Deep rock mass mechanical parameter dynamic perception method based on multi-source while-drilling data fusion

Abstract

The present application relates to the technical field of data processing, and discloses a deep rock mass mechanical parameter dynamic sensing method based on multi-source drilling data fusion, aiming at solving the problem of poor precision and reliability of the existing method, and mainly including: multi-source drilling data acquisition and preprocessing, multi-source data fusion and parameter inversion under physical rule constraint, prediction model construction of AI driven model, data integration and storage. The present application improves the precision and reliability of rock mass mechanical parameter sensing, realizes intelligent sensing of deep engineering rock mass safety, optimizes data storage efficiency and value mining capacity, and is suitable for deep geological engineering.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a deep rock mass mechanical parameter dynamic perception method based on multi-source while-drilling data fusion. BACKGROUND

[0002] In the process of deep geological exploration, oil and gas exploitation, deep tunnel and large underground cavern group construction, accurately obtaining in-situ rock mass mechanical parameters (such as uniaxial compressive strength, elastic modulus, Poisson's ratio, etc.) is crucial for wellbore stability analysis, drilling parameter optimization, support design decision and resource assessment. The traditional method mainly relies on indoor testing after drilling cores, which is tedious, long-period and high-cost, and it is difficult to obtain the real mechanical response of rock mass at engineering scale, especially for broken and soft rock mass which is difficult to core, the traditional method is almost ineffective.

[0003] In recent years, the while-drilling measurement technology has developed rapidly, which indirectly evaluates the rock mass properties by analyzing the parameters (such as torque, rotational speed, drilling speed) during drilling. However, the existing while-drilling parameter interpretation and rock mass mechanical perception method still has significant limitations. First, the existing method often relies on a single data source (such as drilling pressure, torque) or simple empirical formula for parameter inversion, which fails to fully utilize the internal correlation between multi-source heterogeneous data, resulting in low precision and insufficient reliability of the interpretation results. Second, the mainstream data-driven artificial intelligence model lacks physical law constraints, and its prediction results often violate the basic principles of rock mass mechanics, such as Poisson's ratio exceeding the reasonable range, poor interpretability, and difficulty in being trusted and adopted by field engineers. Third, the downhole working conditions are complex and variable, such as sticking, drilling through broken zones, and static models cannot adaptively adjust parameter correlation logic, leading to prediction failure when the working conditions change suddenly.

[0004] Therefore, there is an urgent need in the field for a deep rock mass mechanical parameter dynamic perception method that can fuse multi-source information, embed domain knowledge, and adaptively adjust to dynamic working conditions, to overcome the shortcomings of existing technology and provide reliable technical support for safe and efficient development of deep resources. SUMMARY

[0005] The present application aims to solve the problem of poor precision and reliability in existing while-drilling parameter interpretation and rock mass mechanical perception method, and proposes a deep rock mass mechanical parameter dynamic perception method based on multi-source while-drilling data fusion.

[0006] The technical solution adopted by the present application to solve the above technical problems is:

[0007] The deep rock mass mechanical parameter dynamic perception method based on multi-source while-drilling data fusion comprises:

[0008] Real-time acquisition and preprocessing of multi-source while-drilling data to obtain standardized features corresponding to the multi-source while-drilling data; the multi-source while-drilling data includes while-drilling parameters in the drilling process obtained by a drill pipe built-in sensor array, and auxiliary formation parameters obtained by a monitoring tool arranged in the downhole or around the well;

[0009] Based on the standardized features, a graph neural network combining attention mechanism and physical constraints is constructed; the graph neural network defines different depth points of the borehole as nodes, and the node feature vector contains the while-drilling parameters and the auxiliary formation parameters; the edge weight between nodes is dynamically calculated according to the physical constraints, statistical correlation and spatial distance, and the physical knowledge vector is embedded in the node feature update to guide the update of the node feature through the message passing mechanism; the updated high-dimensional node feature is interpreted for physical properties to output rock mass mechanical parameters and calculate the comprehensive confidence;

[0010] According to the rock mass mechanical parameters and the comprehensive confidence, a physical-data dual-driven mechanism is used to ensure that the prediction of the rock mass mechanical parameters conforms to the rock mass constitutive relation, a dynamic graph topology reconstruction mechanism adaptive to working conditions is used to identify the current drilling working condition in real time, the edge weight and the adjacency matrix of the graph neural network are dynamically adjusted according to the identified working condition, and a multi-scale spatio-temporal alignment federated learning framework is used to realize multi-drilling rig collaborative training and model updating to output risk indicators and decision suggestions;

[0011] Based on the physical indicators, a database index is constructed, a hierarchical early warning is performed according to the risk indicators, and a differential dynamic storage strategy is implemented according to the comprehensive confidence.

[0012] Further, the while-drilling parameters include the torque, rotation speed and drilling speed of the drill bit measured and obtained;

[0013] The auxiliary formation parameters include the longitudinal wave velocity, the transverse wave velocity, the absolute permeability and the compressibility coefficient;

[0014] The rock mass mechanical parameters include the uniaxial compressive strength, the elastic modulus, the Poisson's ratio, the rock strength index and the drilling condition compliance index.

