Well logging permeability prediction method and system based on seepage prior gating

Abstract

The application discloses a well logging permeability prediction method and system based on seepage prior gating, and belongs to the field of well logging interpretation and reservoir permeability evaluation. In order to solve the problems of few permeability labels, missing well logging noise and insufficient prediction credibility in low-permeability and strong-heterogeneous reservoirs, the application realizes the technical effects of continuous permeability interpretable prediction and review label generation by constructing a multi-curve sliding window, seepage prior features, depth features and reliable degree gating fusion, and outputting zero inflation heteroscedasticity calibration results.

Description

Technical Field

[0001] This invention relates to the field of well logging interpretation and reservoir permeability evaluation, and particularly to a well logging permeability prediction method and system based on seepage prior gating. Background Technology

[0002] In the evaluation of tight sandstone, low-permeability, and highly heterogeneous reservoirs, permeability is a crucial parameter reflecting the reservoir's fluid flow capacity. Current well logging interpretation typically relies on empirical formulas, core calibration relationships, or general machine learning models to map resistivity, natural gamma, density, neutron, and sonic logging curves into continuous permeability curves to support reservoir stratification evaluation and development decisions.

[0003] However, the number of core permeability calibration points is usually limited, and permeability can span multiple orders of magnitude and include zero values. Furthermore, logging curves suffer from noise, missing data, depth misregistration, and local anomalies, leading to insufficient applicability of empirical formulas and making general machine learning models prone to overfitting or extrapolation anomalies. In addition, logging responses change rapidly at thin layers and layer boundaries; relying solely on ordinary vertical smoothing or global regression can easily result in poor vertical continuity, weakened boundary abrupt changes, and a lack of expressive reliability in predictions.

[0004] Therefore, there is a need for a well logging permeability prediction method and system based on seepage prior gating that can overcome the shortcomings of the existing technologies. Summary of the Invention

[0005] One objective of this invention is to propose a well logging permeability prediction method and system based on seepage prior gating. Addressing the problem that existing technologies struggle to stably output continuous permeability in low-permeability and highly heterogeneous reservoirs due to factors such as scarce labels, curve anomalies, zero-value skewness, and layer boundary responses, this invention proposes a technical solution that uses a multi-curve sliding window as input, employs seepage prior features and depth features as dual-channel representations, and calibrates the output permeability curve through prior reliability gating and zero-expansion heteroscedasticity calibration. This invention achieves the technical effects of improving prediction continuity, thin-layer response retention capability, and engineering verification reliability.

[0006] This invention provides a well logging permeability prediction method based on seepage prior gating, comprising: S1, acquiring multiple well logging curves of the target well section to be predicted, performing depth benchmark correction, quality control, anomaly identification, missing mask construction, and standardization on the well logging curves, and constructing a multi-curve sliding window input tensor according to continuous depth sampling points; S2, generating seepage prior features based on the resistivity and its depth gradient, natural gamma, density neutron combination, acoustic response, and deep-shallow resistivity ratio in the multi-curve sliding window input tensor, and embedding the seepage prior features to obtain prior embedding; S3, inputting the multi-curve sliding window input tensor into the trained permeability prediction model, and performing curve embedding and multi-scale... S4. Generate deep features using one-dimensional convolution of residuals and curve attention; S5. Generate prior reliability coefficients based on missing mask, anomaly identification results, and seepage prior features, and use the prior reliability coefficients to correct the fusion ratio of the seepage prior gating network to fuse deep features with prior embeddings into gated fusion features; S6. Output zero probability, log permeability mean, variance, and calibration prediction interval based on the gated fusion features, and generate a continuous permeability curve based on the inverse transformation of zero probability and log permeability mean. Generate prior conflict degree based on the response differences obtained by mapping deep features and prior embeddings respectively, and generate credibility and verification marks based on calibration prediction interval, prior conflict degree, and anomaly identification results.

[0007] Optionally, S1 includes:

[0008] A depth correction function is established with well logging curve depth markers and target well section correction benchmarks as constraints, and different well logging curves are resampled to a unified depth grid;

[0009] For each curve, an outlier marker is generated according to the preset range of physical property values ​​and the deviation of the median movement in the same well section;

[0010] Write the missing measurement points, outlier markers, and resampling sources into the missing mask;

[0011] For valid sampling points, standardization was performed based on the statistical data of the training well section;

[0012] The multi-curve sliding window input tensor is constructed using a preset number of sampling points on both sides of the current depth point.

[0013] Optionally, S2 includes:

[0014] Resistivity amplitude characteristics, gradient characteristics, and deep-to-shallow resistivity ratio characteristics are generated from deep resistivity, shallow resistivity, and the first-order difference between the two along the depth direction.

[0015] The mud response characteristics were generated from the natural gamma curve;

[0016] The pore structure response characteristics are generated by the difference and intersection relationship between the density curve and the neutron curve;

[0017] Compaction response characteristics are generated from the acoustic time difference and its depth difference;

[0018] The aforementioned features are arranged in a depth window with the same depth as the input tensor of the multi-curve sliding window, and the prior embedding is obtained by linear mapping and normalization.

[0019] Optionally, S3 includes:

[0020] Set curve embedding vectors for different logging curves, and concatenate the curve embedding vectors with the window sampling values ​​and missing masks of the corresponding curves;

[0021] Longitudinal morphological features at the thin-layer, meso-scale, and segmental scales are extracted by one-dimensional residual convolution branches with different kernel lengths;

[0022] Curve attention weights are generated based on the correlation between the output of each curve branch and the prior embedding;

[0023] The depth feature is obtained by weighting and aggregating the outputs of each curve branch using the curve attention weights.

