A method for predicting rock mechanical parameters in complex strata

Abstract

This invention relates to the field of artificial intelligence technology and discloses a method for predicting rock mechanical parameters in complex strata. The method includes: collecting multi-source data from complex strata; preprocessing the multi-source data; extracting physical and statistical features related to rock mechanical parameters based on the multi-source data; aligning the physical features, statistical features, and preprocessed multi-source data by depth to obtain a high-dimensional feature matrix; constructing an ensemble machine learning model; training the ensemble machine learning model using the high-dimensional feature matrix, wherein the ensemble machine learning model includes a bottom-level model and a top-level model, and the bottom-level model includes a temporal modeling unit and a nonlinear regression unit; optimizing the model output of the ensemble machine learning model based on dynamic weight adjustment and a physical constraint loss function to obtain the predicted rock mechanical parameters. This invention can improve the accuracy of predicting rock mechanical parameters in complex strata.

Description

Technical Field

[0001] This invention relates to the field of integrated learning model technology, and in particular to a method for predicting rock mechanical parameters in complex strata. Background Technology

[0002] Traditional prediction of rock mechanical parameters in complex formations usually relies on a single data source (such as well logging data or core experiments), which has problems such as large data noise interference and incomplete characterization of heterogeneous formations.

[0003] Furthermore, traditional machine learning models lack physical constraints, making their predictions prone to deviating from geomechanical laws and resulting in low reliability. Simultaneously, existing methods are inefficient when processing multi-source heterogeneous data, exhibiting poor dynamic adaptability and failing to meet the demands of real-time drilling conditions. Therefore, there is an urgent need for a rock mechanics parameter prediction method that integrates multi-source data, embeds physical constraints, and possesses dynamic optimization capabilities. Summary of the Invention

[0004] This invention provides a method for predicting the mechanical parameters of rocks in complex formations, the main purpose of which is to solve the problem of low accuracy in predicting the mechanical parameters of rocks in complex formations.

[0005] To achieve the above objectives, the present invention provides a method for predicting the mechanical parameters of rocks in complex strata, comprising:

[0006] S1. Collect multi-source data of complex strata and perform data preprocessing on the multi-source data, wherein the multi-source data includes: well logging data, drilling parameters, seismic inversion data and core test data;

[0007] S2. Based on the multi-source data, extract physical and statistical features related to rock mechanics parameters, and align the physical features, statistical features, and preprocessed multi-source data by depth to obtain a high-dimensional feature matrix;

[0008] S3. Construct an ensemble machine learning model and train the ensemble machine learning model using the high-dimensional feature matrix. The ensemble machine learning model includes a bottom-level model and a top-level model. The bottom-level model includes a time-series modeling unit and a nonlinear regression unit.

[0009] S4. Optimize the model output of the integrated machine learning model based on dynamic weight adjustment and physical constraint loss function to obtain the prediction results of rock mechanics parameters.

[0010] Optionally, the acquisition of multi-source data from complex strata includes:

[0011] Well logging data of the complex formation was collected using logging instruments, including sonic transit time curves, density curves, neutron porosity curves, and natural gamma curves.

[0012] Drilling parameters of the complex formation are acquired in real time using drilling sensors, including drilling pressure, rotation speed, mechanical drilling speed and drill bit diameter;

[0013] Seismic inversion data of the complex strata, including wave impedance and porosity distribution, were obtained using seismic inversion technology.

[0014] Core experimental data of the complex strata were obtained through core experiments, including the laboratory core elastic modulus and Poisson's ratio.

[0015] Optionally, the preprocessing of the multi-source data includes:

[0016] The acoustic time difference curve is subjected to cycle skip noise filtering to obtain the reconstructed and filtered acoustic time difference curve;

[0017] Principal component analysis is used to perform feature dimensionality reduction on the reconstructed filtered acoustic transit time curve, the density curve, the neutron porosity curve, and the natural gamma curve to obtain the dimensionality reduction features of the logging data.

[0018] The drilling parameters are standardized to obtain their mechanical specific energy, wherein the calculation formula for the standardization process is as follows:

[0019]

[0020] Where, It is the mechanical specific energy, It's drilling pressure. It's the rotational speed. It is the mechanical drilling speed. It is the drill bit diameter. It is the statistical standard deviation of the mechanical specific energy of all wells in the work area.

[0021] Optionally, the step of filtering out cycle skip noise from the acoustic time-of-flight curve to obtain the reconstructed filtered acoustic time-of-flight curve includes:

[0022] The acoustic time difference curve is decomposed into high-frequency and low-frequency components by using the Daubechies wavelet basis function.