[0015] Further, the preprocessing includes:

[0016] According to the real-time drilling working condition, a noise reduction algorithm is adaptively switched; the real-time drilling working condition includes a stuck drill pipe, a broken zone and normal drilling; through depth alignment and outlier rejection, the multi-source while-drilling data is aligned in time series and spatial depth; for different rock types, the corresponding compensation algorithm is called for automatic correction.

[0017] Further, the calculation formula of the edge weight between nodes is:

[0018] ;

[0019] in, Indicates adjacent nodes With nodes Edge weights between them This represents the weighting coefficient of the physical constraint term. The weighting coefficients of the statistical correlation terms. This represents the weighting coefficient of the spatial distance term. Represents a node The elastic modulus, Represents a node The elastic modulus, This represents the tolerance threshold for the elastic modulus. Represents a node eigenvectors With nodes eigenvectors The Pearson correlation coefficient between them Represents a node spatial coordinates and nodes spatial coordinates The Euclidean distance between them This represents the maximum neighborhood distance.

[0020] Furthermore, the node feature update employs a physically guided residual connection mechanism, and the update formula is as follows:

[0021] ;

[0022] in, Represents a node In the Feature vectors of a layered neural network Represents a node In the Feature vectors of a layered neural network Represents a node With nodes Edge weights between them Represents a node In the Feature vectors of a layered neural network Let represent the learnable weight matrices, Represents a node The set of adjacent nodes, This represents the GELU activation function. A vector function representing physical knowledge.

[0023] Furthermore, the output of rock mass mechanical parameters and the calculation of their overall confidence level include:

[0024] The rock mass mechanical parameters are interpreted from the output of the physical property interpretation formula, and the physical property interpretation formula is as follows:

[0025] ;

[0026] wherein, represents a node The output predicted physical property vector includes an output vector of a plurality of rock mass mechanical parameters, represents a learnable decoding weight matrix, represents a node The high-dimensional feature vector fused with multi-source data and physical constraints output by the final layer of the graph neural network, represents a bias term;

[0027] After the physical property vector is interpreted, the node-level residual is calculated:

[0028] ;

[0029] wherein, represents the node-level residual under the th physical rule, represents the theoretical value calculated according to the predicted physical property vector by the physical rule formula in the rock mass mechanical knowledge base, represents the true theoretical value corresponding to the physical rule, represents taking the absolute value;

[0030] Gaussian noise is added to the predicted physical property vector for perturbation test, and the perturbation sensitivity coefficient is calculated:

[0031] ;

[0032] wherein, represents the perturbation sensitivity coefficient of the th rock mass mechanical parameter in the predicted physical property vector, represents the perturbed rock mass mechanical parameter interpreted by the graph neural network after Gaussian noise is added to the original input rock mass mechanical parameter, represents the L2 norm of a vector;

[0033] The comprehensive confidence is calculated by the following formula:

[0034] ;

[0035] wherein, represents the comprehensive confidence of a node , and represents the node comprehensive residual vector of a node , representing the node 's overall perturbation sensitivity coefficient, representing the node 's consistency error with its neighbor nodes' predicted values, representing a natural constant.

[0036] Further, the physical-data dual driving mechanism comprises:

[0037] Encode the core law of rock mass mechanics into a graph structure prior constraint, and explicitly define the difference range of the elastic modulus of adjacent nodes:

[0038] ;

[0039] wherein, representing the elastic modulus of the node , representing the elastic modulus of the node , representing a configurable tolerance threshold, representing taking the absolute value;

[0040] Design the edge weight calculation function of the physical constraint:

[0041] ;

[0042] wherein, representing the edge weight between the adjacent node and the node , representing the associated strength of the physical properties between the node and the node , representing the Pearson correlation coefficient between the node and the node 's permeability, representing a weight coefficient.

[0043] Further, the federated learning framework comprises:

[0044] A loss function that fuses the physical rule residual is used for training, and its expression is:

[0045] ;

[0046] wherein, representing the global loss function of federated learning, representing a standard data error term representing a physical rule residual vector, representing a physical constraint weight, representing the modulus of a vector;

[0047] After each drilling rig node is trained based on local data, only encrypted model gradients are uploaded to the cloud server;

[0048] The cloud server allocates aggregation weights according to the data volume and reliability of each drilling rig node, and obtains a global model by weighted average aggregation and issues updates.

[0049] Further, the calculation formula of the risk index is as follows:

[0050] ;

[0051] wherein, represents the risk index, represents the uniaxial compressive strength prediction value, represents the maximum horizontal principal stress;

[0052] The graded early warning includes:

[0053] When the risk index is in a first preset range, it is determined as high risk, triggering an emergency stop drilling instruction;

[0054] When the risk index is in a second preset range, it is determined as medium risk, and a decision suggestion of reducing the rotating speed and starting the support plan is automatically issued;

[0055] When the risk index is in a third preset range, it is determined as low risk, maintaining the current torque, rotating speed and drilling speed, and continuously monitoring.

[0056] Further, a differentiated dynamic storage strategy is implemented according to the comprehensive confidence, including:

[0057] For the rock mass mechanical parameters corresponding to the comprehensive confidence in the fourth preset range, lossless compression storage is adopted, and the complete original physical parameters and feature vectors are retained;

[0058] For the rock mass mechanical parameters corresponding to the comprehensive confidence in the fifth preset range, moderate compression storage is adopted, and only the core physical parameters and residual information are retained;

[0059] For the rock mass mechanical parameters corresponding to the comprehensive confidence in the sixth preset range, only the abnormal identifier and key error parameters are stored.

