A method for identifying fractures in a well based on geophysical prior constraints and conditional hybrid expert small sample logging

Abstract

The application discloses a condition mixed expert small sample well logging fracture identification method based on geophysical prior constraints, S1: a fracture identification basic sample set is constructed; S2: multi-source features are extracted based on the basic sample set; S3: a positive sample-unlabeled learning soft label is constructed; S4: a time sequence sample sequence along a well depth direction is constructed, a condition mixed expert time sequence convolution network constrained by geophysical priors is established, and fracture development probabilities of each depth point are output; S5: the fracture development probabilities are subjected to probability calibration and conformal prediction, and calibrated fracture development probabilities and uncertainty intervals are output; and S6: a fracture development interval identification result is formed. Through the construction of the positive sample-unlabeled learning soft label, the dependence on large-scale artificial labeling samples is reduced, and the time sequence sample sequence along the well depth direction is constructed, the condition mixed expert time sequence convolution network constrained by the geophysical priors is established, and therefore the interval continuity and local abnormal combination features of the fracture type reservoir can be better identified.

Description

Technical Field

[0001] This invention relates to the field of oil and gas exploration and development technology, and in particular to a method for identifying fractures in well logging based on a hybrid expert small sample method with geophysical prior constraints. Background Technology

[0002] In the exploration and development of oil and gas fields, especially in unconventional reservoirs such as low-porosity and low-permeability sandstone, fractured carbonate rocks, and shale oil and gas, the degree of natural fracture development directly affects reservoir permeability, fracturing effect, and single-well productivity. Rapidly and reliably identifying fractured intervals based on well logging data is a key technical step in oil and gas field reservoir evaluation, sweet spot assessment, fracturing interval selection, and well location deployment.

[0003] However, fracture identification in oil and gas fields is highly complex. On the one hand, fracture development is jointly controlled by multiple factors such as tectonic activity, lithological characteristics, rock brittleness, fluid conditions, and diagenesis, resulting in significant differences in logging responses between different blocks and well sections. On the other hand, the appearance of fractures on logging curves is usually not a single attribute anomaly, but a comprehensive response of multiple information such as resistivity, acoustic signature, density, and stratigraphic continuity. Furthermore, in actual oil and gas field production, fracture interpretation data mostly comes from imaging logging, core observation, oil and gas testing, and comprehensive interpretation, which suffers from high annotation costs, small sample sizes, and uneven distribution among wells, further increasing the difficulty of fracture identification and modeling.

[0004] Existing fracture identification methods mainly include empirical threshold discrimination methods, traditional supervised classification methods, and single-attribute anomaly detection methods. Empirical threshold methods are greatly affected by block differences and inter-well differences, and their generalization ability is limited. Traditional supervised classification methods usually rely on a large number of high-quality labeled samples, which is difficult to adapt to the actual situation of scarce fracture samples in oil and gas fields. Although single-attribute anomaly detection methods can reflect local fracture responses, they are difficult to make full use of multi-source logging information and the continuity of the stratigraphic intervals in the well depth direction.

[0005] As machine learning methods are increasingly used in well logging interpretation in oil and gas fields, fracture identification is evolving from experience-based judgment to data-driven modeling. While these methods have the advantage of handling nonlinear and multi-feature problems, they are prone to problems such as overfitting, insufficient cross-well adaptability, and weak interpretability if they lack geological mechanisms and rock physics constraints. This makes it difficult to directly meet the engineering application needs in oil and gas field development. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a conditional hybrid expert small sample well logging fracture identification method based on geophysical prior constraints.

[0007] The objective of this invention is achieved through the following technical solution: a conditional hybrid expert small-sample well logging fracture identification method based on geophysical prior constraints, comprising the following steps:

[0008] S1: Construct a basic sample set for crack identification;

[0009] S2: Extract multi-source features based on the basic sample set from step S1;

[0010] S3: Constructing positive samples—learning soft labels without labeling;

[0011] S4: Construct a time-series sample sequence along the well depth direction, establish a conditional hybrid expert time-series convolutional network constrained by geophysical priors, and output the fracture development probability at each depth point;

[0012] S5: Perform probability calibration and conformal prediction on the crack development probability, and output the calibrated crack development probability and uncertainty range;

[0013] S6: Identification results of fracture-developed layers.

[0014] Preferably, step S1 further includes the following step:

[0015] S11: Follow a uniform depth step size Establish a resampling depth grid:

[0016] ;

[0017] in, For the first Each resampling depth point The starting depth of the target well section. The resampling step size, This represents the total number of sample points after resampling.

