Well logging curve reconstruction method based on ISCSO-TCN-BiGRU-AM architecture

The ISCSO-TCN-BiGRU-AM architecture solves the problems of insufficient accuracy and weak nonlinear mapping capability when well logging data is missing or abnormal, and realizes high-precision well logging curve reconstruction and fracture-cavity anomaly identification, thereby improving the geological interpretation effect.

CN123116115APending Publication Date: 2026-11-17SHAANXI RAILWAY ENG SURVEY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610979762.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-02
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

When logging data is missing or abnormal under complex geological conditions, existing technologies suffer from insufficient accuracy in traditional data completion methods, weak nonlinear mapping capabilities and poor generalization ability in conventional deep learning methods, and a lack of adaptive hyperparameter tuning mechanisms.

Method used

The ISCSO-TCN-BiGRU-AM architecture is adopted. The temporal convolutional network layer TCN extracts local multi-scale features, the bidirectional gated recurrent unit layer BiGRU captures global temporal dependencies, and the multi-head self-attention mechanism layer AM performs feature weighted fusion. Combined with the improved sand cat swarm optimization algorithm, the network hyperparameters are adaptively updated to construct a deep four-layer coupled model.

Benefits of technology

It enables bidirectional reasoning and completion of logging curves with extremely low error, accurately identifies fracture-cavity anomalies, and improves the accuracy and reliability of geological interpretation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116115A_ABST
    Figure CN123116115A_ABST
Patent Text Reader

Abstract

This invention discloses a well logging curve reconstruction method based on the ISCSO-TCN-BiGRU-AM architecture. It acquires multi-source measured well logging data and performs dimensionless preprocessing to construct a time-series tensor. The time-series tensor is input into a temporal convolutional network layer, where local multi-scale feature matrices are extracted through causal convolution and dilatational convolution. The feature matrix is ​​then input into a bidirectional gated recurrent unit layer, outputting a hidden state matrix. This hidden state matrix is ​​further input into a multi-head self-attention mechanism layer to reconstruct and output the target well logging curve prediction vector. During the training phase, an improved sandcat swarm optimization algorithm is used to adaptively update the network hyperparameters, and a physical constraint loss function including a geological evolution penalty term is employed for model optimization. This invention can complete bidirectional inference and completion of well logging curves with extremely low errors and accurately identify fracture-cavity anomalies through residual comparison, demonstrating high geological interpretation value. It is suitable for electronic devices and computer-readable storage media.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of geophysical exploration and artificial intelligence technology, specifically to a well logging curve reconstruction method based on the ISCSO-TCN-BiGRU-AM architecture. Background Technology

[0002] Well logging curve reconstruction and refined geological interpretation are crucial aspects of geological exploration and railway engineering surveys. Under complex geological conditions, well logging data often exhibits missing or anomalies due to factors such as harsh downhole environments and instrument malfunctions. Traditional completion methods rely on empirical formulas, resulting in insufficient accuracy; conventional deep learning methods suffer from weak nonlinear mapping capabilities and poor generalization ability when facing complex geological conditions.

[0003] In addition to the aforementioned patent documents, academia and industry have also proposed a series of technical solutions based on complex neural network architectures. The relevant technical solutions and their advantages and disadvantages are analyzed below:

[0004] 1) Reconstruction method based on multi-head attention: It integrates multi-head attention mechanism, temporal convolutional network and BiGRU, but the parameter scale is huge and the standard network structure lacks an adaptive hyperparameter optimization mechanism;

[0005] 2) Reconstruction algorithm integrating ISCSO: The network model is optimized using the ISCSO algorithm for data extrapolation. Existing ISCSO applications have not been deeply integrated with TCN, and have failed to achieve the four-layer coupling of deep networks. Summary of the Invention

[0006] This invention provides a logging curve reconstruction method based on the ISCSO-TCN-BiGRU-AM architecture to solve the technical problems existing in the prior art, such as local optima, lack of TCN local feature extraction mechanism, insufficient consideration of bidirectional sequence dependency, and difficulty in manual adjustment of model parameters. It constructs a logging network model that achieves deep four-layer coupling and has adaptive global optimization capability.

[0007] According to a first aspect, one embodiment provides a well logging curve reconstruction method based on the ISCSO-TCN-BiGRU-AM architecture, the method comprising:

[0008] Acquire multi-source measured well logging data containing missing or anomalous segments and preprocess them to construct a time series tensor;

[0009] A well logging curve reconstruction model is constructed, which includes a temporal convolutional network layer (TCN), a bidirectional gated recurrent unit layer (BiGRU), and a multi-head self-attention mechanism layer (AM).

[0010] The time series tensor is input into the temporal convolutional network layer TCN, and local multi-scale feature matrices carrying high-frequency geological mutation information are extracted through causal convolution and dilatation convolution.

[0011] The obtained local multi-scale feature matrix is ​​input into the bidirectional gated cyclic unit layer BiGRU. The temporal dependence is calculated synchronously through forward and backward gated cyclic units, and the global feature hidden state matrix that integrates the macroscopic evolution trend of the stratigraphy is output.

[0012] The obtained global feature hidden state matrix is ​​input into the multi-head self-attention mechanism layer AM, and scaled dot product operation and feature weighted fusion are performed to output the final target logging curve reconstruction vector.

[0013] During the training phase, the improved sand cat swarm optimization algorithm is used to adaptively update the network hyperparameters of the temporal convolutional network layer TCN, the bidirectional gated recurrent unit layer BiGRU, and the multi-head self-attention mechanism layer AM.

