2 < bedding density ≤ 0.4, it is determined as layered shale; when the bedding proportion > 40%, it is determined as massive shale. ⑧ When the simple network model determines that the shale bedding is laminated or layered, it needs to be imported into the complex bedding recognition network model for reclassification.

[0094]

[0095] In the formula, H is the height of a single picture, cm; is the total number of cracks in the figure, pieces; hi is the height of the i-th bedding, in cm, F d is the bedding density, F r is the bedding proportion.

[0096] The present invention will be further described below with reference to the accompanying drawings:

[0097] A precise and efficient identification and classification method for shale bedding structures based on deep learning, including:

[0098] (1) Calculation of pre-classification parameters for shale bedding structures:

[0099] ① Collect and sort out the logging data of the target block, obtain the fast shear wave time difference curve and slow shear wave time difference curve in the logging data, and calculate the heterogeneity of the target interval (3410m - 3460m) according to formula (1). ② Select the acoustic wave time difference curve of the target block to calculate the box dimension. The calculation of the box dimension is implemented in Matlab. The analysis window length is set to 11 logging data points. Read the TXT data of the acoustic wave time difference curve through Matlab, calculate the box dimension, and export the calculation result after storing it in the TXT text. ③ Based on the heterogeneity and the box dimension, calculate the pre-classification parameters of the shale bedding structures at different depths of the reservoir according to formula (3). The calculation results are as Figure 1 shown. ④ Coarsen the pre-classification parameters, and reduce the sampling rate of the pre-classification parameters from 12.5 cm to 25 cm according to the calculation method of the mean value.

[0100] (2) Segmentation of the electrical imaging logging map:

[0101] ① Segment the electrical imaging logging map of 3410m - 3460m according to the depth of the electrical imaging logging map corresponding to the pre-classification parameters of the shale bedding structures. According to the calculation results of the pre-classification parameters, there are 20 intervals of laminated shale, 25 intervals of layered shale, and 7 intervals of massive shale in the target block. ② In this embodiment, the minimum unit of the pre-classification window is 25 cm. Set the upper limit of the segmentation height of the electrical imaging map in the laminated and layered shale areas to 50 cm, and do not adjust the segmentation height in the massive shale area. According to the segmentation standard of the height of the electrical imaging map of the laminated and layered shale, a total of 33 laminated shale pictures, 32 layered shale pictures, and 7 massive shale pictures are finally obtained, as Figure 2 shown.

[0102] (3) Construct a shale bedding structure recognition network model based on Swin Transformer deep learning, where the shale bedding structure recognition network model is used to recognize the input segmented image:

[0103] ①Based on Swin Transformer deep learning, a network model for identifying complex bedding is constructed with the Swin-L network model as the benchmark, which is mainly used for identifying the bedding structures of laminated and layered shale. The image feature extraction process of the model is set to four stages, namely: the first stage of image block division and preliminary feature extraction, the second stage of further feature extraction, the third stage of advanced feature extraction, and the fourth stage of refinement and output. Four important components, namely image block division, linear embedding, Transformer encoder, and shifted window attention, are included in these four stages. The Patch Size in the model is set to 3×3; a hybrid strategy is selected to set the window size, that is, the window size in the first two stages of image processing is set to 3×3, and the window size in the last two stages is set to 7×7; the number of Transformer Blocks in the four stages of image processing is set to 2, 2, 18, and 2 respectively, the feature dimensions are set to 192, 384, 768, and 1536 respectively, and the number of attention heads is set to 6, 12, 24, and 48 respectively. ②Based on Swin Transformer deep learning, a network model for identifying simple bedding is constructed with the Swin-T network model as the benchmark, which is mainly used for identifying the bedding structures of massive shale. The image feature extraction process of the model is set to four stages, which is the same as that of the complex bedding identification network model. The Patch Size in the network model is set to 4×4; the window size in the four stages of image processing is set to 7×7; the number of Transformer Blocks in the four stages of image processing is set to 2, 2, 6, and 2 respectively, the feature dimensions are set to 96, 192, 384, and 768 respectively, and the number of attention heads is set to 4, 8, 16, and 32 respectively. ④When training the two Swin Transformer network models, the Adam algorithm is used to construct an adaptive learning rate, with its initial learning rate set to 0.001 and the upper and lower limits of learning rate adjustment set to 0.1 - 0.0001. The loss function adopts a calculation method combining cross-entropy loss and IOU loss, and its total loss function is shown in formula (6). ⑤1000 sample images are made by the method of manually annotating different types of bedding structures in the electrical imaging logging map. The training set and test set are set according to the ratio of 9:1 for model training. The mean intersection over union (MIoU) index, recall rate (Recall), and precision rate (Precision) are used to evaluate the accuracy of the model for bedding classification and recognition. The calculation results of the two models show that the recognition accuracy of bedding is high and meets the application requirements. The evaluation parameters of the network model recognition accuracy are shown in Table 1.

