Well logging data intelligent inversion method fusing attention mechanism-convolutional neural network

Abstract

The invention belongs to the technical field of petroleum and natural gas engineering, and particularly relates to a logging data intelligent inversion method fusing an attention mechanism and a convolutional neural network, and the method comprises the steps: S100, selecting a borehole which is subjected to well drilling and logging as a research well, and obtaining the drilling, logging and logging data of the research well; s200, constructing a unified depth reference of well drilling, well logging and well logging data, and improving the resolution of the well drilling and well logging data by adopting an interpolation method; s300, constructing an attention mechanism module based on a multi-scale adaptive compression excitation network, and improving the modeling capability through triple mechanisms of multi-scale feature perception, adaptive compression proportion adjustment and residual enhancement attention weight; s400, constructing a feature extraction and regression prediction network based on a one-dimensional convolutional neural network; s500, supervised learning training is carried out on the inversion model until comprehensive evaluation index requirements are met; and S600, intelligent prediction and inversion of the logging curve are carried out by using the trained model. According to the invention, intelligent and efficient inversion of the logging data by using the low-cost drilling and logging data is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of oil and gas engineering, and particularly relates to a well logging data intelligent inversion method fusing an attention mechanism and a convolutional neural network. BACKGROUND

[0002] With the continuous growth of global energy demand, shale gas, as an important unconventional oil and gas resource, is gradually becoming an important support for the fossil energy supply system. Thanks to the rapid development of horizontal well drilling technology and large-scale hydraulic fracturing technology, the large-scale development of shale gas has been realized. However, due to the complex heterogeneity and multi-scale geological characteristics of shale reservoirs, shale gas geology engineering sweet spot evaluation is particularly important for horizontal well deployment and fracturing optimization. Generally, conventional logging curves such as gamma (GR), density (DEN) and acoustic time difference (AC) are important data basis for sweet spot evaluation, and the required rock mechanics parameters can be accurately calculated using density and acoustic time difference.

[0003] However, due to economic cost constraints, some shale gas wells do not perform conventional logging, and lack of key mechanical parameter data, making geology engineering sweet spot evaluation face great challenges. Therefore, how to effectively invert gamma, density, acoustic time difference and other logging curves using lower-cost and easily-obtained drilling and logging data has become a core issue for guiding fracturing design of such shale gas wells. SUMMARY

[0004] The present application provides a well logging data intelligent inversion method fusing an attention mechanism and a convolutional neural network, to solve the technical problem of lack of mechanical parameter data when conventional logging is not performed in the prior art.

[0005] In a first aspect, the present application provides a well logging data intelligent inversion method fusing an attention mechanism and a convolutional neural network, comprising the steps of:

[0006] S100: Selecting a wellbore that has performed drilling and logging as a research well, and obtaining drilling, logging and logging data of the research well;

[0007] S200: Constructing a unified depth reference for drilling, logging and logging data, and improving the resolution of drilling and logging data by using an interpolation method;

[0008] S300: Constructing an attention mechanism module based on a multi-scale self-adaptive compression excitation network (MSA-SE), and improving modeling capability through three mechanisms of multi-scale feature perception, self-adaptive compression ratio adjustment and residual enhancement attention weight;

[0009] S400: Constructing a feature extraction and regression prediction network based on a one-dimensional convolutional neural network (1D-CNN);

[0010] S500: Supervised learning training is performed on the inversion model until the comprehensive evaluation index requirement is met.

[0011] S600: The trained MSA-SE-CNN model is used for intelligent prediction and inversion of well logging curves.

[0012] Further, the drilling data in step S100 includes drilling time and total hydrocarbon content, and the logging data includes free hydrocarbon, pyrolysis hydrocarbon, cracking hydrocarbon, and residual organic carbon.

[0013] Further, step S100 further includes: detecting drilling, logging, and well logging data using a detection algorithm, performing physical constraint and wavelet denoising preprocessing on the data after removing outliers.

[0014] Further, the detection algorithm is an improved Z-Score method.

[0015] Further, the improved Z-Score method step includes:

[0016] Calculate the modified Z-Score value:

[0017]

[0018] Wherein, Z modified is the modified Z-Score value, M is the median of the data, MAD is the median absolute deviation; x i is the i-th data point;

[0019] MAD = median(|x i -M|)

[0020] Wherein, median is the median function;

[0021] When the modified Z-Score value is greater than the set threshold value, the detected data point is marked as an outlier.

[0022] Further, the physical constraint condition includes non-negativity constraint and variation amplitude constraint.

[0023] Further, the quality evaluation index is used to evaluate the data after physical constraint and wavelet denoising preprocessing, and if the quality evaluation index meets the preset threshold value, the next step is performed.

[0024] Further, the data interpolation in step S200 is performed using the Kriging interpolation method.

[0025] Further, the unified depth reference method in step S200 is:

[0026]

[0027] Wherein, Htarget h is a target depth sequence min h is a start point depth of data t h is a data node depth t (n) h is the nth target depth, n is a serial number, and N is a maximum serial number.

[0028] Further, step S400 includes:

[0029] S410: constructing the data processed in steps S100 and S200 as time series input;

[0030] S420: constructing a hierarchical one-dimensional convolutional neural network (1D-CNN) feature extractor to improve the feature abstraction level layer by layer;

[0031] S430: performing global feature aggregation and dimension reduction;

[0032] S440: constructing a multi-layer perceptron regression prediction network;

[0033] S450: constructing an MSA-SE-CNN network overall architecture.

