Knowledge and data fusion driven rock strength while drilling prediction method and system

Abstract

The application discloses a knowledge and data fusion driven rock strength while drilling prediction method and system, and belongs to the technical field of rock mass property prediction. In view of the problems of poor adaptability of an empirical formula, strong idealization assumption of a theoretical model, insufficient physical constraint and weak generalization ability of a pure data driven method in the existing rock strength while drilling prediction method, the application constructs a mesoscopic rock breaking knowledge data set containing lithofacies characteristics and drilling parameters through a numerical test, screens to form an input feature vector, pre-trains a data driven model, constructs and calibrates a drilling rock breaking physical model based on a physical relationship revealed by internal rock breaking variables, optimizes network parameters by using a joint loss function after parallel connection of the two models, realizes double embedding of mesoscopic rock breaking knowledge, enhances physical interpretability while maintaining data nonlinear fitting capability, and realizes accurate and stable rock strength while drilling prediction.

Description

Technical Field

[0001] This invention relates to the field of rock mass property prediction technology, and in particular to a knowledge and data fusion-driven method and system for predicting rock strength while drilling. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] In the intelligent construction process of underground engineering such as tunnels, mines, and water conservancy and hydropower, the rapid and accurate acquisition of rock strength is the key to ensuring the stability of the surrounding rock and construction safety. However, traditional core sampling test methods have long cycles and poor representativeness. Existing drilling prediction methods lack adaptability of empirical formulas, are limited by the assumption of homogeneity in theoretical models, and have weak interpretability and generalization ability due to insufficient physical constraints. No single approach can meet the dual requirements of mechanism consistency and learning ability.

[0004] Therefore, existing technologies cannot deeply integrate the microscopic physical mechanisms of drilling and rock breaking with data-driven nonlinear mapping capabilities, nor can they construct a prediction model while drilling that combines physical interpretability and cross-condition generalization. This makes it difficult to achieve continuous, accurate, and stable in-situ sensing of rock strength, which may lead to frequent misjudgments of surrounding rock classification, lack of reliable basis for support parameter design, and lag in construction plan adjustments. This will increase the risk of disasters such as collapses and large deformations, seriously threatening the safety and efficiency of underground engineering construction. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a knowledge- and data-driven method and system for predicting rock strength while drilling. It constructs a microscopic rock breaking knowledge base through numerical experiments and integrates data-driven and physical models through feature selection and physical constraints, thereby achieving accurate and stable prediction of rock strength while drilling.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solution: In a first aspect, the present invention provides a knowledge and data fusion-driven method for predicting rock strength while drilling, comprising: A microscopic rock-breaking knowledge dataset was constructed through numerical experiments; the microscopic rock-breaking knowledge dataset includes lithofacies characteristics, drilling parameters, internal rock-breaking variables, and numerical uniaxial compressive strength. From the raw drilling data of the microscopic rock breaking knowledge dataset, drilling parameters that are sensitive to rock strength and stable under working conditions are selected and combined with lithofacies features to form an input feature vector; a measured dataset is constructed according to the definition of the input feature vector, and the data-driven model is pre-trained to obtain the initial network parameters; A drilling rock-breaking physical model is constructed based on the physical relationships revealed by the internal rock-breaking variables, and the physical model parameters are calibrated using the training set in the measured dataset. By combining the data-driven model with the physical model in parallel, a knowledge and data fusion model is constructed. The field data is used to construct an input feature vector, and the rock strength is predicted through the fusion model.

[0007] Furthermore, a microscopic numerical model of rock is constructed using particle flow discrete element method, and a numerical model of drilling tool is constructed using rigid body element or finite element method. Numerical drilling and uniaxial compression tests are then conducted through contact coupling to obtain the microscopic rock breaking knowledge dataset.

[0008] Furthermore, based on the microscopic rock breaking knowledge dataset, through a combination of statistical analysis and physical correlation verification, drilling condition parameters and drilling response parameters that are sensitive to rock strength and stable under working conditions are selected and combined with lithofacies features to form an input feature vector.

[0009] Furthermore, the measured dataset is constructed according to the input feature vector definition after preprocessing and spatial alignment of the original drilling data; The preprocessing includes outlier handling, filtering, and missing value handling.

[0010] Furthermore, the drilling rock-breaking physical model is constructed based on the relationship between mechanical equilibrium and energy dissipation. After the initial values ​​and ranges of the physical model parameters are determined by microscopic rock-breaking knowledge, they are calibrated by the measured training set.

[0011] Furthermore, by constructing a data fitting term using data-driven predicted values ​​and measured values, and constructing a physical constraint term using physical model predicted values, a joint loss function is formed to optimize network parameters and obtain the knowledge and data fusion model.

[0012] Furthermore, during on-site prediction, the same preprocessing rules, feature definitions, and normalization parameters as in the actual measurement stage are used to construct on-site input feature vectors, which are then input into the fusion model to obtain rock strength results.

