A comprehensive identification method of tight sandstone fracture logging combined with deep learning
By combining deep learning and multiple recognition methods, logging data is used for lithology identification and crack identification, the problem of fracture identification in tight sandstone reservoirs is solved, the identification accuracy and accuracy are improved, and more effective exploration and development are supported.
Patent Information
- Application Number
- CN202210150674.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-02-18
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2042-02-18
AI Technical Summary
The prior art is difficult to accurately identify and predict fractures in tight sandstone reservoirs, especially in environments with strong heterogeneity, small pores and thin throats. Traditional well logging methods are greatly affected by environmental factors, making it difficult to effectively identify sandstone fractures. .
Using a method combined with deep learning, fractures are identified in parallel through logging data preprocessing, lithology recognition based on LightGBM and multiple deep learning models (comprehensive probability index method, wavelet transformation, Gaussian Naive Bayes discrimination method, CNN-LSTM model). Use lithologic constraint-weighted comprehensive identification results to reduce human error and environmental impact.
It improves the accuracy and accuracy of crack identification in tight sandstone reservoirs, and can perform subdivided lithology division and crack identification in the absence of core and well recording data, supporting more effective exploration and development.
Smart Images

Figure CN114638255B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of oil and gas geological exploration, and in particular to a comprehensive identification method for tight sandstone fracture logging combined with deep learning. Background Art
[0002] With the development of the oil industry, tight sandstone reservoirs, as a type of unconventional reservoir with huge resource potential, have become one of the hot spots for exploration and development. Tight sandstone reservoirs are generally characterized by strong heterogeneity, small pores, fine throats, and poor pore-throat connectivity. Since tight sandstone reservoirs are generally dense and brittle, cracks are easily generated under the influence of tectonic stress. The existence of cracks can not only serve as effective oil and gas migration channels, but also communicate isolated sand bodies and improve matrix permeability, thereby forming effective storage space. Therefore, the precise identification and distribution prediction of cracks are the key to the exploration and development of fractured tight sandstone reservoirs.
[0003] Among the existing research methods, the most direct and reliable method is core observation and description, which can directly obtain characteristic parameters such as fracture occurrence, filling, and oil content. However, due to the actual conditions, the drilling coring section is very limited, and it is difficult to generalize the interpretation results vertically and horizontally. At present, the most effective method for logging fracture identification is imaging logging technology. High-resolution imaging data can extract parameters such as fracture inclination, dip, length, and opening. However, due to the high cost of imaging logging and the low coverage of imaging logging in a region, it is necessary to accurately identify and quantitatively evaluate the fracture section on conventional logging, so as to realize regional "point-line-surface" fracture identification and prediction research. However, due to the strong heterogeneity of fracture-type tight sandstone reservoirs, most of which are thin interlayers of sand and mudstone, the logging responses of different lithologies are similar, and the resolution of conventional logging is not high (generally more than 10 cm), which is easily affected by environmental factors (such as fillers, mud, etc.). The fracture response characteristics of a single curve are generally not obvious. Therefore, how to eliminate environmental influences and correctly identify sandstone fractures is a major problem in traditional logging interpretation. Summary of the invention
[0004] The purpose of the present invention is to provide a comprehensive identification method for tight sandstone fractures combined with deep learning in response to the existing technical status, to realize the lithology prediction of tight sandstone reservoirs on the basis of standardized logging curves, and to comprehensively weightedly identify fractures by combining deep learning, wavelet transform and other means under the constraints of lithology prediction.
[0005] In order to achieve the above object, the present invention adopts the following technical scheme:
[0006] A comprehensive identification method for tight sandstone fractures combined with deep learning, comprising the following steps:
[0007] (1) Logging data preprocessing
[0008] Well logging data preprocessing includes denoising and standardizing the original well logging curve. First, adaptive wavelet multi-scale denoising is performed to reduce the noise in the well logging curve while retaining the curve details as much as possible. The mean square error and signal-to-noise ratio parameters are used to evaluate the denoising performance; the well logging curve is standardized on the denoising result. It is necessary to calibrate the "macro standard layer" on the input well logging curve and perform fast Fourier transform on the well logging curve to extract the spectrum information, low-pass filter the spectrum information to obtain the low-frequency component of the curve, and perform translation correction and scale correction on the low-frequency component, and scale correction on the high-frequency component, so as to achieve the standardization of the well logging curve.