[0014] Furthermore, multi-source measured well logging data containing missing or anomalous segments are acquired and preprocessed to construct a time series tensor, specifically including:

[0015] The original measured well logging data were depth aligned, and missing intervals were identified and marked; the local mean method was used to perform preliminary smoothing of obvious outlier noise.

[0016] Based on the aligned multidimensional logging sequence, a standardization transformation is performed on each logging channel, and a time series tensor is constructed. Where T is the number of depth sampling points and C is the number of feature channels.

[0017] Furthermore, the time series tensor is input into the temporal convolutional network (TCN) layer, and local multi-scale feature matrices carrying high-frequency geological abrupt change information are extracted through causal convolution and dilation convolution. Specifically, this includes:

[0018] The time series tensor X is input into multiple stacked dilated convolution residual blocks consisting of causal convolution and dilated convolution for processing, and the local multi-scale feature matrix F is extracted. local Among them, causal convolution ensures that the output at time t depends only on the input at time t and before; dilated convolution expands the receptive field window exponentially by using dilated convolution with dilation coefficients of 1, 2, 4, and 8.

[0019] Furthermore, the obtained local multi-scale feature matrix is ​​input into the bidirectional gated cyclic unit layer BiGRU. Temporal dependencies are calculated synchronously through forward and backward gated cyclic units, outputting a global feature latent state matrix that integrates the macroscopic evolution trend of the stratigraphy. Specifically, this includes:

[0020] The local multi-scale feature matrix F localThe inputs are fed into the forward and backward GRUs respectively, and the forward and backward hidden states are obtained respectively.

[0021] By concatenating the forward and backward hidden states along the feature dimension and fusing the global temporal dependencies between upper and lower strata, we obtain the global feature hidden state matrix F. global .

[0022] Furthermore, the obtained global feature hidden state matrix is ​​input into the multi-head self-attention mechanism layer AM, where scaled dot product operations and feature weighted fusion are performed to output the final target logging curve reconstruction vector, specifically including:

[0023] The input global feature latent state matrix is ​​mapped into query, key and value matrices through three independent linear transformation layers, respectively.

[0024] Calculate the matching similarity between the query matrix and the key matrix, and obtain the attention weights after normalization;

[0025] The attention weights are summed with the value matrix to obtain the multi-head attention fusion feature matrix;

[0026] The fused feature matrix is ​​mapped to the logging curve space through a fully connected layer to obtain the target logging curve prediction vector y.

[0027] Furthermore, during the training phase, the improved sand cat swarm optimization algorithm is used to adaptively update the network hyperparameters of the temporal convolutional network layer TCN, the bidirectional gated recurrent unit layer BiGRU, and the multi-head self-attention mechanism layer AM, specifically including:

[0028] The improved sand cat swarm optimization algorithm executes a variable spiral search strategy based on the standard SSA Producer formula during the iterative optimization process; the population position update follows the joint calculation of the exponential decay formula and the logarithmic spiral formula.

[0029] The formula for exponential decay is as follows:

[0030]

[0031] in, Let i be the position of the j-th dimension of the i-th individual in the t-th iteration; This represents the position in the (t+1)th iteration. This is the parameter for controlling the iteration step size; This represents the maximum number of iterations.

[0032] The logarithmic spiral formula is as follows:

[0033]

[0034] in, This is the helical sensitivity coefficient. This represents the globally optimal position in the t-th iteration. It is a random number within the interval (0,1).

[0035] Furthermore, during the training phase, the improved sand cat swarm optimization algorithm is used to adaptively update the network hyperparameters of the temporal convolutional network layer TCN, the bidirectional gated recurrent unit layer BiGRU, and the multi-head self-attention mechanism layer AM, specifically including:

[0036] Gradient backpropagation is performed using a physical constraint loss function that includes a mean squared error term and a geological evolution penalty term. The loss function formula is as follows:

[0037]

[0038] in, This is the mean square error term. As a penalty item for geological evolution, This refers to the weighting coefficient of the penalty term;

[0039] The formula for calculating the mean square error term is:

[0040]

[0041] in, These are the measured well logging curve values; Predict curve values ​​for the model; This represents the number of depth sampling points;

[0042] The expression for the geological evolution penalty term is:

[0043]

[0044] in For lithological boundary continuity constraints, For the porosity-water saturation Archie equation constraint terms, The constraints are the Fourier series constraints of the stratigraphic sedimentary cycles. The mathematical expressions for each constraint are as follows, where λ1, λ2, and λ3 are the weighting coefficients corresponding to each constraint, used to balance the contribution of different constraint terms to the total loss function.

[0045] Furthermore, the lithological boundary continuity constraint loss term Defined as

[0046]

[0047] in, ∑ represents the number of depth sampling points; ∑ denotes the summation over all depth sampling points.

[0048] The constraints are as follows:

[0049]

[0050] in These are the well logging curve values ​​predicted by the model. For depth variables; corresponding constraints force the second derivative of the prediction curve to approach zero at adjacent depth sampling points, ensuring the smoothness of the reconstructed curve in the transition zone of lithological interfaces;

[0051] Porosity-Water Saturation Archie Equation Constraint Loss Term Defined as:

[0052]

[0053] in, This represents the number of depth sampling points; The values ​​of the water saturation curve predicted by the model; Lithology coefficient; The resistivity of formation water; To predict porosity; The cementation index; Formation resistivity; The saturation index is represented by ∑, which indicates summation over all depth sampling points.