[0034] Further, the comprehensive evaluation indexes in step S500 include a determination coefficient, a mean absolute error, and a root mean square error.

[0035] In a second aspect, the present application provides an electronic device, comprising a memory and a processor; the memory stores computer execution instructions; the processor executes the computer execution instructions stored in the memory, so that the processor executes the method of any one of the above aspects.

[0036] In a third aspect, the present application provides a computer readable storage medium, characterized in that the computer readable storage medium stores computer execution instructions; when the computer execution instructions are executed by a processor, the computer execution instructions are used to implement the method of any one of the above aspects.

[0037] In a fourth aspect, the present application provides a computer program product, characterized in that the computer program product comprises a computer program; when the computer program is executed by a processor, the computer program implements the method of any one of the above aspects.

[0038] The well logging data intelligent inversion method provided by the fusion attention mechanism-convolutional neural network provided by the embodiments of the present application, by capturing the complex nonlinear relationship between the drilling, logging and well logging data of the well which has been subjected to conventional logging, a corresponding fusion attention mechanism-convolutional neural network well logging data intelligent inversion method is constructed, and the well logging data is intelligently inverted by using low-cost drilling and logging data. Compared with the prior art, important breakthroughs are achieved in data preprocessing and spatial modeling. By introducing the fractal Kriging interpolation technology, spatial correlation modeling is performed by constructing a variogram model (exponential, spherical, fractal model), the continuous change rule of the geological horizon at different depths is accurately captured, and the adaptability of the model under complex heterogeneous reservoir conditions is significantly improved. In addition, an adaptive wavelet denoising module containing a physical constraint mechanism is constructed, individualized denoising and non-negative constraint coupling of geochemical parameters and drilling parameters are realized, and the denoising result has high signal-to-noise ratio and geological interpretability.

[0039] In the attention mechanism design aspect, the application innovatively proposes a multi-scale adaptive compression excitation network (MSA-SE), which realizes three technical innovations: a multi-scale feature perception mechanism simultaneously captures global trends and local geological anomaly features through parallel global pooling, multi-scale local pooling and maximum pooling branches; an adaptive compression ratio adjustment mechanism dynamically adjusts network parameters according to the geological horizon characteristics of input data; and a residual enhancement attention weight mechanism prevents key information loss caused by excessive adjustment and maintains the integrity of original geological features.

[0040] The application effectively breaks the isolation of traditional model multi-source data fusion through the deep coupling design of 1D-CNN and MSA-SE, and strengthens the comprehensive expression ability of the model to the correlation between drilling, logging and well logging data through hierarchical feature extraction and adaptive attention weighting. BRIEF DESCRIPTION OF DRAWINGS

[0041] The accompanying drawings, which are incorporated into and form part of the specification, illustrate embodiments consistent with the present application and, together with the specification, serve to explain the principles of the application.

[0042] Figure 1 The flowchart of the well logging data intelligent inversion method provided by the present application is shown in the figure;

[0043] Figure 2 The model structure diagram provided by the present application is shown in the figure;

[0044] Figure 3 The well logging curve denoising comparison diagram provided by the present application is shown in the figure;

[0045] Figure 4 The logging data Kriging interpolation result provided by the present application is shown in the figure;

[0046] Figure 5 The drilling data kriging interpolation result provided for the present application;

[0047] Figure 6 The comparison of the true value and the predicted value of the test set GR, DEN and AC in the embodiments of the present application. DETAILED DESCRIPTION

[0048] The specific embodiments of the present application have been shown by the above-described drawings, and will be described in more detail hereinafter. These drawings and textual descriptions are not intended to limit the scope of the concept of the present application in any way, but to illustrate the concept of the present application to those skilled in the art by referring to specific embodiments.

[0049] The exemplary embodiments will be described in detail herein below with reference to the drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The implementations described in the following exemplary embodiments are not meant to represent all implementations consistent with the present application. Rather, they are merely examples of apparatus and methods consistent with some aspects of the present application as detailed in the appended claims.

[0050] At present, the inversion research of well logging curves mainly focuses on learning the nonlinear mapping relationship between the existing well logging curves (such as gamma, compressional wave slowness) and the target missing curves (such as shear wave slowness) by using deep learning methods such as artificial neural network (ANN), bidirectional recurrent neural network (Bi-RNN) and long short-term memory neural network (LSTM) based on existing conventional well logging data, so as to realize the effective filling of missing data. However, if shale gas wells do not perform conventional well logging, the corresponding inversion model cannot be directly constructed based on existing well logging curves. Compared with conventional well logging, a large amount of low-cost, multi-scale drilling and logging data are generated during drilling and logging operations. These drilling and logging data and well logging data are essentially different dimensions of reservoir physical properties, and there is a highly complex nonlinear correlation. By deeply mining and fusing the potential features of drilling and logging data, a data-driven well logging curve inversion model is constructed, which has important practical significance for realizing comprehensive sweet spot evaluation of geology and engineering and precise fracturing optimization design.