[0009] (2) Lithology identification based on LightGBM framework
[0010] As a preferred solution for the lithology classification method of tight sandstone reservoirs based on LightGBM, it is divided into M0 features according to the coverage of characteristic parameters: spontaneous potential (SP), natural gamma (GR), medium induction resistivity (LIM), deep induction resistivity (ILD), eight lateral resistivity (LL8), acoustic time difference (AC), mud content (SH); M1 features: density logging (DEN), compensated neutron (CNL); M2 features: photoelectric absorption cross-section index (PE), array induction resistivity (HT01, HT02, HT03, HT06, HT09, HT12). The lithology classification types of tight sandstone reservoirs include fine sandstone, siltstone, muddy siltstone, silty mudstone, and mudstone.
[0011] The training process inputs the denoised, standardized and normalized logging series and the true classification labels as the training set, trains and outputs three feature models: M0, M0+M1, M0+M1+M2, and the evaluation indicator of training is multi-classification cross entropy.
[0012] The LightGBM prediction process inputs a single-well logging series that has been denoised, standardized, and normalized with standard deviation, performs lithology prediction based on the input single-well logging series category loading model, and outputs lithology classification results and classification probabilities to constrain subsequent fracture identification results.
[0013] (3) Comprehensive crack identification combined with deep learning
[0014] The comprehensive identification of fractures combined with deep learning includes four parallel methods: fracture identification based on comprehensive probability index method, fracture identification based on one-dimensional multi-scale wavelet decomposition, fracture identification based on Gaussian Naive Bayes discriminant method, and fracture identification based on CNN-LSTM model. To simplify the process, the identification results are divided into two categories: fractures and non-fractures (including filled fracture sections without logging response, fracture-less sections, etc.). In order to ensure that the algorithm has high accuracy and versatility, the identification results of the four methods should be weighted and integrated according to the actual situation. As the preferred solution for comprehensive identification of fractures combined with deep learning, the weight ratio of the four methods is selected as 1:2:2:4.
[0015] First, crack identification is based on the comprehensive probability index method, including the curve change rate method, the resistivity intrusion correction difference ratio method, the crack coefficient method, the skeleton index ratio method, and the three-porosity ratio method. The above prediction results are binarized according to the threshold and weighted averaged to obtain the comprehensive probability index of cracks.
[0016] Second, based on the one-dimensional multi-scale wavelet decomposition for fracture identification, triple wavelet transform is performed on the logging curve to extract high-frequency spectrum features and convert them into high-frequency spectrum energy, and the degree of fracture development is evaluated by the spectrum threshold.
[0017] Third, based on the Gaussian Naive Bayes discriminant method for fracture identification, the training process constructs a training set to obtain a discriminant model. The training samples come from core observation and description results, actual drilling record results, imaging logging interpretation, and manual annotation of fracture segments (coded as 1) and non-fracture segments (coded as 0). As the preferred solution for fracture identification based on the Gaussian Naive Bayes discriminant method, the selected logging curves mainly include acoustic time difference (AC), medium induction resistivity (ILM), deep induction resistivity (ILD), eight lateral resistivity (LL8), density logging (DEN), and compensated neutron (CNL). The training and prediction processes perform range normalization on the data. The prediction process inputs the logging series of a single well and outputs the binary classification probability of the entire target layer segment.
[0018] Fourth, for fracture identification based on the CNN-LSTM model, the training set composition and acquisition method are the same as the above method, the loss function is binary cross entropy, the training process and prediction process use range normalization processing for the well logging series input to the network, and the training process outputs the optimal iterative model. The prediction process inputs a single well logging series and outputs the binary classification probability of the entire target layer.
[0019] (4) Result output
[0020] The four fracture identification probabilities, namely fracture identification based on comprehensive probability index method, fracture identification based on one-dimensional multi-scale wavelet decomposition, fracture identification based on Gaussian Naive Bayes discriminant method and fracture identification based on CNN-LSTM model, are weighted and integrated at a ratio of 1:2:2:4 and multiplied with the classification probability of the sandstone segment of LightGBM to eliminate the influence of the mudstone segment on the prediction results. The comprehensive fracture identification probability based on lithology constraints is obtained, and finally the part with the comprehensive fracture identification probability greater than the set threshold is output as the fracture development segment.
[0021] The beneficial effects of the present invention are:
[0022] 1. The present invention uses fast Fourier transform to extract the spectral features of logging curves for standardization, which can avoid the systematic error of artificially designated standard wells to a certain extent; based on LightGBM, the lithology division of tight sandstone reservoirs is carried out, the training speed is fast and the accuracy is high, and the response model of lithology and multiple logging parameters is obtained to constrain the subsequent identification results; under the constraint of fine lithology division, fracture identification based on comprehensive probability index method, wavelet transform, Gaussian Naive Bayes discrimination, and CNN-LSTM is used, and multiple methods are used in parallel to evaluate the degree of fracture development, and the identification results are weighted to obtain better identification results.