[0013] In a second aspect, the present invention provides a knowledge and data fusion-driven rock strength prediction system for drilling, comprising: Numerical knowledge dataset construction module: used to construct a microscopic rock breaking knowledge dataset through numerical experiments; the microscopic rock breaking knowledge dataset includes lithofacies characteristics, drilling parameters, internal rock breaking variables, and numerical uniaxial compressive strength. Feature-level embedding module: used to filter drilling parameters that are sensitive to rock strength and stable under working conditions from the original drilling data of the microscopic rock breaking knowledge dataset, and combine them with lithofacies features to form an input feature vector; Data-driven pre-training module: used to construct a test dataset according to the input feature vector definition, pre-train the data-driven model, and obtain the initial network parameters; Physical model construction and calibration module: used to construct a drilling rock-breaking physical model based on the physical relationships revealed by the internal rock-breaking variables, and to calibrate the physical model parameters using the training set in the measured dataset; Fusion Model Building Module: Used to connect the data-driven model and the physical model in parallel to build a knowledge and data fusion model; Prediction module: Used to construct input feature vectors from field data and predict rock strength through the fusion model.

[0014] In a third aspect, the present invention also provides a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of the method as described in the first aspect.

[0015] In a fourth aspect, the present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method as described in the first aspect.

[0016] Compared with the prior art, the beneficial effects of the present invention are: This invention deeply integrates microscopic rock-breaking knowledge into a data-driven model through both feature-level embedding (combination of lithofacies and drilling features) and model-level embedding (physical constraint joint loss function). This ensures that the prediction network is always guided by the physical mechanism of drilling rock-breaking while fitting measured data. As a result, while maintaining the nonlinear mapping capability of pure data methods, it improves the model's generalization ability across lithology and working conditions and the physical interpretability of prediction results. This solves the fundamental problem of existing methods having high data fitting accuracy but poor mechanism consistency, and achieves a dual breakthrough in the accuracy and stability of rock strength prediction during drilling. Attached Figure Description

[0017] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0018] Figure 1 This is a flowchart of the method according to Embodiment 1 of the present invention. Detailed Implementation

[0019] The following detailed description is exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0020] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0021] In this embodiment of the invention, the data collection and processing strictly adhere to the requirements of relevant laws and regulations, obtaining informed consent or separate consent from the data subject, and conducting subsequent data use and processing within the scope of laws, regulations, and the data subject's authorization. All data acquisition in this embodiment is based on compliance with laws and regulations and user consent, representing the lawful application of the data.

[0022] Example 1 In existing technologies, empirical formula methods have poor adaptability to complex lithology and variable drilling conditions. Theoretical models, due to idealized assumptions, are difficult to describe the microstructure of rocks and discontinuous failure processes. While pure data-driven methods have strong fitting capabilities, they lack physical constraints, resulting in insufficient generalization ability across engineering projects. In practical applications, this often manifests as a sharp drop in prediction accuracy after changing the formation or drilling parameters of the same model, and the results violate basic mechanical laws and are difficult to interpret, bringing serious uncertainties to on-site support design and disaster identification.

[0023] To address this, the present invention constructs a microscopic rock-breaking knowledge dataset through a numerical experimental system. On the one hand, it combines lithofacies features with drilling features verified by physical correlation into an input vector to eliminate data ambiguity. On the other hand, it constructs a physical model based on the mechanical equilibrium and energy dissipation laws revealed by internal rock-breaking variables. This model is then embedded as a constraint term in the data-driven model training process through a joint loss function. This ensures that the model maintains strong nonlinear learning capabilities while always being guided by the physical mechanism of rock breaking. As a result, even when facing new strata and new working conditions in practical engineering applications, it can output stable prediction results that conform to mechanical principles, thus solving the problem of the separation between mechanism and data in existing methods.

[0024] In a typical embodiment of the present invention, such as Figure 1 As shown, a knowledge- and data-driven method for predicting rock strength while drilling is disclosed, the specific steps of which include: S1: Construct a microscopic rock-breaking knowledge dataset through numerical experiments; the microscopic rock-breaking knowledge dataset includes lithofacies characteristics, drilling parameters, internal rock-breaking variables, and numerical uniaxial compressive strength. S2: From the original drilling data of the microscopic rock breaking knowledge dataset, select drilling parameters that are sensitive to rock strength and stable under working conditions, and combine them with lithofacies features to form an input feature vector; construct a measured dataset according to the definition of the input feature vector, pre-train the data-driven model, and obtain the initial network parameters; S3: Construct a drilling rock-breaking physical model based on the physical relationships revealed by the internal rock-breaking variables, and calibrate the physical model parameters using the training set in the measured dataset; S4: Connect the data-driven model and the physical model in parallel to construct a knowledge and data fusion model; S5: Construct the field data into an input feature vector, and predict the rock strength through the fusion model.

[0025] The following describes in detail a knowledge and data fusion-driven method for predicting rock strength during drilling, using specific implementation methods.

[0026] In S1, the microscopic rock-breaking knowledge dataset includes lithofacies characteristics, drilling parameters, internal rock-breaking variables, and numerical uniaxial compressive strength corresponding to each test sample; wherein, the drilling parameters include drilling condition parameters and drilling response parameters. A microscopic numerical model of rock is constructed using particle flow discrete element method, and numerical drilling and uniaxial compression tests are carried out under different configurations and working conditions through contact coupling with the drill string numerical model. The lithofacies characteristics, drilling conditions, drilling response, internal rock-breaking variables, and numerical strengths corresponding to the same model are matched one by one to form a microscopic rock-breaking knowledge dataset containing lithofacies characteristics, drilling condition parameters, drilling response parameters, internal rock-breaking variables, and numerical uniaxial compressive strength corresponding to each test sample.