[0054] The constraints are as follows:

[0055]

[0056] Implicitly related by the Archie equation; The values ​​of the water saturation curve predicted by the model; Lithology coefficient; The saturation index, The resistivity of formation water; Formation resistivity;

[0057] Stratigraphic Sedimentary Cycle Fourier Series Constrained Loss Term Defined as:

[0058]

[0059] in, y(z) represents the number of sampling points at depth z; y(z) represents the predicted logging curve value of the model at depth z. Let be the amplitude of the m-th harmonic component; m is the harmonic order, m=1,2,…,M, where M is the maximum order of the Fourier expansion; For fundamental frequency parameters; The phase of the m-th harmonic component; ∑ represents the summation over all depth sampling points;

[0060] The constraints are as follows:

[0061]

[0062] in, This represents the predicted logging curve value of the model at depth z; is the constant term of the Fourier series, reflecting the mean level of the formation logging response; m is the harmonic order, m=1,2,…,M, where M is the maximum order of the Fourier expansion; denoted as the amplitude of the m-th harmonic component, and f as the fundamental frequency parameter, reflecting the main cycle frequency of the strata deposition; The phase of the m-th harmonic component is given; the corresponding constraint forces the prediction curve to follow the periodic law of deposition cycles in the depth domain.

[0063] Furthermore, the method also includes:

[0064] The adaptive anomaly identification threshold is dynamically calculated based on the local mean and standard deviation of the reconstructed curve, and the residual between the measured logging curve and the reconstructed logging curve is calculated point by point. When the residual of multiple consecutive depth points exceeds the adaptive anomaly identification threshold, it is determined to be an abnormal area.

[0065] According to a second aspect, one embodiment provides a well logging curve reconstruction system based on the ISCSO-TCN-BiGRU-AM architecture, the system comprising:

[0066] The measured data acquisition and processing module is used to acquire multi-source measured well logging data containing missing or abnormal segments and preprocess it to construct a time series tensor.

[0067] The model building module is used to build a logging curve reconstruction model, which includes a temporal convolutional network layer (TCN), a bidirectional gated recurrent unit layer (BiGRU), and a multi-head self-attention mechanism layer (AM).

[0068] The time series tensor is input into the temporal convolutional network layer TCN, and local multi-scale feature matrices carrying high-frequency geological mutation information are extracted through causal convolution and dilatation convolution.

[0069] The obtained local multi-scale feature matrix is ​​input into the bidirectional gated cyclic unit layer BiGRU. The temporal dependence is calculated synchronously through forward and backward gated cyclic units, and the global feature hidden state matrix that integrates the macroscopic evolution trend of the stratigraphy is output.

[0070] The obtained global feature hidden state matrix is ​​input into the multi-head self-attention mechanism layer AM, and scaled dot product operation and feature weighted fusion are performed to output the final target logging curve reconstruction vector.

[0071] The model training module is used to adaptively update the network hyperparameters of the Temporal Convolutional Network (TCN), the Bidirectional Gated Recurrent Unit (BiGRU) layer, and the Multi-Head Self-Attention Mechanism (AM) layer during the training phase using an improved sand cat swarm optimization algorithm.

[0072] This invention provides a well logging curve reconstruction method based on the ISCSO-TCN-BiGRU-AM architecture. It acquires multi-source measured well logging data and performs dimensionless preprocessing to construct a time-series tensor. The time-series tensor is input into a temporal convolutional network layer, where local multi-scale geophysical feature matrices are extracted through causal convolution and dilatational convolution. The feature matrix is ​​then input into a bidirectional gated recurrent unit layer, where global temporal dependencies are established by concatenating forward and backward state tensors, outputting a hidden state matrix. This hidden state matrix is ​​then input into a multi-head self-attention mechanism layer, where scaling dot product operations and feature weighted fusion are performed to reconstruct and output the target well logging curve prediction vector. During the training phase, an improved sand cat swarm optimization algorithm is used to adaptively update the network hyperparameters, and a physical constraint loss function including a geological evolution penalty term is employed for model optimization. This invention can complete bidirectional inference and completion of well logging curves with extremely low errors and accurately identify fracture-cavity anomalies through residual comparison, demonstrating high geological interpretation value. It is suitable for electronic devices and computer-readable storage media. Attached Figure Description

[0073] Figure 1 A flowchart illustrating a well logging curve reconstruction method based on the ISCSO-TCN-BiGRU-AM architecture, provided as an embodiment of the present invention;

[0074] Figure 2 An overall architecture diagram of a well logging curve reconstruction method based on the ISCSO-TCN-BiGRU-AM architecture provided in one embodiment of the present invention;

[0075] Figure 3 A schematic diagram of the temporal convolutional network TCN in a well logging curve reconstruction method based on the ISCSO-TCN-BiGRU-AM architecture provided in an embodiment of the present invention;

[0076] Figure 4 A comparison diagram of the measured curve and the reconstructed curve in a well logging curve reconstruction method based on the ISCSO-TCN-BiGRU-AM architecture provided in an embodiment of the present invention;

[0077] Figure 5 The flowchart of the ISCSO algorithm in a well logging curve reconstruction method based on the ISCSO-TCN-BiGRU-AM architecture is provided as an embodiment of the present invention. Detailed Implementation

[0078] The present invention will now be described in further detail with reference to specific embodiments and accompanying drawings. Similar elements in different embodiments are referred to by associated similar element reference numerals. In the following embodiments, many details are described to facilitate a better understanding of the invention. However, those skilled in the art will readily recognize that some features may be omitted in different situations, or may be replaced by other elements, materials, or methods. In some cases, certain operations related to the present invention are not shown or described in the specification. This is to avoid obscuring the core parts of the invention with excessive description. For those skilled in the art, detailed description of these related operations is not necessary; they can fully understand the related operations based on the description in the specification and general technical knowledge in the art.