[0104] Table 1

[0105]

[0106] (4) Identification and classification of shale bedding structures:

[0107] ① Use the Filtersim algorithm to complement the blank strips in the electrical imaging logging map. ② Perform threshold segmentation on the image to output a binary image. Filter the binary image using the Sobel operator to enhance the representation of the horizontal direction features of the image. ③ Based on the pre-classification parameters of the shale bedding structure, segment the electrical imaging logging picture. According to the classification of the cut picture, use different deep learning network models to identify the bedding and formation matrix in the electrical imaging logging map, and divide the regional features of the bedding in the image into white and the formation matrix into black. ④ Use the edge detection method based on Python to identify the number of beddings in the picture. Import the OpenCV library and Numpy library in Python, and use the cv2.GaussianBlur function to smooth the image with a 5x5 convolution kernel to reduce the noise in the image. On this basis, use the edge detection function cv2.Canny to enhance the edges of each bedding in the picture, and set the upper and lower limits of the cv2.Canny function threshold to 10 and 220 respectively. Finally, use the Rtre_External retrieval mode in the cv2.findContours function to extract the external contours of the beddings, and obtain the information and number of the external contours of the beddings in the picture. ⑤ Calculate the thickness of each bedding through image profile analysis based on Python. Use the Zhang-Suen algorithm to reduce the thickness of all beddings to 1 pixel to obtain the center line and centroid of each bedding. Based on the centroid, select 5 measurement points at equal interval thicknesses on the center line of the bedding, index the uppermost and lowermost white pixels of the selected 5 measurement points, calculate the distance between them, and obtain the thickness of the bedding at this position. Calculate the average value of the thicknesses of the 5 measurement points of the bedding and set it as the thickness of the bedding. ⑥ Calculate the density of the bedding according to formula (6) and calculate the proportion of the bedding according to formula (7). Compare the proportion of the bedding and the density of the bedding with the second preset classification standard to obtain the final classification result: when the proportion of the bedding < 40%, and the density of the bedding > 0.4, it is determined as laminated shale; when the proportion of the bedding < 40%, and 0.2 ≤ the density of the bedding ≤ 0.4, it is determined as layered shale; when the proportion of the bedding > 40%, it is determined as massive shale. Perform calculation and analysis on the electrical imaging map based on the Swin Transformer deep learning model. The final calculation and determination results of three types of shale bedding structures are selected, as shown in Figure 3 , Figure 4 , Figure 5 shown. As shown in Figure 3 and Table 2 of the geometric data of the laminated shale bedding identified by Swin Transformer, there are 40 beddings in the lithological body with a thickness of 50 cm. The maximum thickness of the bedding is 0.47 cm, the minimum thickness is 0.23 cm, and the average thickness is 0.37 cm; the density of the bedding is 0.80, and the proportion of the bedding is 30.13%; therefore, it is determined as laminated shale. As shown inFigure 4 As shown in Table 3 of the results of the Swin Transformer for identifying the geometric data of the bedding of laminated shale, there are 17 bedding planes in the lithological body with a thickness of 50 cm. The maximum thickness of the bedding plane is 1.52 cm, the minimum thickness is 0.67 cm, and the average thickness is 0.92 cm. The density of the bedding plane is 0.34, and the proportion of the bedding plane is 31.16%. Therefore, it is determined as laminated shale. As Figure 5 As shown in Table 4 of the results of the Swin Transformer for identifying the geometric data of the bedding of massive shale, there are 7 bedding planes in the lithological body with a thickness of 50 cm. The maximum thickness of the bedding plane is 16.36 cm, the minimum thickness is 0.76 cm, and the average thickness is 4.32 cm. The density of the bedding plane is 0.14, and the proportion of the bedding plane is 60.45%. Therefore, it is determined as massive shale. ⑦ When using the model of simple bedding identification to determine the electro-imaging logging map as laminated or laminated shale, the picture needs to be re-imported into the complex bedding identification network model for refined bedding identification and discrimination of shale bedding structure. ⑧ As Figure 1 As shown, through the Swin Transformer deep learning network model, the electro-imaging logging map is analyzed and calculated to divide the types of shale bedding structures at different depths. Among them, there are 12 areas of laminated shale, 9 areas of laminated shale, and 3 areas of massive shale.