[0027] As one embodiment, the construction of the microscopic rock-breaking knowledge dataset is achieved in the following way: 1) Based on the particle flow discrete element method, a microstructure numerical model of rocks is established, considering mineral type, content, grain size, and cementation characteristics. Here, rock microstructure refers to the parameters in the numerical model used to describe the rock's mineral composition, mineral content, grain size, grain morphology, grain gradation, grain contact relationships, and cementation characteristics. Lithofacies characteristics are features selected from the rock microstructure that can be obtained or characterized during the experimental training and field prediction phases, including one or more of the following: mineral composition, mineral content, average grain size, and heterogeneity characterization parameters. During model establishment, grain contact is subdivided into intragranular contact, intergranular contact of the same mineral, and intergranular contact of different minerals to assign differentiated mechanical parameters.

[0028] Specifically: First, a rectangular or cylindrical computational domain is established based on the geometric dimensions of the rock sample to be simulated. A discrete particle set is generated using the radius expansion method, and the size of the discrete particles is determined based on the numerical calculation accuracy. Then, according to the actual grain size distribution of the target rock, multiple discrete particles are divided into grain units, ensuring that the equivalent size distribution of the grain units matches the actual grain size distribution. Next, based on the rock's mineral type (e.g., quartz, feldspar, mica), mineral content, and grain size distribution, each particle in the discrete particle set is sequentially assigned to its corresponding grain unit. Particles within the same grain unit are assigned the same mineral category label and grain identifier. Mechanical transmission occurs between different grain units through particle contact. Based on this, all particle contacts are traversed and classified according to the grain unit and mineral category of adjacent particles: if adjacent particles belong to the same grain unit, it is defined as intragranular contact; if they belong to different grain units but have the same mineral category, it is classified as intragranular contact. Intergranular contact is defined as contact between the same type of mineral, while contact between different types of minerals is defined as contact between different types of minerals. Different contact constitutive models (such as linear parallel bond models) and differentiated micromechanical parameters (including normal stiffness, tangential stiffness, bond strength, and friction coefficient) are assigned to the three types of contact to characterize the differences in intragranular and intergranular mechanical properties. Finally, numerical uniaxial compression tests are conducted on the established model under boundary conditions consistent with those in laboratory tests. The simulated peak strength, elastic modulus, Poisson's ratio, and failure mode are compared with the measured results of laboratory tests on the same batch of rock cores. The micromechanical parameters are iteratively adjusted until the error between the simulation results and the measured results meets the preset requirements (such as controlling the peak strength error within 5%). This completes the model calibration and obtains a rock micromechanical numerical model that can truly reflect the mechanical response characteristics of the target rock, providing a reliable rock sample basis for subsequent numerical drilling tests and the construction of a micromechanical rock breaking knowledge dataset.

[0029] 2) Based on the actual drill bit geometry and cutting tooth arrangement, a numerical model of the drill string is constructed using rigid body elements or finite element meshes, and boundary conditions for motion and load, such as drilling pressure and rotational speed, are set. The numerical model of the drill string sets boundary conditions for drill string motion and load based on the drill bit geometry, cutting tooth arrangement, drilling loading method, and drilling parameter monitoring requirements.

[0030] Specifically, the numerical model of the drill string can be constructed as follows: Based on the diameter of the drill bit used in the actual drilling test, the geometric dimensions of the cutting teeth (including tooth height, tooth width, rake angle, clearance angle, and tooth spacing), and the spatial arrangement of the cutting teeth on the drill bit end face (such as scraper-type or roller cone-type distribution), a three-dimensional geometric model of the drill string is established. The drill bit base and the surface of the cutting teeth are discretized into rigid wall elements. Each rigid element is combined through rigid constraints to form the overall numerical model of the drill string. At the same time, the central axis of the drill string and the loading reference point located at the top of the drill string are determined, so that the entire drill string moves as a single rigid body in the subsequent coupled calculations. On this basis, a constant load is applied along the central axis at the loading reference point. Drilling pressure serves as the propulsion load, and a constant rotational speed or angular velocity is applied around the central axis as the rotational drive, thereby enabling the drill bit to contact and couple with the rock microstructure numerical model under the combined action of axial propulsion and rotational motion. Throughout the numerical drilling process, the axial displacement of the drill bit, the reaction torque, and the contact load (including normal and tangential contact forces) on the drill bit-rock contact surface are recorded in real time. The real-time drilling speed is calculated based on the relationship between axial displacement and time. Simultaneously, drilling pressure, rotational speed, torque, drilling speed, and drilling energy are used as drilling condition parameters and drilling response parameters in the subsequent microstructure rock breaking knowledge dataset, providing numerical input for the construction of the drilling rock breaking physical model.

[0031] 3) Then, the drill string numerical model and the rock microstructure numerical model are coupled by the contact coupling algorithm to form a full-process drilling numerical model (this full-process drilling numerical model is established by the contact coupling between the drill string numerical model and the rock microstructure numerical model). Numerical uniaxial compression tests are carried out on rock models with different microstructures, and numerical drilling tests are carried out under different combinations of drilling pressure and rotation speed.