[0018] S12: Perform depth domain interpolation to complete each logging curve.

[0019] ;

[0020] in, For the first The logging curve at the resampling depth point The value at that location, The original well logging curve, For interpolation operators;

[0021] S13: Establish fracture sample labels based on fracture interpretation layers.

[0022] ;

[0023] in, For the first Crack labels for each sample point To explain the top depth of the fracture section. The depth of the layer is explained to indicate the fracture.

[0024] Preferably, in step S2, the multi-source features include shale content, comprehensive brittleness index, fracture indicator factor, acoustic multi-scale anomaly, reservoir quality index, fluid indicator factor, and fracture prior index.

[0025] Preferably, step S2 further includes the following step:

[0026] S21: Calculate the shale content index based on the natural gamma curve.

[0027] ;

[0028] in, For the first Shale index of each sample point For the first The natural gamma value of each sample point The minimum natural gamma within the target well section. The maximum natural gamma within the target well section. To prevent extremely small positive numbers with a denominator of zero;

[0029] Calculate shale content,

[0030] ;

[0031] in, For the first Shale content at each sample point;

[0032] S22: Calculate the longitudinal wave velocity based on the acoustic time difference.

[0033] ;

[0034] in, For the first P-wave velocity at each sample point For the first Acoustic transit time at each sample point

[0035] The P-wave and S-wave velocity ratio was obtained by weighting based on shale content:

[0036] ;

[0037] Obtain the transverse wave velocity:

[0038] ;

[0039] in, For the first The ratio of P-wave to S-wave velocity at each sample point This is a reference value for the P-wave and S-wave velocity ratio under low mud content conditions. This is a reference value for the ratio of longitudinal and transverse wave velocities under high mud content conditions. For the first Shear wave velocity at each sample point;

[0040] S23: Calculate the dynamic elastic modulus and dynamic Poisson's ratio based on density, P-wave velocity, and S-wave velocity.

[0041] ;

[0042] in, For the first Dynamic elastic modulus of a sample point For the first Volume density of each sample point For the first The dynamic Poisson's ratio of each sample point;

[0043] S24: Construct a comprehensive fragility index.

[0044] ;

[0045] in, For the first The overall fragility index of each sample point This is the normalized dynamic elastic modulus. The normalized dynamic Poisson's ratio;

[0046] S25: Constructing a crack indicator factor based on the separation of deep resistivity and flushing zone resistivity:

[0047] ;

[0048] in, For the first Crack indicator factor for each sample point For the first Deep resistivity of each sample point For the first Resistivity of the rinsing band at each sample point;

[0049] S26: Construct multi-scale acoustic anomalies based on the multi-scale filtering difference of the acoustic curve.

[0050] ;

[0051] in, For the first Multiscale acoustic anomalies at individual sample points The scale parameter is Smoothing operator, The number of scale layers;

[0052] S27: Construct reservoir quality index, fluid indicator factor, and fracture prior index.

[0053] ;

[0054] ;

[0055] ;

[0056] in, For the first Reservoir quality index of each sample point For the first Fluid indicator factor for each sample point For the first Crack prior index for each sample point The normalized overall fragility index, The normalized deep resistivity, The normalized crack indicator factor. The normalized acoustic multiscale anomaly.

[0057] Preferably, step S3 further includes the following step:

[0058] S31: Quantitative thresholds for shale content, fracture indicator factor, and sonic multiscale anomaly were statistically analyzed using training well samples, and denoted as follows: , and ;

[0059] S32: Construct a seed criterion for high-confidence crack samples

[0060] ;

[0061] in, For the first High-confidence crack seed markers for each sample point, The threshold for shale content, The threshold for the crack indicator factor. The threshold for multi-scale anomalies in acoustic waves;

[0062] S33: Constructing a heuristic soft probability prior to the crack.

[0063] ;

[0064] in, For the first Heuristic soft probability of crack prior for each sample point This represents the normalized shale content. The normalized crack indicator factor. The normalized multiscale anomaly of acoustic waves;

[0065] S34: A positive sample-unlabeled learning method based on multiple bootstrapping is used to train the unlabeled samples in multiple rounds to obtain the average crack probability of the unlabeled samples.

[0066] ;

[0067] in, For the first Positive sample-unlabeled learning probability of each sample point For the number of self-service sampling rounds, For the first Wheel classifier for samples The output probability;

[0068] S35: Construct the final soft label.