[0023] High-frequency noise in the high-frequency components is suppressed using a threshold method;

[0024] The high-frequency component after noise suppression and the low-frequency component are reconstructed to obtain the reconstructed filtered acoustic time difference curve.

[0025] Optionally, the network structure of the temporal modeling unit includes: an input layer, a bidirectional LSTM layer, and a multi-head attention layer, wherein the multi-head attention layer is configured with 4 attention heads to calculate the association weights of different depth points in complex strata.

[0026] Optionally, the nonlinear regression unit adopts a limit gradient boosting tree model, wherein the number of random forest decision trees in the limit gradient boosting tree model is dynamically increased according to the cumulative advance.

[0027] Optionally, the top-level model is a dynamic weighted random forest, wherein the weight adjustment formula of the dynamic weighted random forest is as follows:

[0028]

[0029] in, It is the first The weight values ​​of each sample, It is the first Elastic modulus of laboratory core samples. It is the first Elastic modulus of laboratory core samples. It is an adjustment coefficient that is dynamically adjusted based on the cumulative footage of the drill bit. It is the drill bit wear index. and It is a sample index. It represents the total number of samples.

[0030] Optionally, the physical constraint loss function is:

[0031]

[0032] in, It is the total loss value used to optimize the training of the ensemble machine learning model. , , These are weighting coefficients used to control the different constraint terms. It is the first The predicted elastic modulus of a sample It is the first Elastic modulus of laboratory core samples. It is the first The predicted stress gradient for each sample. It is the elastic modulus. It is Poisson's ratio. It is an indicator function. and It is a sample index. It is the total number of samples. This refers to the number of points used in the stress gradient calculation.

[0033] Optionally, the ensemble machine learning model further includes a transfer learning module, wherein the implementation of the transfer learning module includes:

[0034] Based on historical well data from 500 wells in 10 basins, a general rock mechanics prediction model was trained to extract common formation characteristics.

[0035] Using logging data and core experimental data from three wells in the target work area, the parameters of the first three layers of the LSTM network were frozen.

[0036] Using the latest 5-meter well logging data and core data from the target work area, the fully connected layer parameters were updated using the Adam optimizer, with the learning rate adjusted to 0.01 and the number of iterations set to 200.

[0037] Once triggered, a sliding window mechanism is used to perform incremental learning every 1 meter of drilling.

[0038] Optionally, the dynamic adjustment rule for the adjustment coefficient is as follows:

[0039] When the strata lithology is shale, the adjustment coefficient is corrected based on the fracture density, wherein the formula for calculating the coefficient correction is:

[0040]

[0041] in, This is the corrected adjustment coefficient. It is the adjustment coefficient, It refers to the crack density.

[0042] This invention significantly improves data quality and the comprehensiveness of input information through multi-source data fusion and preprocessing. Based on this, an integrated machine learning model architecture is designed, combining a temporal modeling unit (bidirectional LSTM and multi-head attention mechanism) and a nonlinear regression unit (XGBoost) to effectively capture the temporal correlation and static nonlinear characteristics of complex strata, greatly improving prediction accuracy. Simultaneously, to address the problem of data scarcity in the target work area, a dynamic weight adjustment and transfer learning module is introduced: sample weight allocation is optimized through a dynamic weighted random forest (combined with dynamic correction of drill bit wear index), and based on a pre-trained general model and incremental learning of the target work area (sliding window mechanism), the model's adaptability to new strata is enhanced. Furthermore, a physically constrained loss function (integrating mean square error, stress gradient smoothing, and empirical relational constraints) forces the prediction results to conform to geomechanical laws, avoiding non-physical oscillations or invalid outputs in traditional machine learning, significantly improving the reliability of the results. Finally, through real-time incremental learning and dynamic adjustment of the number of decision trees, the model is updated and optimized in real-time during drilling, improving prediction efficiency while completely solving the core problem of low accuracy in predicting rock mechanical parameters in complex strata. Attached Figure Description

[0043] Figure 1 A flowchart illustrating a method for predicting the mechanical parameters of rocks in complex formations according to an embodiment of the present invention;

[0044] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0045] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0046] This application provides a method for predicting rock mechanical parameters in complex formations. The execution entity of this method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the method for predicting rock mechanical parameters in complex formations can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0047] Reference Figure 1 The diagram shown is a flowchart illustrating a method for predicting rock mechanical parameters in complex formations according to an embodiment of the present invention. In this embodiment, the method for predicting rock mechanical parameters in complex formations includes:

[0048] S1. Collect multi-source data of complex strata and perform data preprocessing on the multi-source data, wherein the multi-source data includes: well logging data, drilling parameters, seismic inversion data and core test data.

[0049] In this embodiment of the invention, the acquisition of multi-source data from complex strata includes:

[0050] Well logging data of the complex formation was collected using logging instruments, including sonic transit time curves, density curves, neutron porosity curves, and natural gamma curves.