[0032] Numerical drilling tests record drilling conditions and response parameters in real time, such as drilling pressure, rotational speed, torque, and drilling speed. At the same time, they monitor internal rock-breaking variables such as contact load, crack initiation and propagation, damage accumulation, fracture zone evolution, and energy dissipation. Numerical uniaxial compression tests obtain the corresponding uniaxial compressive strength of the rock.

[0033] Specifically, the contact coupling between the drill bit numerical model and the rock microstructure numerical model can be achieved as follows: the drill bit matrix and the outer surface of the cutting teeth in the drill bit numerical model are defined as the drill bit-rock contact boundary. In each calculation step, a contact search is performed on all rock particles within the coverage area of ​​this contact boundary to determine whether contact occurs between the drill bit surface and the rock particles. For drill bit surface elements and rock particles that are detected to be in contact, the normal contact force, tangential contact force, and frictional force are calculated based on their normal overlap, tangential relative displacement, and relative velocity. These contact forces are then simultaneously applied to the drill bit according to the action-reaction relationship. The system employs a numerical model and rock particles. The contact force applied to the drill bit is equivalent to the loading reference point to update the drill bit's propulsion and rotational motion. The contact force applied to the rock particles acts as an external force to update the particle's velocity, displacement, and inter-particle contact state. Based on this, the contact search and force transmission process within each calculation step is repeatedly executed, causing the drill bit to gradually penetrate the rock under constant drilling pressure and constant rotational speed, initiating particle contact fracture, crack initiation and propagation, fracture zone formation, and energy dissipation, until the drilling depth reaches a preset target value (the preset target value is determined based on the simulated object, the rock's microstructure scale, and the drill bit size, and is generally taken as...). (multiple of the drill bit diameter), and throughout the drilling process, at a preset sampling frequency (e.g., per drilling depth step). (Record once) Real-time recording of drilling response parameters such as drilling pressure, rotation speed, torque, drilling speed and drilling energy, as well as internal rock breaking variables such as contact load distribution, number of cracks, damage variables and fracture zone range, thereby obtaining complete numerical drilling process data to support the subsequent construction of a microscopic rock breaking knowledge dataset.

[0034] 4) Finally, the lithofacies characteristics, drilling condition parameters (at least one of drilling pressure and rotation speed), drilling response parameters (at least one of torque, drilling speed and drilling energy), internal rock-breaking variables (at least one of contact load, crack initiation and propagation, damage accumulation, fracture zone evolution and energy dissipation) and numerical intensity corresponding to the same rock microstructure numerical model are matched and correlated one by one to form a microstructure rock-breaking knowledge dataset covering multiple lithologies and multiple working conditions, which provides a knowledge foundation for subsequent feature selection and physical model construction.

[0035] Specifically, a detailed analysis of the rock-breaking knowledge dataset includes at least: lithofacies characteristics. Drilling operating parameters Including at least one of drilling pressure and rotational speed; drilling response parameters This includes at least one of torque, drilling speed, and drilling energy; numerical internal rock-breaking variables. This includes at least one of contact load, crack initiation and propagation, damage accumulation, fracture zone evolution, and energy dissipation; and the numerical uniaxial compressive strength corresponding to the corresponding rock microstructure numerical model. .

[0036] The detailed rock-breaking knowledge dataset is represented as follows:

[0037] In the formula, To examine the rock-breaking knowledge dataset in detail; For the first The lithofacies feature vector corresponding to each numerical test sample or the numerical parameter vector corresponding to the lithofacies feature; For the first A vector of drilling condition parameters for a numerical test sample; For the first The drilling response parameter vector of a numerical test sample; For the first Numerical internal rock-breaking variable vector of a numerical test sample; In order to be with the first Numerical uniaxial compressive strength corresponding to each numerical test sample; This represents the number of samples in the numerical experiment.

[0038] This invention constructs input features with physical meaning through knowledge screening (feature-level embedding), and pre-trains data-driven models to retain nonlinear fitting capabilities and calibrates physical models to provide mechanical constraints. This lays a dual-pillar foundation for subsequent fusion driven by data learning and physical mechanisms, realizing the dual embedding of microscopic rock breaking knowledge and solving the problems of insufficient physical constraints, weak interpretability, and poor generalization ability of pure data-driven models.

[0039] In S2, S2.1: Based on the microscopic rock breaking knowledge dataset, drilling parameters that are sensitive to rock strength and stable under working conditions are selected through a combination of statistical analysis and physical correlation verification, and combined with lithofacies features to form a unified input feature vector that has both physical meaning and measurability.

[0040] Specifically, the input feature vector is obtained by combining the lithofacies feature vector with the drilling feature vector that has been filtered through both statistical analysis (sensitivity and working condition stability screening) and physical correlation verification (consistency verification of internal rock breaking variable mechanisms), as follows: First, based on the microscopic rock breaking knowledge dataset obtained in step S1, the correlation between lithofacies characteristics, drilling condition parameters, drilling response parameters, numerical internal rock breaking variables and uniaxial compressive strength of rock is established, providing a data foundation for feature selection.