[0079] Furthermore, the features, operations, or characteristics described in the specification can be combined in any suitable manner to form various embodiments. At the same time, the steps or actions in the method description can be rearranged or adjusted in a manner obvious to those skilled in the art. Therefore, the various orders in the specification and drawings are only for the clear description of a particular embodiment and do not imply a necessary order, unless otherwise stated that a particular order must be followed.

[0080] The first embodiment of this invention provides a well logging curve reconstruction method based on the ISCSO-TCN-BiGRU-AM architecture. Its core lies in capturing local abrupt changes and global evolution patterns of formation sequences through a deep learning model, and using an improved intelligent algorithm to achieve adaptive hyperparameter tuning. The overall model architecture consists of a TCN local feature extraction layer, a BiGRU global modeling layer, an AM feature weighting layer, and an ISCSO intelligent optimization layer. The following describes the method in conjunction with... Figure 1 Please provide a detailed explanation.

[0081] The core technical solution of this invention is described in the following table, showing the modules at each layer and the data flow:

[0082]

[0083] The architecture is described below: The ISCSO-TCN-BiGRU-AM architecture used in this invention is a vertically coupled deep learning system (see attached diagram). Figure 2 ):

[0084] 1. Input layer: Processes T×C dimensional time series tensors.

[0085] 2. TCN layer: Composed of multiple stacked dilated convolution residual blocks, it is responsible for capturing local mutations in the geological sequence.

[0086] 3. BiGRU layer: Located after the TCN layer, it uses a bidirectional cyclic structure to establish long-range stratigraphic dependence.

[0087] 4. AM layer: As a feature fusion layer, it automatically identifies the feature dimension that contributes the most to the reconstruction target through a multi-head self-attention mechanism.

[0088] 5. Optimization layer (outer loop): The ISCSO algorithm acts as the outer envelope, responsible for heuristic global optimization of key hyperparameters such as parameter size and learning rate of the above layers.

[0089] like Figure 1 As shown, in step S100, multi-source measured well logging data containing missing or abnormal segments are acquired and preprocessed to construct a time series tensor.

[0090] The above steps specifically include:

[0091] This step aims to eliminate dimensional differences between multi-source logging data, providing a standard tensor input for deep learning networks.

[0092] S110: Raw Data Acquisition and Cleaning

[0093] Input: A collection of raw sequences of multi-source measured well logging data, including missing or anomalous segments (Xraw);

[0094] Intermediate processing: The original data is deeply aligned, and missing intervals (NaN values) are identified and marked; the local mean method is used to perform preliminary smoothing of obvious outliers.

[0095] Output: Aligned multidimensional logging sequence.

[0096] S120: Dimensionless Standardization Transformation

[0097] Input: The logging sequence output in step S110;

[0098] Intermediate processing: Perform a standardization transformation on each logging channel. The calculation formula is as follows:

[0099]

[0100] Where, x i μ is the original observation value of the i-th logging curve; i σ is the arithmetic mean of all data in this channel. i x is the standard deviation of the channel; i ′ represents the dimensionless value after processing.

[0101] Output: The completed time series tensor (Where T is the number of depth sampling points and C is the number of feature channels).

[0102] like Figure 1As shown, in step S200, a logging curve reconstruction model is constructed. The logging curve reconstruction model includes a temporal convolutional network layer (TCN), a bidirectional gated recurrent unit layer (BiGRU), and a multi-head self-attention mechanism layer (AM).

[0103] The above steps specifically include:

[0104] S210 inputs the time series tensor into the temporal convolutional network layer TCN, and extracts the local multi-scale feature matrix carrying high-frequency geological mutation information through causal convolution and dilatation convolution.

[0105] In this embodiment, a temporal convolutional network (TCN) is used to extract local high-frequency response features of the formation (such as thin-layer identification features). Figure 3 As shown, when performing local feature extraction, the temporal convolutional network layer strictly adheres to the principle of causal convolution, ensuring that the prediction of the current depth node depends only on the data nodes of the current and overlying strata. Simultaneously, it exponentially expands the receptive field window using dilation convolutions with dilation coefficient sequences of 1, 2, 4, and 8. Specifically:

[0106] S211: Causal Dilated Convolution Operation

[0107] Input: The time series tensor X output by S100.

[0108] Intermediate processing: Residual blocks composed of causal convolution and dilated convolution are used. Causal convolution ensures that the output at time t depends only on the input at time t and before; dilated convolution expands the receptive field by introducing a dilation coefficient d. The convolution calculation formula is:

[0109]

[0110] in:

[0111] The convolutional output of the l-th layer at depth t;

[0112] The weight parameters of the k-th convolutional kernel in the l-th layer;

[0113] : The expansion coefficient of the l-th layer (in this scheme, it is set to 1, 2, 4, 8, 16, which increases exponentially);

[0114] K: Kernel Size;

[0115] : The bias term of the l-th layer.

[0116] Output: Local multi-scale feature matrix Flocal .

[0117] S220 inputs the obtained local multi-scale feature matrix into the bidirectional gated cyclic unit layer BiGRU, and calculates the temporal dependence synchronously through forward and backward gated cyclic units, outputting a global feature hidden state matrix that integrates the macroscopic evolution trend of the stratigraphy.