[0108] Table 2

[0109]

[0110] Table 3

[0111]

[0112]

[0113] Table 4

[0114]

[0115] First, the present invention calculates the reservoir anisotropy through conventional logging curves, analyzes the fractal characteristics of the curves, and constructs the classification characteristic parameters of the shale bedding structure through these two parameters to preliminarily evaluate and divide the types of shale bedding structures at different depths. On this basis, according to the preliminary classification characteristic parameters, the electro-imaging logging map is cut into pictures of different sizes according to the types of shale bedding structures in the form of a sliding window. Based on the Swin Transformer artificial intelligence technology, different bedding network identification models are designed for the shale bedding structures with complex structures (laminated, laminated) and the massive shale bedding structures with simple structures, so as to realize the efficient and accurate identification of the shape, quantity, and thickness of the bedding in the shale. Finally, based on the discrimination method of the double parameters of bedding density and proportion, the precise division of the shale bedding types is realized.

[0116] Based on traditional logging curves, a new method for calculating characteristic parameters representing shale bedding structure types is proposed, and criteria for the segmentation of electrical imaging logging diagrams are constructed based on the characteristic parameters. For different types of shale bedding structures, a new method for identifying and obtaining shale bedding geometric parameters using different Swin Transformer artificial intelligence network models is proposed, and the types of shale bedding structures are judged by the methods of bedding density and proportion. This method further improves the accuracy and speed of classifying shale bedding structure types, providing theoretical and technical support for the identification of "sweet spots" in shale reservoirs.

[0117] The embodiments described above are only descriptions of the preferred embodiments of the present invention and do not limit the scope of the present invention. Without departing from the design spirit of the present invention, various deformations and improvements made by those of ordinary skill in the art to the technical solutions of the present invention shall fall within the protection scope determined by the claims of the present invention.

Claims

1. A method for accurate and efficient identification and classification of shale bedding structure based on deep learning, characterized in that: include: Calculate the shale bedding structure pre-classification parameters based on the fast shear wave time difference curve, slow shear wave time difference curve and the logging curve related to the shale bedding structure in the target block logging data; According to the shale bedding structure pre-classification parameters, obtaining the segmented electrical imaging logging map and preliminary classification results; According to the preliminary classification results, the segmented electrical imaging logging map is input into different bedding recognition network models to obtain shale bedding structure recognition results, and the recognition results are analyzed using image morphology to obtain final classification results, wherein the bedding recognition network model is constructed based on the Swin Transformer network model and obtained through training based on a training set, and the training set includes electrical imaging logging maps annotated with different types of bedding structures.