[0060] The beneficial effects of this invention are as follows: The dynamic sensing method for deep rock mechanics parameters based on multi-source drilling data fusion provided by this invention overcomes the uncertainty and limitations of single-source data source inversion models by collecting multi-source drilling data and combining it with an attention-based graph neural network (AGNN) with physical constraints for information fusion. Simultaneously, it deeply embeds rock mechanics laws into the data fusion process, forcing the fusion and inversion process to strictly adhere to physical principles such as Hooke's law, effectively avoiding unreasonable outputs from purely data-driven models, and significantly improving the accuracy and reliability of rock mechanics parameter inversion. Furthermore, the adaptive working condition module identifies sudden changes in working conditions such as stuck drill and fractured zones in real time and dynamically adjusts the topology and parameter association logic of the graph neural network, enabling the model to flexibly respond to the complex and ever-changing geological environment downhole, further improving parameter sensing. Accuracy and reliability: A comprehensive confidence scoring mechanism is introduced, which performs multi-dimensional quantitative evaluation of the prediction results for each data point by comprehensively considering physical rule residuals, disturbance sensitivity, and spatial consistency errors, providing quantifiable confidence assessment and decision-making basis; by converting rock mechanics parameters into risk indicators in real time and providing graded early warnings, intelligent perception of rock mass safety in deep engineering is realized, effectively avoiding major safety accidents such as well wall collapse; a confidence-based residual-first dynamic storage strategy achieves the best balance between the integrity and storage efficiency of key data under limited storage resources, optimizing data storage efficiency and value mining capabilities. Attached Figure Description

[0061] Figure 1 This is a flowchart illustrating the dynamic sensing method for deep rock mechanical parameters based on multi-source drilling data fusion, provided as an example. Detailed Implementation

[0062] Current methods for interpreting drilling parameters and sensing rock mechanics rely on a single data source for parameter inversion and prediction, lack physical constraints, and are prone to prediction failure due to sudden changes in working conditions, resulting in poor accuracy and reliability of rock mechanics parameters.

[0063] Based on this, the technical scheme of the present application is proposed. In the present application, first, the real-time drilling parameters directly reflecting the rock-while-drilling interaction are obtained through the sensor array built-in the drill rod; at the same time, the auxiliary parameters characterizing the formation characteristics are obtained through the monitoring tool arranged in the downhole, realizing multi-dimensional perception of the downhole environment, and the collected multi-source, heterogeneous, noisy raw data are preprocessed to provide high-quality standardized feature data input for subsequent analysis. Then, the borehole space is discretized into a graph structure, where each node represents a depth point, and the feature vector of the node is all the preprocessed parameters of the point; the physical law is deeply embedded in the graph neural network, and through the physical constraint edge weight calculation and physical guided feature update, the fusion and inversion process is forced to strictly comply with the physical criteria such as Hooke's law, thereby ensuring the accuracy and reliability of the rock mass mechanical parameter interpretation; after calculating the edge weight between nodes, the graph neural network starts message passing and node feature updating, and then translates the high-dimensional features output by the AGNN into rock mass mechanical parameters that can be directly applied to engineering through the decoding layer, and evaluates the comprehensive confidence of these parameters. Then, dynamically perceive the working condition changes and adjust the connection logic of the graph network accordingly, so that the model can adapt to complex geological environments; at the same time, through the federated learning framework, each drilling machine only shares encrypted model gradients after local training, and aggregates and updates in the cloud, thereby protecting data privacy and fusing multi-party knowledge to improve the generalization ability of the model. Finally, according to its comprehensive confidence, the data is stored differently, and the integrity of high-value data is preferentially guaranteed, while low-value or low-confidence data is compressed or only stored as an abstract, thereby intelligently optimizing the utilization rate of storage resources.

[0064] The technical solutions in the embodiments will be clearly and completely described below with reference to the drawings in the embodiments. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments.

[0065] Figure 1 A flowchart of a deep rock mass mechanical parameter dynamic perception method based on multi-source drilling data fusion is shown, please refer to Figure 1 The method comprises the following steps:

[0066] Step 1, acquisition and preprocessing of multi-source drilling data:

[0067] Real-time acquisition and preprocessing of multi-source drilling data to obtain standardized features corresponding to the multi-source drilling data; the multi-source drilling data includes drilling parameters obtained by the sensor array built-in the drill rod during drilling, and auxiliary formation parameters obtained by the monitoring tool arranged in the downhole or around the well.

[0068] In the embodiment, the while-drilling parameters include measured torque, rotation speed and drilling speed of the drill bit; the auxiliary formation parameters include rock density, longitudinal wave speed, transverse wave speed, absolute permeability, compressibility and maximum horizontal principal stress; and the rock mass mechanical parameters include uniaxial compressive strength, elastic modulus, Poisson's ratio, rock strength index and drilling condition compliance index.

[0069] In practical application, the built-in sensor array is arranged in the TBM (Tunnel Boring Machine) cutter drive part and the main bearing, mainly including torque sensors, rotation speed sensors and the like to collect the while-drilling parameters in the drilling process, and the auxiliary formation parameters are obtained through the monitoring tools arranged in the well or around the well, such as acoustic logging sensors and resistivity sensors.