[0118] In this embodiment, based on local features, a bidirectional gated cyclic unit (BiGRU) is used to capture the macroscopic evolution trend of the stratigraphic sedimentary sequence. The BiGRU layer controls the retention level of historical stratigraphic information through an update gate and controls the amount of current depth node information written through a reset gate. The calculated forward and backward time series tensors are then concatenated and spliced ​​along the feature dimension. Specifically:

[0119] S221: Bidirectional State Parallel Computation

[0120] Input: The feature matrix Flocal output by S210.

[0121] Intermediate processing: The sequence is scanned synchronously through two GRU units, one forward and one backward. For each unit, the internal state update logic is as follows:

[0122] 1. Update Gate: Control the proportion of historical information retained.

[0123] 2. Reset the door: This controls the degree to which the current input information is written.

[0124] 3. Candidate hidden state: .

[0125] 4. Final Hidden State: .

[0126] in, To update the gate output, To reset the gate output, σ is the Sigmoid activation function, ⊙ is the element-wise multiplication operation, and W, U, and b are the corresponding weight matrix, recursive weight matrix, and bias term, respectively.

[0127] Output: Forward state h t With backward state .

[0128] S222: Feature Dimension Cascaded Concatenation

[0129] Input: The bidirectional state tensor output by S221;

[0130] Intermediate processing: Concatenate the bidirectional hidden states along the feature dimension: Ht=h t ;ht .

[0131] Output: Global feature matrix Fglobal, which incorporates bidirectional temporal information.

[0132] S230 inputs the obtained global feature hidden state matrix into the multi-head self-attention mechanism layer AM, performs scaling dot product operation and feature weighted fusion, and outputs the final target logging curve reconstruction vector.

[0133] High-precision reconstruction is achieved by assigning importance to features at different depths through an attention mechanism. The multi-head self-attention mechanism layer transforms the hidden state matrix into a query space matrix, a key space matrix, and a value space matrix through linear mapping, and calculates the point integral distribution in the query space and key space to determine the attention weights. Specifically:

[0134] S231: Linear Mapping and Score Calculation

[0135] Input: The global feature matrix Fglobal output by S220;

[0136] Intermediate processing: The input is mapped to the query space (Q), key space (K), and value space (V) through three independent linear transformation layers:

[0137]

[0138] Calculate the scaled dot product attention weights:

[0139]

[0140] in, Let d be the mapping weight matrix to be trained; k is the dimension of the key vector, used to prevent gradient vanishing.

[0141] Output: Attention weight distribution matrix A.

[0142] S232: Feature Weighting and Vector Output

[0143] Input: Weight matrix A and value matrix V.

[0144] Intermediate processing: Perform weighted summation It is then mapped back to the target logging space through a fully connected layer.

[0145] Output: y, the predicted vector of the target logging curve.

[0146] like Figure 1 As shown, in step S300, during the training phase, the network hyperparameters of the Temporal Convolutional Network (TCN), the Bidirectional Gated Recurrent Unit (BiGRU) layer, and the Multi-Head Self-Attention Mechanism (AM) layer are adaptively updated using an improved sand cat swarm optimization algorithm.

[0147] The above steps specifically include:

[0148] In this embodiment, the improved Sand Cat Group Algorithm (ISCSO) is used to globally optimize the model's hyperparameters, and geophysical constraints are introduced, such as... Figure 5 As shown. The ISCSO algorithm is built on the standard Sparrow Search Algorithm (SSA) framework and introduces a Logistic chaotic mapping to initialize the sand cat (sparrow) population.

[0149] S310: Construction of Physical Constraint Loss Function

[0150] Input: Predicted vector y and measured target sequence y.

[0151] Intermediate processing: Constructing a total loss function that includes a geological evolution penalty term:

[0152]

[0153] in, This is the mean square error term. As a penalty item for geological evolution, This refers to the weighting coefficient of the penalty term;

[0154] The formula for calculating the mean square error term is:

[0155]

[0156] in, These are the measured well logging curve values; Predict curve values ​​for the model; This represents the number of depth sampling points;

[0157] The expression for the geological evolution penalty term is:

[0158]

[0159] in For lithological boundary continuity constraints, For the porosity-water saturation Archie equation constraint terms, The constraints are the Fourier series constraints of the stratigraphic sedimentary cycles. The mathematical expressions for each constraint are as follows, where λ1, λ2, and λ3 are the weight coefficients corresponding to each constraint, used to balance the contribution of different constraint terms to the total loss function.

[0160] Lithological boundary continuity constraint loss term Defined as:

[0161]

[0162] in, ∑ represents the number of depth sampling points; ∑ denotes the summation over all depth sampling points.

[0163] The constraints are as follows:

[0164]

[0165] in These are the well logging curve values ​​predicted by the model. For depth variables; corresponding constraints force the second derivative of the prediction curve to approach zero at adjacent depth sampling points, ensuring the smoothness of the reconstructed curve in the transition zone of lithological interfaces;

[0166] Porosity-Water Saturation Archie Equation Constraint Loss Term Defined as:

[0167]

[0168] in, This represents the number of depth sampling points; The values ​​of the water saturation curve predicted by the model; Lithology coefficient; The resistivity of formation water; To predict porosity; The cementation index; Formation resistivity; The saturation index is represented by ∑, which indicates summation over all depth sampling points.