[0041] Based on this, the RReliefF algorithm combined with Pearson correlation analysis is used to comprehensively screen candidate drilling features. Specifically, the RReliefF algorithm is used to rank the importance of all candidate drilling features, such as torque, drilling speed, drilling energy, and their derived features, to obtain the relative ranking of the contribution of each feature to the prediction of uniaxial compressive strength of rock. At the same time, the microscopic rock breaking knowledge dataset is divided into multiple drilling condition groups according to different combinations of drilling pressure and rotational speed. Within each condition group, the Pearson correlation coefficient between each candidate drilling feature and the numerical uniaxial compressive strength is calculated. The magnitude of the absolute value of the correlation coefficient represents the sensitivity of the corresponding feature to rock strength (for example, an absolute value of the correlation coefficient greater than or equal to 0.3 is considered sensitive). The stability of the condition is evaluated by comparing the directional consistency and absolute value fluctuation of the correlation coefficient of the same feature in different condition groups. In addition, by analyzing the Pearson correlation coefficients between candidate features, highly redundant feature combinations are identified and eliminated. By combining the importance ranking results of RReliefF with the Pearson correlation analysis results, the characteristics that are ranked high in importance, have consistent correlation coefficients with high absolute values ​​in different drilling condition groups, and have low redundancy among each other are retained to form a set of candidate drilling features after screening.

[0042] Then, the candidate drilling features selected through statistical analysis were physically correlated and verified using numerical internal rock-breaking variables (contact load transfer, crack initiation and propagation, damage accumulation, fracture zone evolution, and energy dissipation). Specifically, the variation process of candidate drilling features in the same numerical drilling test was correlated with the aforementioned internal rock-breaking variables. The variation trends, key change locations, and correspondences between different rock-breaking stages were compared. Furthermore, the consistency of this correspondence under different rock facies and drilling conditions was examined to verify whether the candidate features could reasonably reflect the physical process of drilling rock breaking. The numerical internal rock-breaking variables... It is used only for mechanism analysis and physical correlation verification, and is not used as a direct input to the final prediction model.

[0043] Finally, constrained by what can be obtained or calculated on-site, the drilling condition parameters that have been confirmed as valid after the above statistical analysis and physical correlation verification are selected. (such as drilling pressure, rotation speed) and drilling response parameters (Such as torque, drilling speed, drilling energy, etc.) together constitute the drilling characteristic sub-vector. ; lithofacies characteristics (Including mineral composition, mineral content, average grain size, and heterogeneity characterization parameters, etc.) as essential components, combined with the drilling feature sub-vectors to form a unified input feature vector. X :

[0044] In the formula, Lithofacies characteristics; This is a drilling feature vector formed by filtering or combining drilling condition parameters and drilling response parameters based on microscopic rock breaking knowledge; X This is the input feature vector shared by the data-driven model and the drilling rock-breaking physics model. This step clarifies the name, dimensions, calculation method, size, and order of each input feature, for use in subsequent construction of the measured dataset and on-site prediction.

[0045] S2.2: Based on the input feature vector definition, preprocess and spatially align the synchronously acquired lithofacies features, drilling data and measured intensity to construct a measured dataset, and use this pre-trained data to drive the model to obtain network parameters with initial nonlinear mapping capabilities.

[0046] Specifically, the construction of the experimental dataset and the pre-training of the data-driven model are implemented in the following ways: First, indoor or field rock drilling tests covering different lithologies and drilling conditions were conducted. During drilling, raw drilling data such as drill pressure, rotational speed, torque, and drilling rate were collected simultaneously. Meanwhile, the lithofacies characteristics of the corresponding sections were obtained through core or drill cuttings image analysis and X-ray diffraction analysis. These lithofacies characteristics included mineral content, average grain size, and heterogeneity characterization parameters. Mineral content was obtained through X-ray diffraction analysis, while average grain size and heterogeneity characterization parameters were obtained through core or drill cuttings image analysis. Indoor uniaxial compression tests were conducted on core specimens from the same source as the drilling test rock samples or those corresponding to the spatial location of the drilling section to obtain the corresponding measured values ​​of the rock's uniaxial compressive strength, which were used as output labels.

[0047] Then, the raw drilling data is preprocessed. Specifically, statistical methods (such as the Laida criterion or box plot method) are used to identify and remove outliers; wavelet analysis is used to filter and reduce noise in the raw drilling data to eliminate high-frequency noise interference during drilling; and missing data is filled using linear interpolation or spline interpolation methods. After preprocessing, derived characteristics such as drilling rock-breaking energy are calculated based on drill pressure, rotational speed, torque, and drilling speed.

