Rock mass integrity identification method based on gradient boosting decision tree and while-drilling parameters

Abstract

A rock mass integrity identification method based on gradient boosting decision tree and drilling parameters. The drilling parameters and borehole wall image information are obtained, the statistical analysis method is applied, and the classic rock mass quality classification technology is referred to, the database containing rock mass drilling parameters and integrity grade is constructed, and the correlation analysis is carried out; based on the established database, the modeling work is carried out, the gradient boosting decision tree GBDT classification algorithm is introduced; with the help of grey wolf optimization algorithm GWO, the intelligent configuration of the core parameters in the GBDT classifier is completed, and the hybrid classification model GWO-GBDT with reliable performance and stable generalization is formed; a variety of evaluation indexes are applied to screen the optimal hybrid classification model, and the conversion precision of drilling parameters and rock mass integrity is ensured to reach the applicable standard; the importance analysis result of input parameters is obtained, and the corresponding two-dimensional prediction analysis graph is derived from the GWO-GBDT model. The method can scientifically and reasonably realize the high-precision identification of rock mass integrity through an integrated process.

Description

Technical Field

[0001] This invention belongs to the field of geotechnical engineering data processing technology, specifically relating to a rock mass integrity identification method based on gradient boosting decision tree and drilling parameters. Background Technology

[0002] In the fields of mining and rock engineering, identifying underground rock mass information using surveying techniques is a crucial guarantee for safe construction. For underground rock masses with ambiguous information, the reliability of design schemes and decision-making cannot be guaranteed without a reasonable understanding of rock strata distribution, basic structural features, the location of key weak interlayers, and especially the integrity of the rock mass. Geophysical exploration, core sampling, and logging while drilling are conventional techniques for identifying this type of geological information. Core sampling and logging experiments, in particular, are widely used in underground mining, tunnel construction, and the construction of large buildings and structures. However, while these techniques provide complete and accurate geological information, they are often complex, time-consuming, and labor-intensive, requiring significant production time and economic costs, and relying heavily on the professional skills and experience of the surveyors. Furthermore, in many engineering practices, core samples are difficult to obtain, or only qualitative information about the lithology of the study area is needed, posing a significant challenge to traditional identification techniques. Numerous studies have shown a strong correlation between drilling parameters such as drilling rate, rotational speed, torque, and pressure during drilling and rock strength and structural parameters of the rock mass. Numerous scholars have conducted correlation analyses between drilling parameters and rock mass properties using statistical analysis, empirical criteria, and numerical simulations, and have proposed corresponding prediction models and schemes. However, these methods still suffer from limitations such as subjective experience interference, poor prediction accuracy, and limited adaptability. Further research and development of more reliable prediction models are needed. Therefore, there is an urgent need to provide a method that can overcome the shortcomings of traditional methods and identify rock mass integrity based on drilling parameters. Summary of the Invention

[0003] To address the problems existing in the prior art, this invention provides a rock mass integrity identification method based on gradient boosting decision tree and drilling parameters. This method can scientifically and rationally identify rock mass integrity through an integrated process, which is conducive to significantly reducing production costs and has good economic benefits. At the same time, it can significantly simplify the exploration process, shorten the exploration cycle, and greatly improve the prediction accuracy, which helps to predict rock mass integrity information in advance.

[0004] To achieve the above objectives, this invention provides a rock mass integrity identification method based on gradient boosting decision trees and drilling parameters, specifically including the following steps:

[0005] Step 1: Sample collection and database setup;

[0006] S11: Determine the borehole location and use a drilling instrument to record and collect various key drilling parameters during the drilling process;

[0007] S12: Use an imager to acquire image data of the entire borehole wall;

[0008] S13: Conduct logging work based on core samples, analyze drilling parameter data and borehole wall image data, understand the variation law of drilling parameters throughout the borehole, complete the core integrity classification, and construct a drilling database;

[0009] S14: Conduct correlation analysis on the constructed database to explore the degree of correlation between various parameters;

[0010] Step 2: Construct the GBDT basic model for rock mass integrity prediction;

[0011] The gradient boosting decision tree (GBDT) algorithm was selected as the basic classifier for rock mass integrity identification, and the GBDT basic model for rock mass integrity prediction was completed.

[0012] Step 3: Modeling and optimization of the hybrid classification model GWO-GBDT;

[0013] The iterative optimization mechanism of GWO is adopted to extensively search for possible candidate solutions within the set search space to find the optimal GBDT parameter configuration. The optimization strategy GWO is combined with the classification algorithm GBDT, and different hybrid classification models GWO-GBDT are constructed using different population sizes. The best hybrid classification model GWO-GBDT is then selected through comparative analysis.