[0069] ;

[0070] in, For the first The final soft label for each sample point This is a truncation function.

[0071] Preferably, step S4 further includes the following step:

[0072] S41: Using the current depth point and its predecessors. Construct a time-series sample sequence from each depth point.

[0073] ;

[0074] in, For the first The time-series input matrix corresponding to each sample point For the first Feature vectors of depth points The time window length, The dimension of the input features;

[0075] S42: Multiple temporal convolutional expert networks are used to encode the temporal sample sequence to obtain the feature representations of each expert.

[0076] ;

[0077] in, For the first The feature representation output by the expert network For the number of expert networks;

[0078] S43: Construct a conditional vector based on shale content, brittleness normalized value, fluid indicator factor, lithofacies code, and reservoir quality index.

[0079] ;

[0080] in, For the first The condition vector of each sample point The normalized overall fragility index, As a fluid indicator, For lithofacies coding, This is a reservoir quality index. For conditional mapping functions;

[0081] S44: Construct gating weights based on the output of the auxiliary task.

[0082] ;

[0083] in, For the first The expert gating weight vector for each sample point For the weight matrix of the gated network, For the gated network bias term, For vector concatenation, Output vectors for auxiliary tasks;

[0084] S45: Weighted fusion of multiple expert features based on gating weights.

[0085] ;

[0086] in, For the fused feature representation, For the first The gating weights corresponding to each expert;

[0087] S46: Perform attention calculation on the weighted fused features and output the crack development probability.

[0088] ;

[0089] in, For the first The probability of crack development at each sample point For multi-head attention operators, For classification mapping function, This is the Sigmoid activation function.

[0090] Preferably, step S5 further includes the following step:

[0091] S51: Constructing a weighted cross-entropy loss based on soft labels

[0092] ;

[0093] in, For weighted cross-entropy loss, The total number of samples, It is a soft tag. To predict the probability of cracks;

[0094] S52: Construct auxiliary task loss,

[0095] ;

[0096] in, To help average the loss of the task, Losses due to lithofacies auxiliary missions. For fragile auxiliary tasks, Loss due to fluid-assisted tasks;

[0097] S53: Construct physical saliency constraints, let the set of physical features be... Its gradient proportion is,

[0098] ;

[0099] And define physical constraint penalty terms,

[0100] ;

[0101] in, The proportion of physical feature gradient. For the first Single-sample loss for each sample point For the first The first sample point Each input feature For physical constraint loss, Set a preset physical dependency threshold;

[0102] S54: Construct the total loss function.

[0103] ;

[0104] in, For the total loss, , and These are the weighting coefficients. For gated constraint loss;

[0105] S55: The output probability is calibrated using isotonic regression to obtain the calibrated crack development probability. ;

[0106] S56: Constructing conformal prediction boundaries based on calibration set residuals

[0107] ;

[0108] And output the probability range of crack development.

[0109] ;

[0110] in, To predict the boundary width in a conformal manner, for Quantile operator, This represents the calibrated crack development probability. For the first The confidence interval for the probability of cracks at each sample point.

[0111] Preferably, in step S6, the fracture-developed segments are screened by constructing a fracture sweetness index, whereby the fracture sweetness index is expressed as:

[0112] ;

[0113] in, For the first Cracked dessert index for each sample point This is the normalized calibration crack probability. The normalized crack indicator factor. The normalized overall fragility index, , and Let be the weight coefficient, and satisfy... .

[0114] Preferably, step S6 further includes constructing a decision score and a risk index, which are expressed as follows:

[0115] ;

[0116] ;

[0117] in, To calculate the overall engineering decision score, As a risk index, To normalize uncertainty, Shale content, The normalized overall fragility index, This is a reservoir quality index. As a fluid indicator, This represents the crack prior index.

[0118] The present invention has the following advantages: By constructing positive sample-unlabeled learning soft labels, the present invention reduces the dependence on large-scale manually labeled samples, and constructs a time-series sample sequence along the well depth direction to establish a conditional hybrid expert time-series convolutional network constrained by geophysical priors. This enables better identification of the segment continuity and local anomaly combination characteristics of fractured reservoirs. Furthermore, the present invention constructs a fracture sweet spot index and a risk index, so that the fracture identification results can directly serve sweet spot evaluation, fracturing segment selection, and well location deployment. Attached Figure Description

[0119] Figure 1 This is a schematic diagram of the process for a conditional hybrid expert small-sample well logging fracture identification method based on geophysical prior constraints. Detailed Implementation

[0120] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0121] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0122] It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined with each other.