[0051] Drilling parameters of the complex formation are acquired in real time using drilling sensors, including drilling pressure, rotation speed, mechanical drilling speed and drill bit diameter;

[0052] Seismic inversion data of the complex strata, including wave impedance and porosity distribution, were obtained using seismic inversion technology.

[0053] Core experimental data of the complex strata were obtained through core experiments, including the laboratory core elastic modulus and Poisson's ratio.

[0054] In detail, well logging data reflects the continuous physical properties of the formation, but is affected by noise; drilling parameters reflect the mechanical response of the drilling process in real time; seismic inversion data provides the regional macroscopic structure; and core experimental data serves as a direct calibration benchmark.

[0055] In detail, the sonic transit time curve (Δt) reflects the propagation speed of sound waves in rocks and is used to calculate the elastic modulus; the density curve (DEN) measures the formation density through gamma rays, which affects the rock strength; the neutron porosity curve (CNL) characterizes the properties of pore fluids and indirectly reflects the degree of rock cementation; and the natural gamma curve (GR) indicates the clay content and distinguishes lithology.

[0056] In detail, the weight on drill bit (WOB) refers to the axial force applied by the drill bit, which affects the rock breaking efficiency; the rotation speed (RPM) refers to the drill bit's rotational speed, which is related to the mechanical drilling rate; the mechanical drilling rate (ROP) refers to the footage per unit time, reflecting the rock's drillability; and the drill bit diameter (D) affects the calculation of drilling energy consumption.

[0057] In detail, acoustic impedance (AI) refers to the product of acoustic velocity and density, reflecting the characteristics of the formation interface; porosity distribution is obtained through seismic attribute inversion, supplementing the regional deficiencies of well logging data.

[0058] In detail, the elastic modulus (E) refers to the stress-strain relationship of rock; Poisson's ratio (PR) is the ratio of transverse strain to longitudinal strain, reflecting the brittleness of rock.

[0059] For example: the data for a certain depth point includes:

[0060] Well logging data: Δt=85μs / ft, DEN=2.65g / cm³, CNL=15%, GR=45API.

[0061] Drilling parameters: MSE_norm=0.92.

[0062] Seismic inversion data: AI=12,000 (m / s·g / cm³), porosity=8%.

[0063] Core data: E=35GPa, PR=0.25.

[0064] In detail, multi-source heterogeneous data from complex strata were collected, including well logging data, drilling parameters, seismic inversion data, and core experimental data, and data preprocessing was performed to improve data quality. Complex strata are characterized by heterogeneity and high levels of data noise, and a single data source cannot fully reflect the strata characteristics. Multi-source data fusion can provide more comprehensive input information, while preprocessing eliminates noise interference and ensures the reliability of subsequent model training.

[0065] In detail, seismic inversion technology is used to obtain wave impedance and porosity distribution, providing information on the macroscopic physical properties of the formation and supplementing the local limitations of well logging and drilling data.

[0066] In detail, the elastic modulus and Poisson's ratio of the laboratory core provide realistic labels for model training and guide the design of physical constraints.

[0067] In this embodiment of the invention, the preprocessing of the multi-source data includes:

[0068] The acoustic time difference curve is subjected to cycle skip noise filtering to obtain the reconstructed and filtered acoustic time difference curve;

[0069] Principal component analysis is used to perform feature dimensionality reduction on the reconstructed filtered acoustic transit time curve, the density curve, the neutron porosity curve, and the natural gamma curve to obtain the dimensionality reduction features of the logging data.

[0070] The drilling parameters are standardized to obtain their mechanical specific energy, wherein the calculation formula for the standardization process is as follows:

[0071]

[0072] Where, It is the mechanical specific energy, It's drilling pressure. It's the rotational speed. It is the mechanical drilling speed. It is the drill bit diameter. It is the statistical standard deviation of the mechanical specific energy of all wells in the work area.

[0073] In detail, standardizing the drilling parameters can eliminate dimensional differences and make different drilling parameters comparable.

[0074] In detail, standardizing the drilling parameters means converting parameters such as drilling pressure and rotation speed into normalized mechanical specific energy. The original drilling parameters lack comparability due to differences in drill bit size and operating conditions. Mechanical specific energy reflects the energy consumption per unit footage and can uniformly characterize rock drillability.

[0075] In this embodiment of the invention, the step of filtering out cycle jump noise from the acoustic time difference curve to obtain the reconstructed filtered acoustic time difference curve includes:

[0076] The acoustic time difference curve is decomposed into high-frequency and low-frequency components by using the Daubechies wavelet basis function.