[0169] The constraints are as follows:

[0170]

[0171] Implicitly related by the Archie equation; The values ​​of the water saturation curve predicted by the model; Lithology coefficient; The saturation index, The resistivity of formation water; Formation resistivity;

[0172] Stratigraphic Sedimentary Cycle Fourier Series Constrained Loss Term Defined as:

[0173]

[0174] in, y(z) represents the number of sampling points at depth z; y(z) represents the predicted logging curve value of the model at depth z. Let be the amplitude of the m-th harmonic component; m is the harmonic order, m=1,2,…,M, where M is the maximum order of the Fourier expansion; For fundamental frequency parameters; The phase of the m-th harmonic component; ∑ represents the summation over all depth sampling points;

[0175] The constraints are as follows:

[0176]

[0177] in, This represents the predicted logging curve value of the model at depth z; is the constant term of the Fourier series, reflecting the mean level of the formation logging response; m is the harmonic order, m=1,2,…,M, where M is the maximum order of the Fourier expansion; denoted as the amplitude of the m-th harmonic component, and f as the fundamental frequency parameter, reflecting the main cycle frequency of the strata deposition; The phase of the m-th harmonic component is given; the corresponding constraint forces the prediction curve to follow the periodic law of deposition cycles in the depth domain.

[0178] Output: Fitness rating.

[0179] S320: ISCSO Variable Spiral Strategy Iteration

[0180] Input: A set of hyperparameters to be optimized (such as learning rate, number of convolution kernels, number of hidden units).

[0181] Intermediate processing: The sand cat population performs a variable spiral search based on the current best individual position. The position update formula follows a joint calculation:

[0182] Strategy 1 (Exponential Decline):

[0183] Strategy 2 (Logarithmic Spiral):

[0184] in, For the t-th generation position; T max is the maximum number of iterations; S is the spiral sensitivity coefficient; u is a random number in the interval (0,1); This represents the current globally optimal parameter combination.

[0185] Output: Optimal network hyperparameters, final reconstructed curve, and anomaly identification results based on the residuals between the measured logging curve and the reconstructed logging curve.

[0186] Furthermore, the method also includes:

[0187] Step S400: Dynamically calculate the adaptive anomaly identification threshold based on the local mean and standard deviation of the reconstructed curve, and calculate the residual between the measured logging curve and the reconstructed logging curve point by point. When the residual of multiple consecutive depth points exceeds the adaptive anomaly identification threshold, it is determined to be an abnormal area.

[0188] In this embodiment, the adaptive anomaly detection threshold function is set to ,in and These represent the local mean and standard deviation of the reconstructed residual sequence, respectively. The residuals at three consecutive depth sampling points break through... This is considered an abnormal interface. For example... Figure 4 As shown.

[0189] After testing, the root mean square error (RMSE) of this invention is 0.032, the mean absolute error (MAE) is 0.024, and the coefficient of determination is... The score reached 0.965, which represents a significant improvement over existing technologies in core evaluation indicators.

[0190] Corresponding to the aforementioned well logging curve reconstruction method based on the ISCSO-TCN-BiGRU-AM architecture, this invention also discloses a well logging curve reconstruction system based on the ISCSO-TCN-BiGRU-AM architecture, which specifically includes:

[0191] The measured data acquisition and processing module is used to acquire multi-source measured well logging data containing missing or abnormal segments and preprocess it to construct a time series tensor.

[0192] The model building module is used to build a logging curve reconstruction model, which includes a temporal convolutional network layer (TCN), a bidirectional gated recurrent unit layer (BiGRU), and a multi-head self-attention mechanism layer (AM).

[0193] The time series tensor is input into the temporal convolutional network layer TCN, and local multi-scale feature matrices carrying high-frequency geological mutation information are extracted through causal convolution and dilatation convolution.

[0194] The obtained local multi-scale feature matrix is ​​input into the bidirectional gated cyclic unit layer BiGRU. The temporal dependence is calculated synchronously through forward and backward gated cyclic units, and the global feature hidden state matrix that integrates the macroscopic evolution trend of the stratigraphy is output.

[0195] The obtained global feature hidden state matrix is ​​input into the multi-head self-attention mechanism layer AM, and scaled dot product operation and feature weighted fusion are performed to output the final target logging curve reconstruction vector.

[0196] The model training module is used to adaptively update the network hyperparameters of the Temporal Convolutional Network (TCN), the Bidirectional Gated Recurrent Unit (BiGRU) layer, and the Multi-Head Self-Attention Mechanism (AM) layer during the training phase using an improved sand cat swarm optimization algorithm.

[0197] Furthermore, the system also includes:

[0198] The anomaly identification module is used to dynamically calculate the adaptive anomaly identification threshold based on the local mean and standard deviation of the reconstructed curve, and to calculate the residual between the measured logging curve and the reconstructed logging curve point by point. When the residual of multiple consecutive depth points exceeds the adaptive anomaly identification threshold, it is determined to be an abnormal area.

[0199] It should be noted that for a detailed description of the logging curve reconstruction system based on the ISCSO-TCN-BiGRU-AM architecture provided in the embodiments of the present invention, please refer to the relevant description of the logging curve reconstruction method based on the ISCSO-TCN-BiGRU-AM architecture provided in the embodiments of the present invention, which will not be repeated here.

[0200] In addition, embodiments of the present invention also provide an electronic device, the device comprising: a processor and a memory; the memory being used to store one or more program instructions; the processor being used to execute one or more program instructions to perform the steps of a well logging curve reconstruction method based on the ISCSO-TCN-BiGRU-AM architecture as described in any of the preceding embodiments.