[0123] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0124] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product of this invention is in use, or the orientation or positional relationship commonly understood by those skilled in the art. They are only used for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention. In addition, the terms "first," "second," etc., are only used to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0125] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0126] In this embodiment, as Figure 1 As shown, a conditional hybrid expert small-sample well logging fracture identification method based on geophysical prior constraints includes the following steps:

[0127] S1: Construct a basic sample set for fracture identification; specifically, firstly, read multi-well logging data, stratigraphic data, and a small amount of existing fracture interpretation data, and standardize the field names. The depth field is standardized to DEPTH, and the natural gamma, sonic transit time, density, deep resistivity, and flushed zone resistivity fields are standardized to GR, AC, DEN, RT, and RXO, respectively. Then, the following steps are also included:

[0128] S11: Follow a uniform depth step size Establish a resampling depth grid:

[0129] ;

[0130] in, For the first Each resampling depth point The starting depth of the target well section. The resampling step size, This represents the total number of sample points after resampling.

[0131] S12: Perform depth domain interpolation to complete each logging curve.

[0132] ;

[0133] in, For the first The logging curve at the resampling depth point The value at that location, The original well logging curve, For interpolation operators;

[0134] S13: Establish fracture sample labels based on fracture interpretation layers.

[0135] ;

[0136] in, For the first Crack labels for each sample point To explain the top depth of the fracture section. This step interprets the bottom depth of the fracture interval. Specifically, in exploration and development practice, fracture samples are usually derived from a small number of oil testing, gas testing, imaging logging, core observation, or comprehensive interpretation results. Therefore, this step forms a training basis of a small number of known fracture samples and a large number of unlabeled samples, rather than a complete large-sample fully supervised dataset.

[0137] S2: Extract multi-source features based on the basic sample set in step S1;

[0138] S3: Constructing positive samples—learning soft labels without labeling;

[0139] S4: Construct a time-series sample sequence along the well depth direction, establish a conditional hybrid expert time-series convolutional network constrained by geophysical priors, and output the fracture development probability at each depth point;

[0140] S5: Perform probability calibration and conformal prediction on the crack development probability, and output the calibrated crack development probability and uncertainty range;

[0141] S6: Forming fracture development segment identification results. By constructing positive sample-unlabeled learning soft labels, the dependence on large-scale manually labeled samples is reduced, and a time-series sample sequence along the well depth direction is constructed. A conditional hybrid expert time-series convolutional network constrained by geophysical priors is established, which can better identify the segment continuity and local anomaly combination characteristics of fractured reservoirs. Furthermore, a fracture sweet spot index and risk index are constructed so that the fracture identification results can directly serve sweet spot evaluation, fracturing segment selection, and well location deployment.

[0142] Furthermore, in step S2, the multi-source characteristics include shale content, comprehensive brittleness index, fracture indicator factor, acoustic multi-scale anomaly, reservoir quality index, fluid indicator factor, and fracture prior index. Further still, step S2 also includes the following steps:

[0143] S21: Calculate the shale content index based on the natural gamma curve.

[0144] ;

[0145] in, For the first Shale index of each sample point For the first The natural gamma value of each sample point The minimum natural gamma within the target well section. The maximum natural gamma within the target well section. To prevent extremely small positive numbers with a denominator of zero;

[0146] Calculate shale content,

[0147] ;

[0148] in, For the first The shale content of each sample point; this index is used to describe the constraint effect of clay content on the development and preservation conditions of fractures. It is a basic background parameter in fracture identification. High shale content usually indicates strong clay infill and weak brittleness, which is not conducive to keeping natural fractures open.

[0149] S22: Calculate the longitudinal wave velocity based on the acoustic time difference.

[0150] ;

[0151] in, For the first P-wave velocity at each sample point For the first Acoustic transit time at each sample point

[0152] The P-wave and S-wave velocity ratio was obtained by weighting based on shale content:

[0153] ;

[0154] Obtain the transverse wave velocity:

[0155] ;

[0156] in, For the first The ratio of P-wave to S-wave velocity at each sample point This is a reference value for the P-wave and S-wave velocity ratio under low mud content conditions. This is a reference value for the ratio of longitudinal and transverse wave velocities under high mud content conditions. For the first Shear wave velocity at each sample point;

[0157] S23: Calculate the dynamic elastic modulus and dynamic Poisson's ratio based on density, P-wave velocity, and S-wave velocity.