[0077] High-frequency noise in the high-frequency components is suppressed using a threshold method;

[0078] The high-frequency component after noise suppression and the low-frequency component are reconstructed to obtain the reconstructed filtered acoustic time difference curve.

[0079] In detail, because the acoustic transit time curve is prone to cycle jumps (high-frequency noise) in fractured formations, which affects the subsequent calculation of elastic parameters, the acoustic transit time curve is filtered to remove cycle jump noise, retain the true acoustic response of the formation, and improve the accuracy of logging data.

[0080] In detail, the Db4 wavelet basis is selected to decompose the acoustic time difference curve (Δt curve) into three levels to obtain high-frequency components (D1-D3) and low-frequency components (A3); a soft threshold (such as the VisuShrink threshold) is applied to the high-frequency components to suppress noise; the noise-reduced high-frequency components and low-frequency components are merged to obtain a smooth Δt curve.

[0081] For example, standardize the four curves (mean is 0, standard deviation is 1); calculate the covariance matrix, extract the first two principal components (cumulative contribution rate > 85%), reduce the feature dimension from 4 dimensions to 2 dimensions, and reduce the information loss to < 15%.

[0082] In detail, the original logging curves contain redundancy. Dimensionality reduction can reduce computational complexity and avoid overfitting.

[0083] S2. Based on the multi-source data, extract physical and statistical features related to rock mechanical parameters, and align the physical features, statistical features, and preprocessed multi-source data by depth to obtain a high-dimensional feature matrix.

[0084] In this embodiment of the invention, physical features (such as elastic modulus and Poisson's ratio) and statistical features (such as mean and variance) are extracted from multi-source data, and a high-dimensional feature matrix is ​​constructed by aligning the data by depth.

[0085] In detail, the elastic modulus is calculated using the sonic transit time curve (Δt) and density curve (DEN) to characterize the rock’s ability to resist deformation. A high elastic modulus indicates a hard stratum (such as flint).

[0086] In detail, Poisson's ratio (PR) is calculated as the ratio of lateral strain to axial strain. It reflects the lateral deformation capacity of a rock, and shale typically has a high Poisson's ratio (0.3-0.4).

[0087] In detail, statistical characteristics are extracted from the preprocessed logging curves to reflect the data distribution characteristics, including: mean, which refers to the average value of logging curves within a certain depth range, reflecting the overall level of formation physical properties (e.g., a high-density mean indicates a tight formation); variance, which refers to the dispersion of logging curve values, with a large variance indicating strong formation heterogeneity (e.g., a significantly increased variance in the natural gamma curve of fractured sections); skewness, which refers to the asymmetry of logging curve distribution, with positive skewness indicating that the data is concentrated in low-value areas (e.g., the porosity curve is right-skewed in tight layers); and kurtosis, which refers to the sharpness of logging curve distribution, with high kurtosis indicating concentrated data distribution (e.g., the neutron porosity curve of homogeneous sandstone has a high kurtosis).

[0088] For example: Perform statistical calculations on the preprocessed natural gamma curve (GR):

[0089] Mean = 60 API (indicating mudstone as the dominant strata)

[0090] Variance = 120 (reflects dramatic changes in lithology within the layer)

[0091] Skewness = 0.8 (Data is right-skewed, indicating the presence of a high-gamma anomaly layer)

[0092] Kurtosis = 4.2 (Concentrated distribution, with relatively strong homogeneity in some local segments)

[0093] In detail, complex stratigraphic features are multi-scale, requiring the integration of features from different sources to comprehensively characterize rock mechanical properties.

[0094] In detail, physical characteristics include direct mechanical parameters such as Young's modulus, Poisson's ratio, and compressive strength; statistical characteristics refer to the mean, variance, skewness, and kurtosis of the calculated logging curves, reflecting the data distribution characteristics.

[0095] In detail, the well logging data (after dimensionality reduction), standardized MSE, seismic inversion data, and core data are aligned by depth to form an N×M matrix (N is the number of sampling points, and M is the total number of features).

[0096] In detail, all data are aligned by depth sampling points (e.g., one point every 0.1 meters) to ensure spatiotemporal consistency. Drilling parameters, seismic inversion data, and core interpolation data are aligned based on the well logging data sampling interval (e.g., 0.1 meters).

[0097] In detail, assume that each depth point contains the following features: dimensionality-reduced logging features (2D, from PCA); normalized mechanical energy (MSE_norm, 1D); seismic inversion data (wave impedance, porosity, 2D); physical features (E, PR, UCS, 3D); statistical features (mean, variance, skewness, kurtosis, 4D), so the total number of features is 2+1+2+3+4=12 dimensions. If there are N sampling points in the work area, an N×12 high-dimensional feature matrix is ​​formed.