[0201] It should be noted that for a detailed description of the electronic device provided in the embodiments of the present invention, please refer to the relevant description of a well logging curve reconstruction method based on the ISCSO-TCN-BiGRU-AM architecture provided in the embodiments of this application, which will not be repeated here.

[0202] In addition, embodiments of the present invention also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of a well logging curve reconstruction method based on the ISCSO-TCN-BiGRU-AM architecture as described in any of the preceding claims.

[0203] It should be noted that for a detailed description of a computer-readable storage medium provided in the embodiments of the present invention, please refer to the relevant description of a well logging curve reconstruction method based on the ISCSO-TCN-BiGRU-AM architecture provided in the embodiments of this application, which will not be repeated here.

[0204] Those skilled in the art will understand that all or part of the functions of the various methods in the above embodiments can be implemented by hardware or by computer programs. When all or part of the functions in the above embodiments are implemented by computer programs, the program can be stored in a computer-readable storage medium, which may include: read-only memory, random access memory, disk, optical disk, hard disk, etc., and the program is executed by a computer to achieve the above functions. For example, the program can be stored in the memory of a device, and when the program in the memory is executed by the processor, all or part of the above functions can be achieved. In addition, when all or part of the functions in the above embodiments are implemented by computer programs, the program can also be stored in a server, another computer, disk, optical disk, flash drive, or external hard drive, etc., and can be downloaded or copied to the memory of a local device, or the system of the local device can be updated. When the program in the memory is executed by the processor, all or part of the functions in the above embodiments can be achieved.

[0205] The above examples illustrate the present invention only to aid in understanding it and are not intended to limit the scope of the invention. Those skilled in the art can make various simple deductions, modifications, or substitutions based on the principles of this invention.

Claims

1. A well logging curve reconstruction method based on the ISCSO-TCN-BiGRU-AM architecture, characterized in that, The method includes: Acquire multi-source measured well logging data containing missing or anomalous segments and preprocess them to construct a time series tensor; A well logging curve reconstruction model is constructed, which includes a temporal convolutional network layer (TCN), a bidirectional gated recurrent unit layer (BiGRU), and a multi-head self-attention mechanism layer (AM). The time series tensor is input into the temporal convolutional network layer TCN, and local multi-scale feature matrices carrying high-frequency geological mutation information are extracted through causal convolution and dilatation convolution. The obtained local multi-scale feature matrix is ​​input into the bidirectional gated cyclic unit layer BiGRU. The temporal dependence is calculated synchronously through forward and backward gated cyclic units, and the global feature hidden state matrix that integrates the macroscopic evolution trend of the stratigraphy is output. The obtained global feature hidden state matrix is ​​input into the multi-head self-attention mechanism layer AM, and scaled dot product operation and feature weighted fusion are performed to output the final target logging curve reconstruction vector. During the training phase, the improved sand cat swarm optimization algorithm is used to adaptively update the network hyperparameters of the temporal convolutional network layer TCN, the bidirectional gated recurrent unit layer BiGRU, and the multi-head self-attention mechanism layer AM.

2. The well logging curve reconstruction method based on the ISCSO-TCN-BiGRU-AM architecture as described in claim 1, characterized in that, Acquire multi-source measured well logging data containing missing or anomalous segments and preprocess them to construct time series tensors, specifically including: The original measured well logging data were depth aligned, and missing intervals were identified and marked; the local mean method was used to perform preliminary smoothing of obvious outlier noise. Based on the aligned multidimensional logging sequence, a standardization transformation is performed on each logging channel, and a time series tensor is constructed. Where T is the number of depth sampling points and C is the number of feature channels.

3. The well logging curve reconstruction method based on the ISCSO-TCN-BiGRU-AM architecture as described in claim 1, characterized in that, The time series tensor is input into the temporal convolutional network (TCN) layer. Local multi-scale feature matrices carrying high-frequency geological abrupt change information are extracted through causal convolution and dilation convolution. Specifically, this includes: The time series tensor X is input into multiple stacked dilated convolution residual blocks consisting of causal convolution and dilated convolution for processing, and the local multi-scale feature matrix F is extracted. local Among them, causal convolution ensures that the output at time t depends only on the input at time t and before; dilated convolution expands the receptive field window exponentially by using dilated convolution with dilation coefficients of 1, 2, 4, and 8.

4. The well logging curve reconstruction method based on the ISCSO-TCN-BiGRU-AM architecture as described in claim 1, characterized in that, The obtained local multi-scale feature matrix is ​​input into the BiGRU bidirectional gated cyclic unit layer. The temporal dependency is calculated synchronously through forward and backward gated cyclic units, and the global feature latent state matrix that integrates the macroscopic evolution trend of the stratigraphy is output, specifically including: The local multi-scale feature matrix F local The inputs are fed into the forward and backward GRUs respectively, and the forward and backward hidden states are obtained respectively. By concatenating the forward and backward hidden states along the feature dimension and fusing the global temporal dependencies between upper and lower strata, we obtain the global feature hidden state matrix F. global .

5. The well logging curve reconstruction method based on the ISCSO-TCN-BiGRU-AM architecture as described in claim 1, characterized in that, The obtained global feature hidden state matrix is ​​input into the multi-head self-attention mechanism layer AM, where scaling dot product operations and feature weighted fusion are performed to output the final target logging curve reconstruction vector, specifically including: The input global feature latent state matrix is ​​mapped into query, key and value matrices through three independent linear transformation layers, respectively. Calculate the matching similarity between the query matrix and the key matrix, and obtain the attention weights after normalization; The attention weights are summed with the value matrix to obtain the multi-head attention fusion feature matrix; The fused feature matrix is ​​mapped to the logging curve space through a fully connected layer to obtain the target logging curve prediction vector y.