[0158] ;

[0159] in, For the first Dynamic elastic modulus of a sample point For the first Volume density of each sample point For the first The dynamic Poisson's ratio of a sample point; in exploration and development, the higher the elastic modulus and the lower the Poisson's ratio, the better the rock brittleness, which is more conducive to the development of natural fractures and the formation of complex fracture networks by subsequent hydraulic fracturing.

[0160] S24: Construct a comprehensive fragility index.

[0161] ;

[0162] in, For the first The overall fragility index of each sample point This is the normalized dynamic elastic modulus. The normalized dynamic Poisson's ratio;

[0163] S25: Constructing a crack indicator factor based on the separation of deep resistivity and flushing zone resistivity:

[0164] ;

[0165] in, For the first Crack indicator factor for each sample point For the first Deep resistivity of each sample point For the first The resistivity of the flushing zone at each sample point; this index is used to reflect the disturbance of electrical distribution caused by cracks, microcracks and flow channels.

[0166] S26: Construct multi-scale acoustic anomalies based on the multi-scale filtering difference of the acoustic curve.

[0167] ;

[0168] in, For the first Multiscale acoustic anomalies at individual sample points The scale parameter is Smoothing operator, The scale is the number of layers; this index can be used to capture local acoustic response disturbances and changes in layer heterogeneity caused by cracks.

[0169] S27: Construct reservoir quality index, fluid indicator factor, and fracture prior index.

[0170] ;

[0171] ;

[0172] ;

[0173] in, For the first Reservoir quality index of each sample point For the first Fluid indicator factor for each sample point For the first Crack prior index for each sample point The normalized overall fragility index, The normalized deep resistivity, The normalized crack indicator factor. This represents the normalized multi-scale acoustic anomalies. Specifically, through the above feature construction, the model input is no longer limited to a single logging anomaly, but integrates multi-source information closely related to fracture development, such as shale background, rock brittleness, electrical separation, acoustic disturbance, reservoir quality, and fluid response.

[0174] Furthermore, due to the limited number of crack interpretation samples, this invention does not directly employ traditional fully supervised training. Instead, it first constructs a positive sample-unlabeled sample learning soft label system. The specific steps are as follows:

[0175] S31: Quantitative thresholds for shale content, fracture indicator factor, and sonic multiscale anomaly were statistically analyzed using training well samples, and denoted as follows: , and ;

[0176] S32: Construct a seed criterion for high-confidence crack samples

[0177] ;

[0178] in, For the first High-confidence crack seed markers for each sample point, The threshold for shale content, The threshold for the crack indicator factor. The threshold for acoustic multi-scale anomalies is used; this step reflects the geological-logging response characteristics of fracture-developed strata, which typically have low shale content, high resistivity separation, or obvious acoustic anomalies.

[0179] S33: Constructing a heuristic soft probability prior to the crack.

[0180] ;

[0181] in, For the first Heuristic soft probability of crack prior for each sample point This represents the normalized shale content. The normalized crack indicator factor. The normalized multiscale anomaly of acoustic waves;

[0182] S34: A positive sample-unlabeled learning method based on multiple bootstrapping is used to train the unlabeled samples in multiple rounds to obtain the average crack probability of the unlabeled samples.

[0183] ;

[0184] in, For the first Positive sample-unlabeled learning probability of each sample point For the number of self-service sampling rounds, For the first Wheel classifier for samples The output probability;

[0185] S35: Construct the final soft label.

[0186] ;

[0187] in, For the first The final soft label for each sample point This is a cutoff function used to limit the probability between 0.02 and 0.98. Specifically, this step addresses the problems of scarce fracture samples, high labeling costs, and a large number of unlabeled samples in actual exploration and development work, enabling model training to shift from relying on a large number of manual labels to using a small number of known fracture segments to drive learning from a large number of unlabeled samples.

[0188] In this embodiment, step S4 further includes the following step:

[0189] S41: Using the current depth point and its predecessors. Construct a time-series sample sequence from each depth point.

[0190] ;

[0191] in, For the first The time-series input matrix corresponding to each sample point For the first Feature vectors of depth points The time window length, The input feature dimension is used; specifically, fracture development has obvious vertical segment continuity, so in this embodiment, a time series sample sequence is constructed along the well depth direction.