[0051] The data-driven model fully integrates the drilling, logging and logging data of the wells that have been routinely logged, and provides an effective path for logging inversion of low-cost drilling and logging data of shale gas well site by capturing the complex nonlinear relationship between multiple data sources. At present, various deep learning models such as artificial neural network, recurrent neural network, long short-term memory network, convolutional neural network and Transformer have been widely used in the field of petroleum geology. Since the drilling, logging and logging data are essentially characterized by obvious time sequence and spatial correlation, the selected deep learning model not only needs to consider the correlation between data features, but also needs to consider the law of geological sequence changing with depth. Compared with the defects of gradient disappearance, gradient explosion and difficulty in model interpretation in the training process of other network models, the application adopts one-dimensional convolutional neural network (1D-CNN), which is naturally suitable for sequence data and can effectively extract joint features of data in spatial and depth dimensions, and has training stability and model interpretability, and shows obvious advantages in shale gas logging curve inversion application.

[0052] Figure 1 The flowchart of the intelligent logging data inversion method provided by the application is shown in the figure. Figure 1 As shown in the figure, the method comprises:

[0053] S100: selecting a wellbore that has been drilled and logged as a research well, and obtaining drilling, logging and logging data of the research well;

[0054] S200: constructing a unified depth reference of drilling, logging and logging data, and improving the resolution of drilling and logging data by interpolation method;

[0055] S300: constructing an attention mechanism module based on multi-scale adaptive compression excitation network (MSA-SE), and improving modeling ability through three mechanisms of multi-scale feature perception, adaptive compression ratio adjustment and residual enhancement attention weight;

[0056] S400: constructing a feature extraction and regression prediction network based on one-dimensional convolutional neural network (1D-CNN);

[0057] S500: supervised learning training of the inversion model until the comprehensive evaluation index requirement is met;

[0058] S600: intelligent prediction and inversion of logging curve by using the trained MSA-SE-CNN model.

[0059] The specific method of step S100 can include the following sub-steps:

[0060] S110: selection of data type;

[0061] Due to the large difference in geological horizon between the pilot hole section and the horizontal well section, the pilot hole section and the horizontal well section data are respectively analyzed by Spearman correlation, and the features most related to the target curve are selected as the input data.

[0062] Preferably, the drilling data selects drilling time and total hydrocarbon content, and the logging data selects free hydrocarbon, pyrolysis hydrocarbon, cracking hydrocarbon, residual organic carbon,

[0063] S120: detecting the drilling, logging and well logging data by using a detection algorithm, and performing physical constraint and wavelet denoising pretreatment on the data after removing outliers;

[0064] Specifically, step S120 includes the following steps:

[0065] S121: performing outlier detection by using an improved Z-Score method, and calculating a modified Z-Score value:

[0066]

[0067] Wherein, M is the median of the data, x i is the i-th data point, and MAD is the median absolute deviation:

[0068] MAD = median(|x i -M|)

[0069] In this application, when Z modified (x i ) > θ (the threshold is usually 3.0), the data point is marked as an outlier;

[0070] S122: for the selected geochemical logging parameters S0 (free hydrocarbon), S1 (pyrolysis hydrocarbon), S2 (cracking hydrocarbon), S4 (residual organic carbon) and drilling parameters (drilling time, total hydrocarbon content), the physical constraint conditions are established as follows:

[0071] For geochemical logging parameters: S0≥0, S1≥0, S2≥0, S4≥0

[0072] For drilling parameters: drilling time > 0, total hydrocarbon content ≥ 0.

[0073] S123: wavelet denoising processing.

[0074] (a) Adaptive wavelet basis function selection and parameter optimization

[0075] An adaptive parameter selection mechanism based on signal characteristics is established, and the wavelet basis function selection follows the following criteria:

[0076] For drilling, logging and well logging data, the Symlets8 wavelet is preferred, and its reconstruction error is minimum:

[0077]

[0078] wherein: ψ opt is the optimal wavelet basis function, ψ is the candidate wavelet basis function, L is the decomposition level, ε (j) is the reconstruction error of the jth level.

[0079] The decomposition level L is determined by the signal length N:

[0080] L = floor (log2 (N / 16)) + 1, L≤5

[0081] wherein N is the signal length, floor(·) is the floor function

[0082] A differentiated threshold strategy is established for different parameters, and the scaling factor α i is defined as:

[0083] α i = α base × β i

[0084] wherein α i is the scaling factor of the ith parameter, α base is the basic scaling factor, β i is the parameter-specific adjustment coefficient.

[0085] (b) Physical constraint wavelet transform algorithm

[0086] Perform multi-level wavelet decomposition:

[0087] {c (0) ,d (1) ,d (2) ,...,d (L)} = DWT (x)

[0088] wherein c (0) is the approximation coefficient, d (j) is the jth level detail coefficient, DWT is the discrete wavelet transform, x: input signal

[0089] Soft threshold processing is performed on the detail coefficients:

[0090]

[0091] wherein,

[0092] is the kth detail coefficient of the jth level after processing, is the kth detail coefficient of the jth level before processing, λ j is the threshold value of the jth level.

[0093] The threshold value λ jThe calculation is based on robust estimation:

[0094]

[0095] (c) Physical constraint reconstruction and correction mechanism

[0096] Wavelet reconstruction obtains a preliminary denoising signal:

[0097]

[0098] Wherein, is the preliminary denoising signal, and IDWT is the discrete wavelet inverse transform.

[0099] Apply double physical constraint correction:

[0100] (1) Non-negativity constraint:

[0101]

[0102] Wherein, x constrained (n) is the nth data point after constraint, x min is the minimum physical value.