[0024] Optionally, S4 includes:

[0025] Based on the anomaly identification results and the seepage prior features, and combined with the missing mask, the anomaly value marker formed by the anomaly identification results, the resistivity validity marker determined by the resistivity missing state and resistivity value range, the deep-shallow resistivity ratio feature, the mud response feature, the density neutron combination feature, and the acoustic wave mutation feature generated by the acoustic wave response depth difference in the seepage prior features, the prior reliability coefficient with a value in the closed interval of 0 to 1 is output through the prior reliability branch;

[0026] The multi-curve joint gradient is generated based on the depth direction gradient of the multiple logging curves on a unified depth grid, and the multi-curve joint gradient is input into the boundary protection branch, and the boundary protection coefficient with a value in the closed interval of 0 to 1 is output.

[0027] The fusion ratio of the percolation prior gating network is modified by increasing the weight of the prior embedding in the fusion ratio when the prior reliability coefficient increases, and by increasing the weight of the deep feature in the fusion ratio when the prior reliability coefficient decreases.

[0028] Furthermore, the permeability prediction model is trained using training samples, which include multiple logging curves and core permeability labels for the training well section.

[0029] The core permeability K is converted to y=ln(K+1) and used as a label for logarithmic permeability;

[0030] Based on the training samples, a training multi-curve sliding window input tensor, a training prior embedding, a training deep feature, a training gated fusion feature, and the layer boundary protection coefficient are generated.

[0031] The joint loss consists of log permeability prediction error, seepage consistency loss, boundary-aware longitudinal smoothing loss, non-negative constraint loss, heteroscedastic negative log-likelihood loss, and regularization loss.

[0032] The seepage consistency loss is determined based on the consistency between the logarithmic mean of the permeability output during the training phase and the corresponding response of the prior embedding used for training.

[0033] The boundary-aware longitudinal smoothing loss determines the smoothing penalty weight based on the layer boundary protection coefficient. When the layer boundary protection coefficient is not less than a preset protection threshold, the smoothing penalty weight of the difference between adjacent depth points is set as a first preset weight. When the layer boundary protection coefficient is less than the preset protection threshold, the smoothing penalty weight of the difference between adjacent depth points is set as a second preset weight, and the first preset weight is less than the second preset weight.

[0034] Optionally, S5 includes:

[0035] The zero-value probability, the mean log permeability, the variance, and the calibration prediction interval are generated by the zero-expansion heteroscedasticity calibration output head, respectively.

[0036] When the zero-value probability is not less than the preset zero-value determination threshold, the permeability of the corresponding depth point will be output as zero.

[0037] When the zero-value probability is less than the preset zero-value determination threshold, the permeability of the corresponding depth point is obtained according to exp(log average permeability)-1;

[0038] Connect the permeability at each depth point in order of depth to form the continuous permeability curve;

[0039] Furthermore, the calibration prediction interval is generated in the following way: based on the true log penetration rate and the mean log penetration rate of the calibration sample, a non-consistent score is formed; the quantile value of the non-consistent score is obtained according to the preset target coverage rate; and the quantile value and the variance are jointly mapped to the lower bound and upper bound of the interval for the corresponding depth point.

[0040] The lower bound of the interval is truncated to zero when it is less than zero after inverse transformation;

[0041] Furthermore, the credibility and verification mark are generated in the following way: the interval width is generated based on the difference between the upper bound and the lower bound of the interval;

[0042] The deep features are input into the first response mapping layer to obtain the model response, the prior embedding is input into the second response mapping layer to obtain the prior response, and the prior conflict degree is generated based on the difference between the model response and the prior response.

[0043] Input the interval width, the prior conflict degree, the outlier label formed by the anomaly identification result, and the missing mask into the review rule mapping table, and output the confidence level and review label;

[0044] The review rule mapping table includes a marking triggering condition determined by a preset interval width threshold, a preset conflict degree threshold, an outlier marking status, and a missing mask status, and maps the marking triggering condition to either a non-reviewing status or a reviewing status.

[0045] On the other hand, the present invention also provides a well logging permeability prediction system based on seepage prior gating, comprising:

[0046] The data preprocessing module is used to acquire multiple logging curves of the target well section to be predicted, and generate a multi-curve sliding window input tensor, anomaly identification results, and a missing mask; the prior feature module is used to generate seepage prior features and prior embeddings; the depth feature module is used to generate depth features through the trained permeability prediction model; the gated fusion module is used to generate a prior reliability coefficient based on the anomaly identification results, the seepage prior features, and the missing mask, use the prior reliability coefficient to correct the fusion ratio, and fuse the depth features with the prior embeddings into a gated fusion feature; the prediction calibration module is used to output continuous permeability curves, confidence levels, and verification markers.

[0047] The beneficial effects of this invention are:

[0048] 1. By constructing a multi-curve sliding window input tensor after well logging curve preprocessing, and combining resistivity, natural gamma, density neutron combination, acoustic response and deep-shallow resistivity ratio to form seepage prior features, the prediction process can simultaneously utilize longitudinal morphology information and curve responses related to seepage mechanism, thereby reducing the dependence of a simple data-driven model on a small number of wall core calibration points.

[0049] 2. By setting a prior reliability branch and using the prior reliability coefficient to correct the fusion ratio of depth features and prior embedding, the fusion results can be dynamically adjusted when logging curves are missing, abnormal, or the reliability of prior features changes, making the model output more suitable for well sections where noise and local anomalies coexist.

[0050] 3. By generating zero-value probability, logarithmic permeability mean, variance, calibration prediction interval, confidence level, and verification mark through the zero-expansion heteroscedasticity calibration output head, it can uniformly express zero-value skewness, prediction interval width, and prior conflict, thus facilitating the verification of continuous permeability evaluation results. Attached Figure Description

[0051] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0052] Figure 1 The flowchart shows a well logging permeability prediction method based on seepage prior gating.