[0070] In the embodiment, the preprocessing includes: adaptively switching the noise reduction algorithm according to the real-time drilling conditions, the real-time drilling conditions including sticking, broken zone and normal drilling; aligning the depth and removing the outliers, so that the multi-source while-drilling data are aligned in time sequence and spatial depth; and calling the corresponding compensation algorithm for automatic correction for different rock types.

[0071] Specifically, the preprocessing part adopts a three-layer architecture design, including:

[0072] Signal layer processing: the working condition adaptive noise reduction directly improves the calculation accuracy of the uniaxial compressive strength, rock strength index and drilling condition compliance index, the Kalman filter eliminates the torque jump when sticking, thereby reducing the uniaxial compressive strength calculation error; the notch filter preserves the acoustic wave main frequency when in the broken zone, ensuring more accurate Poisson's ratio inversion; and the wavelet noise reduction optimizes the vibration signal when in normal drilling, which is beneficial to more stable drilling condition compliance index calculation.

[0073] Data layer cleaning: through depth alignment and outlier removal, the spatio-temporal consistency of the parameters is ensured.

[0074] Engineering layer compensation: for different rock types, the parameter deviation caused by environmental interference is automatically corrected.

[0075] Step 2, multi-source data fusion and parameter inversion under physical rule constraints:

[0076] Based on the standardized features, a graph neural network combining attention mechanism and physical constraints is constructed; the graph neural network defines different depth points of the borehole as nodes, and the node feature vector contains the while-drilling parameters and auxiliary formation parameters; the edge weight between nodes is dynamically calculated according to the physical constraints, statistical correlation and spatial distance, and the physical knowledge vector is embedded in the node feature update for guidance, and the node features are updated through the message passing mechanism; the updated high-dimensional node features are interpreted for physical properties, and the rock mass mechanical parameters are output, and the comprehensive confidence is calculated.

[0077] It can be understood that this step is used to adopt a graph neural network (AGNN) combined with attention mechanism and physical constraints, dynamically calculate the edge weight between nodes by using the rock mass mechanics knowledge base, and embed the physical rule residual in the feature update to force the fusion process to strictly comply with the physical laws such as Hooke's law. The final interpretation output is a high-credibility rock mass mechanical parameter and a real-time risk index, realizing accurate and interpretable dynamic perception from multi-source noisy data to rock mass mechanical parameters.

[0078] Specifically, the embodiment discretizes the drilling space into a graph structure, where each node represents a depth point, and the node feature vector is all the preprocessed parameters of the point.

[0079] Subsequently, the edge weight between nodes is calculated, and the calculation formula is as follows:

[0080] ;

[0081] wherein, represents the edge weight between adjacent nodes and node , represents the weight coefficient of the physical constraint term, represents the weight coefficient of the statistical correlation term, represents the weight coefficient of the spatial distance term, represents the elastic modulus of node , represents the elastic modulus of node , represents the tolerance threshold of the elastic modulus, represents the Pearson correlation coefficient between the feature vector of node and the feature vector of node , represents the Euclidean distance between the spatial coordinates of node and the spatial coordinates of node , represents the maximum neighborhood distance.

[0082] The greater the edge weight, the more "similar" or "correlated" the two nodes are in the physical and data statistical sense, the stronger the information transmission, that is, the greater the result of the formula, indicating that the reasonable connection strength of the two nodes is higher.

[0083] In the calculation formula of the above edge weight:

[0084] represents the physical constraint term, which has the highest proportion and the highest weight, such as =0.6; wherein, and The "intrinsic stiffness" of the rock mass, which directly represents the integrity of the rock mass, is most representative as a physical constraint, and the difference in elastic modulus of the two nodes helps to avoid the model treating the nodes on both sides of the fault as continuous rock mass; in normal rock mass without faults and fracture zones, the change in elastic modulus within a 5m range usually does not exceed 2GPa, so it can be taken that = 2GPa.

[0085] The statistical correlation term is represented, and the weight is lower than the physical constraint, which can be taken as = 0.3; Quantifies the degree of linear correlation of node observation data in statistics, the value range is [-1, 1], positive value represents positive correlation, negative value represents negative correlation, and the absolute value is greater, the correlation is stronger.

[0086] The spatial distance penalty term is represented, and the weight is the lowest, which can be taken as = 0.1; In drilling, the rock mass properties change little within a 5m range, and the continuity of the rock mass will be significantly weakened beyond 5m, so it can be taken that = 5m.

[0087] In the calculation formula of the above edge weight, the three rules of "physical constraint (elastic modulus continuity), statistical correlation (permeability correlation), and spatial continuity (distance)" are integrated, and each parameter has a clear rock mechanics basis and engineering experience support, avoiding unreasonable connections driven by pure data from the source. The "isolated data points" in step 1 are translated into "structured relationship network consistent with geological laws", which provides a "high-quality data basis" for subsequent feature fusion, parameter interpretation, and AI prediction.