[0023] 2. For well sections lacking core and logging data, the present invention can use the logging data as the basis and the lithology classification model to subdivide the lithology of tight sandstone reservoirs, and realize comprehensive identification of fractures based on this constraint, thereby improving the accuracy of fracture identification and facilitating the next step of exploration and development. BRIEF DESCRIPTION OF THE DRAWINGS
[0024] Figure 1 It is a technical flow diagram of the comprehensive identification method of tight sandstone fracture logging of the present invention;
[0025] Figure 2 This is a flowchart of well logging curve denoising based on multi-scale wavelet transform;
[0026] Figure 3 Flowchart for Fast Fourier Transform-based Well Log Normalization
[0027] Figure 4 It is a schematic diagram of the CNN-LSTM network model structure of the present invention.
[0028] Notes: DETAILED DESCRIPTION
[0029] In order to make the technical methods and advantages of the present invention more clearly understood, a detailed description is given below based on specific embodiments in conjunction with the accompanying drawings.
[0030] As we all know, the Ordos Basin spans five provinces in China, namely Shaanxi, Gansu, Ningxia, Inner Mongolia and Shanxi. The basin stretches from Yinshan Mountain and Daqing Mountain in the north to Qinling Mountain in the south, from Liupan Mountain and Helan Mountain in the west to Lvliang Mountain in the east, with a total area of 33×104km 2 . The basin is located in the central and western part of the North China Craton Basin. It is a large composite craton basin with overall stable settlement, multiple tectonic systems, multiple evolutionary stages, multiple sedimentary systems, and multiple prototype basins. The Mesozoic strata in the Ordos Basin are from bottom to top the Triassic, the middle and lower Jurassic, and the lower Cretaceous. Among them, the Triassic strata are river-delta-lake deposits. The lake and delta deposits of the Upper Triassic Yanchang Formation are the thickest and the most widely distributed. The lithology is interlayered with gray, grayish white powder-fine sandstone and black, grayish black mudstone. The Upper Triassic Yanchang Formation is divided into 10 oil layer groups from bottom to top according to the changes in lithology, and is divided into Chang 10-Chang 1 from bottom to top. Among them, due to tectonic movement, the western edge of the basin, especially the southwestern edge, has been uplifted, and the top surface of the Yanchang Formation has different degrees of erosion. The southwestern edge of the basin is generally missing the Chang 1-Chang 3 strata. The Chang 8 oil layer group is the target layer section in this embodiment, with an average thickness of about 100m. The Chang 8 Member was deposited in a lake and delta environment. The lithology is mainly composed of interbedded combinations of gray-white and gray-green medium-fine sandstone, siltstone, and black mudstone. Its permeability is less than 1×10 -3 μm 2 , which is a low-permeability to ultra-low-permeability reservoir; the stratum is overlyed by the Zhangjiatan Shale of Chang 7 Member (KT marker layer), which is mainly black shale and oil shale.
[0031] The embodiment of the present invention is implemented on 20 exploration wells and 5 horizontal wells of the Triassic Yanchang Formation in the Binchang area on the southwestern edge of the Ordos Basin, including core, logging, mud logging and other data. The exploration well logging data includes natural gamma (GR), acoustic time difference (AC), spontaneous potential (SP), density logging (DEN), compensated neutron (CNL), medium induction resistivity (ILM), deep induction resistivity (ILD), eight lateral resistivity (LL8), and shale content (SH) curve data. Some well sections contain photoelectric absorption cross-section index (PE), array induction resistivity (HT01, HT02, HT03, HT06, HT09, HT12), deep lateral resistivity (LLD), and shallow lateral resistivity (LLS) curve data; horizontal well logging data includes GR, AC, SP, ILM, ILD, LL8, and SH curve data.
[0032] 1. Well logging data preprocessing
[0033] Well logging data preprocessing refers to the denoising and standardization of well logging curves. Specifically:
[0034] 1. Denoising of well logging curves based on multi-scale wavelet transform
[0035] like Figure 2The figure shows the flow chart of logging curve denoising based on multi-scale wavelet transform. The input logging curve is decomposed by triple wavelet, and wavelet reconstruction is performed after the nonlinear soft threshold is applied to the high-frequency wavelet coefficients. The reconstructed curve is output as the denoising result. The calculation formula is as follows:
[0036] η λ (ω) = sgn(ω)(|ω| - λ);
[0037]
[0038] Where: ω represents the high-frequency wavelet coefficient, η λ (ω) represents the result after filtering, sgn(ω) is the sign function, indicating the positive or negative, indicating that the sign of the value is the same as ω, λ is the soft threshold, n is the total number of signal sampling points, and σ is the standard deviation of the signal.