[0103] (2) Change amplitude constraint:

[0104] Relative change rate:

[0105]

[0106] Wherein: r(n) is the relative change rate of the nth point, and ε is a small positive number to prevent division by zero.

[0107] When , ρ is the change rate threshold (usually 2.0), and the weighted correction is adopted:

[0108]

[0109] Wherein, x final (n) is the nth data point after final correction, and γ is the original signal protection weight

[0110] (d) Boundary effect control and quality evaluation

[0111] Periodic boundary processing mode is adopted to reduce boundary artifacts, and multi-dimensional quality evaluation indexes are established, specifically signal-to-noise ratio improvement, signal fidelity, and physical constraint satisfaction rate:

[0112] Signal-to-noise ratio improvement:

[0113]

[0114] Wherein, ΔSNR: signal-to-noise ratio improvement, P signalP is signal power noise,before P is noise power before noise reduction noise,after P is noise power after noise reduction.

[0115] Signal fidelity:

[0116]

[0117] Wherein, p is signal fidelity, cov is covariance, var is variance.

[0118] Physical constraint satisfaction rate:

[0119]

[0120] Wherein, η is physical constraint satisfaction rate, N total N is total data points, N violation N is the number of data points that violate constraints.

[0121] Specifically, the quality control threshold can be set as η≥95%, p≥0.85; when the index is not up to standard, automatically adjust α i And re-execute the noise reduction process until the quality evaluation index meets the preset threshold, then proceed to the next step.

[0122] Through the above physical constraint wavelet denoising preprocessing, the signal-to-noise ratio of drilling and logging data is significantly improved, and at the same time, all output data strictly meet the physical meaning of geological parameters, providing high-quality, physically reasonable data basis for subsequent Kriging spatial interpolation and neural network training.

[0123] The specific method of step S200 includes the following sub-steps:

[0124] S210: Construct a unified depth reference of drilling, logging and logging data to align the depths of multi-source data

[0125] In view of the problem that the sampling intervals of drilling, logging and logging data are inconsistent, the application constructs a unified depth reference as follows:

[0126]

[0127] Wherein, H target is a target depth reference set, is the nth target depth point, h min is the minimum depth value, n is the depth point serial number, and N is the total depth point number.

[0128] Quality assessment and outlier rejection are performed on each data source to ensure the reliability of the input data.

[0129] S220: Spatial variation function modeling

[0130] Establish the semi-variogram function to quantify the spatial correlation of the geological parameters:

[0131]

[0132] Wherein, γ(h) is the semi-variogram value of the lag distance h, N(h) is the number of data pairs of the lag distance h, Z(x i ) is the attribute value at the position x i , and h is the lag distance.

[0133] Construct three theoretical variogram function models:

[0134] Exponential type: γ(h) = C0+C1[1-exp(-3h / a)]

[0135] Spherical type:

[0136] Fractal type: γ(h) = C0+C1·h α

[0137] Wherein, C0 is the nugget value, C1 is the base value, a is the range, and α is the fractal index.

[0138] In this application, the spatial correlation modeling is performed by constructing the variogram function model (exponential type, spherical type, fractal model), which can accurately capture the continuous change rule of the geological horizon at different depths, and significantly improve the adaptability of the model under the condition of complex heterogeneous reservoirs.

[0139] S230: Trend detection and optimal model selection

[0140] Establish an adaptive trend detection, linear trend T1(h) = β0+β1·h, and quadratic trend T2(h) = β0+β1·h+β2·h 2 .

[0141] Wherein, T1(h) is the linear trend function, T2(h) is the quadratic trend function, β0, β1, β2 are the parameters of the trend function.

[0142] Select the optimal variogram function by the least square criterion:

[0143]

[0144] Wherein, θ opt is the optimal parameter combination, θ is the variogram parameter, M is the number of observation point pairs, γ obs (h k ) is the observed semi-variogram value, and γ model (h k ; θ) is the semi-variogram value of the theoretical model.

[0145] S240: Improve the resolution of drilling and logging data by using interpolation method

[0146] Preferably, the present application uses Kriging interpolation method for data interpolation.

[0147] Firstly, based on the covariance function C(h) = C(0) - γ(h), the Kriging equation set is constructed:

[0148]

[0149] Where C is the covariance matrix, 1 is the all-one vector, λ is the Kriging weight vector, μ is the Lagrange multiplier, and c0 is the covariance vector between the estimated point and the known point.

[0150] Kriging interpolation estimate and uncertainty:

[0151]

[0152] Where, is the Kriging estimate value at position x0, λ i is the weight of the i-th known point, and n is the number of known data points.

[0153]

[0154] Where, is the ordinary Kriging variance, C(0) is the zero distance covariance (base value), C(h i0 ) is the covariance between the i-th known point and the estimated point, and h i0 is the distance between the i-th known point and the estimated point.

[0155] S250: Quality control and degradation mechanism

[0156] In this step, the quality evaluation index is established:

[0157] Matrix condition number

[0158] cond(A) < 10 11

[0159] Where cond(A) is the condition number of matrix A, and A is the Kriging coefficient matrix.

[0160] Interpolation rationality

[0161]

[0162] Where, is the data mean, and σ Z is the data standard deviation.