[0053] Figure 2 This is a flowchart of step S4 of the present invention, which dynamically gates and fuses deep features and prior embeddings based on prior reliability coefficients.

[0054] Figure 3 This is an overlay diagram of the permeability curves of multiple methods and the measured permeability data points of the wall core in the embodiments of the present invention;

[0055] Figure 4 This is a superimposed image of the final recommended permeability curve after gated fusion correction and the measured wall core permeability point in this embodiment of the invention;

[0056] Figure 5 This is a comparison chart of the prediction curve of a single depth model and the prediction curve after gating fusion correction in an embodiment of the present invention. Detailed Implementation

[0057] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0058] refer to Figures 1-2A well logging permeability prediction method based on seepage prior gating includes: S1. Acquiring multiple well logging curves of the target well section to be predicted, performing depth benchmark correction, quality control, anomaly identification, missing mask construction, and standardization on the well logging curves, and constructing a multi-curve sliding window input tensor according to continuous depth sampling points; S2. Generating seepage prior features based on resistivity and its depth gradient, natural gamma, density neutron combination, acoustic response, and deep-shallow resistivity ratio in the multi-curve sliding window input tensor, and embedding the seepage prior features to obtain the prior embedding; S3. Inputting the multi-curve sliding window input tensor into the trained permeability prediction model, and performing curve embedding and multi-scale residual analysis. One-dimensional convolution and curve attention generate deep features; S4, generate prior reliability coefficients based on missing mask, anomaly identification results and seepage prior features, and use the prior reliability coefficients to correct the fusion ratio of the seepage prior gating network to fuse deep features and prior embeddings into gated fusion features; S5, output zero probability, log permeability mean, variance and calibration prediction interval based on gated fusion features, generate continuous permeability curves based on inverse transformation of zero probability and log permeability mean, generate prior conflict degree based on the response differences obtained by mapping deep features and prior embeddings respectively, and generate confidence and verification marks based on calibration prediction interval, prior conflict degree and anomaly identification results.

[0059] In this specific embodiment, S1 includes:

[0060] Acquire the deep resistivity curve, shallow resistivity curve, natural gamma curve, density curve, neutron curve and sonic transit time curve within the target well section, and align the original sampling depth of each curve with the correction reference depth of the target well section. The correction reference depth of the target well section is taken from the depth marking system of the natural gamma curve and recorded as the reference depth axis.

[0061] For each logging curve other than the reference curve, the depth markers given with the well depth are read, and each depth marker is matched with the corresponding marker on the reference depth axis to obtain the depth offset of the curve at each marker. Then, the depth correction function of the curve is established by piecewise linear interpolation between the markers using the depth offset. The original sampling depth of the curve is converted into the corrected sampling depth using the depth correction function, so that all curves are expressed under the same depth reference.

[0062] After depth correction is completed, a uniform depth grid is constructed from the top depth to the bottom depth of the target well section. The sampling interval of the uniform depth grid is set to [value missing]. Each curve is resampled to a uniform depth grid at the corrected sampling depth. The resampling uses linear interpolation, and when a depth point on the uniform depth grid is outside the original sampling depth range, the depth point is recorded as a missing measurement.

[0063] Quality control and anomaly identification were performed on each resampled curve. Quality control included unit consistency checks and removal of non-numerical points. Anomaly identification employed two rules: property value range determination and moving median deviation determination for the same well section. Property value range determination was applied to natural gamma curves. API, for density curves Using neutron curves (Decimal porosity units), using the acoustic transit time curve For deep resistivity curves and shallow resistivity curves, the following methods are used: Sampling points that fall outside the corresponding range are marked as outliers;

[0064] The moving median deviation of the same well section is determined by using a uniform depth grid as the sequence. For each curve, the moving median is calculated by a sliding window with a length of 21 sampling points, and the absolute deviation of the sampling point from the moving median is calculated. Then, a statistic of 5 times the absolute deviation of the median is used as the threshold. Sampling points with an absolute deviation exceeding the threshold are marked as outliers.

[0065] Missing points, outlier markers, and resampling sources are written together into the missing mask. The missing mask records three types of binary states at each depth sampling point of each curve: missing state indicates whether the point is missing or invalid; outlier state indicates whether the point is hit by the outlier identification rule; and source state indicates whether the point is obtained by interpolation, with the interpolation point recorded as 1 and the original sampling point recorded as 0.

[0066] Sampling points that passed quality control and were not marked as missing were standardized. Standardization used the mean and standard deviation of each curve pre-statistically calculated and fixed in the training well section as standardization parameters, and the sampling points at each depth of each curve were standardized according to the formula. Calculate the standardized value, where Indicates the first The logging curves are in the same depth grid. Standardized sampled values ​​at each depth sampling point This represents the original sampled value after resampling. Indicates the first training well section The mean of the logging curves, Indicates the first training well section The standard deviation of the logging curves The depth sampling points are indexed and increment from top to bottom along a uniform depth grid. The curves are indexed and numbered in a fixed order among the six curves mentioned above;

[0067] Construct a multi-curve sliding window input tensor centered on each current depth sampling point on a unified depth grid, and take samples from the left and right sides of the sliding window. There are 1 sampling point, so the window length is 1 The system takes 10 sampling points and stacks the standardized sampled values ​​of the six curves within the window according to the curve dimension and the depth dimension to form an input tensor. At the same time, the missing mask aligned with the window is clipped and stacked to form the corresponding window missing mask. Thus, the system outputs a multi-curve sliding window input tensor, anomaly identification results and missing mask that can be directly used in subsequent steps.