[0039] The embodiment uses mean square error and signal-to-noise ratio to evaluate the denoising capabilities of different wavelet bases. The embodiment of the present invention selects the sym6 wavelet as the optimal denoising wavelet base based on the comprehensive results of the evaluation indicators.
[0040]
[0041]
[0042] Where: MSE and SNR represent mean square error and signal-to-noise ratio, N is the total number of signal sampling points, y i Signal before denoising, is the signal after denoising.
[0043] 2. Logging curve standardization based on fast Fourier transform
[0044] like Figure 3 The figure shows a flow chart of the standardization of logging curves based on fast Fourier transform. In the embodiment of the present invention, the Zhangjiatan shale at the bottom of Chang 7 section is characterized by obvious and widespread distribution, so the mud shale at the bottom of Chang 7 section is included in the de-noising and the standardized characteristic parameters are obtained. In the embodiment of the present invention, the logging curve is first subjected to fast Fourier transform to obtain the spectrum, and then low-pass filtering is performed to obtain the low-frequency component of the spectrum. The low-frequency component is subjected to inverse Fourier transform to obtain the low-frequency component in the depth domain of the logging data, which is subtracted from the original curve to obtain the high-frequency component in the depth domain of the logging data; the low-frequency component in the depth domain is subjected to proportional and translation correction, and the standardized curve is obtained after the proportional correction of the high-frequency component in the depth domain. The correction formula is:
[0045]
[0046]
[0047] Where: c l(z) is the low-frequency component to be standardized, c′ l (z) is the standardized low-frequency component, c h (z) is the high frequency component to be standardized, c′ h (z) is the normalized high frequency component. C′ max and C′ min are the maximum and minimum standard values of the low-frequency component of the logging parameter, C max and C min are the maximum and minimum values of the low-frequency components to be standardized.
[0048] The embodiment screened multiple well logging data evenly distributed, and set the low-pass filter cutoff frequency of the embodiment to 0.2 according to the spectrum characteristics to ensure that the acquired low-frequency components are not mixed with too many high-frequency components as much as possible and contain sufficient characteristics. The maximum and minimum average values of the low-frequency components are used as standard values, as shown in Table 1.
[0049] Table 1 Standard values of main logging curves in the embodiment
[0050]
[0051] 2. Lithology Identification Based on LightGBM Framework
[0052] LightGBM is a new boosting framework model released by Microsoft in 2017. The algorithm introduces two new technologies (gradient unilateral sampling technology and independent feature merging technology) and one improvement (histogram algorithm) based on the traditional gradient boosting decision tree (GBDT): the gradient unilateral sampling technology removes data with too small gradient to avoid the influence of the long tail of low gradient; the independent feature merging technology bundles mutually exclusive features to achieve feature dimensionality reduction; the histogram algorithm finds the optimal segmentation point to reduce memory consumption.
[0053] The GBDT method uses a greedy algorithm to define leaf nodes during model construction. The optimization goal is to determine the tree so that the leaf node gain of the entire tree is maximized after segmentation. There are iterative processes and optimization goals:
[0054]
[0055]
[0056] Where: represents the gap between the predicted value and the actual label, λ and ω are regularization parameters, T is the number of leaf nodes in each tree, and γ is the complexity cost of introducing new nodes; f t (x i) represents the weight of the leaf where the i-th sample is classified in the t-th tree, To optimize the objective function.
[0057] Ignore the constant term and perform Taylor expansion on the optimization objective function:
[0058]
[0059]
[0060] Where: g i and h i is the partial derivative, w i is the weight of the leaf node, I j is the sample instance on the leaf node j.
[0061] w for the upper form i Find the partial derivative and set it to 0, then we have Substituting back the optimization objective function, we get:
[0062]
[0063]
[0064] Considering the loss function of segmenting leaf nodes, the maximum expectation should be achieved after segmentation:
[0065]
[0066] Where: represents the expected segmentation, γ represents the segmentation cost of adding a new node, and the three items in the formula represent the expectation of the left subtree, the expectation of the right subtree, and the expectation when not segmented.
[0067] Preferably, the lithology classification types of the embodiment of the present invention include fine sandstone, siltstone, muddy siltstone, silty mudstone, and mudstone, and the characteristic parameters are divided into M0 characteristics according to the coverage of the logging series: spontaneous potential (SP), natural gamma (GR), deep induction resistivity (ILD), medium induction resistivity (ILM), eight lateral resistivity (LL8), acoustic time difference (AC), mud content (SH); M1 characteristics: density logging (DEN), compensated neutron (CNL); M2 characteristics: photoelectric absorption cross-section index (PE), array induction resistivity (HT01, HT02, HT03, HT06, HT09, HT12).