[0163] Linear interpolation degradation is started when the quality is not up to standard, and the spatial correlation fusion of multi-source data under the unified depth reference is realized through the fractal Kriging interpolation technology, which fully utilizes the spatial structure information of geological data to provide high-quality, spatially continuous multi-dimensional feature data for the SE-CNN neural network.

[0164] S300: An attention mechanism module based on a multi-scale adaptive compression excitation network (MSA-SE) is constructed to improve modeling capability through three mechanisms of multi-scale feature perception, adaptive compression ratio adjustment and residual enhancement attention weight

[0165] In view of the limitation that the traditional SE module only uses global average pooling and ignores multi-scale local features, the application proposes a multi-scale adaptive compression excitation network (Multi-Scale Adaptive SE, MSA-SE) module.

[0166] S310: Multi-scale feature perception mechanism

[0167] A multi-branch parallel pooling strategy is used to replace single global average pooling:

[0168] (1) Global pooling branch:

[0169] z global =GlobalAvgPool(X)

[0170] Wherein, z global is a global pooling feature vector, and X is an input feature matrix.

[0171] (2) Multi-scale local pooling branch:

[0172]

[0173] Wherein, is the local pooling feature vector of the kth scale, k is the size of the pooling window, and X is the input feature matrix.

[0174] (3) Significant feature pooling branch:

[0175] z max =GlobalMaxPool(X)

[0176] Wherein, z max is a global maximum pooling feature vector, and X is an input feature matrix.

[0177] Multi-scale feature fusion:

[0178]

[0179] Wherein, z multiis the fused multi-scale feature vector, a is the fusion weight of the global pooling branch, z global is the global pooling feature vector, β is the overall weight of the local pooling branch, w k is the weight coefficient of the kth local pooling branch, is the local pooling feature vector of the kth scale, γ is the fusion weight of the max pooling branch, z max is the global max pooling feature vector.

[0180] S320: Adaptive compression ratio of geologic horizon perception

[0181] Design a dynamic compression ratio adjustment mechanism to adaptively adjust the network parameters according to the geologic horizon characteristics:

[0182] (1) Horizon feature extraction:

[0183] f layer = Conv1D 3×1 (X)

[0184] where f layer is the extracted horizon feature vector, and X is the input feature matrix.

[0185] (2) Adaptive compression ratio calculation:

[0186] r adaptive = r base + Δr·sigmoid(W layer ·GAP(f layer )+ b layer )

[0187] where r adaptive is the adaptive compression ratio, r base is the basic compression ratio, Δr is the compression ratio adjustment range, W layer is the weight matrix of the horizon perception network, f layer is the horizon feature vector, and b layer is the bias vector of the horizon perception network.

[0188] S330: Residual enhanced attention weight generation

[0189] Generate channel attention weights through a two-layer fully connected network and introduce a residual mechanism:

[0190] (1) Weight generation:

[0191] SE weight = Sigmoid(Dense C (ReLU(Dense C / r (z multi ))))

[0192] where SE weight is the generated attention weight vector, z multi is the multi-scale fusion feature vector, C is the number of feature channels, and r is the current compression ratio.

[0193] (2) Residual enhancement output:

[0194] Output=X⊙(λ·SE weight +(1-λ)·1 C )+X

[0195] where Output is the final output feature of the module, X is the input feature matrix, λ is the balance parameter, SE weight is the attention weight vector, and 1C is a C-dimensional unit vector. The operator ⊙ is the element-wise multiplication operator.

[0196] The MSA-SE module significantly improves the modeling ability of different geological horizon features in logging data through the triple mechanisms of multi-scale perception, adaptive adjustment, and residual enhancement.

[0197] S400: Constructing a feature extraction and regression prediction network based on a one-dimensional convolutional neural network (1D-CNN);

[0198] Based on the MSA-SE attention mechanism module constructed in S300, a one-dimensional convolutional neural network architecture is designed to realize deep feature extraction of drilling and logging multi-source data and regression prediction of logging curves.

[0199] S410: Input data preprocessing and sequence construction

[0200] The multi-source fusion data processed by S100 and S200 is constructed as a time series input:

[0201] (1) Input feature matrix: where B is the batch size, T is the time series length (sliding window size 11), and D is the feature dimension (drilling and logging 6-dimensional features: S0, S1, S2, S4, drilling time, and total hydrocarbon)

[0202] (2) Target output matrix: corresponding to 3 logging curves (GR, DEN, AC)

[0203] (3) Data standardization: Z-score standardization is used to ensure feature scale consistency

[0204] S420: Multi-level one-dimensional convolutional feature extraction network

[0205] A hierarchical 1D-CNN feature extractor is constructed to improve the feature abstraction level layer by layer:

[0206] First convolutional block:

[0207] F1 = BatchNorm(LeakyReLU(Conv1D 64×3 (X)))

[0208] Second convolutional block:

[0209] F2 = BatchNorm(LeakyReLU(Conv1D 128×3 (F1)))

[0210] First MSA-SE enhancement:

[0211]

[0212] Third convolutional block:

[0213]

[0214] Second MSA-SE enhancement:

[0215]

[0216] where Conv1D C×K represents a one-dimensional convolution operation with C output channels and a kernel size of K.