2. The method for accurately and efficiently identifying and classifying shale bedding structure based on deep learning according to claim 1 is characterized in that: Calculation of shale bedding structure pre-classification parameters include: Calculating the anisotropy coefficients of the shale reservoir at different depths according to the fast shear wave time difference curve and the slow shear wave time difference curve; According to the logging curves correlated with the shale bedding structure, the box counting dimensions of the logging curves at different depths of the reservoir are calculated, wherein the logging curves correlated with the shale bedding structure are acoustic time difference curves, natural gamma curves or neutron curves; The shale bedding structure pre-classification parameters are obtained according to the anisotropy coefficient and the box counting dimension.

3. The method for accurate and efficient identification and classification of shale bedding structure based on deep learning according to claim 2 is characterized in that: The anisotropy coefficient of shale reservoir at different depths is calculated as: The box-counting dimension of the logging curve at different depths of the reservoir is calculated as: The shale bedding structure pre-classification parameters are obtained as follows: P p =D f ANI Where ANI is the anisotropy coefficient of shear wave delay; s1 is the fast shear wave delay; s2 is the slow shear wave delay; D f is the box counting dimension; r is the length of the box; N(r) is the minimum number of boxes covering the well logging curve, P p Pre-classification parameters for shale bedding structure.

4. The method for accurate and efficient identification and classification of shale bedding structure based on deep learning according to claim 1 is characterized in that: The segmented electrical imaging logging images and preliminary classification results include: Normalizing the shale bedding structure pre-classification parameters to obtain normalized pre-classification parameters; Comparing the normalized pre-classification parameter with a first preset classification standard to obtain the preliminary classification result; According to the shale bedding structure pre-classification parameters, the electrical imaging logging map is cut in a sliding window manner to obtain the segmented electrical imaging logging map.

5. The method for accurate and efficient identification and classification of shale bedding structure based on deep learning according to claim 1 is characterized in that: The bedding identification network model includes: a complex bedding identification network model and a simple bedding identification network model; The complex bedding recognition network model is constructed based on the Swin Transformer network model and is used for the recognition of laminar and layered shale bedding structures; The simple bedding recognition network model is constructed based on the Swin Transformer network model and is used for the recognition of massive shale bedding structure; Among them, the Swin Transformer network model uses Swin-L as the benchmark network. The Swin-L network model for constructing the complex bedding recognition network model selects high feature dimension and number of attention heads, and sets the size of the attention layer window according to the hybrid strategy. The Swin-L network model for constructing the simple bedding recognition network model selects low feature dimension and number of attention heads, and the window for attention layer processing is set to a medium window.

6. The method for accurate and efficient identification and classification of shale bedding structure based on deep learning according to claim 1 is characterized in that: The layer recognition network model adopts a total loss function including cross entropy loss and IOU loss; The total loss function is: TotalLoss=α·CE+β·IoU Loss In the formula, y is the true label; is the predicted output of the model; y i is the true label of the i-th sample; is the predicted probability of the i-th sample; N is the number of samples; A is the true area; B is the predicted area; α is the weight of the cross loss function; β is the weight of the IoU loss function.

7. The method for accurate and efficient identification and classification of shale bedding structure based on deep learning according to claim 1 is characterized in that: Before obtaining the segmented electrical imaging log map, the following steps are required: Filling the blank strips in the electrical imaging logging image, performing threshold segmentation, and outputting a binary image; The binary image is filtered using a Sobel operator to obtain the pre-processed electrical imaging logging image.

8. The method for accurate and efficient identification and classification of shale bedding structure based on deep learning according to claim 1 is characterized in that: The recognition results are analyzed using image morphology to obtain the final classification results, including: Image morphology is used to detect the edges of the identified beddings, calculate the thickness of each bedding and count the number; According to the thickness and number of each bedding, the bedding ratio and bedding density are calculated, wherein the bedding ratio is the ratio of the total thickness of the bedding in the lithology area to the thickness of the lithology, and the bedding density is the ratio of the number of beddings in the lithology area to the thickness of the lithology; The bedding ratio and bedding density are compared with a second preset classification standard to obtain the final classification result.

Images