A rock mass quality grade deduction and evaluation method, system, device and medium

Abstract

The present application provides a rock mass quality grade deduction and evaluation method, system, device and medium, the method comprises: obtaining a plurality of real historical slope samples, generating synthetic samples based on the real historical slope samples using a neighborhood feature resampling strategy; constructing a data set based on the real historical slope samples and the synthetic samples; based on the constructed data set, an improved NRBO optimization algorithm is used to train the rock mass quality grade prediction model, output the optimal model hyperparameter combination, and obtain the trained prediction model; collecting the original image data set of the to-be-measured slope, and extracting the feature parameters of the to-be-measured slope based on the original image data set and the field investigation data; the feature parameters do not include the uniaxial compressive strength parameter; the feature parameters are input into the trained rock mass quality grade prediction model to predict the rock mass quality grade of the to-be-measured slope. The present application solves the problem that it is difficult to directly predict due to the absence of key geological parameters in the existing rock mass quality grade evaluation process.

Description

Technical Field

[0001] This invention relates to the field of geotechnical engineering technology, and in particular to a method, system, equipment and medium for deducing and evaluating rock mass quality grades. Background Technology

[0002] Rock mass quality classification is a crucial foundation for geotechnical engineering investigation, evaluation of rock mass quality, and slope stability analysis. Among these, RMR (Rock Mass Rating) is a common rock mass quality classification system, widely used in engineering practice due to its clear physical meaning. The RMR system comprehensively considers six geological parameters, including uniaxial compressive strength (UCS), rock quality index (RQD), structural plane spacing, structural plane conditions, groundwater conditions, and structural plane attitude, and achieves rock mass quality classification through quantitative scoring.

[0003] However, in areas with steep slopes, dangerous rock masses, and complex exposed rock masses, traditional contact methods relying on geological compasses and transects to obtain structural surface information often suffer from low acquisition efficiency, limited sample quantities, and high operational risks. Furthermore, many structural surfaces are located inside the rock mass or are obscured, making direct observation difficult and hindering the accurate determination of key geometric parameters. Simultaneously, obtaining accurate UCS data requires destructive laboratory testing of intact core samples, a process that is not only time-consuming but also costly. In areas with steep slopes, dangerous rock masses, fractured rock masses, or scarce samples, standard sample preparation and testing are often limited by terrain and safety conditions, potentially leading to unavailable UCS data and thus restricting the further application of the RMR system in practical engineering.

[0004] In recent years, UAV oblique photography, 3D point cloud reconstruction, and semi-automatic structural surface recognition methods have been widely applied. Current research on these technologies mainly focuses on improving the accuracy of image acquisition, 3D point cloud reconstruction algorithms, and the accuracy of structural surface recognition. With these visual technologies, existing UAV systems can efficiently extract geometric parameters such as joint orientation, trace length, and spacing, and can achieve non-contact exploration on steep slopes and in areas inaccessible to personnel. However, these methods can essentially only extract surface geometric information and cannot penetrate the rock mass to obtain core mechanical indicators such as UCS (Underlying Coefficient of Motion). Therefore, when UCS data is missing, RMR (Rock Mass Quality Assessment) rock mass grading is difficult to complete directly.