[0068] Model training is performed on Python 3. The main packages and libraries used include sklearn, lightgbm, pandas, and numpy. The training process inputs the well logging series and true classification labels of the training samples, and outputs three tree models containing M0, M0+M1, and M0+M1+M2 features. The key parameters of the model are shown in Table 2:
[0069] Table 2 Main parameters of LightGBM model in Example
[0070]
[0071] The evaluation index of the model is the multi-classification cross entropy loss function:
[0072]
[0073] Where: M is the number of categories; y ic is a sign function, which takes 1 if the true category of sample i is equal to c, otherwise it takes 0; p ic is the predicted probability that sample i belongs to category c.
[0074] The prediction process inputs the single well logging curve series that has been preprocessed, selects the model according to the coverage of the logging curve series, and outputs the lithology classification results and classification probability.
[0075] The training and prediction process uses standard deviation normalization (Z-score) to balance the contribution of each parameter to the weight. The processed data follows a distribution with a mean of 0 and a standard deviation of 1.
[0076]
[0077] Where: σ is the standard deviation of the original data, is the original data mean.
[0078] 3. Comprehensive prediction of cracks combined with deep learning
[0079] The present invention simplifies fracture identification into two categories, namely fracture zones and non-fracture zones. The fracture zones are effective fracture zones with obvious logging responses, while the non-fracture zones include closed fracture zones without obvious logging responses and zones without fracture development.
[0080] 1. Crack identification based on comprehensive probability index method
[0081] The present invention refers to the previous research results of crack identification parameters and constructs a comprehensive probability index using the curve change rate method, resistivity intrusion correction difference ratio method, crack coefficient method, three-porosity ratio method, and skeleton index ratio method.
[0082] The curve change rate method generally selects logging methods with higher vertical resolution, such as AC, DEN, CNL, GR, etc., and generally uses the three-point or five-point method to calculate the curve change rate. The presence of fractures will increase the curve change rate. The calculation formula is as follows:
[0083] ΔX i =(|X i-i -X i |+|X i+1 -X i |) / 2;
[0084]
[0085] Where: ΔX i Indicates the rate of change of the curve, X i is the i-th sampling point of a logging curve.
[0086] Resistivity intrusion correction difference ratio method, the calculation formula is as follows:
[0087]
[0088] R t =2.589R LLD -1.589R LLs :
[0089] Where: R t is the true resistivity of the formation after invasion correction, R LLD , R LLS is the deep and shallow lateral resistivity. When the formation is a fractured oil and gas layer, R TC >0; when the stratum is a water layer or a dense stratum, R TC ≈0.
[0090] The calculation formula of crack coefficient method is as follows:
[0091]
[0092] Where: V p 、V pma Indicates the longitudinal wave velocity of the stratum and rock skeleton, Δt ma , Δt p Indicates the acoustic time difference between the rock skeleton and the stratum. The larger the S, the better the integrity of the rock; the smaller the S, the more serious the rock damage and the more cracks are developed.
[0093] Three porosity ratio method, the calculation formula is as follows:
[0094]
[0095]
[0096]
[0097] Where: Δt ma , Δt f , Δt are the acoustic time differences of rock skeleton, fluid and formation respectively; φ S ,φ D ,φ N ,φ t are acoustic wave, density, neutron, and total porosity respectively; ρ ma , f , b are the densities of rock skeleton, stratum and fluid respectively; Φ ma , Φ Nf , Φ N are the neutron fluxes of rock skeleton, fluid and stratum respectively. The more developed the secondary porosity of the stratum, the higher the R p The bigger.
[0098] Skeleton index ratio method: through the response of the fracture to density logging and acoustic logging, the intersection of acoustic skeleton index and density skeleton index is calculated to identify the fracture. The calculation formula is as follows:
[0099]
[0100]
[0101]
[0102] Where: φ s ,φ D is the acoustic wave, density porosity, Δt f , Δt is the time difference between fluid and formation sound waves; S ma , D ma is the acoustic skeleton index and density skeleton index. When cracks develop in the stratum, S ma >D ma , that is, X ma >0.