[0014] Step 4: Identify the integrity of the rock mass and obtain a predictive analysis chart reflecting the rock mass integrity classification;

[0015] The drilling database constructed in Step 1 is input into the optimal hybrid classification model GWO-GBDT, and the complex nonlinear relationship between input and output parameters is captured to achieve non-contact identification of rock mass integrity by referring only to the drilling parameters. Parameter importance analysis is performed using the optimal hybrid classification model GWO-GBDT to select the two drilling parameters with the highest contribution to integrity prediction. Only the two highest-contribution drilling parameters are retained in the drilling database constructed in Step 1, and obvious outliers are removed to obtain a new drilling database. The hybrid classification model GWO-GBDT is modeled again based on the new drilling database. A three-class database is constructed using the two highest-ranking drilling parameters and integrity as the data basis for graphical output, and then input into the new hybrid classification model GWO-GBDT to obtain and display the corresponding predictive analysis graph of rock mass integrity zoning results.

[0016] As an alternative, several key drilling parameters in S11 include depth, feed rate, thrust pressure, pump pressure, torque pressure, and rotational speed.

[0017] As a preferred option, an intelligent drilling 3D television imager is used in S12 to acquire image data of the entire borehole wall.

[0018] Furthermore, in order to obtain a hybrid classification model GWO-GBDT with high recognition accuracy, the modeling and optimization process of the hybrid classification model GWO-GBDT in step three is as follows:

[0019] S31: Data preprocessing and parameter initialization;

[0020] The drilling database was constructed by splitting it in a 4:1 ratio. 80% of the samples were used to form a training set to train the model, enabling it to acquire predictive capabilities after autonomously learning data information. 20% of the samples were used to form a test set to verify whether the generalization ability of the constructed model met the standard. The built-in parameters of the GWO-GBDT model were also set.

[0021] S32: Fitness evaluation and iterative optimization;

[0022] A fitness function is set based on the cross-validation strategy to evaluate the quality of all candidate solutions generated by the model in each iteration. The optimization rules set based on the GWO strategy are used for iterative loops to continuously improve the quality of candidate solutions given by the hybrid model, promote the solution set to develop in a better direction, and seek the optimal GBDT parameter configuration.

[0023] S33: Model Optimization and Performance Evaluation;

[0024] Different GWO-GBDT models were constructed with population sizes of 10, 20, 30, 40, 50, and 100, and a thorough comparative analysis was conducted. The training and testing results of the models were measured by accuracy, precision, recall, F1 score, confusion matrix, and receiver operating characteristic curve. Each GWO-GBDT model was scored, and the best model was selected. The generalization performance of the best model was comprehensively analyzed to ensure that it has the advantages of high accuracy and high stability in predicting rock mass integrity by combining drilling parameters.

[0025] As a preferred option, the built-in parameters in S31 include the number of iterations, population size, and the range of values ​​for the parameter to be optimized.

[0026] This invention combines drilling experiments, in-hole television image measurement, and core logging operations to obtain sufficient and accurate drilling parameters and borehole wall image data. It then processes this data with existing information to determine the integrity level of different core segments, constructing a reliable drilling database that provides robust data support. Gradient Boosting Decision Tree (GBDT) is used as the foundation for rock mass integrity identification. Leveraging the high-performance Gray Wolf Optimization Algorithm (GWO), the core hyperparameters of GBDT are autonomously configured to improve the performance of the prediction model. This ensures superior classification performance of the hybrid classification model on the drilling database, leading to a high-precision and highly stable hybrid classification model, GWO-GBDT. Comparative analysis of multiple hybrid models selects the optimal hybrid classification model as the basis for rock mass integrity prediction based on drilling parameters. This ensures that the intelligent conversion accuracy between drilling parameters and rock mass integrity meets engineering application standards, demonstrating significant application potential in related fields. Finally, parameter importance analysis identifies the two most important types of drilling parameters for the prediction process and derives corresponding prediction analysis charts, providing an intuitive classification of integrity categories. This offers a convenient, intuitive, and accurate visualization tool for the engineering exploration field. Considering the current state of existing exploration and prediction technologies, such as complex processes, high costs, limited accuracy, and difficulty in expansion, this method proposes a rock mass integrity identification method based on gradient boosting decision trees and drilling parameters, combining emerging digital drilling testing and intelligent prediction technologies. This method can scientifically and rationally identify rock mass integrity through an integrated process, which is conducive to significantly reducing production costs and achieving good economic benefits. At the same time, it can significantly simplify the exploration process, shorten the exploration cycle, and greatly improve the prediction accuracy, which helps to predict rock mass integrity information in advance. This provides a brand-new research idea and practical means for related fields, and can assist in design schemes and on-site decision feedback.

[0027] This invention overcomes the shortcomings of traditional technologies in obtaining underground rock mass information, such as being time-consuming, labor-intensive, and lacking accuracy. It effectively reduces the workload and economic costs of the exploration process and provides a fast and accurate tool for determining the integrity of rock masses on-site. It has excellent innovation and practical value, as detailed below:

[0028] 1. Based on the principles of statistics and machine learning, a high-precision rock mass integrity prediction model is constructed. Combining the novel prediction technology GBDT, it delves into the complex nonlinear relationship between drilling parameters and rock mass integrity, overcoming the limitations of traditional prediction techniques such as empirical criteria, statistical analysis, and numerical simulation, which suffer from poor prediction accuracy and limited applicability. The data-driven model can efficiently process large amounts of data and intelligently make judgments based on learned predictive intelligence.