[0098] S3. Construct an ensemble machine learning model and train the ensemble machine learning model using the high-dimensional feature matrix. The ensemble machine learning model includes a bottom-level model and a top-level model. The bottom-level model includes a time-series modeling unit and a nonlinear regression unit.

[0099] In this embodiment of the invention, the network structure of the temporal modeling unit includes: an input layer, a bidirectional LSTM layer, and a multi-head attention layer, wherein the multi-head attention layer is configured with 4 attention heads to calculate the association weights of different depth points in complex strata.

[0100] In detail, the temporal modeling unit is used to capture the temporal correlation of complex strata (such as interbedded sandstone and mudstone, fracture development sequence). The input data of the input layer are well logging curves sorted by depth (such as sonic transit time Δt, density DEN, neutron porosity CNL, natural gamma ray GR) and preprocessed features (such as PCA dimensionality reduction results).

[0101] In detail, assuming that each depth point contains 4 logging features (2-dimensional after dimensionality reduction), the input dimension of the time series modeling unit is (sequence length, 2), and the sequence length is defined by a sliding window (e.g., every 10 depth points constitute one sequence).

[0102] In detail, the bidirectional long short-term memory network (Bi-LSTM) contains forward and backward LSTM units, each with 128 hidden units. The forward LSTM is used to capture temporal dependencies from shallow to deep (such as the influence of the upper strata on the lower strata); the backward LSTM is used to capture reverse correlations from deep to shallow (such as the feedback of the lower fractures on the stress distribution in the upper strata).

[0103] In detail, the output of the Bidirectional Long Short-Term Memory (Bi-LSTM) network is the hidden state of each depth point, where the output dimension is 256.

[0104] In detail, there are 4 independent attention heads (Multi-HeadAttention), each with a query (Q), key (K), and value (V) matrix dimension of 64. The outputs of the 4 heads are concatenated and fused through a linear layer, resulting in a final output dimension of 256.

[0105] In detail, the weights of different depth points are dynamically allocated, for example, the attention weights of fracture development segments (high Δt variance) or dense layers (high density) are higher; this enhances the model's ability to extract features from key layers.

[0106] In this embodiment of the invention, the nonlinear regression unit adopts the extreme gradient boosting tree model, wherein the number of random forest decision trees in the extreme gradient boosting tree model increases dynamically according to the cumulative advance.

[0107] In detail, an integrated model with a stacking architecture is constructed. The bottom-level model processes time-series and static features, while the top-level model dynamically fuses and predicts the results. Complex formation data has both time-series characteristics (well logging curves) and nonlinear characteristics (geological features), so it is necessary to combine the advantages of different models.

[0108] In detail, rock mechanical parameters are depth-dependent (e.g., interbedded sandstone and mudstone), LSTM is good at processing time-series data, and the attention mechanism can dynamically focus on key depth points.

[0109] In detail, a bidirectional LSTM layer (128 hidden units) captures the temporal correlation of logging curves; a multi-head attention layer (4 attention heads) calculates the correlation weights of different depth points, enhancing the ability to extract key features.

[0110] In detail, the Extreme Gradient Boosting Tree (XGBoost) model handles static geological features (such as density and porosity) with parameters set to maximum depth = 6 and learning rate = 0.1.

[0111] In detail, the number of decision trees in the Extreme Gradient Boosting (XGBoost) method dynamically increases with the cumulative drilling footage. In the early stages of new well drilling, the sample size is small, and a fixed number of trees is prone to underfitting; as the drilling footage increases, the number of trees is gradually increased to improve the fitting ability.

[0112] Specifically, when encountering fractured zones (fracture density ≥ 3 fractures / meter), the XGBoost model weights are set to 0.75, and the cumulative drilling footage (meters) increases dynamically with the number of random forest decision trees. The number of random forest decision trees is... ,in, It is the cumulative advance (in meters).

[0113] In detail, the basic parameters of Extreme Gradient Boosting Tree (XGBoost) include: learning rate (ηη) = 0.1, which controls the contribution weight of each tree; maximum depth = 6, which limits the tree complexity to avoid overfitting; and an initial number of 100 trees (to ensure basic fitting ability).

[0114] In detail, cumulative footage refers to the real-time recording of the cumulative drilling depth of the drill bit, which serves as the basis for adjusting the number of trees; the model capacity is dynamically expanded as drilling progresses, avoiding underfitting in the early stage and overfitting in the later stage.

[0115] In this embodiment of the invention, the top-level model is a dynamic weighted random forest, wherein the weight adjustment formula of the dynamic weighted random forest is as follows:

[0116]

[0117] in, It is The weight values ​​of each sample, It is Elastic modulus of laboratory core samples. It is Elastic modulus of laboratory core samples. It is an adjustment coefficient that is dynamically adjusted based on the cumulative footage of the drill bit. It is the drill bit wear index. and It is a sample index. It represents the total number of samples.