[0217] S430: Global feature aggregation and dimension reduction

[0218] Global average pooling is used to convert the variable-length sequence features into a fixed-length vector:

[0219]

[0220] Global average pooling calculation formula:

[0221]

[0222] S440: Multi-layer perceptron regression prediction network

[0223] A two-layer fully connected network is designed to implement nonlinear regression mapping:

[0224] First regression layer:

[0225] H1 = ReLU(Dense 256 (G) + b1)

[0226] Second regression layer:

[0227] H2 = ReLU(Dense 128 (H1) + b2)

[0228] Output layer:

[0229]

[0230] where Dense n represents a fully connected layer with output dimension of n, b i is the bias term.

[0231] S450: MSA-SE-CNN network overall architecture

[0232] The complete SE-CNN network can be represented as the following mapping relationship:

[0233]

[0234] where: Conv1D stack is a multi-layer 1D convolution feature extractor, MSA-SE is a multi-scale adaptive compression excitation attention module, GAP is a global average pooling, MLP is a multi-layer perceptron regression network, and Θ is all trainable network parameters

[0235] The model structure is shown in Figure 2 .

[0236] S500: Supervised learning training of the inversion model until the comprehensive evaluation index requirements are met

[0237] In this step, based on the constructed SE-CNN network architecture, the model parameters are trained using supervised learning to realize the nonlinear mapping relationship learning from drilling and logging multi-source data to logging curves.

[0238] S510: Loss function and optimization objective

[0239] The mean square error (MSE) is used as the loss function to measure the difference between the predicted logging curve and the real logging curve:

[0240]

[0241] where, is the loss function, N is the number of training samples, i is the sample index, j is the logging curve index (corresponding to GR, DEN, AC), y i,j is the true value, is the predicted value, and Θ is the trainable network parameter.

[0242] S520: Optimization algorithm and parameter setting

[0243] The Adam adaptive optimization algorithm is used to update the network parameters:

[0244] m t = β1m t-1 +(1-β1)g t

[0245]

[0246] where m t is the first moment estimate, v t is the second moment estimate, g t is the gradient at step t, is the bias-corrected first moment, is the bias-corrected second moment, a is the learning rate, b1, b2 are the exponential decay rates, and e is the numerical stability term.

[0247] S530: Training strategy and regularization mechanism

[0248] (1) Batch training settings:

[0249] Maximum number of training rounds: 600 epochs

[0250] Batch size: 128

[0251] Validation set monitoring: Calculate validation loss every round

[0252] (2) Early stopping and learning rate scheduling:

[0253] Early stopping mechanism: Stop training when validation loss does not improve for 100 consecutive rounds

[0254] Learning rate decay: Multiply learning rate by 0.5 when validation loss does not improve for 20 rounds

[0255] Minimum learning rate: 1 x 10 -6

[0256] (3) Model checkpointing:

[0257] Automatically save model weights with optimal validation loss to ensure model generalization performance.

[0258] S540: Multi-index evaluation system

[0259] Construct a comprehensive evaluation index system to monitor training effectiveness in real time:

[0260] (1) Coefficient of determination (R 2 ):

[0261]

[0262] (2) Mean absolute error (MAE):

[0263]

[0264] (3) Root mean square error (RMSE):

[0265]

[0266] where y i is the true value, is the predicted value, is the true value mean, and N is the sample number.

[0267] The above indicators are calculated for the GR, DEN, and AC curves after each round of training.

[0268] S600: Intelligent prediction and inversion of well logging curves using the trained MSA-SE-CNN model

[0269] Based on the trained MSA-SE-CNN model, intelligent prediction and inversion of well logging curves for new drilling data are realized.

[0270] S610: Model inference and prediction process

[0271] (1) Input data preprocessing: normalize the drilling and logging data of the new well according to the standardization parameters during training:

[0272]

[0273] (2) Sliding window sequence construction: use the same sliding window strategy (window size 11) as in the training phase to construct the time series input sequence.

[0274] (3) Model forward inference:

[0275]

[0276] where y is the standardized prediction result, Xnew is the new input data, and Θ is the optimal network parameter obtained by training.

[0277] S620: De-standardization and output generation

[0278] Convert the standardized prediction result back to the original well logging curve value:

[0279]

[0280] where μ target and σ target are the mean and standard deviation of the training set well logging curves, respectively.

[0281] S630: Inversion performance verification

[0282] Quantitatively evaluate the inversion results:

[0283] (1) Single well continuity evaluation:

[0284] Compute the gradient change of the prediction value of the adjacent depth point, evaluate the curve continuity:

[0285]

[0286] wherein Continuity is the continuity evaluation index, M is the number of depth points, k is the depth point index, and are the prediction values of adjacent depth points.

[0287] (2) Geological interpretation consistency verification:

[0288] Compare the inverted GR, DEN, AC curves with the geological horizon information to verify the rationality of the geological interpretation.

[0289] S640: Model deployment and application

[0290] (1) Model persistent storage:

[0291] Save the trained network weight file (.h5 format) and the standardization parameter (.pkl format) to ensure that the model can be reused.

[0292] (2) Batch inversion processing:

[0293] Support batch processing of drilling and logging data of multiple wells to improve the efficiency of well logging inversion.

[0294] (3) Real-time inversion capability:

[0295] Combined with the real-time acquisition of drilling and logging data, the real-time prediction of logging curves and geosteering while drilling are realized. Through the end-to-end deep learning framework, the intelligent inversion system realizes high-precision and high-efficiency automatic inversion from drilling and logging data to logging curves, providing strong technical support for drilling engineering and geological interpretation.