[0068] In this specific embodiment, S2 includes:

[0069] Based on the input tensor of a multi-curve sliding window, the first curve is indexed according to the fixed curve index. The standardized window sampling values ​​of the curves are denoted as follows: , will the The standardized window sample values ​​of the curves used as shallow resistivity curves are denoted as follows: , will the The standardized window sample values ​​of the curves are denoted as follows: , will the The standardized window sample values ​​of the curves used as density curves are denoted as follows: , will the The standardized window sample values ​​of the curves are denoted as follows: , will the The standardized window sample value of the curve is denoted as . ,in Indicates the index of the current depth sampling point. Indicates the relative offset index within the window and And set the window length to and along a uniform depth grid at sampling intervals arrangement;

[0070] Generate seepage prior feature vectors point by point within the window. The resistivity amplitude characteristic is composed of and Directly constituted to characterize the strength of resistivity response, resistivity gradient characteristics are formed by and Divide the first-order difference in the depth direction by This is used to characterize the abrupt change in resistivity with depth. The deep-to-shallow resistivity ratio is calculated from the ratio of deep resistivity to shallow resistivity, and to avoid zero denominator, a formula is introduced. And the shallow resistivity is compared with the calculation before the calculation Take the larger value;

[0071] mud response characteristics The composition is used to characterize changes in clay content;

[0072] The response characteristics of the pore structure are determined by the density neutron difference. Together with the intersection relationship, it constitutes the intersection relationship, which is represented by a binary indicator and in The value is 1 if the condition is met and 0 otherwise, to characterize the intersection of the density and neutron curves at that depth point.

[0073] Compaction response characteristics are and its first difference along the depth direction divided by The structure is designed to characterize the acoustic wave response and its abrupt changes;

[0074] When the source curve sample value of any depth point within the missing mask indication window is missing or abnormal, the relevant prior feature component corresponding to that depth point is set to 0, and the missing or abnormal state of that depth point in the missing mask remains unchanged for the purpose of step [further steps]. It participates in the generation of prior reliability coefficients;

[0075] The seepage prior feature vectors at each depth point are stacked in order of window depth to form a seepage prior feature window matrix. ,in The prior feature dimension of each depth point is represented by the resistivity amplitude, resistivity gradient, depth-to-shallow resistivity ratio, mud response, density-neutron difference, intersection indicator, acoustic response and acoustic gradient components.

[0076] Subsequently, embedding is performed on the seepage prior features to obtain the prior embedding, specifically for each depth point within the window. Using the same set of linear mapping parameters and performing normalization, we obtain ,in Representing depth point The prior embedding vector, Indicates the prior embedding dimension. This represents the linear mapping weight matrix, which consists of model parameters learned during the training phase and fixed for use during the prediction phase. Representing depth point The seepage prior eigenvector, This represents the linear mapping bias vector, which is a model parameter learned during the training phase and fixed for use during the prediction phase. This indicates that layer normalization is performed on the last dimension of the vector to make the vector have zero mean and unit variance in that dimension and include learnable scaling and translation parameters;

[0077] All in the window The prior embedding window tensor is formed by stacking the elements in depth order. and the As a step Prior inputs for intermediate seepage prior gating fusion.

[0078] In this specific embodiment, S3 includes:

[0079] Let the input tensor of the multi-curve sliding window be denoted as... ,in This represents the index of the current depth sampling point, which increments from top to bottom along a uniform depth grid. Indicates the window length and Indicates half the width of the window and Indicates the number of curves and with steps The curve indices are consistent;

[0080] At the same time, the missing mask window is recorded as The three binary components of the third dimension represent the missing state, the abnormal state, and the interpolation source state, respectively.

[0081] Let the prior embedded window be denoted as ,in Represents the prior embedding dimension and ;

[0082] The penetration rate prediction model includes a curve embedding layer, a multi-scale residual one-dimensional convolutional module, and a curve attention module. After training, all parameters are fixed for prediction.

[0083] The curve embedding layer is for each curve Maintain a learnable curve embedding vector ,in For curve index and with The second dimension is consistent. Embedded dimension for curve and ;

[0084] For each relative offset within the window With each curve Take standardized sample values Corresponding missing test status Abnormal state Interpolation source status ,Will With three types of binary states and The vectors are concatenated to form the input vector at each point and then mapped to the feature channels via a linear mapping layer. The weight matrix of the linear mapping layer is denoted as And the bias vector is denoted as The mapped window sequence is then arranged according to the depth direction to obtain the curve-level input feature sequence;

[0085] The multi-scale residual one-dimensional convolution module sets three convolutional branches in parallel for each curve's curve-level input feature sequence to extract longitudinal morphological features at different scales. The kernel lengths of the three branches are set sequentially as follows: Each branch contains two residual convolutional blocks, and each residual convolutional block is composed of a first-layer one-dimensional convolution, a non-linear activation function, and a second-layer one-dimensional convolution, which are then added to the block input as residuals. The one-dimensional convolutions are all calculated along the depth direction and zero-padding is used to maintain the output length. The nonlinear activation function adopts The residuals are summed and then normalized to the channel dimension execution layer to stabilize training and inference.