[0029] 2. By introducing cross-validation and metaheuristic strategies, the generalization performance of the GBDT model is further improved. The GWO algorithm introduced in this invention is a widely acclaimed optimization technique in the field. Combined with swarm intelligence strategies, it can better configure the parameter combination of classification algorithms, giving it better predictive ability on the database and ensuring the reliability of prediction results and the stability of the model.

[0030] 3. The proposed hybrid classification model GWO-GBDT can identify the integrity of unknown underground rock masses based on the drilling parameters collected during the drilling process, and provide qualitative outputs such as complete, relatively complete, relatively broken, or broken. It helps to solve the problem of knowing the basic condition of underground rock masses in engineering practice, overcomes the drawbacks of traditional exploration technology such as complicated operation process, high time and economic cost, and difficulty in real-time prediction, and provides more convenient and accurate practical technology for on-site technicians.

[0031] 4. The predictive analysis diagrams exported by the hybrid classification model GWO-GBDT can provide intuitive and convenient reference. The practical tools based on the diagram format can greatly improve design efficiency, improve the operational complexity presented by previous empirical formulas, index criteria or field measurements, realize the rapid and accurate identification of the integrity classification of unknown underground rock masses, and provide alternative scientific decision support for engineering sites.

[0032] In summary, the method proposed in this invention constructs a reliable drilling database based on extensive field data accumulated in the early stages. It utilizes digital drilling testing and intelligent prediction technologies to explore the correlation mechanism between drilling parameters and rock mass integrity, constructs a stable and reliable identification model GWO-GBDT, and achieves non-contact identification of rock mass integrity based on this model. It also irregularly derives corresponding visual analysis charts, greatly simplifying the lithology identification process. The prediction results are accurate and reliable, overcoming the shortcomings of traditional methods and possessing good practical value and application potential. Attached Figure Description

[0033] Figure 1 This is a flowchart illustrating the present invention. Figure 1 ;

[0034] Figure 2 This is a flowchart illustrating the present invention. Figure 2 ;

[0035] Figure 3 This is a schematic diagram of the database creation process during drilling in this invention;

[0036] Figure 4 This is a correlation analysis diagram of the drilling database in this invention;

[0037] Figure 5 This is a diagram illustrating the modeling principle and overall framework of the GWO-GBDT model in this invention.

[0038] Figure 6 This is a comparison chart of the confusion matrix and ROC curve of the GWO-GBDT and GBDT models in this invention;

[0039] Figure 7 This is the feature importance map output by the GWO-GBDT model in this invention;

[0040] Figure 8 This is a diagram showing the rock mass integrity prediction analysis of the six simple classification models in this invention;

[0041] Figure 9 This is a rock mass integrity prediction and analysis diagram output by the GWO-GBDT model in this invention. Detailed Implementation

[0042] The present invention will be further described below with reference to the embodiments.

[0043] like Figures 1 to 9 As shown, this invention provides a rock mass integrity identification method based on gradient boosting decision trees and drilling parameters, specifically including the following steps:

[0044] Step 1: Sample collection and database setup;

[0045] S11: Determine the borehole location and use a drilling instrument to record and collect various key drilling parameters during the drilling process;

[0046] Based on the original geological data of the drilling test area, the experimental locations, hole depths, and diameters were determined, and a drilling parameter acquisition scheme was designed to ensure the authenticity and completeness of the acquired drilling parameters. Three vertical boreholes were drilled in this experiment, and the PocketLIM multi-functional data acquisition unit in the drilling rig was used to record drilling rig operating data, acquiring various key drilling parameters during the drilling process. These key parameters included depth (in meters), feed rate (in meters per hour), thrust pressure (in MPa), pump pressure (in MPa), torque pressure (in MPa), and rotational speed (in rpm). Complete core samples were obtained from the entire drilling process. These core samples were then analyzed using geological expertise to obtain information on rock type, integrity, basic structure, weathering degree, and core recovery rate. The drilled core samples were then cataloged, recording information such as rock type, integrity, basic structure, weathering degree, and core recovery rate, resulting in cataloged data.

[0047] S12: Use an imager to acquire image data of the entire borehole wall;

[0048] After the borehole is completed, it is cleaned with high-pressure water to ensure that there is no water or dust inside. Then, the water inside the borehole is removed, and after the fog dissipates, the JL-IDOI(C) intelligent borehole 3D television imager is used to record all the borehole wall image data. By analyzing the borehole wall image data, the distribution and development of the strata lithology and geological structure inside the borehole can be detected. This helps to classify the stratigraphic structure, identify the formation fractures and karst development in the well wall, and provide important data for compiling key information such as strata lithology, geological structure, weak interlayers, and fracture development. This provides auxiliary reference for subsequent work.