[0118] Specifically, high-modulus formations (such as flint) are sensitive to weighting, requiring enhanced training weights; drilling parameters become distorted after drill bit wear, which can be addressed by… Decrease the weight of old samples.

[0119] In detail, The higher the elastic modulus, the greater the sample weight, and the more the model focuses on the prediction accuracy of high-strength strata (such as flint and dense sandstone).

[0120] In detail, logical functions Will Mapped to the (0,1) interval, used for dynamically decaying or enhancing sample weights.

[0121] In detail, when the drill bit is severely worn ( When k > 0, the value of the logic function approaches 0. Weight allocation is subject to Control; if k < 0, the logistic function value approaches 0, and the weight of old samples is suppressed (applicable to data distortion scenarios after drill bit wear);

[0122] In detail, when the drill bit is in good condition ( When the value of the logical function approaches 0.5, the weight distribution is relatively balanced.

[0123] In this embodiment of the invention, the dynamic adjustment rule of the adjustment coefficient is as follows:

[0124] When the strata lithology is shale, the adjustment coefficient is corrected based on the fracture density, wherein the formula for calculating the coefficient correction is:

[0125]

[0126] in, This is the corrected adjustment coefficient. It is the adjustment coefficient, It refers to the crack density.

[0127] In detail, assuming a cumulative advance of s = 100 meters in a certain work area, a fracture density ρ = 4 fractures / meter (shale formation), and an adjustment coefficient... The corrected adjustment coefficient Drill bit wear index ,but If a certain sample The total elastic modulus of the work area is ,but

[0128] Furthermore, compare with the uncorrected case. hour, Therefore, after the crack density correction, the sample weights are slightly reduced, and the model pays more attention to the current drilling status.

[0129] In this embodiment of the invention, the ensemble machine learning model further includes a transfer learning module, wherein the implementation of the transfer learning module includes:

[0130] Based on historical well data from 500 wells in 10 basins, a general rock mechanics prediction model was trained to extract common formation characteristics.

[0131] Using logging data and core experimental data from three wells in the target work area, the parameters of the first three layers of the LSTM network were frozen.

[0132] Using the latest 5-meter well logging data and core data from the target work area, the fully connected layer parameters were updated using the Adam optimizer, with the learning rate adjusted to 0.01 and the number of iterations set to 200.

[0133] Once triggered, a sliding window mechanism is used to perform incremental learning every 1 meter of drilling.

[0134] In detail, the parameters of the first three layers of the LSTM in the pre-trained model are frozen, and the common temporal features extracted by them are retained (such as stratigraphic sequence patterns and lithological transition laws); only the parameters of the fully connected layers are fine-tuned to make the model adapt to the local characteristics of the target work area (such as specific shale fracture density and regional stress field).

[0135] In detail, the shallow layers (first 3 layers) of the pre-trained model typically learn basic features (such as curve shape and noise filtering), while the deeper layers (later layers) learn advanced abstract features (such as crack correlation). Experiments show that freezing the first 3 layers can balance generality and adaptability, and avoid overfitting to small sample data of the target work area.

[0136] In detail, a general model pre-trained on 10 basins and 500 wells was transferred to the target work area; the dependence of the target work area on core samples was reduced, and the new strata were quickly adapted.

[0137] In detail, the first 3 layers of the LSTM were frozen, and the fully connected layers were fine-tuned using data from 3 wells in the target work area; the model was updated with the latest 5-meter data (including core samples) for every 1 meter of drilling, with a learning rate of 0.01 and 200 iterations.

[0138] In detail, 5 meters can cover typical thin interbedded sandstone and mudstone (1-3 meters single layer thickness); too short (e.g., 1 meter) is easily affected by noise, and too long (e.g., 10 meters) reduces the model response speed.

[0139] Furthermore, the window moves forward dynamically as drilling progresses, retaining the latest 5-meter data and discarding old data; if the core data is lagging, it is temporarily filled with logging data, and replaced with core data once it is available.

[0140] For example: In a shale gas target area, the cumulative drilling footage is 200 meters, and the latest 5-meter section encountered high-fracture density shale ( (strips / meter).

[0141] Furthermore, the well logging data showed Δt = 95 μs / ft (significant cycle skipping) and GR = 80 API; the core data showed E = 28 GPa (lower than the initial model prediction of 30 GPa); the elastic modulus prediction formula for the fully connected layer was fine-tuned to reduce the E estimation deviation of high GR formations.