[0086] The outputs of the three branches are concatenated along the channel dimension and then compressed using a one-dimensional convolution with a kernel length of 1 to obtain the multi-scale output of each curve. ,in Represents the number of deep feature channels and ;

[0087] The curve attention module generates curve attention weights based on the correlation between the multi-scale outputs of each curve and the prior embedding. First, it... Perform average pooling along the depth dimension to obtain the curve representation vector and then... The prior representation vector is obtained by first performing a linear mapping and then average pooling along the depth dimension, where the linear mapping weight matrix of the prior representation vector is denoted as... And the bias vector is denoted as ;

[0088] Then, the attention score for each curve is obtained using dot product correlation and normalized in the curve dimension to obtain the curve attention weights. Perform weighted aggregation to output deep features The calculation relationship is expressed by the following formula:

[0089] ;

[0090] in Indicates the current window is on the curve The attention weights on the curves are set to values ​​between 0 and 1, and the summation over all curves is equal to 1. Represents curve The attention score is obtained by the dot product of the curve representation vector and the prior representation vector. This indicates the curve index used for normalized summation. Represents curve Multi-scale output feature sequences Indicates the steps Gated fusion deep feature output sequence.

[0091] In this specific embodiment, S4 includes:

[0092] For depth feature windows With prior embedded window Perform prior reliability gating fusion, where Indicates the index of the current depth sampling point. Indicates the window length and Represents the number of deep feature channels and Represents the prior embedding dimension and ;

[0093] First, align the prior embeddings to the same number of channels as the depth features, specifically by aligning each relative offset index within the window. Take the prior embedding vector And obtained through linear mapping The linear mapping parameters are fixed as the weight matrix. With bias vector After mapping, layer normalization is performed to ensure that the prior amplitudes of different well sections are at a stable scale;

[0094] Simultaneously extract the depth feature vector And obtained through identity mapping To preserve the vertical morphological information extracted by convolution and attention;

[0095] Based on this, a priori reliability branch is constructed to output the priori reliability coefficient. The input vector of the prior reliability branch at each depth point is denoted as... Its seven components are generated in a fixed order from the missing mask and anomaly identification results: missing measurement indicator, outlier indicator, resistivity validity marker, deep-shallow resistivity ratio characteristic, mud response characteristic, density neutron combination characteristic, and acoustic wave mutation characteristic. The missing measurement indicator is set to 1 if the missing measurement state of any of the six curves is 1, otherwise it is set to 0. The outlier indicator is set to 1 if the outlier state of any of the six curves is 1, otherwise it is set to 0. The resistivity validity marker is set if both the deep and shallow resistivity curves simultaneously satisfy the conditions of being neither missing nor outlier and are located in the step. The resistivity property value is taken as 1 when the range is defined, and 0 otherwise. The characteristics of the deep-shallow resistivity ratio, mud response, density neutron combination, and acoustic wave abrupt change are directly taken from the steps. Prior characteristic components of seepage at the same depth point;

[0096] The prior reliability branch is implemented using a two-layer fully connected network, with the first layer being a weight matrix. With bias vector Second layer weight matrix With bias scalar Interlayer activation function takes The output layer adopts Compress the output to a closed interval between 0 and 1 to obtain ;

[0097] Further construct boundary protection branches to output boundary protection coefficients. The boundary protection branch first calculates the joint gradient of multiple curves on a unified depth grid and takes the joint gradient scalar of the corresponding depth point as input within the window. The joint gradient of multiple curves is obtained by summing the absolute values ​​of the first-order differences of the standardized sample values ​​of the six curves in the depth direction. When the differences on both sides of the depth point are unavailable, the single-sided difference is used to maintain the completeness of the definition. Then, the joint gradient scalar is input into a two-layer fully connected network to obtain the desired result. The hidden layer width of this two-layer fully connected network is 8 and the activation function is... The output layer is ;

[0098] In a gated fusion network, a base gate value is calculated for the fusion ratio at each depth point. The basic gating values ​​are determined by a two-layer fully connected network. The mapping is performed to obtain the first layer, which has a width of 16 and an activation function of . The output layer is ;

[0099] The prior reliability coefficient is used to adjust the fusion ratio and complete gated fusion. The relationship is as follows:

[0100] ;

[0101] in Indicates relative offset as Gated fusion feature vectors of depth points This represents the corrected fusion ratio, and a larger value indicates a higher prior embedding weight. This represents the prior feature vector after channel alignment of the prior embedding. Represents a deep feature vector. This represents the prior reliability coefficient output by the prior reliability branch, and the larger the value, the more reliable the prior feature. This represents the basic gating value output by the gating network. and Used to trim the fusion ratio to a closed range of 0 to 1;

[0102] All in the window Gated fusion feature windows are obtained by stacking them in depth order. As a step Input;

[0103] During the model training phase, the prior reliability branch, the boundary protection branch, and the gated fusion network and steps are discussed. The deep feature modules are trained end-to-end and their parameters are jointly fixed, including the layer boundary protection coefficient. The smoothing penalty weights used to construct the boundary-aware longitudinal smoothing loss are protected by a threshold value of [value missing]. When the adjacent depth points correspond Not less than The smoothing penalty weight for the difference between adjacent depth points is set to the first preset weight. ,when Less than The smoothing penalty weight for the difference between adjacent depth points is set to the second preset weight. Together with the log permeability prediction error, seepage consistency loss, non-negative constraint loss, heteroscedasticity negative log likelihood loss, and regularization loss, the joint loss is formed to ensure that gating fusion maintains abrupt changes at the layer boundary and maintains longitudinal stability within the layer.