[0049] S13: Conduct logging work based on core samples, analyze drilling parameter data and borehole wall image data, understand the variation law of drilling parameters throughout the borehole, complete the core integrity classification, and construct a drilling database;

[0050] Statistical analysis methods were applied to comprehensively analyze the drilling parameter data of three boreholes, identifying the range of output values ​​for torque, pressure, pump pressure, and thrust throughout the borehole, as well as the patterns of significant changes in each parameter. A comprehensive analysis of the borehole cores was conducted, combining core logging data, drilling parameter data, and borehole television images, to understand the overall drilling parameter variation patterns, identify core segments with distinct uniform characteristics, and extract a certain length of drilling data (preferably 30-40 cm). After removing obviously abnormal data, the average value was taken as the input value for a sample in the database. Following the classic rock mass quality grading technique and combining the core sample analysis results, logging data, and borehole wall image data analysis results, the core section was qualitatively classified to determine the corresponding integrity level, which was used as the output value for this sample. (If the core is clearly intact, and the borehole television image shows that the core section has almost no or very few joints and fractures, the RQD value for this core sampling cycle is close to 90%–100%, and it is judged as intact; if the core is clearly broken, and the borehole television image shows that the core section has well-developed joints and fractures, the RQD value for this core sampling cycle is low (e.g., below 40% or below 30%, it is judged as broken). Figure 3As shown in Table 1, a drilling database can be constructed with drilling speed (drilling rate), thrust (pump pressure), pump pressure, torque pressure, and rotational speed as input variables and integrity classification as output variable. It should be noted that in the three boreholes drilled in this experiment, relatively few core samples could be classified as either very intact or very broken; the vast majority were either relatively intact or relatively broken. Simply integrating them into a four-category database could lead to a severe imbalance in sample sizes across different categories, affecting the stability and reliability of the model's predictions. Therefore, the target variables were further integrated, ultimately dividing the data into two integrity categories to ensure a more reliable dataset. Core segments originally defined as intact or relatively intact were classified as category 0, indicating good or average integrity, low joint and fracture development, and a high RQD value for this core sampling run. Core segments originally defined as relatively broken or broken were classified as category 1, indicating poor or very poor integrity, high joint and fracture development, and a low RQD value for this core sampling run.

[0051] Table 1. Sample examples of the Drilling Database

[0052]

[0053] S14: Conduct correlation analysis on the constructed database to explore the degree of correlation between various parameters;

[0054] After organizing the drilling database, correlation analysis was performed on the input and output parameters to ensure that the parameter combinations would not cause overfitting due to multicollinearity, and to verify the usability of the database itself. Figure 4 As shown, the five input parameters exhibit a low degree of correlation. The correlation coefficients between torque pressure and propulsion speed and thrust are 0.36 and 0.32, respectively, indicating a certain positive correlation, but the linear correlation is not strong. Therefore, the input parameters remain effective overall. Regarding the output parameters, the correlation coefficients between torque pressure and propulsion speed are 0.67 and 0.37, respectively, meaning that to a certain extent, higher torque pressure and propulsion speed correspond to a higher integrity level (less complete). The other three parameters have lower correlation coefficients.

[0055] Step 2: Construct the GBDT basic model for rock mass integrity prediction;

[0056] The gradient boosting decision tree (GBDT) algorithm was selected as the basic classifier for rock mass integrity identification, and the GBDT basic model for rock mass integrity prediction was completed.

[0057] Simply analyzing the novel model proposed in this invention is not convincing enough. Therefore, to verify whether the new model is sufficient to meet engineering requirements, various classification algorithms were introduced, including Gradient Boosting Decision Tree (GBDT), Random Forest (RF), Support Vector Machine (SVM), Gaussian Naive Bayes (GNB), k-Nearest Neighbors (KNN), and Multilayer Perceptron (MLP). These six algorithms basically cover the classic classification techniques in the learning field. The constructed drilling database was input into the above six algorithms, and the corresponding test accuracy was obtained. Based on the test accuracy, Gradient Boosting Decision Tree (GBDT) was selected as the basic classifier for rock mass integrity identification, as shown in Table 2.

[0058] Table 2 Comparison of test accuracy of six simple classification algorithms

[0059]

[0060] The data in Table 2 show that the GBDT algorithm has the highest performance across all metrics, with a test accuracy of 74.58%, indicating that introducing this integrated algorithm as the basic model for rock mass integrity prediction is the optimal choice.

[0061] The modeling method for GBDT is as follows:

[0062] GBDT can be viewed as a combination of a simple decision tree model and the boosting algorithm. Decision trees are weak classifiers with average performance. When dealing with prediction problems independently, they are prone to overfitting and have very limited accuracy. Boosting technology can effectively connect a certain number of decision trees to combine them into a more powerful ensemble model.

[0063] In the Boosting ensemble method, the dependencies between weak classifiers are strong; different weak classifiers cannot operate independently. Subsequent decision tree models must be built upon the previous model, continuously adjusting the modeling approach based on prior modeling experience to achieve more efficient ensemble integration and drive the classification model towards higher accuracy. The core mechanism of GBDT is to continuously reduce the residuals generated during training, using the negative gradient of the loss function in the current model as an approximation of the residuals in the algorithm to find the optimal solution. During the training phase, each iteration follows the principle of residual reduction to generate a new decision tree; that is, the residual between the training result of the first decision tree and the true value is the target for optimization during the training of the second decision tree.