[0088] In this embodiment, the deterministic rules of rock mechanics (such as wave velocity-elastic modulus relationship, Poisson's ratio value range) are converted into model recognizable vector signals through node feature updating, that is, "physical knowledge vectors" are used to guide model learning, avoiding deviation from mechanical criteria. The node feature update adopts a physically guided residual connection mechanism, and the update formula is:

[0089] ;

[0090] Wherein, represents the feature vector of node in the first layer neural network, represents the feature vector of node in the first layer neural network, represents the feature vector of node and node edge weight between nodes, representing nodes In the first feature vector of the layer neural network, respectively represent the learnable weight matrix, representing nodes adjacent node set of the node, representing GELU activation function, representing the physical knowledge vector function.

[0091] It should be noted that a purely data-driven model may learn features that violate the mechanical criteria, such as a Poisson's ratio greater than 0.5, while will directly embed the correct physical parameters into the feature update, guide the model learning, and ensure that each layer of feature update does not deviate from the rock mass mechanics law.

[0092] It can be understood that the high-dimensional feature vector obtained through node feature update is not directly convenient to use, and the updated high-dimensional node feature needs to be physically interpreted to output rock mass mechanical parameters and calculate the comprehensive confidence, that is, to map abstract features to physical attribute vectors containing rock mass mechanical parameters such as uniaxial compressive strength and elastic modulus, to provide directly applicable mechanical parameters for subsequent rock mass stability prediction, parameter inversion and other downstream tasks, and realize the conversion from fused features to practical parameters. Specifically, it includes:

[0093] The rock mass mechanical parameters are interpreted and output through the physical attribute interpretation formula, and the physical attribute interpretation formula is as follows:

[0094] ;

[0095] wherein, representing nodes output predicted physical attribute vector, including output vector of multiple rock mass mechanical parameters, representing the learnable decoding weight matrix, representing nodes high-dimensional feature vector fused with multi-source data and physical constraints output by the final layer of the graph neural network, representing the bias term;

[0096] After the physical attribute vector is interpreted, the node-level residual error is calculated:

[0097] ;

[0098] wherein, representing the node-level residual error under the physical rule, representing the predicted physical attribute vector Theoretical value calculated by physical rule formula in rock mass mechanics knowledge base, represents the real theoretical value corresponding to the physical rule, represents taking the absolute value;

[0099] Add Gaussian noise to the predicted physical property vector for perturbation test, and calculate the perturbation sensitivity coefficient:

[0100] ;

[0101] wherein, represents the perturbation sensitivity coefficient of the th rock mass mechanics parameter in the predicted physical property vector, represents the perturbed rock mass mechanics parameter obtained by interpreting the original input rock mass mechanics parameter with Gaussian noise through the graph neural network, represents the L2 norm of the vector;

[0102] The comprehensive confidence is calculated by the following formula:

[0103] ;

[0104] wherein, represents the comprehensive confidence of node , represents the node comprehensive residual vector of node , represents the node comprehensive perturbation sensitivity coefficient of node , represents the consistency error of node with the predicted value of its neighbor nodes, represents the natural constant.

[0105] The comprehensive confidence is the comprehensive score of the reliability of the rock mass mechanics parameter:

[0106] ≥0.8 indicates that the parameter is reliable and can be directly used for drilling pressure optimization and other decisions;

[0107] 0.6≤ <0.8 indicates low confidence, which needs to be combined with downhole camera verification;

[0108] <0.6 indicates that it is unreliable and needs to suspend operation and reacquire data.

[0109] Step 3, prediction model construction of AI driven model:

[0110] According to the rock mass mechanical parameters and the comprehensive confidence, based on a physical-data dual driving mechanism to ensure that the prediction of the rock mass mechanical parameters conforms to the rock mass constitutive relation, based on a dynamic graph topology reconstruction mechanism adaptive to working conditions, the current drilling working condition is identified in real time, the edge weight and the adjacency matrix of the graph neural network are dynamically adjusted according to the identified working condition in real time, and through a multi-scale space-time alignment federated learning framework, multi-drilling machine collaborative training and model updating are realized, and risk indicators and decision suggestions are output.

[0111] It can be understood that step 3 receives the rock mass mechanical parameters and the comprehensive confidence output by step 2, ensures that the parameters conform to the mechanical criteria through the physical-data dual driving mechanism, responds to drilling mutations through the dynamic graph topology reconstruction adaptive to working conditions, and realizes the safety cooperation of multiple drilling machines through the multi-scale space-time alignment federated learning, forming a closed loop from perception to decision.

[0112] Among them, the physical-data dual driving mechanism mainly solves the problem that the traditional pure data driven model ignores the rock mass constitutive relation, leading to the parameter inversion result violating the basic mechanical criteria, and through the deep fusion of physical rules and data learning, the reliability of the predicted parameters is ensured. The physical-data dual driving mechanism comprises:

[0113] The core law of rock mass mechanics is coded as a graph structure prior constraint, and the difference range of the elastic modulus of adjacent nodes is determined:

[0114] ;

[0115] Among them, represents the elastic modulus of node , represents the elastic modulus of node , represents a configurable tolerance threshold, represents taking an absolute value;

[0116] The edge weight calculation function of the physical constraint is designed as:

[0117] ;

[0118] Among them, represents the edge weight between adjacent nodes and node , represents the association strength of the physical properties between node and node , represents the Pearson correlation coefficient between the permeability of node and node , represents a weight coefficient.