[0104] In this specific embodiment, S5 includes:

[0105] Gating fusion feature window Input the zero-expansion heteroscedasticity calibration output head and generate a continuous permeability curve, calibration prediction interval, confidence level, and verification markers. This represents the index of the current depth sampling point, which increments from top to bottom along a uniform depth grid. Indicates the window length and The number of channels representing the gating fusion feature and ;

[0106] Relative offset of window center The gated fusion feature vector is denoted as and with As the input to the output head at that depth point;

[0107] The zero-inflation heteroscedasticity calibration output head consists of a shared backbone and four output branches. The shared backbone is a two-layer fully connected network, with the first layer weight matrix being... And the bias vector is The second layer weight matrix is And the bias vector is Both the two layers and the output of the second layer use Activate and perform layer normalization on the second layer output to obtain the shared latent vector. ;

[0108] The zero-value probability branch is added through a fully connected layer. Output zero probability Its weight vector is And the bias is ;

[0109] The log-penetration mean branch outputs the log-penetration mean through a fully connected layer. Logarithmic penetration rate refers to and For penetration rate;

[0110] The variance branch outputs a scalar through a fully connected layer and then... Transformation yields logarithmic permeability variance And let the standard deviation of the logarithmic penetration rate be ;

[0111] The calibration prediction interval branch does not directly learn the interval endpoints but instead reads the non-consistent fractional quantile values ​​fixed during the calibration phase. and combined Generate log-permeability interval endpoints, where the calibration phase calculates inconsistency scores using calibration samples independent of the training phase and according to the target coverage. Take its quantile value The non-consistency score is defined by dividing the absolute value of the error between the true logarithmic penetration rate and the mean prediction of the calibrated sample by the corresponding standard deviation, and is calculated after calibration. Solidified as a constant for prediction;

[0112] Regarding the point estimation output, let the zero-value determination threshold be... , according to the following formula Inverse transformation to permeability point values Connect them in order of depth to form a continuous permeability curve. ,Right now:

[0113] ;

[0114] in Indicates depth sampling point The penetration rate point value, Indicates depth sampling point The zero probability, This indicates the threshold for determining zero values. Indicates depth sampling point The logarithmic mean penetration rate Represents an exponential function;

[0115] Regarding the calibration of the prediction interval output, the lower bound of the log permeability interval is first generated in the logarithmic domain. Upper bound of the logarithmic permeability interval And then respectively according to The lower and upper bounds of the permeability intervals are obtained by inverse transformation to the permeability domain and truncating the lower bound of the intervals less than 0 after the inverse transformation to 0. The endpoints of the logarithmic domain intervals are determined by the following formula. ,in Indicates depth sampling point The lower bound of the logarithmic penetration range, Indicates depth sampling point The upper bound of the logarithmic penetration range, This indicates that the calibration samples are based on the target coverage. Obtained and solidified non-consistent fractional quantile values, Indicates depth sampling point The standard deviation of the logarithmic penetration rate Indicates depth sampling point The logarithmic mean penetration rate;

[0116] Regarding the generation of credibility and verification markers, the interval width is first calculated based on the difference between the upper and lower bounds of the permeability interval and denoted as... Then calculate the prior conflict degree and record it as . The prior conflict degree is generated by two response mapping branches, the first response mapping branch will divide the steps Window center depth feature vector Input linear mapping yields the model response scalar The second response mapping branch will take steps Window center prior embedding vector Input linear mapping yields prior response scalar and with Define the degree of prior conflict;

[0117] Steps Missing and abnormal states in the missing mask at the center of the window are summarized into missing markers. Outlier markers and will Input verification rule mapping table output credibility level With review mark The review rule mapping table uses a fixed threshold. and As a trigger condition and output according to the following deterministic rules, if or Then let low and ,like and And satisfy or Then let and ,like and And satisfy and Then let High and and all depth points Output in depth order to form a continuous permeability curve and its corresponding calibration prediction interval, confidence level, and verification mark.

[0118] Example 1:

[0119] refer to Figures 3 to 5 The method of this invention was verified by selecting actual logging data and measured core permeability data of the target well section. Deep resistivity curves, shallow resistivity curves, natural gamma curves, density curves, neutron curves, and sonic transit time curves were obtained within the target well section. Following step S1, depth benchmark correction, quality control, anomaly identification, missing mask construction, and standardization were performed to ensure that all logging curves were represented on a unified depth grid. Simultaneously, measured core permeability data points were projected into this unified depth grid according to their corresponding coring depths, serving as the basis for model training, calibration, and prediction result verification.

[0120] After data preprocessing, resistivity amplitude features, resistivity gradient features, shallow-depth resistivity ratio features, clay response features, density neutron combination features, and acoustic compaction response features are extracted according to step S2, and the above-mentioned seepage prior features are embedded as prior embeddings; depth features are extracted by inputting tensors through a multi-curve sliding window according to step S3; then, the fusion ratio of depth features and prior embeddings is corrected by prior reliability coefficient according to step S4 to obtain gated fusion features; and finally, continuous permeability curves, calibration prediction intervals, confidence levels, and verification marks are output according to step S5.

[0121] In this embodiment, to facilitate the explanation of the engineering effects of the output results of the present invention, the fusion recommendation curve generated by the present invention is compared with the prediction curves of candidate methods such as deep learning DNN, random forest RF, gradient boosting GBDT, extreme random tree ET, KNN regression, and PSO-SVM. The above candidate methods are used to form control results or auxiliary candidate responses in engineering verification, and are not intended to limit the present invention to the use of the above specific models; the final prediction result of the present invention is still a continuous permeability curve jointly generated by multi-curve sliding window, seepage prior features, depth features, prior reliability gating, and zero-expansion heteroscedasticity calibration output head.

[0122] like Figure 3 As shown, each method provides a continuous permeability curve along the depth direction in the target well section, and the measured core permeability data points are superimposed on the corresponding curves according to the matching depth. Figure 3 It can be seen that different candidate methods have differences in amplitude response, peak response and low-permeability background response in local segments. The fusion recommendation curve of the present invention can maintain the correspondence between the overall curve and the measured wall core permeability point while maintaining continuous change in the depth direction, thereby providing a continuous curve basis for subsequent credibility evaluation and manual verification.