[0103] The probability comprehensive index CWP is obtained by comprehensive evaluation of these crack parameters. The calculation formula is as follows:
[0104]
[0105]
[0106] Where: f(x) represents the activation function, a is the critical value, x < a means no abnormality, no cracks, and the critical values of different crack parameters need to be selected according to regional characteristics. f(x) normalizes the results of the above five methods to 0 or 1, and the probability comprehensive index CWP weights the results to [0, 1].
[0107] The method inputs the preprocessed logging series: AC, LLD, LLS, DEN, CNL, etc., and outputs the probability comprehensive index according to the preset critical value.
[0108] 2. Crack identification based on one-dimensional multi-scale wavelet decomposition
[0109] Taking acoustic time difference logging as an example, the presence of fractures leads to enhanced formation heterogeneity, and the acoustic time difference curve shows a small oscillation, which is usually not significant on the original curve. Therefore, the present invention uses the db4 wavelet basis to perform triple multi-scale wavelet decomposition on the logging curve, extracts and reconstructs the high-frequency signal D3, and at this time, the fracture-developed layer and other layers with strong heterogeneity show high-frequency oscillations on the reconstructed curve. This method can be implemented on python3, and the main packages and libraries used include pywavelets and numpy.
[0110] This method inputs a single well logging curve (generally an AC curve) and outputs a high-frequency signal segment with an amplitude greater than the threshold as a fracture development segment according to a preset threshold.
[0111] 3. Crack identification based on Gaussian Naive Bayes discriminant method
[0112] Naive Bayes is to calculate the posterior probability distribution of a given sample to be classified and then select the class with the maximum posterior probability as the class to which the feature belongs. Naive Bayes is based on two basic assumptions: the features are independent of each other and each feature is equally important.
[0113] The Bayesian classifier algorithm assumes that the sample priors of each label follow a standard normal distribution. Its principle is relatively simple to implement, with low computational complexity, and can obtain better classification results. However, its disadvantages are that it is sensitive to the input training data, and the classifier assumes that the features are independent of each other, which may not be applicable to scenarios with strong correlation between features; assuming that the features satisfy the Gaussian normal distribution will result in a loss in accuracy.
[0114] The present invention implements the method on python3, and the main packages and libraries used include sklearn and numpy. The logging curve parameters used include: AC, LL8, ILD, ILM, CNL, DEN. The training samples come from core observation and description results, actual drilling record results, imaging logging interpretation, and manual annotation of fracture segments (coded as 1) and non-fracture segments (coded as 0).
[0115] The training process inputs the well logging series and true labels of the training samples and outputs the classification model; the prediction process inputs the preprocessed single well logging series and outputs the binary classification probability of the entire target layer segment.
[0116] In order to balance the contribution of each parameter to the weight, both the training and prediction processes use range normalization to normalize each parameter to [0, 1]. The calculation formula is as follows:
[0117]
[0118] Where: x max 、x min Indicates the maximum and minimum values of the sample data.
[0119] In actual situations, it is easy to have extremely high peak values in LL8 logging due to the influence of crude oil. Directly using the above formula will cause most of the data to be compressed to near 0. Therefore, the maximum and minimum values can be restricted. Take the acoustic time difference logging as an example:
[0120]
[0121] Where: AC max The average AC value of the most developed section of the tight sandstone fracture zone in the sample, AC min It is the average AC value of typical non-fracture zone of tight sandstone.
[0122] 4. Crack identification based on CNN-LSTM model
[0123] The basic structure of the network is as follows Figure 4 As shown in the figure, the CNN network includes three one-dimensional convolutional layers (Conv1D) with a step size of 3 and one maximum pooling layer (MaxPooling1D). Its main purpose is to use different convolution templates to extract effective features from the original data and use the maximum pooling layer to reduce the parameters that need to be trained.
[0124] The LSTM network consists of a basic LSTM layer and a subsequent fully connected layer, whose main function is to use the extracted features to achieve prediction.
[0125] The CNN network performs convolution kernel and pooling operations to extract data features. Its weight sharing effectively reduces the number of weights and simplifies the model. The AC, LL8, ILD, ILM, CNL, and DEN data are processed into sequence data and input into the convolution network to extract feature maps and perform maximum pooling operations on each feature map. The feature maps after pooling operations are directly sent to the subsequent LSTM network.