[0192] S42: Multiple temporal convolutional expert networks are used to encode the temporal sample sequence to obtain the feature representations of each expert.

[0193] ;

[0194] in, For the first The feature representation output by the expert network The number of expert networks; by setting up multiple expert networks, the variation characteristics of fracture development under different layer thickness scales, different response frequency bands and different well sections can be focused on separately.

[0195] S43: Construct a conditional vector based on shale content, brittleness normalized value, fluid indicator factor, lithofacies code, and reservoir quality index.

[0196] ;

[0197] in, For the first The condition vector of each sample point The normalized overall fragility index, As a fluid indicator, For lithofacies coding, This is a reservoir quality index. This is a conditional mapping function; the purpose of this conditional vector is to explicitly input geological background information into the model, so that different well sections can call more suitable expert branches under different lithology and reservoir conditions.

[0198] S44: Construct gating weights based on the output of the auxiliary task.

[0199] ;

[0200] in, For the first The expert gating weight vector for each sample point For the weight matrix of the gated network, For the gated network bias term, For vector concatenation, Output vectors for auxiliary tasks;

[0201] S45: Weighted fusion of multiple expert features based on gating weights.

[0202] ;

[0203] in, For the fused feature representation, For the first The gating weights corresponding to each expert;

[0204] S46: Perform attention calculation on the weighted fused features and output the crack development probability.

[0205] ;

[0206] in, For the first The probability of crack development at each sample point For multi-head attention operators, For classification mapping function, The activation function is Sigmoid. Specifically, this modeling approach enables the model to utilize the temporal continuity along the well depth direction and dynamically call the most suitable expert features according to different reservoir backgrounds, achieving coupled modeling of geological background conditions, logging feature patterns, and fracture probability output.

[0207] Furthermore, step S5 also includes the following steps:

[0208] S51: Constructing a weighted cross-entropy loss based on soft labels

[0209] ;

[0210] in, For weighted cross-entropy loss, The total number of samples, It is a soft tag. To predict the probability of cracks, this loss reduces the risk of positive classes being overwhelmed by negative classes or unlabeled samples under small sample conditions by increasing the weight of crack samples.

[0211] S52: Construct auxiliary task loss,

[0212] ;

[0213] in, To help average the loss of the task, Losses due to lithofacies auxiliary missions. For fragile auxiliary tasks, Loss due to fluid-assisted tasks;

[0214] S53: To enhance the model's dependence on key geophysical features, a physical significance constraint is introduced. This constraint is constructed by setting the set of physical features as follows: Its gradient proportion is,

[0215] ;

[0216] And define physical constraint penalty terms,

[0217] ;

[0218] in, The proportion of physical feature gradient. For the first Single-sample loss for each sample point For the first The first sample point Each input feature For physical constraint loss, Set a preset physical dependency threshold;

[0219] S54: Construct the total loss function.

[0220] ;

[0221] in, For the total loss, , and These are the weighting coefficients. For gated constraint loss;

[0222] S55: The output probability is calibrated using isotonic regression to obtain the calibrated crack development probability. ;

[0223] S56: Constructing conformal prediction boundaries based on calibration set residuals

[0224] ;

[0225] And output the probability range of crack development.

[0226] ;

[0227] in, To predict the boundary width in a conformal manner, for Quantile operator, This represents the calibrated crack development probability. For the first The crack probability confidence interval for each sample point can be used to interpret reliability assessments and engineering risk warnings.

[0228] Furthermore, in step S6, the fracture-developed segments are screened by constructing a fracture sweetness index, which is expressed as:

[0229] ;

[0230] in, For the first Cracked dessert index for each sample point This is the normalized calibration crack probability. The normalized crack indicator factor. The normalized overall fragility index, , and Let be the weight coefficient, and satisfy... ;

[0231] Further construct decision scores and risk indices, which are expressed as follows:

[0232] ;

[0233] ;

[0234] in, To calculate the overall engineering decision score, As a risk index, To normalize uncertainty, Shale content, The normalized overall fragility index, This is a reservoir quality index. As a fluid indicator, This is the crack prior index. Specifically, in practical applications, it can be... high, High and Low-depth continuous segments are identified as priority fracture sweet spots for fracturing segment selection, well location optimization, and development deployment.

[0235] In this embodiment, the present invention solves the problems of few fracture samples, large inter-well differences, and insufficient engineering applicability in exploration and development scenarios by constructing multi-source features from well logging, learning soft labels from small samples, identifying geophysical prior constraints in time series, probabilistic calibration and conformal prediction, and evaluating fracture sweet spots. It achieves robustness, interpretability, and engineering feasibility of well logging fracture identification under small sample conditions.