[0119] In addition, the module adopts a physically guided residual learning mechanism, which explicitly embeds the conversion formula of wave velocity and elastic parameters and other physical knowledge in the network update process of each layer, and realizes the convergence of the prediction result to the theoretical value through the following update rule:

[0120] ;

[0121] wherein, represents the node in the feature vector of the first layer neural network, represents the node in the feature vector of the first layer neural network, represents the edge weight between the node and the node , represents the node in the feature vector of the first layer neural network, respectively represent the learnable weight matrix, represents the adjacent node set of the node , represents the GELU activation function, represents the physical knowledge vector function.

[0122] The working condition adaptive dynamic graph topology reconstruction is used to cope with the limitations of static models that cannot deal with sudden working conditions (such as sticking, entering broken zones, etc.) in the drilling process. By real-time sensing of working condition changes, dynamically adjusting parameter correlation logic and reconstructing graph topology structure, the model can adapt to complex geological environments.

[0123] Specifically, first, based on the vertical acceleration, torque rate of change and horizontal principal stress gradient and other data collected by the IMU sensor on the drill rod, a working condition classifier is constructed to realize real-time identification of the current drilling state (normal drilling, sticking, broken zone, etc.). Then, according to the identified working condition label, the parameter correlation logic is dynamically adjusted, for example, in the sticking working condition, to strengthen the correlation between uniaxial compressive strength and related parameters, the weight coefficient of the statistical correlation term in the edge weight formula is increased from 0.3 to 0.7; in the broken zone working condition, to enhance the spatial continuity of rock mass parameters, the weight coefficient of the spatial distance term is reduced from 0.1 to 0.01; and in the normal working condition, the default parameters are maintained. At the same time, the adjacency matrix of the graph is updated in real time according to the working condition label, allowing connections to be established across fault nodes to adapt to the discontinuity of the broken zone, ensuring that the model still maintains stable prediction performance when the working condition changes.

[0124] ​​The multi-scale space-time alignment federated learning framework is used to solve the problem of difficult secure sharing of downhole and cloud data in multi-drill cooperative operation, and the limited generalization ability of single-drill model. Through hierarchical feature compression, physical constraint federated learning and distributed model updating, the module realizes safe and efficient cooperation of multiple devices.

[0125] Specifically, first, a rock mass physical feature extractor is designed to compress the features output by the final layer of AGNN, reducing the high-dimensional feature vector to 30% of the original dimension. This preserves core physical information such as elastic modulus and permeability while reducing data transmission and ensuring data security.

[0126] In the federated learning process, a loss function that combines physical rule residuals is used for training, and its expression is:

[0127] ;

[0128] where, represents the global loss function of federated learning, represents the standard data error term represents the physical rule residual vector, represents the physical constraint weight (default 0.5), which ensures that the distributed learning process does not violate the rock mass strength theory, represents the norm of the vector.

[0129] In terms of model updating, after each drill node is trained based on local data, it only uploads encrypted model gradients to the cloud server. The cloud server allocates aggregation weights based on the data volume and reliability of each drill node, and aggregates the global model through weighted averaging and updates it. This mechanism protects data privacy while integrating the geological experience of multiple drills, significantly improving model generalization ability and reducing generalization error.

[0130] The above three mechanisms work together. The physical-data dual-driven mechanism ensures the physical reasonableness of parameter prediction, the working condition self-adaptive reconstruction ensures the stable output of the model in complex environments, the federated learning framework improves the efficiency of multi-device cooperation, and together they form a complete closed loop from perception to decision-making, providing strong support for safe and efficient construction of deep rock mass engineering.

[0131] Step 4, data integration and storage:

[0132] Based on the physical indicators, a database index is constructed, a hierarchical early warning is performed according to the risk indicators, and a differentiated dynamic storage strategy is implemented according to the comprehensive confidence.

[0133] It can be understood that after obtaining the rock mass mechanical parameters, the risk index and the decision suggestion, a physical constraint driven space-time database architecture is constructed to realize the whole process data management from while-drilling perception to intelligent decision-making, and to provide data support and real-time response capability for deep engineering rock mass safety intelligent management.

[0134] Specifically, to realize efficient data retrieval and intuitive presentation of three-dimensional geological state, a hierarchical index tree structure based on physical indicators is adopted to break through the index limitation of traditional single dimension of time or space. The first-level index takes wave velocity ratio as the core indicator, and divides the lithology category according to the wave velocity ratio characteristics of different lithologies to quickly locate the lithology attribute of the target formation; the second-level index is based on elastic modulus curvature, and identifies the stress concentration area by calculating the second-order derivative of the elastic modulus of adjacent depth.

[0135] In this embodiment, the calculation formula of the risk index is as follows:

[0136] ;

[0137] wherein, represents the risk index, represents the uniaxial compressive strength prediction value, represents the maximum horizontal principal stress;

[0138] The graded early warning includes:

[0139] When the risk index is in a first preset range (such as ), it is determined as high risk, and an emergency stop drilling instruction is triggered to avoid safety accidents such as well wall collapse;

[0140] When the risk index is in a second preset range (such as ), it is determined as medium risk, and a decision suggestion of reducing the rotating speed and starting the support plan is automatically issued;

[0141] When the risk index is in a third preset range (such as ), it is determined as low risk, and the current torque, rotating speed and drilling speed are maintained, and continuous monitoring is carried out.