[0126] LSTM is an improvement on RNN network. Its basic units include forget gate, input gate and output gate. The forget gate accepts the current input x t and the previous hidden output h t-1 And output the memory state of [0, 1] to update the cell memory information C t-1 ;The input gate accepts the current input x t and the previous hidden output ht-1 The new candidate cell information is obtained through the sigmoid and tanh layers, and the new candidate cell information is added to the cell memory information updated by the forget gate to obtain the current cell memory information C t ; After updating the cell state, the output gate needs to be based on the input h t-1 and x t To determine the state characteristics of the output cells, the calculation formula is as follows:
[0127] f t =σ(W f *[h t-1 , x t ]+b f );
[0128] i t =σ(W i *[h t-1 , x t ]+b i ):
[0129]
[0130]
[0131] o t =σ(W o *[h t-1 , x t ]+b o );
[0132] h t =o t *tanh(C t );
[0133] Where: f t 、i t , o t 、h t , C t are the forget gate, input gate, output gate, input node, output gate, implicit output, and state of the state unit respectively; W f , W i , W c , W o is the corresponding weight matrix; b f , b i , b c , b o is the bias term; σ represents the sigmoid activation function.
[0134] The subsequent fully connected layer (Dense) maps the 64-dimensional LSTM output to a single output. Since the problem is simplified to a two-classification problem, only one neuron is used to represent the output.
[0135] In order to balance the contribution of each parameter to the weight, both the training process and the prediction process need to use range normalization on the samples to normalize each parameter to [0, 1].
[0136] The training process inputs the well logging series of training samples: acoustic time difference (AC), deep induction resistivity (ILD), medium induction resistivity (ILM), eight lateral resistivity (LL8), density logging (DEN), compensated neutron (CNL) and true classification labels: fracture (coded as 1), non-fracture (coded as 0), and the loss function is binary cross entropy, and the calculation formula is as follows:
[0137] Loss=-(y true *log(y pred )+(1-y true )*log(1-y pred );
[0138] Where: y true is the true label, y pred is the predicted probability.
[0139] The prediction process inputs the preprocessed single well logging series and outputs the binary classification probability of fractures in all target layers.
[0140] 4. Result Output
[0141] The four fracture identification probabilities, namely fracture identification based on comprehensive probability index method, fracture identification based on one-dimensional multi-scale wavelet decomposition, fracture identification based on Gaussian Naive Bayes discriminant method and fracture identification based on CNN-LSTM model, are weighted and integrated at a ratio of 1:2:2:4 and multiplied with the classification probability of the sandstone segment of LightGBM to eliminate the influence of the mudstone segment on the prediction results. The comprehensive fracture identification probability based on lithology constraints is obtained, and finally the part with the comprehensive fracture identification probability greater than the set threshold is output as the fracture development segment.
[0142] In the embodiment of the present invention, the threshold is set to 0.75.
[0143] Of course, the above are only preferred embodiments of the present invention, and are not intended to limit the scope of use of the present invention. Therefore, any equivalent changes made to the principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A comprehensive identification method for tight sandstone fracture logging combined with deep learning, characterized by: The following steps are involved: Well logging data preprocessing, including well logging curve denoising based on multi-scale wavelet transform and well logging curve normalization based on fast Fourier transform; The steps of logging curve denoising based on multi-scale wavelet transform are as follows: The input logging curve is decomposed into triple wavelets using the sym6 wavelet basis, and the high-frequency wavelet coefficients are acted on by nonlinear soft threshold and then the wavelet reconstruction is performed. The reconstructed curve is output as the denoising result. The calculation formula is as follows: or λ (ω)=sgn(ω)(|ω|-λ); Where: ω represents the high-frequency wavelet coefficient, η λ (ω) represents the result after filtering, sgn(ω) is the sign function, indicating the positive or negative, indicating that the sign of the value is the same as ω, λ is the soft threshold, n is the total number of signal sampling points, and σ is the standard deviation of the signal; The steps of logging curve standardization based on fast Fourier transform are: Input the denoised logging curve, first perform fast Fourier transform on the logging curve to obtain the spectrum, then perform low-pass filtering to obtain the low-frequency component of the spectrum, and obtain the low-frequency component of the depth domain of the logging data after inverse Fourier transform. Subtract it from the original curve to obtain the high-frequency component of the depth domain of the logging data; perform proportional and translation correction on the low-frequency component in the depth domain, and perform proportional correction on the high-frequency component in the depth domain. The correction results are added to obtain the standardized curve. The correction formula is as follows: Where: c l (z) is the low-frequency component to be standardized, c′ l (z) is the standardized low-frequency component, c h (z) is the high frequency component to be standardized, c′ h (z) is the normalized high frequency component, C′ max and C′ min are the maximum and minimum standard values of the low-frequency component of the logging parameter, C max and C min are the maximum and minimum values of the low-frequency components to be standardized; Lithology identification, based on the LightGBM framework; The steps to implement lithology identification based on the LightGBM framework are: The training process divides the characteristic