[0064] The mathematical model of the GBDT algorithm can be understood as follows: the input dataset is regarded as D={(x1,y1),(x2,y2),…,(xN,yN)}, and the input space is divided into J leaf node regions R1,R2,…RJ. First, the classifier is initialized by formula (1), that is, a decision tree with only one root node;

[0065]

[0066] When the model iterates to the mth generation, the negative gradient value γm,i of the i-th input sample is calculated using formula (2) and used as an approximation of the residual.

[0067]

[0068] For each sample in a leaf node region, the optimal output value γ for that node region is calculated using formula (3). jm This minimizes the loss function value;

[0069]

[0070] The decision tree fitting function after this iteration is obtained through formula (4).

[0071]

[0072] To complete the modeling, M iterations are required, and finally a strong classifier composed of multiple basic classifiers is obtained through formula (5);

[0073]

[0074] When tackling prediction tasks, the GBDT model needs to have a corresponding loss function set. For classification problems, the logarithmic function is chosen as the loss function. The specific steps of GBDT operation are as follows:

[0075] Step 1: Initialize using formula (6);

[0076]

[0077] In the formula,

[0078] Step 2: Calculate the negative gradient of the loss function for the i-th input sample in the m-th iteration using formula (2):

[0079] Step 3: Select the k-th feature variable of the i-th sample using formula (7) and its values To serve as the optimal splitting variable and splitting point;

[0080]

[0081] In the formula, for The number of samples, t = 1, 2, l = 1, 2, ..., N. If j = J, it means the tree structure has been grown in this iteration, and we proceed to step four; otherwise, we proceed to step four for the sub-region. The third step of the recursive call;

[0082] Step 4: Calculate the optimal output value γ after the m-th iteration using formula (8). jm ;

[0083]

[0084] Step 5: Update If the final number of iterations is reached, the modeling process is complete, and the model will output the calculation results. Otherwise, iteration must continue, proceeding to the second step.

[0085] In addition, the GBDT model has two core parameters: max_depth and n_estimators. The former is the maximum depth of the generated decision tree, which limits the number of nodes in the decision tree, while the latter refers to the number of iterations of the algorithm modeling, which is equivalent to the number of decision trees in the ensemble model.

[0086] Step 3: Modeling and optimization of the hybrid classification model GWO-GBDT;

[0087] Hybrid modeling was conducted on the Drilling-While-Drilling Database. To further improve model performance and stability, the Grey Wolf Optimization (GWO) algorithm was applied to iteratively tune the GBDT model. The specific principle of the GWO strategy is as follows:

[0088] The core principle of GWO is to achieve iterative optimization of the target problem through biomimetic simulation of gray wolf behavior. In the algorithm, each gray wolf represents a candidate solution, the prey position is the optimal solution, and the wolf pack has a strict hierarchical order, named α, β, δ, and ω wolves in descending order of status. The first three levels of wolves are the leaders, while the remaining wolves (ω wolves) must obey the commands of α, β, and δ wolves. Based on this hierarchical order, GWO constructs a corresponding mathematical model, which consists of three parts:

[0089] Step 1 (Surrounding the prey): The gray wolf pack identifies the location of the prey and surrounds it. The surrounding process is modeled according to formula (9).

[0090]

[0091] In the formula, X represents the current position vector of the gray wolf. P The vector represents the prey's position, and D represents the distance between the prey and the gray wolf. t represents the current iteration number, and A and C are both coefficient vectors. Here, r1 and r2 are both random vectors in [0,1], and a will decrease from 2 to 0 as the iteration proceeds.

[0092] Second (hunting process): This part simulates the hunting behavior of gray wolves, which determines the location update and optimization direction of the gray wolf pack. Since the location of the prey is actually unknown, the algorithm assumes that the gray wolves in the first three levels (α, β and δ wolves) have priority in knowing the potential information of the prey, while the remaining wolves (ω wolves) need to continuously adjust their orientation according to the positions of α, β and δ wolves through formula (10) to complete the optimization and migration of the entire population;

[0093]

[0094] In the formula, X(t+1) represents the position of the lower-ranking wolf ω after the next iteration. Xα, Xβ, and Xδ are the position vectors of α, β, and δ wolves, respectively.

[0095] Step 3 (Development and Exploration): This step determines the balance between global and local optimization in the GWO algorithm, aiming to prevent the algorithm from prematurely falling into local optima and affecting the reliability of the output results. Once the gray wolf has determined the potential location of its prey and initiated hunting behavior, the model development process begins. To avoid the model prematurely identifying the global optimum, the algorithm needs to ensure that the gray wolf does not move in a single direction but rather updates sufficiently within a broad search space, allowing the algorithm to capture as many possible solutions as possible.

[0096] By combining the optimization strategy GWO with the classification algorithm GBDT, a hybrid model GWO-GBDT can be constructed. The model framework is as follows: Figure 5 As shown, its modeling and optimization methods are as follows:

[0097] S31: Data preprocessing and parameter initialization;

[0098] The drilling database was constructed by splitting it in a 4:1 ratio. 80% of the samples were used to form a training set to train the model, which then learned the data information autonomously and acquired predictive ability. 20% of the samples were used to form a test set to verify whether the generalization ability of the constructed model met the standard.