[0048] Subsequently, using rock samples or actual drilling sections as data sample units, the preprocessed drilling data is spatially aligned with the petrographic characteristics and measured values ​​of the uniaxial compressive strength of the corresponding sections, according to the drilling depth or pre-defined drilling sections (several continuous depth intervals pre-divided in the drilling direction based on engineering requirements or geological conditions, such as each section being 1.0 m). After alignment, based on the input feature vector definition determined in S2.1 (including the name, dimensions, calculation method, and arrangement order of each feature), the data in each data sample unit is transformed one by one into an input feature vector of a unified format.X j And compared with the corresponding measured values ​​of uniaxial compressive strength of rock. These together constitute a single test sample. All test samples are aggregated to form the test training dataset. , represented as:

[0049]

[0050] In the formula, This is the actual training dataset; For the first The input feature vector of a single actual drilled data segment; For the first The lithofacies feature subvectors corresponding to each actual drilling data segment; For the first A vector of drilling condition parameters for each actual drilling data segment; For the first A vector of drilling response parameters for each actual drilling data segment; In order to be with the first Measured values ​​of uniaxial compressive strength of rock corresponding to each actual drilling data segment; This represents the actual number of samples measured. For the first Drilling feature vectors corresponding to each actual drilling data segment.

[0051] Based on this, the measured training dataset is grouped and divided according to rock sample number, borehole number, or project, resulting in a training set, validation set, and test set. This ensures that data from the same rock sample, the same borehole, or adjacent drilling sections do not appear in different sets simultaneously. Statistical parameters used for normalization or standardization (such as mean and standard deviation) are calculated only based on the training set and applied to the training set, validation set, and test set.

[0052] With the input feature vector of the training set Input: Measured value of uniaxial compressive strength of rock To output the label, a pure data-driven prediction model is constructed using artificial neural networks, long short-term memory networks, one-dimensional convolutional neural networks, or Transformer networks. When using artificial neural networks, the model includes an input layer, at least one hidden layer, and an output layer. When using long short-term memory networks, one-dimensional convolutional neural networks, or Transformer networks, the model includes a feature extraction layer adapted to the corresponding sequence or local feature processing, and a fully connected layer that outputs the uniaxial compressive strength of the rock. The pure data-driven prediction model is pre-trained using a data fitting loss function. The network weights are iteratively updated by minimizing the fitting error between the predicted and measured values ​​until the error stabilizes on the validation set, thus obtaining the initial network parameters. The pre-trained network with the following parameters will be completed. The model was determined to be a purely data-driven rock strength prediction model during drilling, denoted as... And it serves as the foundational model for constructing knowledge and data fusion-driven prediction models.

[0053] In S3, based on the mechanical equilibrium and energy dissipation laws revealed by the internal rock-breaking variables in the microscopic rock-breaking knowledge, a physical model is constructed with the input feature vector as input, and its parameters are calibrated through the measured training set, so that the model can quantitatively correlate drilling response and rock strength from the physical mechanism.

[0054] Specifically, based on the internal rock-breaking variables obtained in step S1, such as contact load transfer, crack propagation, damage accumulation, fracture zone evolution, and energy dissipation, the physical relationship between drilling conditions and drilling response, characterized by rock-breaking energy consumption per unit volume, is determined. Lithofacies features and corresponding feature components such as drill pressure, rotational speed, and drilling rate, which correspond to the physical relationship, are extracted from the input feature vector. Considering the differences in the conversion relationship between rock-breaking energy consumption and rock strength under different mineral compositions, grain sizes, and heterogeneous conditions, a correction coefficient determined by the lithological features is introduced to construct a physical model for predicting rock strength while drilling, driven by mesoscopic rock-breaking knowledge.

[0055] For the The predicted values ​​of the physical model for each training sample are expressed as follows:

[0056] In the formula, This is the lithofacies characteristic number. The total number of lithofacies features involved in the physical model correction. For the first The training sample was dimensionless. The lithofacies characteristics include one or more of the following: mineral content, average grain size, and heterogeneity characterization parameters; , and These are the drilling pressure, rotation speed, and drilling rate for the corresponding samples; Where is the drill bit radius. This refers to the width of the cutting teeth; and These are the angle parameters during the cutting process; The basic energy-intensity conversion factor, For the first Correction coefficients corresponding to the facies characteristics of the phyllite.

[0057] Determine the physical model parameter vector based on the microscopic rock-breaking knowledge data obtained in step S1. The initial values ​​and ranges of the parameters are determined, and the training set in the actual training dataset is used as calibration data to solve for the parameters, thereby obtaining the calibrated physical model parameter vector. .

[0058] For ease of explanation, the above physical model will be abbreviated as:

[0059] In the formula, For the first j The input feature vector of each training sample Extract the petrographic features and drilling features corresponding to the physical relationships from them; The training will remain consistent throughout subsequent joint training sessions.

[0060] The calibration iteration is terminated when the root mean square error between the physical model prediction and the measured value decreases below a preset threshold (e.g., 1%) within a preset number of iterations (e.g., 50 iterations).

[0061] The calibrated drilling rock-breaking physical model is used to calculate the predicted values ​​of the physical model and to construct physical constraint terms.

[0062] In step S4, after obtaining the initial network parameters of the pre-trained data-driven model and the fixed parameters of the calibrated physical model, the data-driven model and the calibrated physical model are set up in parallel during the training phase. The data fitting term is constructed by the fitting error between the data-driven predicted value and the measured value, and the physical constraint term is constructed by the residual between the data-driven predicted value and the physical model predicted value. Together, a joint loss function is constructed to optimize the network parameters, thereby constructing a knowledge and data fusion model constrained by physical mechanisms.