[0005] With the development of artificial intelligence technology, machine learning methods, due to their strong nonlinear mapping capabilities, have been attempted to be applied to rock mass quality grade prediction. For example, artificial neural networks, random forests, decision trees, and other ensemble learning models are used to deduce RMR or related rock mass classification results from relevant rock mass parameters, thereby improving evaluation efficiency and reducing reliance on human experience. However, most of these methods still rely on relatively complete parameters or high-quality samples. In actual engineering, when parameters such as UCS are missing, or when the sample size is limited or the data distribution is unbalanced, the model stability and generalization ability remain insufficient, and the interpretability for engineers is limited. However, such data-driven methods still face limitations in practical engineering applications: First, most existing models are built on the premise of relatively complete parameters or high sample quality. Once mechanical parameters such as UCS are missing, or when faced with limited sample size and extremely uneven data distribution in extreme rock mass conditions, the models are prone to blurred and distorted classification boundaries. Rock mass quality grade prediction models built based on the remaining RMR parameters will face problems such as incomplete input information, degradation of parameter-related structures, and decreased stability of classification and discrimination boundaries, thus affecting the reliability and consistency of RMR grade determination results. Second, conventional machine learning algorithms often lack a strict global hyperparameter optimization mechanism (parameter tuning). When dealing with complex hyperspace problems, they are prone to getting trapped in local optima, resulting in poor convergence and easy overfitting, and insufficient model stability and generalization ability. Finally, existing intelligent prediction models mostly have strong "black box" attributes, lack effective interpretation mechanisms, and are difficult to quantitatively reveal the contribution of each remaining feature parameter to the final classification result, which has limited interpretability for engineers.

[0006] In summary, there is an urgent need for a rock mass quality grading method that is designed for complex geological environments, uses non-contact data acquisition by UAVs as the front end, machine learning grading as the back end, and can automatically trigger inference when UCS is missing, so as to achieve rapid and accurate evaluation of rock mass quality. Summary of the Invention

[0007] This invention provides a method, system, equipment, and medium for rock mass quality grade estimation and evaluation, which addresses the technical problem of difficulty in rapidly and accurately evaluating rock mass quality when field testing is limited or UCS data is missing under complex slope conditions.

[0008] According to one aspect of the present invention, a method for deducing and evaluating rock mass quality grades is provided, comprising: Several real historical slope samples are obtained, and synthetic samples are generated based on the real historical slope samples using a neighborhood feature resampling strategy. A dataset is constructed based on the real historical slope samples and the synthetic samples. Each sample in the constructed dataset includes characteristic parameters other than the uniaxial compressive strength parameter, as well as RMR value and rock mass quality grade label. Based on the constructed dataset, the improved NRBO optimization algorithm is used to train the rock mass quality grade prediction model, output the optimal hyperparameter combination, and obtain the trained rock mass quality grade prediction model. The original image dataset of the slope to be measured is collected, and the characteristic parameters of the slope to be measured are extracted based on the original image dataset and the field survey data; the characteristic parameters do not include the uniaxial compressive strength parameter; The feature parameters are input into the trained rock mass quality grade prediction model to predict the rock mass quality grade of the slope to be tested.

[0009] According to another aspect of the present invention, a rock mass quality grade estimation and evaluation system is provided, comprising: The sample acquisition unit is used to acquire several real historical slope samples and generate synthetic samples based on the real historical slope samples using a neighborhood feature resampling strategy. The dataset construction unit is used to construct a dataset based on the real historical slope samples and the synthetic samples. Each sample in the constructed dataset includes feature parameters other than the uniaxial compressive strength parameter, as well as RMR value and rock mass quality grade label. The model training unit is used to train the rock mass quality grade prediction model based on the constructed dataset using the improved NRBO optimization algorithm, output the optimal hyperparameter combination, and obtain the trained rock mass quality grade prediction model. The parameter extraction unit is used to collect the original image dataset of the slope to be measured, and extract the characteristic parameters of the slope to be measured based on the original image dataset and the field survey data; the characteristic parameters do not include the uniaxial compressive strength parameter; The prediction unit is used to input the feature parameters into the trained rock mass quality grade prediction model to predict the rock mass quality grade of the slope to be measured.

[0010] According to another aspect of the present invention, an electronic device is provided, the electronic device comprising: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, which enables the at least one processor to perform the rock mass quality grade derivation and evaluation method according to any embodiment of the present invention.

[0011] According to another aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing computer instructions for causing a processor to execute and implement the rock mass quality grade derivation and evaluation method according to any embodiment of the present invention.