[0099] As an example, the drilling database contains 291 samples, with 199 cases in category 0 and 92 cases in category 1. The database is divided into a training set (232 samples) and a test set (59 samples) for GWO-GBDT modeling and performance validation. Furthermore, considering the uneven distribution of samples between the two classes in the drilling database, stratified random sampling is used to control the proportion of samples from the three integrity classes in the training and test sets, preventing excessive sampling bias that could compromise model reliability. Additionally, the training and test sets used for each model are completely identical.

[0100] Before performing optimization iterations, set the built-in parameters of the GWO-GBDT model; as a preferred option, the built-in parameters include the number of iterations, population size, and the range of values ​​for the parameters to be optimized.

[0101] Table 3 Built-in parameters of the hybrid model GWO-GBDT

[0102]

[0103] S32: Fitness evaluation and iterative optimization;

[0104] A fitness function is set based on a cross-validation strategy to evaluate the quality of all candidate solutions generated by the hybrid classification model in each iteration. The training set is divided into five equal subsets. One subset is used as the test subset, and the remaining four are used as training subsets. The corresponding test accuracy is obtained. This process is repeated five times, with each of the five subsets taking turns as the test set. The average of the test accuracy after five-fold cross-validation is calculated to obtain the fitness value of the corresponding candidate solution, which serves as the basis for model iteration. This effectively improves the model's optimization ability and stability. After setting the fitness function, iterative optimization is performed. Based on the optimization rules set by the GWO strategy, iterative loops are performed to broadly search for possible candidate solutions within the set search space. This continuously improves the quality of candidate solutions provided by the hybrid model, prompting the solution set to evolve towards a better direction until the set maximum number of iterations is reached, at which point the optimal solution is output, thus finding the optimal GBDT parameter configuration.

[0105] S33: Model Optimization and Performance Evaluation;

[0106] Population size represents the number of search agent gray wolves in the GWO algorithm. Setting only one population size will result in unrepresentative prediction results. Therefore, for the hybrid classification model, multiple population sizes are set. Specifically, the optimization strategy GWO is combined with the classification algorithm GBDT, and different GWO-GBDT models are constructed with population sizes of 10, 20, 30, 40, 50, and 100. Accuracy, precision, recall, and F1 score are used to evaluate each GWO-GBDT model, and the training results are scored (scores 1-6, with higher evaluation index values ​​resulting in higher scores). Confusion matrix and receiver operating characteristic (ROC) curves are used to measure the training and testing results of the models. In this way, the various GWO-GBDT models are compared and analyzed to select the best model. The generalization performance of the best model is comprehensively analyzed. Through sufficient comparative analysis, the random bias in modeling can be effectively prevented, ensuring that it has the advantages of high accuracy and high stability in predicting rock mass integrity in combination with drilling parameters, reflecting its engineering application value. Thus, the best GWO-GBDT model can be comprehensively selected, as shown in Tables 4 and 5.

[0107] Table 4. Training set performance and scores of six GWO-GBDT models

[0108]

[0109] Table 5. Test set performance of the six GWO-GBDT models

[0110]

[0111]

[0112] Data shows that models with population sizes of 20 and 50 perform better during the training phase. Therefore, the GWO-GBDT model with a population size of 50 was selected as the best model. For this model, all indicators during the training process exceeded 98%, meaning that almost all 232 original training samples were correctly classified, indicating that the model has a very strong learning ability during the modeling process. Compared with the unprocessed simple GBDT, its test accuracy improved by 8.47%, while precision, recall, and F1 coefficient improved by 12.57%, 15.79%, and 14.22%, respectively, indicating that the intervention of optimization strategies effectively improved the classifier's learning ability. In addition to quantitative indicators such as accuracy, the confusion matrix and receiver operating characteristic (ROC) curve are also important tools for evaluating the model, such as... Figure 6 As shown, out of 59 test samples, 49 samples were correctly classified into the hybrid model, and the AUC value of GWO-GBDT reached 0.8566, proving that it has considerable classification efficiency.

[0113] Step 4: Identify the integrity of the rock mass and obtain a predictive analysis chart reflecting the rock mass integrity classification;

[0114] Traditional methods are complex, unintuitive, and inconvenient. Therefore, by combining the selected optimal GWO-GBDT model, a user-friendly predictive analysis chart is derived, providing a clear and convenient aid to rock mass integrity analysis. First, the drilling database constructed in step one is input into the optimal hybrid classification model GWO-GBDT, capturing the complex nonlinear relationships between input and output parameters to achieve contactless identification of rock mass integrity by referring only to drilling parameters. Then, parameter importance analysis is performed using the optimal hybrid classification model GWO-GBDT to select the two parameters that have a dominant contribution to the integrity prediction process (ranked highest in contribution), such as... Figure 7As shown, in the drilling database constructed in Step 1, only the top two contributing parameters (ranked first) are retained, and obvious outliers are removed to obtain a new drilling database. Analysis reveals that the parameter contribution rankings are, in order: propulsion speed (24.78%), pump pressure (21.66%), rotational speed (20.69%), torque pressure (17.44%), and thrust (15.42%). Therefore, a three-category database consisting of the two top-ranked drilling parameters (propulsion speed and pump pressure) and completeness is constructed as the data basis for the graphical output.