[0123] like Figure 4As shown, the final recommended permeability curve, after correction by prior reliability gating and calibration output head, is overlaid with the measured core permeability point. This final recommended permeability curve can continuously output low-permeability background, locally higher permeability response, and abrupt change response along the well depth direction. When the measured point deviates significantly from the predicted curve, a verification mark can be generated by combining the calibration prediction interval, prior conflict degree, missing mask, and anomaly identification results to indicate that further core, logging, or interpretation result verification is needed for that depth point or adjacent segment.

[0124] like Figure 5 As shown, compared with the original single DNN model prediction curve, the curve corrected by gated fusion in this invention reduces a large number of zero-value outputs and long-distance constant segments, making the continuous permeability curve have a more stable vertical variation shape within the target well section. Simultaneously, the curve corrected by gated fusion still retains local high-permeability responses and abrupt responses near layer boundaries, avoiding the weakening of thin-layer or layer boundary characteristics by ordinary global smoothing. These results demonstrate that by correcting the fusion ratio with a priori reliability coefficient and combining it with zero-expansion heteroscedasticity calibration output, this invention can improve the interpretability and engineering verification value of continuous permeability prediction results in low-permeability and highly heterogeneous reservoirs.

[0125] Furthermore, in Figures 3 to 5 In the illustrated embodiment, when the logging curve at a certain depth point has missing or abnormal data, the prior reliability branch reduces the weight of the prior embedding in the fusion ratio corresponding to that depth point, allowing the depth feature to make a higher contribution to the prediction at that location. When the depth-to-shallow resistivity ratio, clay response, density neutron combination, and acoustic abrupt change characteristics are all within the reliable range, the prior reliability branch increases the fusion weight of the prior embedding, making the prediction results more constrained by the seepage mechanism-related responses. Therefore, this invention can adaptively adjust the prediction process based on the logging curve quality and the prior reliability of seepage.

[0126] Furthermore, in Figures 3 to 5 In the illustrated embodiment, the multi-curve joint gradient is used to output the boundary protection coefficient. When multiple logging curves exhibit synchronous abrupt changes between adjacent depth points, the boundary protection coefficient increases and the smoothing penalty between the corresponding adjacent depth points decreases, enabling the final recommended permeability curve to retain abrupt changes at the boundary or thin layer. When the curve changes smoothly within the same layer segment, the boundary protection coefficient decreases and the vertical smoothing constraint increases, ensuring that the predicted curve remains continuous and stable within the layer.

[0127] Furthermore, Figures 3 to 5The measured wall core permeability points can be used as calibration or validation samples to determine the inconsistency fractional quantiles, and a calibration prediction interval can be generated accordingly. When the prediction interval width of a certain depth point exceeds a preset interval width threshold, or the prior conflict degree between the model response and the prior response exceeds a preset conflict degree threshold, the prediction calibration module outputs the depth point as medium or low confidence and generates a verification mark; when both the prediction interval width and the prior conflict degree are less than the corresponding thresholds and there are no missing or abnormal marks, the prediction calibration module outputs the depth point as high confidence and does not trigger the verification state.

[0128] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0129] This invention integrates the depth window representation of multiple logging curves, seepage prior features, depth feature extraction, and reliability gating into a continuous processing chain. This allows the local morphology of the input curve, seepage-related responses, and abnormal missing states to all participate in the prediction process. Thus, it forms a corresponding technical processing mechanism to address the problems of few permeability calibration points, unstable curve quality, and thin-layer response sensitivity in low-permeability and highly heterogeneous reservoirs.

[0130] This invention employs a priori reliability branch, a layer boundary protection branch, and a zero-inflation heteroscedasticity calibration output head. The priori reliability branch is used to adjust the fusion ratio of prior embedding and deep features, the layer boundary protection branch is used for boundary-aware longitudinal smoothing during the training phase, and the zero-inflation heteroscedasticity calibration output head is used to generate zero-value probabilities and calibration prediction intervals. This enables the method to better balance zero-value handling, layer boundary mutation preservation, and engineering verification credibility expression.

Claims

1. A well logging permeability prediction method based on seepage prior gating, characterized in that, include: S1. Obtain multiple logging curves of the target well section to be predicted, perform depth benchmark correction, quality control, anomaly identification, missing mask construction and standardization on the logging curves, and construct a multi-curve sliding window input tensor according to continuous depth sampling points; S2. Generate seepage prior features based on the resistivity and its depth gradient, natural gamma, density neutron combination, acoustic response, and deep-shallow resistivity ratio in the multi-curve sliding window input tensor, and embed the seepage prior features to obtain the prior embedding; S3. Input the multi-curve sliding window input tensor into the permeability prediction model trained, and generate deep features through curve embedding, multi-scale residual one-dimensional convolution, and curve attention. S4. Generate prior reliability coefficients based on missing mask, anomaly identification results and seepage prior features, and use the prior reliability coefficients to correct the fusion ratio of the seepage prior gating network so as to fuse deep features and prior embeddings into gating fusion features. S5. Output zero probability, log penetration mean, variance and calibration prediction interval based on gated fusion features. Generate continuous penetration curve based on inverse transformation of zero probability and log penetration mean. Generate prior conflict degree based on response differences obtained by mapping deep features and prior embedding respectively. Generate credibility and verification mark based on calibration prediction interval, prior conflict degree and anomaly identification results.

2. The method according to claim 1, characterized in that, S1 includes: A depth correction function is established with well logging curve depth markers and target well section correction benchmarks as constraints, and different well logging curves are resampled to a unified depth grid; For each curve, outlier markers are generated according to the preset physical property value range and the moving median deviation of the same well section; missing measurement points, outlier markers, and resampling sources are written into the missing mask; valid sampling points are standardized according to the statistics of the training well section; and the multi-curve sliding window input tensor is constructed with a preset number of sampling points on both sides of the current depth point.