[0063] Specifically, to obtain the initial network parameters of the data-driven model. Using initial values, the data-driven model will be used. Compared with the calibrated drilling rock breaking physical model The parallel configuration is applied during the training phase. For any training sample in the training set... Using the same input feature vector Input the data-driven model and the drilling rock-breaking physics model respectively, and obtain the predicted values ​​from the data-driven model. and physical model predictions ;according to Measured values ​​of uniaxial compressive strength of rock The error between them is used to construct a data fitting term, based on and The residuals between the data are used to construct physical constraint terms, and the joint loss function is used to fine-tune the physical constraints of the data-driven model.

[0064] The data-driven model prediction is expressed as follows:

[0065] The data fitting term, physical constraint term, and joint loss function are as follows:

[0066]

[0067]

[0068] In the formula, For parameters Data-driven model; For data fitting terms; For physical constraints; The number of samples in the training set; For the joint loss function; These are the weighting coefficients for the physical constraint terms, used to balance the fitting accuracy of the measured data with the constraint strength of the physical relationship. Based on the physical model calibration error, training sample size, and validation set prediction results, optimization methods such as grid search, cross-validation, Bayesian optimization, or adaptive weight adjustment can be used to balance data fitting accuracy and consistency with physical constraints.

[0069] The network parameters of the prediction network are iteratively updated by minimizing the joint loss function. calibrated physical model parameters By keeping the prediction network constant, the prediction results can be made to fit the nonlinear relationship between the actual input features and the measured value of the uniaxial compressive strength of the rock, while also tending to satisfy a definite physical relationship.

[0070] When the joint loss function reaches the preset convergence condition, or when the prediction error on the validation set no longer decreases within a preset number of consecutive training epochs (e.g., the change in the joint loss function is less than 10 for 10 consecutive epochs), the joint loss function will converge. -4 Or the prediction error of the validation set is continuous The number of epochs no longer decreases.

[0071] Stop training and obtain the optimized network parameters. Set the network parameters as follows: The prediction network, optimized through dual embedding at the feature level and model level of rock breaking knowledge, was identified as a knowledge- and data-driven rock strength prediction model during drilling, denoted as [model name missing]. This is the final model for predicting rock strength during drilling.

[0072] Therefore, the knowledge of rock breaking is embedded at both the feature level and the model level through the construction of input features and physical constraints, and together with the measured training data, it drives the formation of the final prediction model.

[0073] In step S5, the lithofacies features and drilling parameters collected on-site are constructed into an input feature vector according to the same definition, dimensions, calculation methods and arrangement order as in the training phase. After normalization of the statistical parameters of the training set, the input knowledge and data are fused to drive the prediction model, thereby obtaining the predicted value of the uniaxial compressive strength of the rock at the corresponding drilling location.

[0074] Specifically, on-site rock strength prediction during drilling is achieved in the following manner: First, during the drilling process of the target project, the original drilling parameters such as drilling pressure, rotation speed, torque, drilling speed, drilling displacement, drilling depth and drilling time are collected in real time through the drilling measurement system. At the same time, the lithofacies characteristics of the drilling section are obtained by means of core or drill cuttings image analysis, mineral spectral analysis, X-ray diffraction analysis, thin section identification, intelligent drill cuttings identification or borehole wall imaging, including mineral composition, mineral content, average grain size and heterogeneity characterization parameters.

[0075] Then, based on the field drilling data, a depth window or preset drilling section is determined. Following the same preprocessing rules as in the actual training phase, outlier removal, filtering, smoothing, and missing value imputation are performed on the original drilling data. The processed drilling data is then spatially aligned with the lithofacies features at the corresponding depth. According to the input feature names, dimensions, calculation methods, dimensions, and arrangement order determined in step S2.1, the spatially aligned data is transformed into a unified field input feature vector. Using the statistical parameters such as the mean and standard deviation calculated and saved only from the training set in step S2.2, the field input feature vector undergoes normalization or standardization processing consistent with the training phase, ensuring that the field data and training data have consistent feature representation and numerical scale.

[0076] Finally, the constructed field input feature vector is input into the trained knowledge and data fusion-driven prediction model, which then outputs the predicted value of the uniaxial compressive strength of the rock at the drilling location. This is achieved by continuously sliding the calculation along a preset depth window or depth interval, for example, during continuous drilling, at a fixed length (e.g., ...). Extract drilling data within a segment as an evaluation unit, and then process it in fixed step sizes (e.g., ...). The window is gradually moved backward, generating a prediction sample with each movement, thus obtaining a rock strength curve continuously distributed along the entire borehole depth. Furthermore, the same on-site input feature vector can be simultaneously input into the drilling rock-breaking physical model calibrated in step S3. By comparing the deviation between the fused model's predictions and the physical model's predictions, abnormal prediction results that may significantly deviate from the physical relationship can be quickly identified, providing continuous rock strength perception information with both data support and physical interpretability for on-site construction decisions.