[0115] Based on the new drilling database, the hybrid classification model GWO-GBDT was re-developed. To avoid a lack of comparative standards during discussions, six algorithms—GBDT, RF, SVM, GNB, KNN, and MLP—were introduced as references. Modeling and computation were performed on the same training set (from a three-class classification database). The output uses the two drilling parameters with the top two contributions as the x and y axes, and displays predictive analysis charts of different integrity level partitioning results, providing a convenient and practical visualization tool for engineering sites. Figure 8 As shown, the predictive analysis graphs are compared, with black and white areas representing category 0 (relatively complete and complete areas) and category 1 (relatively fragmented and fragmented areas), as defined by the classification algorithm. The test accuracies of the six classifiers are GBDT = 77.19%, RF = 77.19%, SVM = 75.44%, GNB = 75.44%, KNN = 73.68%, and MLP = 75.44%. It is evident that GBDT performs best, outperforming SVM, GNB, and MLP by 1.75%, and outperforming the worst-performing KNN by 3.51%. Therefore, using GBDT as the base classifier better reflects the data characteristics and demonstrates the core role of data samples in the prediction process.

[0116] As for the hybrid model GWO-GBDT, its derived predictive analysis graph is as follows: Figure 9 Compared to other classifiers, it boasts the highest test accuracy, achieving a 3.51% improvement over GBDT. It enables better data mining and predictive classification for samples of different categories. Within an advance speed of 20-40 m / h and a pump pressure of 0.3-0.6 MPa, GWO-GBDT defines the integrity of this rock segment as Category 1, meaning that the core sample was relatively fragmented during the drilling experiment. Outside this coordinate range, the model defines it as Category 0, indicating that the core sample was relatively intact.

[0117] This invention combines drilling experiments, in-hole television image measurement, and core logging operations to obtain sufficient and accurate drilling parameters and borehole wall image data. It then processes this data with existing information to determine the integrity level of different core segments, constructing a reliable drilling database that provides robust data support. Gradient Boosting Decision Tree (GBDT) is used as the foundation for rock mass integrity identification. Leveraging the high-performance Gray Wolf Optimization Algorithm (GWO), the core hyperparameters of GBDT are independently configured to improve the performance of the prediction model. This ensures superior classification performance of the hybrid classification model on the drilling database, leading to a high-precision and high-stability hybrid classification model, GWO-GBDT. Comparative analysis of multiple hybrid models selects the optimal hybrid classification model as the basis for rock mass integrity prediction based on drilling parameters. This ensures accurate intelligent conversion between drilling parameters and rock mass integrity, demonstrating significant application potential in related fields. Finally, parameter importance analysis identifies the two most important types of drilling parameters for the prediction process and derives corresponding prediction analysis charts, providing an intuitive classification of integrity categories and offering a convenient, intuitive, and accurate visualization tool for engineering exploration. Considering the current state of existing exploration and prediction technologies, such as complex processes, high costs, limited accuracy, and difficulty in expansion, this method proposes a rock mass integrity identification method based on gradient boosting decision trees and drilling parameters, combining emerging digital drilling testing and intelligent prediction technologies. This method can scientifically and rationally identify rock mass integrity through an integrated process, which is conducive to significantly reducing production costs and achieving good economic benefits. At the same time, it can significantly simplify the exploration process, shorten the exploration cycle, and greatly improve the prediction accuracy, which helps to predict rock mass integrity information in advance. This provides a brand-new research idea and practical means for related fields, and can assist in design schemes and on-site decision feedback.

[0118] This invention overcomes the shortcomings of traditional technologies in obtaining underground rock mass information, such as being time-consuming, labor-intensive, and lacking accuracy. It effectively reduces the workload and economic costs of the exploration process and provides a fast and accurate tool for determining the integrity of rock masses on-site. It has excellent innovation and practical value, as detailed below:

[0119] 1. Based on the principles of statistics and machine learning, a high-precision rock mass integrity prediction model is constructed. Combining the novel prediction technology GBDT, it delves into the complex nonlinear relationship between drilling parameters and rock mass integrity, overcoming the limitations of traditional prediction techniques such as empirical criteria, statistical analysis, and numerical simulation, which suffer from poor prediction accuracy and limited applicability. The data-driven model can efficiently process large amounts of data and intelligently make judgments based on learned predictive intelligence.

[0120] 2. By introducing cross-validation and metaheuristic strategies, the generalization performance of the GBDT model is further improved. The GWO algorithm introduced in this invention is a widely acclaimed optimization technique in the field. Combined with swarm intelligence strategies, it can better configure the parameter combination of classification algorithms, giving it better predictive ability on the database and ensuring the reliability of prediction results and the stability of the model.

[0121] 3. The proposed hybrid classification model GWO-GBDT can identify the integrity of unknown underground rock masses based on the drilling parameters collected during the drilling process, and provide qualitative outputs such as complete, relatively complete, relatively broken, or broken. It helps to solve the problem of knowing the basic condition of underground rock masses in engineering practice, overcomes the drawbacks of traditional exploration technology such as complicated operation process, high time and economic cost, and difficulty in real-time prediction, and provides more convenient and accurate practical technology for on-site technicians.