[0142] In summary, LSTM and attention mechanisms enhance the ability to model temporal features, XGBoost strengthens the nonlinear fitting of static features, dynamic weight allocation optimizes model fusion based on lithology and drill bit status, and transfer learning addresses the problem of data scarcity in the target work area.

[0143] S4. Optimize the model output of the integrated machine learning model based on dynamic weight adjustment and physical constraint loss function to obtain the prediction results of rock mechanics parameters.

[0144] In this embodiment of the invention, the physical constraint loss function is:

[0145]

[0146] in, It is the total loss value used to optimize the training of the ensemble machine learning model. , , These are weighting coefficients used to control the different constraint terms. It is the first The predicted elastic modulus of a sample It is the first Elastic modulus of laboratory core samples. It is the first The predicted stress gradient for each sample. It is the elastic modulus. It is Poisson's ratio. It is an indicator function. and It is a sample index. It is the total number of samples. This refers to the number of points used in the stress gradient calculation.

[0147] In detail, the mean squared error term is used to minimize the difference between the predicted elastic modulus and the laboratory value, ensuring that the model output is consistent with the real data. The larger the value, the more the model tends to fit the experimental data (usually set to 0.5~0.7).

[0148] In detail, the stress gradient smoothing term is the sum of the absolute values ​​of the constrained stress gradients to avoid non-physical oscillations (such as stress abrupt changes) in the prediction results. In actual formations, stress gradients usually change continuously, and drastic fluctuations may cause model overfitting or data noise. The larger β is, the smoother the stress field (usually set to 0.2~0.3).

[0149] In detail, the empirical relation constraint term forces the ratio of elastic modulus to Poisson's ratio E / PR ≤ 50, which conforms to the empirical law of geomechanics; in shale formations, E / PR is usually 30~40. If the model predicts E / PR=60, a penalty is applied; γ is relatively small (usually set to 0.1~0.2) to avoid excessive interference with the main loss term.

[0150] In detail, the model output is optimized through a physical constraint loss function to ensure that the prediction results conform to the laws of geomechanics; traditional machine learning may output prediction values ​​that violate physical laws, so constraints are needed to enhance the rationality of the results.

[0151] In detail, the physical constraint loss function consists of three parts: mean square error term, stress gradient smoothing term, and empirical relation constraint term. The mean square error term (α term) is used to minimize the difference between the predicted elastic modulus and the laboratory value; the stress gradient smoothing term (β term) is used to constrain the predicted stress gradient smoothing to avoid non-physical oscillations; and the empirical relation constraint term (γ term) is used to enforce the geomechanical empirical relation that E / PR≤50.

[0152] In detail, the stress gradient is calculated using the finite difference method or a physical simulation model.

[0153] For example, with α=0.7, β=0.2, and γ=0.1, the weighting coefficients can be optimized through cross-validation to balance the importance of different constraints, improve the physical rationality of the prediction results, and avoid model overfitting or outputting invalid values.

[0154] In detail, the dataset is divided into training and validation sets, and the optimal α, β, and γ are determined through grid search or Bayesian optimization. Example: If the mean squared error of the validation set is minimized when α=0.7, β=0.2, and γ=0.1, then this combination is used.

[0155] Furthermore, the weights are adjusted according to the geological characteristics of the work area. For example, in fractured strata, β is increased to strengthen stress smoothing constraints; in strata with high elastic modulus, γ is increased to strictly limit E / PR anomalies.

[0156] For example: In the prediction of a tight sandstone stratum, the model output... Laboratory values stress gradient Then the mean square error term is The stress gradient term is ; Experience constraints are

[0157] Furthermore, set The total loss is:

[0158]

[0159] Furthermore, the optimization direction is to reduce backpropagation. Up to approximately 40 GPa, while adjusting make .

[0160] In summary, adjusting model weights based on lithological classification and drill bit condition improves adaptability to complex formations, while embedding geomechanical experience into model training enhances the reliability of prediction results.

[0161] In the several embodiments provided by this invention, it should be understood that the disclosed method can be implemented in other ways.