[0142] It should be noted that the risk index is not a fixed value, but is dynamically updated with the statistical results of multiple drilling machines working together, and the threshold interval is continuously optimized by fusing geological data of different drill holes to improve the adaptability and accuracy of risk assessment.

[0143] In this embodiment, a differentiated dynamic storage strategy is implemented according to the comprehensive confidence, including:

[0144] For the rock mass mechanical parameters corresponding to the comprehensive confidence in a fourth preset range (such as ≥0.8), lossless compression storage is adopted to retain the complete original physical parameters and feature vectors;

[0145] For the comprehensive confidence in the fifth preset range (such as 0.6≤ <0.8) corresponding to the rock mass mechanical parameter, moderate compression storage is adopted, and only the core physical parameter and residual information are reserved;

[0146] For the comprehensive confidence in the sixth preset range (such as <0.6) corresponding to the rock mass mechanical parameter, only the abnormal identifier and key error parameter are stored.

[0147] Specifically, in order to guarantee the integrity of key data under limited storage resources, a differentiated storage strategy is implemented according to the comprehensive confidence:

[0148] For high-confidence data (≥0.8), lossless compression storage is adopted, and the original physical parameter and feature vector are reserved, which are used for model iteration and geological rule analysis; for medium-confidence data (0.6≤ <0.8), moderate compression storage is adopted, and the core parameter and residual information are reserved, which support subsequent verification and correction; for low-confidence data (0.6 <0.6), only the abnormal identifier and key error parameter are stored, which reduces the space occupied by redundant data. This dynamic storage strategy not only ensures the complete retention of high-risk and high-value data, but also reduces the storage cost through targeted compression, achieving the balance between data storage efficiency and security.

[0149]

[0150] ​To sum up, the deep rock mass mechanical parameter dynamic perception method based on multi-source drilling data fusion provided by the embodiment overcomes the uncertainty and limitations of single data source inversion model by collecting multi-source drilling data and combining AGNN for information fusion, at the same time, the depth of rock mass mechanical law is embedded in the data fusion process, forcing the fusion and inversion process to strictly comply with physical criteria such as Hooke's law, effectively avoiding the unreasonable output of pure data-driven models, and greatly improving the accuracy and reliability of rock mass mechanical parameter inversion; through the working condition adaptive module, the sudden working conditions such as sticking and broken zone are identified in real time, and the topological structure and parameter correlation logic of the graph neural network are dynamically adjusted, so that the model can flexibly respond to the complex and changeable geological environment underground, further improving the accuracy and reliability of parameter perception; the comprehensive confidence score mechanism is introduced, the prediction results of each data point are quantitatively evaluated in multiple dimensions by comprehensively considering the physical rule residual, disturbance sensitivity and spatial consistency error, and a quantifiable confidence evaluation and decision basis is provided; by converting the rock mass mechanical parameters into risk indicators in real time and grading the warning, the intelligent perception of the safety of deep engineering rock mass is realized, and major safety accidents such as well wall collapse are effectively avoided; based on the confidence residual priority dynamic storage strategy, the best balance between the integrity and storage efficiency of key data under limited storage resources is realized, and the data storage efficiency and value mining ability are optimized.

Claims

1. A deep rock mass mechanical parameter dynamic perception method based on multi-source while-drilling data fusion, characterized in that, The method comprises: Real-time acquisition and preprocessing of multi-source while-drilling data to obtain standardized features corresponding to the multi-source while-drilling data; the multi-source while-drilling data includes while-drilling parameters obtained by a drill pipe built-in sensor array during drilling, and auxiliary formation parameters obtained by a monitoring tool arranged in a well or around a well; Based on the standardized features, a graph neural network combining attention mechanism and physical constraints is constructed; the graph neural network defines different depth points of a borehole as nodes, and a node feature vector includes the while-drilling parameters and the auxiliary formation parameters; an edge weight between nodes is dynamically calculated according to physical constraints, statistical correlation and spatial distance, and a physical knowledge vector is embedded in node feature updating to guide, and node features are updated through a message passing mechanism; high-dimensional node features after updating are interpreted for physical properties, rock mass mechanical parameters are output, and a comprehensive confidence is calculated; According to the rock mass mechanical parameters and the comprehensive confidence, a physical-data dual driving mechanism is used to ensure that the prediction of the rock mass mechanical parameters conforms to the rock mass constitutive relation, a dynamic graph topology reconstruction mechanism adaptive to working conditions is used to identify a current drilling working condition in real time, an edge weight and an adjacency matrix of the graph neural network are dynamically adjusted according to the identified working condition, and a multi-scale space-time aligned federated learning framework is used to realize collaborative training and model updating of multiple drilling rigs, and risk indicators and decision suggestions are output; A database index is constructed based on physical indicators, a hierarchical early warning is performed according to the risk indicators, and a differential dynamic storage strategy is implemented according to the comprehensive confidence.

2. The method according to claim 1, wherein, The while-drilling parameters include measured torque, rotational speed and drilling speed of a drill bit; The auxiliary formation parameters include longitudinal wave velocity, transverse wave velocity, absolute permeability and compressibility; The rock mass mechanical parameters include uniaxial compressive strength, elastic modulus, Poisson's ratio, rock strength index and drilling condition conformity index.