[0012] The technical solution of this invention involves acquiring several real historical slope samples, generating synthetic samples based on these samples using a neighborhood feature resampling strategy. This expands the effective sample size while preserving the original local correlation structure and distribution characteristics as much as possible, thereby improving the reliability of model training and evaluation. An improved NRBO optimization algorithm is used to train the rock mass quality grade prediction model, outputting the optimal hyperparameter combination and obtaining the trained rock mass quality grade prediction model. This enables global optimization of the model's hyperparameters, improving model stability, classification performance, and generalization ability. The original image dataset of the slope to be tested is collected, and feature parameters of the slope are extracted based on this dataset. Since the feature parameters do not include uniaxial compressive strength parameters, this invention can quickly and accurately predict the rock mass quality grade of the slope to be tested without employing complex experimental testing methods to obtain UCS data.

[0013] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a flowchart of a method for deducing and evaluating rock mass quality grades according to Embodiment 1 of the present invention; Figure 2 This is a flowchart of a method for deducing and evaluating rock mass quality grades according to Embodiment 2 of the present invention; Figure 3 This is a schematic diagram comparing the distribution of real and synthetic datasets in one embodiment of the present invention; Figure 4 This is a heatmap of the Pearson correlation coefficients of various feature parameters in a real dataset and a synthetic dataset in one embodiment of the present invention; Figure 5 This is a flowchart of a process for hyperparameter optimization of an XGBoost model using an improved NRBO optimization algorithm in one embodiment of the present invention; Figure 6 This is a schematic diagram showing the change in fitness values ​​of the improved NRBO optimization algorithm and the comparison algorithm as the number of iterations changes in one embodiment of the present invention. Figure 7 This is a confusion matrix diagram of the model of the present invention and the comparative model in one embodiment of the present invention; Figure 8 This is a schematic diagram showing the comparison of evaluation metrics between the mNRBO-XGBoost model of the present invention and other models in one embodiment of the present invention; Figure 9 This is a schematic diagram of the rock mass structural surface orientation extracted using the Facet Detect program in a real application example of the present invention; Figure 10 This is an architecture diagram of a rock mass quality grade estimation and evaluation system provided in Embodiment 1 of the present invention; Figure 11 This is a schematic diagram of the structure of an electronic device for implementing the rock mass quality grade deduction and evaluation method of this invention. Detailed Implementation

[0016] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0017] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0018] Example 1 Figure 1 This document provides a flowchart for a method of deducing and evaluating rock mass quality grades, as described in Embodiment 1 of the present invention. Figure 1 As shown, the method includes: S101. Obtain several real historical slope samples, and generate synthetic samples based on the real historical slope samples using a neighborhood feature resampling strategy.

[0019] The real historical slope samples can be obtained from case data in actual engineering projects and related literature, covering both tunnel and slope engineering scenarios. These samples include the core parameters of the RMR system and their corresponding label values. The core parameters of the RMR system include uniaxial compressive strength (UCS), rock quality index (RQD), structural plane spacing, discontinuity conditions (assessed by five sub-items: trace length, opening, roughness, infill material, and weathering degree), groundwater conditions, and structural plane attitude correction terms. Label values ​​can include the RMR value and rock mass quality grade. The RMR value can be calculated using the rating of the core parameters of the RMR system, and the rock mass quality grade label can be determined based on the numerical range of the RMR value.

[0020] In this embodiment, a neighborhood feature resampling strategy can be used to generate synthetic samples from real historical slope samples, thereby performing data augmentation on the sample data. For example, the Manhattan distance between a current sample data and other sample data can be calculated. Based on the magnitude of the Manhattan distance value, the k sample data with the smallest Manhattan distance value from the current sample data can be selected as the neighborhood sample set of the current sample data. Then, parameter values ​​are randomly extracted from the current sample data and the neighborhood sample set, and the RMR value is calculated based on the extracted parameter values. The rock mass quality grade label corresponding to each sample is determined based on the RMR value to synthesize a new sample.