[0236] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A conditional hybrid expert small-sample well logging fracture identification method based on geophysical prior constraints, characterized in that: Includes the following steps: S1: Construct a basic sample set for crack identification; S2: Extract multi-source features based on the basic sample set in step S1; S3: Constructing positive samples—learning soft labels without labeling; S4: Construct a time-series sample sequence along the well depth direction, establish a conditional hybrid expert time-series convolutional network constrained by geophysical priors, and output the fracture development probability at each depth point; S5: Perform probability calibration and conformal prediction on the crack development probability, and output the calibrated crack development probability and uncertainty range; S6: Identification results of fracture-developed layers.

2. The fracture identification method based on geophysical prior constraints and conditional hybrid expert small-sample logging according to claim 1, characterized in that: Step S1 further includes the following steps: S11: Follow a uniform depth step size Establish a resampling depth grid: ； in, For the first Each resampling depth point The starting depth of the target well section. The resampling step size, This represents the total number of sample points after resampling. S12: Perform depth domain interpolation to complete each logging curve. ； in, For the first The logging curve at the resampling depth point The value at that location, The original well logging curve, For interpolation operators; S13: Establish fracture sample labels based on fracture interpretation layers. ； in, For the first Crack labels for each sample point To explain the top depth of the fracture section. The depth of the layer is explained to indicate the fracture.

3. The method for fracture identification based on conditional hybrid expert small-sample logging with geophysical prior constraints according to claim 2, characterized in that: In step S2, the multi-source characteristics include shale content, comprehensive brittleness index, fracture indicator factor, acoustic multi-scale anomaly, reservoir quality index, fluid indicator factor, and fracture prior index.

4. The fracture identification method based on geophysical prior constraints and conditional hybrid expert small-sample logging according to claim 3, characterized in that: Step S2 further includes the following steps: S21: Calculate the shale content index based on the natural gamma curve. ； in, For the first Shale index of each sample point For the first The natural gamma value of each sample point The minimum natural gamma within the target well section. The maximum natural gamma within the target well section. To prevent extremely small positive numbers with a denominator of zero; Calculate shale content, ； in, For the first Shale content at each sample point; S22: Calculate the longitudinal wave velocity based on the acoustic time difference. ； in, For the first P-wave velocity at each sample point For the first Acoustic transit time at each sample point The P-wave and S-wave velocity ratio was obtained by weighting based on shale content: ； Obtain the transverse wave velocity: ； in, For the first The ratio of P-wave to S-wave velocity at each sample point This is a reference value for the P-wave and S-wave velocity ratio under low mud content conditions. This is a reference value for the ratio of longitudinal and transverse wave velocities under high mud content conditions. For the first Shear wave velocity at each sample point; S23: Calculate the dynamic elastic modulus and dynamic Poisson's ratio based on density, P-wave velocity, and S-wave velocity. ； in, For the first Dynamic elastic modulus of a sample point For the first Volume density of each sample point For the first The dynamic Poisson's ratio of each sample point; S24: Construct a comprehensive fragility index. ； in, For the first The overall fragility index of each sample point This is the normalized dynamic elastic modulus. The normalized dynamic Poisson's ratio; S25: Constructing a crack indicator factor based on the separation of deep resistivity and flushing zone resistivity: ； in, For the first Crack indicator factor for each sample point For the first Deep resistivity of each sample point For the first Resistivity of the rinsing band at each sample point; S26: Construct multi-scale acoustic anomalies based on the multi-scale filtering difference of the acoustic curve. ； in, For the first Multiscale acoustic anomalies at individual sample points The scale parameter is Smoothing operator, The number of scale layers; S27: Construct reservoir quality index, fluid indicator factor, and fracture prior index. ； ； ； in, For the first Reservoir quality index of each sample point For the first Fluid indicator factor for each sample point For the first Crack prior index for each sample point The normalized overall fragility index, The normalized deep resistivity, The normalized crack indicator factor. The normalized acoustic multiscale anomaly.