6. The well logging curve reconstruction method based on the ISCSO-TCN-BiGRU-AM architecture as described in claim 1, characterized in that, During the training phase, the improved sand cat swarm optimization algorithm is used to adaptively update the network hyperparameters of the temporal convolutional network layer TCN, the bidirectional gated recurrent unit layer BiGRU, and the multi-head self-attention mechanism layer AM. Specifically, this includes: The improved sand cat swarm optimization algorithm executes a variable spiral search strategy based on the standard SSA Producer formula during the iterative optimization process; the population position update follows the joint calculation of the exponential decay formula and the logarithmic spiral formula. The formula for exponential decay is as follows: in, Let i be the position of the j-th dimension of the i-th individual in the t-th iteration; This represents the position in the (t+1)th iteration. This is the parameter for controlling the iteration step size; This represents the maximum number of iterations. The logarithmic spiral formula is as follows: in, This is the helical sensitivity coefficient. This represents the globally optimal position in the t-th iteration. It is a random number within the interval (0,1).

7. The well logging curve reconstruction method based on the ISCSO-TCN-BiGRU-AM architecture as described in claim 1, characterized in that, During the training phase, the improved sand cat swarm optimization algorithm is used to adaptively update the network hyperparameters of the temporal convolutional network layer TCN, the bidirectional gated recurrent unit layer BiGRU, and the multi-head self-attention mechanism layer AM. Specifically, this includes: Gradient backpropagation is performed using a physical constraint loss function that includes a mean squared error term and a geological evolution penalty term. The loss function formula is as follows: in, This is the mean square error term. As a penalty item for geological evolution, This refers to the weighting coefficient of the penalty term; The formula for calculating the mean square error term is: in, These are the measured well logging curve values; Predict curve values ​​for the model; This represents the number of depth sampling points; The expression for the geological evolution penalty term is: in For lithological boundary continuity constraints, For the porosity-water saturation Archie equation constraint terms, The constraints are the Fourier series constraints of the stratigraphic sedimentary cycles. The mathematical expressions for each constraint are as follows, where λ1, λ2, and λ3 are the weighting coefficients corresponding to each constraint, used to balance the contribution of different constraint terms to the total loss function.

8. The well logging curve reconstruction method based on the ISCSO-TCN-BiGRU-AM architecture as described in claim 7, characterized in that, Lithological boundary continuity constraint loss term Defined as: in, ∑ represents the number of depth sampling points; ∑ indicates summation over all depth sampling points. These are the well logging curve values ​​predicted by the model. For depth variables; Porosity-Water Saturation Archie Equation Constraint Loss Term Defined as: in, This represents the number of depth sampling points; The values ​​of the water saturation curve predicted by the model; Lithology coefficient; The resistivity of formation water; To predict porosity; The cementation index; Formation resistivity; The saturation index is represented by ∑, which indicates summation over all depth sampling points. Stratigraphic Sedimentary Cycle Fourier Series Constrained Loss Term Defined as: in, y(z) represents the number of sampling points at depth z; y(z) represents the predicted logging curve value of the model at depth z. For the constant term of the Fourier series; Let be the amplitude of the m-th harmonic component; m is the harmonic order, m=1,2,…,M, where M is the maximum order of the Fourier expansion; For fundamental frequency parameters; ∑ represents the phase of the m-th harmonic component; ∑ represents the summation over all depth sampling points.

9. The well logging curve reconstruction method based on the ISCSO-TCN-BiGRU-AM architecture as described in claim 1, characterized in that, The method further includes: The adaptive anomaly identification threshold is dynamically calculated based on the local mean and standard deviation of the reconstructed curve, and the residual between the measured logging curve and the reconstructed logging curve is calculated point by point. When the residual of multiple consecutive depth points exceeds the adaptive anomaly identification threshold, it is determined to be an abnormal area.

10. A well logging curve reconstruction system based on the ISCSO-TCN-BiGRU-AM architecture, characterized in that, The system includes: The measured data acquisition and processing module is used to acquire multi-source measured well logging data containing missing or abnormal segments and preprocess it to construct a time series tensor. The model building module is used to build a logging curve reconstruction model, which includes a temporal convolutional network layer (TCN), a bidirectional gated recurrent unit layer (BiGRU), and a multi-head self-attention mechanism layer (AM). The time series tensor is input into the temporal convolutional network layer TCN, and local multi-scale feature matrices carrying high-frequency geological mutation information are extracted through causal convolution and dilatation convolution. The obtained local multi-scale feature matrix is ​​input into the bidirectional gated cyclic unit layer BiGRU. The temporal dependence is calculated synchronously through forward and backward gated cyclic units, and the global feature hidden state matrix that integrates the macroscopic evolution trend of the stratigraphy is output. The obtained global feature hidden state matrix is ​​input into the multi-head self-attention mechanism layer AM, and scaled dot product operation and feature weighted fusion are performed to output the final target logging curve reconstruction vector. The model training module is used to adaptively update the network hyperparameters of the Temporal Convolutional Network (TCN), the Bidirectional Gated Recurrent Unit (BiGRU) layer, and the Multi-Head Self-Attention Mechanism (AM) layer during the training phase using an improved sand cat swarm optimization algorithm.