[0021] S102. Construct a dataset based on the real historical slope samples and the synthetic samples; each sample in the constructed dataset includes characteristic parameters other than the uniaxial compressive strength parameter, as well as RMR value and rock mass quality grade label.

[0022] In order to train a prediction model that does not rely on uniaxial compressive strength (UCS) as input, in this embodiment, the uniaxial compressive strength parameter can be removed from real historical slope samples and synthetic samples to construct a dataset, and the constructed dataset can be used for training the subsequent rock mass quality grade prediction model.

[0023] S103. Based on the constructed dataset, the improved NRBO optimization algorithm is used to train the rock mass quality grade prediction model, output the optimal hyperparameter combination, and obtain the trained rock mass quality grade prediction model.

[0024] In this embodiment, the improved NRBO optimization algorithm can improve the initial population distribution with Sobol sequences and generate dual candidate solutions with standard Newton-Raphson search rule paths and golden sine update paths; it can also optimize the XGBoost hyperparameter combination by competing for fitness based on hierarchical ten-fold cross-validation and greedily updating individuals.

[0025] S104. Collect the original image dataset of the slope to be measured, and extract the characteristic parameters of the slope to be measured based on the original image dataset and the field survey data; the characteristic parameters do not include the uniaxial compressive strength parameter.

[0026] Appropriate drone models and corresponding photography equipment can be selected for oblique photogrammetry to collect raw image datasets of the slope to be measured. The raw image dataset includes multi-view images obtained by drones or other equivalent image acquisition equipment. The raw image dataset of the slope to be measured obtained by this non-contact acquisition method can be converted into point cloud data, 3D reality models, or other data that can characterize the spatial morphology of the slope surface through 3D reconstruction processing.

[0027] Furthermore, based on the converted slope point cloud data, 3D reality model, or other data that can characterize the spatial morphology of the slope surface, further point cloud processing can be used to extract each structural surface. Then, some characteristic parameters of the structural surfaces, such as surface roughness and surface opening, can be extracted from the feature point sets of each surface. Additionally, some characteristic parameters of the slope under test, such as groundwater conditions, can be extracted from field survey data.

[0028] It should be noted that data obtained through this non-contact acquisition method cannot be used to determine uniaxial compressive strength parameters.

[0029] S105. Input the feature parameters into the trained rock mass quality grade prediction model to predict the rock mass quality grade of the slope to be tested.

[0030] This embodiment uses a rock mass quality grade prediction model constructed using machine learning methods to predict the input feature parameters, thereby enabling rapid and accurate prediction of the rock mass quality grade of the slope to be measured.

[0031] It should be noted that the rock mass quality grade prediction model can use the XGBoost classification model, and the improved NRBO method can be used to optimize the parameters of the XGBoost classification model in order to improve the accuracy and stability of the predicted rock mass quality grade as much as possible when UCS data is missing in the input parameters.

[0032] The technical solution of this invention involves acquiring several real historical slope samples, generating synthetic samples based on these samples using a neighborhood feature resampling strategy. This expands the effective sample size while preserving the original local correlation structure and distribution characteristics as much as possible, thereby improving the reliability of model training and evaluation. An improved NRBO optimization algorithm is used to train the rock mass quality grade prediction model, outputting the optimal hyperparameter combination and obtaining the trained rock mass quality grade prediction model. This enables global optimization of the model's hyperparameters, improving model stability, classification performance, and generalization ability. The original image dataset of the slope to be tested is collected, and feature parameters of the slope are extracted based on this dataset. Since the feature parameters do not include uniaxial compressive strength parameters, this invention can quickly and accurately predict the rock mass quality grade of the slope to be tested without employing complex experimental testing methods to obtain UCS data.

[0033] Example 2 Figure 2 This is a flowchart illustrating a method for deducing and evaluating rock mass quality grades according to Embodiment 2 of the present invention. Figure 2 As shown, the method includes: S201. Obtain several real historical slope samples, and generate synthetic samples based on the real historical slope samples using a neighborhood feature resampling strategy.