3. The method according to claim 1, characterized in that, S2 includes: generating resistivity amplitude features, gradient features, and deep-to-shallow resistivity ratio features from deep resistivity, shallow resistivity, and their first-order difference along the depth direction; generating clay response features from natural gamma curves; generating pore structure response features from the difference and intersection relationship between density curves and neutron curves; generating compaction response features from acoustic time difference and its depth difference; arranging the above features according to the same depth window as the multi-curve sliding window input tensor, and obtaining the prior embedding through linear mapping and normalization.

4. The method according to claim 1, characterized in that, S3 includes: setting curve embedding vectors for different logging curves, and concatenating the curve embedding vectors with the windowed sample values ​​and missing masks of the corresponding curves; extracting vertical morphological features at the thin-layer, medium-layer, and segmental scales through one-dimensional residual convolution branches with different kernel lengths; generating curve attention weights based on the correlation between the outputs of each curve branch and the prior embeddings; and using the curve attention weights to weight and aggregate the outputs of each curve branch to obtain the depth features.

5. The method according to claim 1, characterized in that, S4 includes: Based on the anomaly identification results and the seepage prior features, and combined with the missing mask, the anomaly value marker formed by the anomaly identification results, the resistivity validity marker determined by the resistivity missing state and resistivity value range, the depth-to-shallow resistivity ratio feature, the clay response feature, the density neutron combination feature, and the acoustic abrupt change feature generated by the acoustic response depth difference in the seepage prior features, the prior reliability branch outputs a prior reliability coefficient with a value in the closed interval of 0 to 1; a multi-curve joint gradient is generated based on the depth direction gradient of the multiple logging curves on a unified depth grid, and the multi-curve joint gradient is input into the boundary protection branch, outputting a boundary protection coefficient with a value in the closed interval of 0 to 1; the fusion ratio of the seepage prior gating network is corrected by increasing the weight of the prior embedding in the fusion ratio when the prior reliability coefficient increases, and increasing the weight of the depth feature in the fusion ratio when the prior reliability coefficient decreases.

6. The method according to claim 1, characterized in that, S5 includes: generating the zero-value probability, the logarithmic mean permeability, the variance, and the calibration prediction interval through a zero-expansion heteroscedasticity calibration output head; when the zero-value probability is not less than a preset zero-value judgment threshold, outputting the permeability of the corresponding depth point as zero; when the zero-value probability is less than the preset zero-value judgment threshold, obtaining the permeability of the corresponding depth point according to exp(logarithmic mean permeability)-1; and connecting the permeability of each depth point in depth order to form the continuous permeability curve.

7. The method according to claim 5, characterized in that, The permeability prediction model is trained using training samples, which include multiple logging curves and core permeability labels for the training well section; the core permeability K is converted to y=ln(K+1) and used as the logarithmic permeability label; Based on the training samples, a training multi-curve sliding window input tensor, training prior embedding, training depth features, training gated fusion features, and the layer boundary protection coefficient are generated. A joint loss is constructed using logarithmic permeability prediction error, seepage consistency loss, boundary-aware longitudinal smoothing loss, non-negativity constraint loss, heteroscedasticity negative log-likelihood loss, and regularization loss. The seepage consistency loss is determined based on the consistency between the mean logarithmic permeability output during the training phase and the corresponding response of the training prior embedding. The boundary-aware longitudinal smoothing loss determines the smoothing penalty weight based on the layer boundary protection coefficient. When the layer boundary protection coefficient is not less than a preset protection threshold, the smoothing penalty weight of the difference between adjacent depth points is set as a first preset weight. When the layer boundary protection coefficient is less than the preset protection threshold, the smoothing penalty weight of the difference between adjacent depth points is set as a second preset weight, and the first preset weight is less than the second preset weight.

8. The method according to claim 6, characterized in that, The calibration prediction interval is generated as follows: a non-consistent score is formed based on the true log penetration rate and the mean log penetration rate of the calibration sample; the quantile value of the non-consistent score is obtained according to the preset target coverage rate; and the quantile value and the variance are mapped together to the lower bound and upper bound of the interval for the corresponding depth point; the lower bound of the interval is truncated to zero when it is less than zero after inverse transformation.

9. The method according to claim 8, characterized in that, The credibility and verification mark are generated as follows: the interval width is generated based on the difference between the upper and lower bounds of the interval; the deep features are input into the first response mapping layer to obtain the model response, the prior embedding is input into the second response mapping layer to obtain the prior response, and the prior conflict degree is generated based on the difference between the model response and the prior response; the interval width, the prior conflict degree, the outlier mark formed by the anomaly identification result, and the missing mask are input into the verification rule mapping table, and the credibility level and verification mark are output; the verification rule mapping table contains the mark triggering conditions determined by the preset interval width threshold, the preset conflict degree threshold, the outlier mark status, and the missing mask status, and maps the mark triggering conditions to either not triggering the verification state or triggering the verification state.

10. A well logging permeability prediction system based on seepage prior gating, used to execute the method according to any one of claims 1 to 9, characterized in that, include: The data preprocessing module is used to acquire multiple logging curves of the target well section to be predicted, and generate multi-curve sliding window input tensors, anomaly identification results, and missing mask. The prior feature module is used to generate seepage prior features and prior embeddings; The deep feature module is used to generate deep features from the trained penetration prediction model. The gated fusion module is used to generate a priori reliability coefficient based on the anomaly identification result and the seepage prior features, combined with the missing mask, to correct the fusion ratio using the priori reliability coefficient, and to fuse the depth features with the prior embedding into a gated fusion feature; The prediction calibration module is used to output continuous permeability curves, confidence levels, and verification markers.

Images