parameters into M0 characteristics according to the coverage of the logging series: natural potential, natural gamma, deep induction resistivity, medium induction resistivity, eight lateral resistivity, acoustic time difference, and mud content; M1 features: density logging, compensated neutron; M2 features: photoelectric absorption cross-section index, array induction resistivity, and the types of tight sandstone reservoir lithology are divided into fine sandstone, siltstone, muddy siltstone, silty mudstone, and mudstone; the training process inputs the logging series and true classification labels of the training samples, and outputs three tree models containing M0, M0+M1, and M0+M1+M2 features; the evaluation index of the training is the cross entropy function, and the formula is as follows: Where: M is the number of categories, N is the number of samples; y ic is a sign function, which takes 1 if the true category of sample i is equal to c, otherwise it takes 0; p ic is the predicted probability that sample i belongs to category c; The prediction process inputs a single well logging curve series, performs lithology prediction based on the coverage loading model of the single well logging curve series, and outputs lithology classification results and classification prediction probability; In order to balance the contribution of each parameter to the weight, the training and prediction processes need to normalize the samples by standard deviation. The processed data obeys the distribution with a mean of 0 and a standard deviation of 1. The calculation formula is as follows: Where: σ is the standard deviation of the original data, is the original data mean; Comprehensive crack prediction combined with deep learning, including crack identification based on comprehensive probability index method, crack identification based on one-dimensional multi-scale wavelet decomposition, crack identification based on Gaussian Naive Bayes discriminant method and crack identification based on CNN-LSTM model; The steps of crack identification based on the comprehensive probability index method are as follows: The logging curve is decomposed by triple multi-scale wavelet using db4 wavelet basis, and the high-frequency signal D3 is extracted and reconstructed. The high-frequency oscillation part on the output reconstructed curve is regarded as the fracture development section. The steps of crack identification based on Gaussian Naive Bayes discriminant method are: The training process inputs the well logging series of training samples: acoustic time difference, deep induction resistivity, medium induction resistivity, eight lateral resistivity, density logging, compensated neutron and true classification labels: fractures coded as 1, non-fractures coded as 0; iteratively calculates the posterior probability of the samples based on the basic assumptions of Gaussian Bayes, and outputs the classification model; The prediction process inputs the logging series of a single well and outputs the binary classification probability of the entire target layer interval; In order to balance the contribution of each parameter to the weight, both the training process and the prediction process need to first use range normalization on the samples to normalize each parameter to [0, 1]. The calculation formula is as follows: Where: x max 、x min Indicates the maximum and minimum values of sample data; The steps of crack identification based on the CNN-LSTM model are: The basic structure of the model includes three one-dimensional convolutional layers with a step size of 3 to extract features and one maximum pooling layer for feature dimensionality reduction, one basic LSTM unit and two fully connected layers to achieve binary classification; In order to balance the contribution of each parameter to the weight, both the training process and the prediction process need to first use range normalization on the samples to normalize each parameter to [0, 1]; The training process inputs the well logging series of training samples: acoustic time difference, deep induction resistivity, medium induction resistivity, eight lateral resistivity, density logging, compensated neutron and true classification labels: fractures coded as 1 and non-fractures coded as 0. The loss function is the binary cross entropy, and the calculation formula is as follows: Loss=-(and true *log(y pred )+(1-and true )*log(1-y pred ); Where: y true is the true label, y pred is the predicted probability; The prediction process inputs the logging series of a single well and outputs the binary classification probability of the entire target layer interval; Result output: output the comprehensive identification probability of tight sandstone fractures based on lithology constraints; The steps for outputting the results are: The four fracture identification probabilities, namely fracture identification based on comprehensive probability index method, fracture identification based on one-dimensional multi-scale wavelet decomposition, fracture identification based on GaussianNaiveBayes discriminant method and fracture identification based on CNN-LSTM model, are weighted and integrated at a ratio of 1:2:2:4 and multiplied with the classification probability of the sandstone segment of LightGBM to eliminate the influence of the mudstone segment on the prediction results. The comprehensive fracture identification probability based on lithology constraints is obtained, and finally the part with the comprehensive fracture identification probability greater than the set threshold is output as the fracture development segment.
Citation Information
Patent Citations
Method for identifying compact sandstone cracks by wavelet high-frequency attributes
CN109781862A
Systems and Methods for the Determination of Lithology Porosity from Surface Drilling Parameters
US20210293139A1
Method and apparatus for identifying concrete crack based on video semantic segmentation technology
US20210319547A1
Method for predicting reservoir lithogenous phase using geology and logging information
WO2016161914A1
Cited By
Multi-scale crack identification method fusing logging data and wavelet transform
CN122260525A