[0034] This embodiment selects actual engineering cases and relevant literature to compile an initial dataset containing multiple real historical slope samples, covering both tunnel and slope engineering scenarios. The collected data is organized according to a unified rock mass quality evaluation index system to ensure consistency in variable definitions, measurement methods, and evaluation standards across samples from different sources.

[0035] Data cleaning can be performed on the obtained real historical slope samples to remove samples with obvious logical errors, duplicate records, abnormal fluctuations, or serious missing key information. For outliers that deviate significantly from the actual engineering scope, manual verification and removal can be carried out in conjunction with the original records and engineering background.

[0036] For each real historical slope sample after preprocessing, a score value for the core parameters of the RMR system is included, including uniaxial compressive strength (UCS), rock quality index (RQD), structural plane spacing, discontinuity conditions (comprehensively evaluated by five sub-items: trace length, opening, roughness, infill material, and weathering degree), groundwater conditions, structural plane attitude correction term, RMR value, and corresponding rock mass quality grade.

[0037] It should be noted that the RMR system includes six key parameters: (1) uniaxial compressive strength; (2) rock quality indicators; (3) structural plane spacing; (4) structural plane conditions; (5) groundwater conditions; and (6) structural plane attitude correction. The structural plane conditions consist of five sub-parameters: discontinuity length, opening, roughness, infill material properties, and weathering / alteration degree. The final RMR value is the algebraic sum of the scores for the six parameters.

[0038] Among them, UCS rating ( R 1) RQD rating (R2), discontinuity spacing rating (R3), discontinuity condition rating (R) 4a –R 4e The groundwater condition rating (R5) and orientation correction rating (R6) are calculated according to the specific scoring criteria in Table 1. Then, according to Table 2, the rock mass quality is divided into five grades: Class I, Class II, Class III, Class IV and Class V. The larger the RMR value, the smaller the rock mass grade, the better the rock mass quality, and the more favorable it is for construction; conversely, the smaller the RMR, the less favorable the rock mass is for engineering construction.

[0039] Table 1. RMR Classification Method

[0040] Table 2 Rock mass quality grades determined based on scoring

[0041] Specifically, the process of generating synthetic samples based on real historical slope samples using a neighborhood feature resampling strategy includes: S2011. Obtain several real historical slope samples, select a portion of real historical slope samples from the real historical slope samples, and normalize the parameters in the real historical slope samples.

[0042] Since the number of real historical slope samples obtained above is relatively small, this embodiment can construct synthetic data to expand the sample size and improve the stability of model training.

[0043] Specifically, real historical samples can be divided into training, validation, and test sets. Neighborhood feature resampling is performed only on real historical slope samples from the training set to generate synthetic samples. The validation and test sets do not participate in neighborhood sample search, synthetic sample generation, or model parameter updates. After the model hyperparameters are fixed, the test set, which was not involved in training or hyperparameter tuning, is used for final evaluation. For example, 30% of the real historical slope samples are taken as the test set, and 20% as the validation set. The test set is retained for final model evaluation, and neither the validation nor the test set participates in any subsequent synthesis or training steps. The remaining 50% is used as the training set to generate synthetic samples.

[0044] The parameters in the real historical slope samples can be normalized. For example, all input parameters of the real historical slope samples can be normalized using Min-Max, mapping all parameter values ​​to the [0,1] interval.

[0045] S2012. In the normalized feature space, calculate the Manhattan distance between a current sample data and other sample data.

[0046] S2013. Based on the Manhattan distance value, select the k sample data with the smallest Manhattan distance value from the current sample data as the neighborhood sample set of the current sample data.

[0047] In the normalized feature space, for each real historical slope sample, calculate its Manhattan distance to all other real historical slope samples. Sort the samples by distance from smallest to largest, and select the k nearest samples (including the sample itself). These k samples constitute the neighborhood sample set of the original sample.