3. The method according to claim 1, wherein, The preprocessing comprises: According to a real-time drilling working condition, a noise reduction algorithm is adaptively switched, the real-time drilling working condition includes a stuck drill pipe, a broken zone and normal drilling, multi-source while-drilling data are aligned in time series and spatial depth through depth alignment and abnormal value elimination, and different rock types are corrected automatically by calling corresponding compensation algorithms.

4. The method according to claim 1, wherein, A calculation formula of the edge weight between the nodes is: ； wherein, denotes the edge weight between adjacent nodes and nodes , denotes the weight coefficient of the physical constraint term, denotes the weight coefficient of the statistical correlation term, denotes the weight coefficient of the spatial distance term, denotes the elastic modulus of node , denotes the elastic modulus of node , denotes the tolerance threshold of the elastic modulus, denotes the feature vector of node , denotes the Pearson correlation coefficient between the feature vector of node and the feature vector of node , denotes the spatial coordinates of node , and the Euclidean distance between the spatial coordinates of node and node , denotes the maximum neighborhood distance.

5. The method according to claim 1, wherein, The node feature updating adopts a physically guided residual connection mechanism, and an updating formula is: ； wherein, representing a node In a first feature vector of a layer neural network, representing a node In a first feature vector of a layer neural network, representing a node edge weight between nodes , representing a node In a first feature vector of a layer neural network, respectively representing a learnable weight matrix, representing a set of adjacent nodes of a node representing a GELU activation function, representing a physical knowledge vector function.

6. The method according to claim 1, wherein, The output rock mass mechanical parameters and the calculation of the comprehensive confidence comprise: The output rock mass mechanical parameters are interpreted by a physical property interpretation formula as follows: ； wherein, representing a node an output predicted physical property vector, comprising an output vector of a plurality of rock mass mechanical parameters, representing a learnable decoding weight matrix, representing a node a high-dimensional feature vector fused with multi-source data and physical constraints output by a final layer of the graph neural network, representing a bias term; After the physical property vector is interpreted, a node-level residual is calculated: ； wherein, represents the node-level residual under the i-th physical rule, represents the predicted physical attribute vector the theoretical value calculated by the formula of the physical rule in the rock mass mechanics knowledge base, represents the real theoretical value corresponding to the physical rule, represents taking the absolute value;​ A Gaussian noise is added to the predicted physical property vector for perturbation test, and a perturbation sensitivity coefficient is calculated: ； wherein, represents a perturbation sensitivity coefficient of the i-th rock mass mechanical parameter in the predicted physical property vector, represents the perturbed rock mass mechanical parameter obtained by interpreting the original input rock mass mechanical parameter added with Gaussian noise through the graph neural network, represents the L2 norm of a vector;​ The comprehensive confidence is calculated by the following formula: ； wherein, represents the node comprehensive confidence, represents the node comprehensive residual vector of the node, represents the node comprehensive perturbation sensitivity coefficient of the node, represents the node consistency error with the predicted value of its neighbor node, represents the natural constant.

7. The method according to claim 1, wherein, The physical-data dual driving mechanism comprises: A graph structure prior constraint is coded as a rock mass mechanical core law to clearly define a difference range of elastic moduli of adjacent nodes: ； wherein, represents the modulus of elasticity of the node , represents the modulus of elasticity of the node , represents the configurable tolerance threshold, represents taking the absolute value; A physical constraint edge weight calculation function is designed: ； wherein, denotes the edge weight between adjacent nodes and nodes , denotes the associated strength of the physical property between nodes and nodes , denotes the Pearson correlation coefficient between the permeability of nodes and nodes , denotes the weight coefficient.

8. The method according to claim 1, wherein, The federated learning framework comprises: A loss function fusing a physical rule residual is used for training, and an expression is as follows: ； wherein, represents the federated learning global loss function, represents the standard data error term represents the physical rule residual vector, represents the physical constraint weight, represents the norm of a vector; After each drilling rig node is trained based on local data, only an encrypted model gradient is uploaded to a cloud server. The cloud server allocates aggregation weights according to the data volume and reliability of each drilling rig node, and obtains a global model by weighted average aggregation and issues an update.

9. The method according to claim 1, wherein, The calculation formula of the risk index is as follows: ； wherein, denotes a risk indicator, denotes a uniaxial compressive strength prediction value, denotes the maximum horizontal principal stress; The hierarchical early warning comprises: When the risk index is in a first preset range, it is determined as high risk, and an emergency drilling stop instruction is triggered; When the risk index is in a second preset range, it is determined as medium risk, and a decision suggestion of reducing the rotating speed and starting a support plan is automatically issued; When the risk index is in a third preset range, it is determined as low risk, and the current torque, rotating speed and drilling speed are maintained, and continuous monitoring is performed.

10. The method according to claim 1, wherein the method is characterized in that, According to the comprehensive confidence, a differentiated dynamic storage strategy is implemented, comprising: For the rock mass mechanical parameters corresponding to the comprehensive confidence in a fourth preset range, lossless compression storage is adopted, and complete original physical parameters and feature vectors are retained; For the rock mass mechanical parameters corresponding to the comprehensive confidence in a fifth preset range, moderate compression storage is adopted, and only core physical parameters and residual information are retained; For the rock mass mechanical parameters corresponding to the comprehensive confidence in a sixth preset range, only an abnormal identifier and key error parameters are stored.

Images