[0122] 4. The predictive analysis diagrams exported by the hybrid classification model GWO-GBDT can provide intuitive and convenient reference. The practical tools based on the diagram format can greatly improve design efficiency, improve the operational complexity presented by previous empirical formulas, index criteria or field measurements, realize the rapid and accurate identification of the integrity classification of unknown underground rock masses, and provide alternative scientific decision support for engineering sites.

[0123] In summary, the method proposed in this invention constructs a reliable drilling database based on extensive field data accumulated in the early stages. It utilizes digital drilling testing and intelligent prediction technologies to explore the correlation mechanism between drilling parameters and rock mass integrity, constructs a stable and reliable identification model GWO-GBDT, and achieves non-contact identification of rock mass integrity based on this model. It also irregularly derives corresponding visual analysis charts, greatly simplifying the lithology identification process. The prediction results are accurate and reliable, overcoming the shortcomings of traditional methods and possessing good practical value and application potential.

Claims

1. A method for identifying rock mass integrity based on gradient boosting decision trees and drilling parameters, characterized in that, Specifically, the following steps are included: Step 1: Sample collection and database setup; S11: Determine the borehole location and use a drilling instrument to record and collect various key drilling parameters during the drilling process; S12: Use an imager to acquire image data of the entire borehole wall; S13: Conduct logging work based on core samples, analyze drilling parameter data and borehole wall image data, understand the variation law of drilling parameters throughout the borehole, complete the core integrity classification, and construct a drilling database; S14: Conduct correlation analysis on the constructed database to explore the degree of correlation between various parameters; Step 2: Construct the GBDT basic model for rock mass integrity prediction; The gradient boosting decision tree (GBDT) algorithm was selected as the basic classifier for rock mass integrity identification, and the GBDT basic model for rock mass integrity prediction was completed. Step 3: Modeling and optimization of the hybrid classification model GWO-GBDT; The iterative optimization mechanism of GWO is adopted to extensively search for candidate solutions within the set search space to find the optimal GBDT parameter configuration. The optimization strategy GWO is combined with the classification algorithm GBDT, and different hybrid classification models GWO-GBDT are constructed using different population sizes. The best hybrid classification model GWO-GBDT is then selected through comparative analysis. S31: Data preprocessing and parameter initialization; The drilling database was constructed by splitting it in a 4:1 ratio. 80% of the samples were used to form a training set to train the model, enabling it to acquire predictive capabilities after autonomously learning from the data. 20% of the samples were used to form a test set to verify whether the generalization ability of the constructed model met the standard. The built-in parameters of the GWO-GBDT model were also set. The built-in parameters include the number of iterations, population size, and the range of values ​​for the parameters to be optimized. S32: Fitness evaluation and iterative optimization; A fitness function is set based on the cross-validation strategy to evaluate the quality of all candidate solutions generated by the model in each iteration. The optimization rules set based on the GWO strategy are used for iterative loops to continuously improve the quality of candidate solutions given by the hybrid model, promote the solution set to develop in a better direction, and seek the optimal GBDT parameter configuration. S33: Model Optimization and Performance Evaluation; Different GWO-GBDT models were constructed with population sizes of 10, 20, 30, 40, 50, and 100, and a thorough comparative analysis was conducted. The training and testing results of the models were measured by accuracy, precision, recall, F1 score, confusion matrix, and receiver operating characteristic curve. Each GWO-GBDT model was scored, and the best model was selected. The generalization performance of the best model was comprehensively analyzed to ensure that it has the advantages of high accuracy and high stability in predicting rock mass integrity by combining drilling parameters. Step 4: Identify the integrity of the rock mass and obtain a predictive analysis chart reflecting the rock mass integrity classification; The drilling database constructed in Step 1 is input into the optimal hybrid classification model GWO-GBDT, and the complex nonlinear relationship between input and output parameters is captured to achieve non-contact identification of rock mass integrity by referring only to the drilling parameters. Parameter importance analysis is performed using the optimal hybrid classification model GWO-GBDT to select the two drilling parameters with the highest contribution to integrity prediction. Only the two highest-contribution drilling parameters are retained in the drilling database constructed in Step 1, and obvious outliers are removed to obtain a new drilling database. The hybrid classification model GWO-GBDT is modeled again based on the new drilling database. A three-class database is constructed using the two highest-ranking drilling parameters and integrity as the data basis for graphical output, and then input into the new hybrid classification model GWO-GBDT to obtain and display the corresponding predictive analysis graph of rock mass integrity zoning results.

2. The rock mass integrity identification method based on gradient boosting decision tree and drilling parameters according to claim 1, characterized in that, Key drilling parameters in S11 include depth, feed rate, thrust pressure, pump pressure, torque pressure, and rotational speed.

3. A rock mass integrity identification method based on gradient boosting decision tree and drilling parameters according to claim 1 or 2, characterized in that, In S12, an intelligent borehole 3D television imager is used to acquire borehole wall image data for the entire borehole.

Images