[0296] Example 1

[0297] The sample well data in the embodiment 1 of the present application is derived from a shale gas reservoir, and the logging data includes key parameters such as acoustic time difference (AC), gamma ray (GR) and density (DEN) (Table 1). Each 0.125 meter records a logging point position. The drilling and logging data includes gas logging data (Table 2), and geochemical logging data (Table 3).

[0298] Table 1 Partial logging data

[0299]

[0300]

[0301] Table 2 Partial gas logging data

[0302]

[0303]

[0304] Table 3 Partial Geochemical Logging Data

[0305]

[0306]

[0307] Taking multi-source data from the 1580-2115m well section as an example, Symlets8 wavelet basis functions were used to uniformly denoise logging parameters gamma (GR), acoustic transit time (AC), and density (DEN), geochemical logging parameters S0 (free hydrocarbons), S1 (pyrolytic hydrocarbons), S2 (cracking hydrocarbons), and S4 (residual organic carbon), as well as drilling parameters (drilling time, total hydrocarbon content). Differential decomposition layers were set for different parameter characteristics: GR was denoised using 3-4 standard layers to balance noise removal and formation identification accuracy; AC was denoised using 2-3 layers of light denoising to preserve subtle changes in rock mechanical properties; and DEN was denoised using 4-5 layers of enhanced denoising to remove statistical noise from density measurements. The same adaptive wavelet denoising strategy was used for both geochemical logging and drilling parameters, with differentiated processing based on the signal characteristics and physical constraints of each parameter. Threshold scaling factors are adjusted according to the physical characteristics of the parameters: For logging parameters, GR uses a standard noise reduction factor of 1.0 to maintain radioactive anomaly characteristics, AC uses a conservative noise reduction factor of 0.8 to maintain acoustic continuity, and DEN uses an enhanced noise reduction factor of 1.2 to improve density curve smoothness. During the noise reduction process, the physical constraints of various parameters are strictly enforced to ensure that logging parameters GR ≥ 0, AC > 0, DEN > 0, and geochemical logging parameter S... i The data is ≥0, and the non-negative physical meaning of drilling parameters is established. Variation amplitude constraints are used to prevent excessive smoothing from causing formation interface blurring. After processing, the signal-to-noise ratio of multi-source data is improved by an average of 15-30 dB, and the signal fidelity reaches above 0.92. Random noise and environmental interference from various instruments are effectively removed, while maintaining the true response characteristics of formation lithology changes. The high-quality denoising in this application provides a reliable multi-source data foundation for subsequent Kriging spatial interpolation, significantly improving the accuracy of drilling and logging data fusion and the stability of SE-CNN network training, ultimately improving the accuracy of drilling and logging data inversion to logging curves. The logging data denoising effect in this embodiment is as follows: Figure 3 As shown.

[0308] Based on the drilling and logging data after noise reduction in step S100, spatial correlation fusion was performed on geochemical logging parameters (S0, S1, S2, S4) and drilling parameters (drilling time, total hydrocarbon content). Multi-source data with different sampling intervals were uniformly interpolated to a 0.125m fine grid: geochemical logging data with intervals ranging from 2-8m and drilling data with intervals ranging from 1-5m were precisely aligned to the logging data depth benchmark. Three variogram models were constructed to quantify the spatial correlation structure of each parameter, and the optimal model was automatically selected using the least squares criterion: in geochemical logging, S0 and S4 adopted a spherical model (range a = 25-40m) to reflect the layered characteristics of organic matter distribution; S1 and S2 adopted a fractal model (α = 0.7-0.9) to reflect the nonlinear changes in pyrolysis characteristics; in drilling, drilling time data adopted a fractal model (α = 0.9), and total hydrocarbon data adopted an exponential model (range a = 35m) to reflect the spatial continuity of oil and gas displays. Significant trend characteristics were detected in multiple parameters, automatically initiating detrended kriging processing. The kriging interpolation success rate reached 80-90%. Compared to traditional linear interpolation, the kriging method fully utilizes the spatial correlation information of drilling and logging data, accurately capturing changes in organic matter content, hydrocarbon gradients, and formation drilling response patterns, avoiding the distortion problem of linear interpolation in geological abrupt change zones. The interpolation results not only achieve deep unification of multi-source heterogeneous data but also quantify interpolation uncertainty through kriging variance, providing data quality weights for SE-CNN network training. This spatial correlation modeling significantly improves the geological adaptability of multi-source data fusion, enabling precise matching of drilling and logging multidimensional features with logging responses in depth. This lays a high-quality data foundation for neural networks to learn complex geological parameter mapping relationships, ultimately improving the accuracy and reliability of logging curve inversion. The multi-source data kriging interpolation fusion effect in this embodiment is as follows: Figure 4 , Figure 5 As shown.

[0309] Based on multi-source fusion data processed by S1 denoising and S2 interpolation, an end-to-end SE-CNN intelligent inversion network for logging curves was constructed. First, the six-dimensional feature data of the drilling and logging (S0, S1, S2, S4, drilling time, and total hydrocarbons) were used to construct the time-series input using a sliding window strategy, with the window size set to 11 sampling points, corresponding to approximately 1 meter of well section. The network architecture adopted a hierarchical design, including four core modules: feature extraction, attention enhancement, global aggregation, and regression prediction. The model parameters are shown in Table 4.