5. The fracture identification method based on geophysical prior constraints and conditional hybrid expert small-sample logging according to claim 4, characterized in that: Step S3 further includes the following steps: S31: Quantitative thresholds for shale content, fracture indicator factor, and sonic multiscale anomaly were statistically analyzed using training well samples, and denoted as follows: , and ; S32: Construct a seed criterion for high-confidence crack samples ； in, For the first High-confidence crack seed markers for each sample point, The threshold for shale content, The threshold for the crack indicator factor. The threshold for multi-scale anomalies in acoustic waves; S33: Constructing a heuristic soft probability prior to the crack. ； in, For the first Heuristic soft probability of crack prior for each sample point This represents the normalized shale content. The normalized crack indicator factor. The normalized multiscale anomaly of acoustic waves; S34: A positive sample-unlabeled learning method based on multiple bootstrapping is used to train the unlabeled samples in multiple rounds to obtain the average crack probability of the unlabeled samples. ； in, For the first Positive sample-unlabeled learning probability of each sample point For the number of self-service sampling rounds, For the first Wheel classifier for samples The output probability; S35: Construct the final soft label. ； in, For the first The final soft label for each sample point This is a truncation function.

6. The fracture identification method based on geophysical prior constraints and conditional hybrid expert small-sample logging according to claim 5, characterized in that: Step S4 also includes the following steps: S41: Using the current depth point and its predecessors. Construct a time-series sample sequence from each depth point. ； in, For the first The time-series input matrix corresponding to each sample point For the first Feature vectors of depth points The time window length, The dimension of the input features; S42: Multiple temporal convolutional expert networks are used to encode the temporal sample sequence to obtain the feature representations of each expert. ； in, For the first The feature representation output by the expert network For the number of expert networks; S43: Construct a conditional vector based on shale content, brittleness normalized value, fluid indicator factor, lithofacies code, and reservoir quality index. ； in, For the first The condition vector of each sample point The normalized overall fragility index, As a fluid indicator, For lithofacies coding, This is a reservoir quality index. For conditional mapping functions; S44: Construct gating weights based on the output of the auxiliary task. ； in, For the first The expert gating weight vector for each sample point For the weight matrix of the gated network, For the gated network bias term, For vector concatenation, Output vectors for auxiliary tasks; S45: Weighted fusion of multiple expert features based on gating weights. ； in, For the fused feature representation, For the first The gating weights corresponding to each expert; S46: Perform attention calculation on the weighted fused features and output the crack development probability. ； in, For the first The probability of crack development at each sample point For multi-head attention operators, For classification mapping function, This is the Sigmoid activation function.

7. The fracture identification method based on geophysical prior constraints and conditional hybrid expert small-sample logging according to claim 6, characterized in that: Step S5 also includes the following steps: S51: Constructing a weighted cross-entropy loss based on soft labels ； in, For weighted cross-entropy loss, The total number of samples, It is a soft tag. To predict the probability of cracks; S52: Construct auxiliary task loss, ； in, To help average the loss of the task, Losses due to lithofacies auxiliary missions. For fragile auxiliary tasks, Loss due to fluid-assisted tasks; S53: Construct physical saliency constraints, let the set of physical features be... Its gradient proportion is, ； And define physical constraint penalty terms, ； in, The proportion of physical feature gradient. For the first Single-sample loss for each sample point For the first The first sample point Each input feature For physical constraint loss, Set a preset physical dependency threshold; S54: Construct the total loss function. ； in, For the total loss, , and These are the weighting coefficients. For gated constraint loss; S55: The output probability is calibrated using isotonic regression to obtain the calibrated crack development probability. ; S56: Constructing conformal prediction boundaries based on calibration set residuals ； And output the probability range of crack development. ； in, To predict the boundary width in a conformal manner, for Quantile operator, This represents the calibrated crack development probability. For the first The confidence interval for the probability of cracks at each sample point.

8. The method for fracture identification based on conditional hybrid expert small-sample logging with geophysical prior constraints according to claim 7, characterized in that: In step S6, the fracture development segments are screened by constructing a fracture sweetness index, which is expressed as: ； in, For the first Cracked dessert index for each sample point This is the normalized calibration crack probability. The normalized crack indicator factor. The normalized overall fragility index, , and Let be the weight coefficient, and satisfy... .

9. The method for fracture identification based on conditional hybrid expert small-sample logging with geophysical prior constraints according to claim 8, characterized in that: Step S6 further includes constructing a decision score and a risk index, which are expressed as follows: ； ； in, To calculate the overall engineering decision score, As a risk index, To normalize uncertainty, Shale content, The normalized overall fragility index, This is a reservoir quality index. As a fluid indicator, This represents the crack prior index.

Images