[0077] Example 2 In a typical embodiment of the present invention, a knowledge and data fusion-driven rock strength prediction system for drilling is provided, comprising: Numerical knowledge dataset construction module: used to construct a microscopic rock breaking knowledge dataset through numerical experiments; the microscopic rock breaking knowledge dataset includes lithofacies characteristics, drilling parameters, internal rock breaking variables, and numerical uniaxial compressive strength. Feature-level embedding module: used to filter drilling parameters that are sensitive to rock strength and stable under working conditions from the original drilling data of the microscopic rock breaking knowledge dataset, and combine them with lithofacies features to form an input feature vector; Data-driven pre-training module: used to construct a test dataset according to the input feature vector definition, pre-train the data-driven model, and obtain the initial network parameters; Physical model construction and calibration module: used to construct a drilling rock-breaking physical model based on the physical relationships revealed by the internal rock-breaking variables, and to calibrate the physical model parameters using the training set in the measured dataset; Fusion Model Building Module: Used to connect the data-driven model and the physical model in parallel to build a knowledge and data fusion model; Prediction module: Used to construct input feature vectors from field data and predict rock strength through the fusion model.

[0078] Example 3 The purpose of this embodiment is to provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the steps of the method of Embodiment 1.

[0079] Example 4 The purpose of this embodiment is to provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method of Embodiment 1.

[0080] The steps and methods involved in Embodiments 3 and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0081] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0082] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0083] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A knowledge and data fusion-driven method for predicting rock strength while drilling, characterized in that, include: A knowledge dataset on microscopic rock breaking was constructed through numerical experiments; The microscopic rock-breaking knowledge dataset includes lithofacies features, drilling parameters, internal rock-breaking variables, and numerical uniaxial compressive strength. From the raw drilling data of the microscopic rock breaking knowledge dataset, drilling parameters that are sensitive to rock strength and stable under working conditions are selected and combined with lithofacies features to form an input feature vector; a measured dataset is constructed according to the definition of the input feature vector, and the data-driven model is pre-trained to obtain the initial network parameters; A drilling rock-breaking physical model is constructed based on the physical relationships revealed by the internal rock-breaking variables, and the physical model parameters are calibrated using the training set in the measured dataset. By combining the data-driven model with the physical model in parallel, a knowledge and data fusion model is constructed. The field data is used to construct an input feature vector, and the rock strength is predicted through the fusion model.

2. The knowledge and data fusion-driven rock strength prediction method for drilling as described in claim 1, characterized in that, A microscopic numerical model of rock is constructed using granular flow discrete element method, and a numerical model of drilling tool is constructed using rigid body element method or finite element method. Numerical drilling and uniaxial compression tests are then conducted through contact coupling to obtain the microscopic rock breaking knowledge dataset.

3. The knowledge and data fusion-driven rock strength prediction method for drilling as described in claim 1, characterized in that, Based on the microscopic rock breaking knowledge dataset, through a combination of statistical analysis and physical correlation verification, drilling condition parameters and drilling response parameters that are sensitive to rock strength and stable under working conditions are selected and combined with lithofacies features to form an input feature vector.

4. The knowledge and data fusion-driven rock strength prediction method for drilling as described in claim 1, characterized in that, The measured dataset is constructed by preprocessing and spatially aligning the original drilling data, and then defining it according to the input feature vector. The preprocessing includes outlier handling, filtering, and missing value handling.

5. The knowledge and data fusion-driven rock strength prediction method for drilling as described in claim 1, characterized in that, The drilling and rock-breaking physical model is constructed based on the relationship between mechanical equilibrium and energy dissipation. After the initial values ​​and ranges of the physical model parameters are determined by microscopic rock-breaking knowledge, they are calibrated by the measured training set.

6. The knowledge and data fusion-driven rock strength prediction method as described in claim 1, characterized in that, The knowledge and data fusion model is obtained by constructing a data fitting term by combining data-driven predicted values ​​and measured values, constructing a physical constraint term by combining the predicted values ​​of the physical model, and forming a joint loss function to optimize network parameters.

7. The knowledge and data fusion-driven rock strength prediction method for drilling as described in claim 1, characterized in that, During on-site prediction, the same preprocessing rules, feature definitions, and normalization parameters as in the actual measurement stage are used to construct on-site input feature vectors, which are then input into the fusion model to obtain rock strength results.

8. A knowledge and data fusion-driven rock strength prediction system during drilling, characterized in that, include: Numerical knowledge dataset construction module: used to construct a microscopic rock breaking knowledge dataset through numerical experiments; The microscopic rock-breaking knowledge dataset includes lithofacies features, drilling parameters, internal rock-breaking variables, and numerical uniaxial compressive strength. Feature-level embedding module: used to filter drilling parameters that are sensitive to rock strength and stable under working conditions from the original drilling data of the microscopic rock breaking knowledge dataset, and combine them with lithofacies features to form an input feature vector; Data-driven pre-training module: used to construct a test dataset according to the input feature vector definition, pre-train the data-driven model, and obtain the initial network parameters; Physical model construction and calibration module: used to construct a drilling rock-breaking physical model based on the physical relationships revealed by the internal rock-breaking variables, and to calibrate the physical model parameters using the training set in the measured dataset; Fusion Model Building Module: Used to connect the data-driven model and the physical model in parallel to build a knowledge and data fusion model; Prediction module: Used to construct input feature vectors from field data and predict rock strength through the fusion model.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1-7.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method as described in any one of claims 1-7.

Images