[0310] Table 4 Model Parameters

[0311]

[0312] The network was trained using the Adam optimizer with an initial learning rate of 1×10⁻⁶. -4The batch size was 128. The training strategy included an early stopping mechanism (patience=100) and adaptive learning rate decay (factor=0.5, patience=20), with a maximum training epoch of 600 epochs. The loss function used was mean squared error (MSE), while mean absolute error (MAE) was monitored as an auxiliary indicator. During training, data from four wells were used as the training set, and one well was used as the validation set to optimize model parameters. The network triggered the early stopping mechanism at the 295th training epoch, at which point the validation loss converged to 0.0045. The MSA-SE attention module effectively captured the geological feature variation patterns at different scales through multi-scale pooling (window sizes 3, 5, and 7) and adaptive compression ratio adjustment.

[0313] After training, data from one well not included in the model was used as the test set to test the model's performance. The inversion results show: GR curve R... 2 =0.84, RMSE=9.27; DEN curve R 2 =0.71, RMSE=0.02; AC curve R 2 =0.80, RMSE=2.84. Compared with traditional methods, this network architecture fully verifies the effectiveness of the MSA-SE attention mechanism and 1D-CNN feature extraction, with an average R0.80 for well logging curve inversion accuracy. 2 The model achieved a performance score of 0.783, fully validating the advanced nature and practical value of the technical solution. After training, the optimal weight file (best_well_model2.h5) and standardized parameter files (feature_scaler.pkl, target_scaler.pkl) are automatically saved, enabling the engineering deployment of intelligent well logging curve inversion. Model performance metrics are shown in Table 5, and the test set inversion results are as follows: Figure 6 As shown.

[0314] Table 5 Performance Indicators

[0315] Prediction target [R 2 ]]> MSE RMSE MAE GR 0.84 85.90 9.27 6.73 DEN 0.71 0.00 0.02 0.01 AC 0.80 8.07 2.84 2.05

Claims

1. A method for intelligent inversion of well logging data integrating attention mechanism and convolutional neural network, characterized in that, Includes the following steps: S100: Select a wellbore that has been drilled and logged as a study well, and obtain drilling, logging, and well logging data of the study well; S200: Construct a unified depth benchmark for drilling, logging, and well logging data, and use interpolation methods to improve the resolution of drilling and logging data; S300: Construct an attention mechanism module based on a multi-scale adaptive compressed excitation network (MSA-SE), which improves modeling capabilities through a triple mechanism of multi-scale feature perception, adaptive compression ratio adjustment, and residual-enhanced attention weights; S400: Construct a feature extraction and regression prediction network based on a one-dimensional convolutional neural network (1D-CNN); S500: Supervised learning training of the inversion model until it meets the comprehensive evaluation index requirements; S600: Utilizes a trained MSA-SE-CNN model for intelligent prediction and inversion of well logging curves.

2. The intelligent inversion method for well logging data fusion using attention mechanism-convolutional neural network as described in claim 1, characterized in that, Step S100 also includes: using a detection algorithm to detect drilling, logging, and well logging data, removing outliers, and then performing physical constraints and wavelet noise reduction preprocessing on the data.

3. The intelligent inversion method for well logging data fusion based on attention mechanism and convolutional neural network as described in claim 1, characterized in that, The detection algorithm is an improved Z-Score method; The improved Z-Score method includes the following steps: Calculate the corrected Z-Score: Among them, Z modified The corrected Z-Score value, where M is the median and MAD is the median absolute deviation; x i Let i be the i-th data point; MAD=median(|x i -M|) Where median is the median function; When the corrected Z-Score value exceeds the set threshold, the detected data point is marked as an outlier.

4. The intelligent inversion method for well logging data fusion using attention mechanism-convolutional neural network as described in claim 1, characterized in that, The physical constraints and the data after wavelet denoising preprocessing are evaluated using quality evaluation indicators. If the preset threshold of the quality evaluation indicators is met, the next step is performed.

5. The intelligent inversion method for well logging data fusion using attention mechanism-convolutional neural network as described in claim 1, characterized in that, Physical constraints include nonnegativity constraints and variation range constraints.

6. The intelligent inversion method for well logging data fusion using attention mechanism-convolutional neural network as described in claim 1, characterized in that, The method for unifying the depth benchmark in step S200 is as follows: Among them, H target For the target depth sequence, h min h represents the starting depth of the data. t h represents the depth of the data node. t (n) Let N be the depth of the nth target, where n is the sequence number and N is the maximum sequence number.

7. The intelligent inversion method for well logging data fusion using attention mechanism-convolutional neural network as described in claim 1, characterized in that, Step S400 includes: S410: Construct the time-series input from the data processed by steps S100 and S200; S420: Construct a hierarchical one-dimensional convolutional neural network (1D-CNN) feature extractor to progressively improve the level of feature abstraction; S430: Perform global feature aggregation and dimensionality reduction; S440: Construct a multilayer perceptron regression prediction network; S450: Construct the overall architecture of the MSA-SE-CNN network.

8. An electronic device, comprising: Memory, processor; The memory stores computer-executable instructions; the processor executes the computer-executable instructions stored in the memory, causing the processor to perform the method according to any one of claims 1-7.

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

10. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method described in any one of claims 1-7.

Images