[0162] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0163] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, and technology that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0164] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for predicting the mechanical parameters of rocks in complex strata, characterized in that, The method includes: S1. Collect multi-source data of complex strata and perform data preprocessing on the multi-source data, wherein the multi-source data includes: well logging data, drilling parameters, seismic inversion data and core test data; S2. Based on the multi-source data, extract physical and statistical features related to rock mechanics parameters, and align the physical features, statistical features, and preprocessed multi-source data by depth to obtain a high-dimensional feature matrix; S3. Construct an ensemble machine learning model and train it using the high-dimensional feature matrix. The ensemble machine learning model includes a bottom-level model and a top-level model. The bottom-level model includes a temporal modeling unit and a nonlinear regression unit. The top-level model is a dynamic weighted random forest, and the weight adjustment formula for the dynamic weighted random forest is as follows: ； in, It is the first The weight values ​​of each sample, It is the first Elastic modulus of laboratory core samples. It is the first Elastic modulus of laboratory core samples. It is an adjustment coefficient that is dynamically adjusted based on the cumulative footage of the drill bit. It is the drill bit wear index. and It is a sample index. The total number of samples is given, and the dynamic adjustment rule for the adjustment coefficient is as follows: When the strata lithology is shale, the adjustment coefficient is corrected based on the fracture density, wherein the formula for calculating the coefficient correction is: ； in, This is the corrected adjustment coefficient. It is the adjustment coefficient, It is the crack density; S4. Optimize the model output of the integrated machine learning model based on dynamic weight adjustment and physical constraint loss function to obtain the prediction results of rock mechanics parameters.

2. The method for predicting rock mechanical parameters in complex strata as described in claim 1, characterized in that, The acquisition of multi-source data from complex strata includes: Well logging data of the complex formation was collected using logging instruments, including sonic transit time curves, density curves, neutron porosity curves, and natural gamma curves. Drilling parameters of the complex formation are acquired in real time using drilling sensors, including drilling pressure, rotation speed, mechanical drilling speed and drill bit diameter; Seismic inversion data of the complex strata, including wave impedance and porosity distribution, were obtained using seismic inversion technology. Core experimental data of the complex strata were obtained through core experiments, including the laboratory core elastic modulus and Poisson's ratio.

3. The method for predicting rock mechanical parameters in complex strata as described in claim 2, characterized in that, The preprocessing of the multi-source data includes: The acoustic time difference curve is subjected to cycle skip noise filtering to obtain the reconstructed and filtered acoustic time difference curve; Principal component analysis is used to perform feature dimensionality reduction on the reconstructed filtered acoustic transit time curve, the density curve, the neutron porosity curve, and the natural gamma curve to obtain the dimensionality reduction features of the logging data. The drilling parameters are standardized to obtain their mechanical specific energy, wherein the calculation formula for the standardization process is as follows: ； In the formula, It is the mechanical specific energy, It's drilling pressure. It's the rotational speed. It is the mechanical drilling speed. It is the drill bit diameter. It is the statistical standard deviation of the mechanical specific energy of all wells in the work area.

4. The method for predicting rock mechanical parameters in complex strata as described in claim 2, characterized in that, The step of filtering out cycle skip noise from the acoustic time-of-flight curve to obtain the reconstructed filtered acoustic time-of-flight curve includes: The acoustic time difference curve is decomposed into high-frequency and low-frequency components by using the Daubechies wavelet basis function. High-frequency noise in the high-frequency components is suppressed by using a threshold method; The high-frequency component after noise suppression and the low-frequency component are reconstructed to obtain the reconstructed filtered acoustic time difference curve.

5. The method for predicting rock mechanical parameters in complex strata as described in claim 1, characterized in that, The network structure of the temporal modeling unit includes an input layer, a bidirectional LSTM layer, and a multi-head attention layer. The multi-head attention layer is configured with four attention heads to calculate the association weights of different depth points in complex strata.

6. The method for predicting rock mechanical parameters in complex strata as described in claim 1, characterized in that, The nonlinear regression unit adopts the extreme gradient boosting tree model, wherein the number of random forest decision trees in the extreme gradient boosting tree model increases dynamically according to the cumulative advance.

7. The method for predicting rock mechanical parameters in complex strata as described in claim 1, characterized in that, The physical constraint loss function is: ； in, It is the total loss value used to optimize the training of the ensemble machine learning model. , , These are weighting coefficients used to control the different constraint terms. It is The predicted elastic modulus of a sample It is Elastic modulus of laboratory core samples. It is The predicted stress gradient for each sample. It is the elastic modulus. It is Poisson's ratio. It is an indicator function. and It is a sample index. It is the total number of samples. This refers to the number of points used in the stress gradient calculation.

8. The method for predicting rock mechanical parameters in complex strata as described in any one of claims 1 to 7, characterized in that, The ensemble machine learning model further includes a transfer learning module, wherein the implementation of the transfer learning module includes: Based on historical well data from 500 wells in 10 basins, a general rock mechanics prediction model was trained to extract common formation characteristics. Using logging data and core experimental data from three wells in the target work area, the parameters of the first three layers of the LSTM network were frozen. Using the latest 5-meter well logging data and core data from the target work area, the fully connected layer parameters were updated using the Adam optimizer, with the learning rate adjusted to 0.01 and the number of iterations set to 200. Once triggered, a sliding window mechanism is used to perform incremental learning every 1 meter of drilling.

Images