[0048] S2014. Randomly extract parameter values ​​from the current sample data and the neighboring sample set, calculate the RMR value based on the extracted parameter values, and determine the rock mass quality grade label corresponding to each sample according to the RMR value to synthesize a new sample; the real historical slope sample includes uniaxial compressive strength parameters.

[0049] Iterate through all samples in the neighborhood sample set, assuming each sample includes parameters 1, 2, and 3. For parameter 1, randomly select a sample from the neighborhood sample set and extract parameter 1 from it as parameter 1 in the newly synthesized sample. Similarly, for parameter 2, randomly select a sample from the neighborhood sample set and extract parameter 2 from it as parameter 2 in the newly synthesized sample. For parameter 3, randomly select a sample from the neighborhood sample set and extract parameter 3 from it as parameter 3 in the newly synthesized sample. Based on the extracted parameters, obtain parameters 1, 2, and 3 in the synthesized sample.

[0050] For parameters 1, 2, and 3 in the synthetic samples, the RMR value can be calculated according to the RMR grading method in Table 1, that is, the RMR value is calculated based on the UCS and other geological parameters used for RMR scoring; and the rock mass quality grade label is determined based on the RMR value according to the method in Table 2. It should be noted that real historical slope samples include uniaxial compressive strength parameters, so the RMR value can be calculated based on parameters 1, 2, and 3 in the synthetic samples. However, when constructing training and testing data, the uniaxial compressive strength parameters, RMR values, and rock mass quality grade labels in the samples need to be removed before being used as input to the model, with the rock mass quality grade used only as the supervised learning objective. This ensures that even when UCS data is unavailable, the trained prediction model can still predict the RMR rock mass quality grade.

[0051] S2015. Repeat the above steps until the required number of synthetic samples are synthesized.

[0052] For synthesized sample data, such as Figure 3 As shown, the kernel density estimation (KDE) curve further demonstrates that the synthetic data and the actual RMR data exhibit high consistency in terms of the main peak shapes and overall distribution trends, with only slight deviations in the density peak positions of a few parameters. Simultaneously, the proportions of samples from each category in the original dataset are maintained, making the distribution structure of the synthetic data similar to the original data and effectively reducing data distribution shift. Furthermore, the two-sample Kolmogorov–Smirnov test (KS statistic) and Wasserstein distance (WD) are used to quantitatively evaluate the distribution consistency between the synthetic and real data, thus providing a more rigorous analysis of the above results. The KS statistic measures the maximum absolute difference between the cumulative distribution functions of the two sample sets, while the Wasserstein distance describes the overall morphological deviation between the two distributions. The results show that, except for the R6 parameter representing the relationship between the direction and dip angle, the KS statistics for the other parameters are all below 0.10, and no significant distribution differences were observed between the real and synthetic data (p>0.05), indicating a high degree of consistency. The heatmap analysis results based on the Pearson correlation coefficient ( Figure 4 This is used to evaluate the degree to which synthetic data preserves the correlation structure between parameters. In summary, the construction of synthetic datasets not only expands the effective sample size, but more importantly, it generates synthetic samples that conform to real engineering constraints in the multidimensional rock mass physical feature space while maintaining the main statistical features, local correlation structure, and variable coupling relationships of the original RMR data. This alleviates the problems of classifier bias, boundary ambiguity, and generalization distortion caused by small samples, and can be applied to the needs of subsequent machine learning model construction and analysis.

[0053] Table 3 presents a descriptive statistical comparison of the mean, standard deviation, and range of values ​​for each parameter in the synthetic and measured data. The results show that... R The mean deviations of 1 and RMR were both controlled within 1%, of which R The mean deviation of 1 is 1.0%, and the standard deviation difference is 1.6%; the mean deviation of RMR is 0.8%, and the standard deviation difference is 1.75%. For most parameters, synthetic data can effectively maintain the statistical characteristics of the original dataset.

[0054] Table 3. Descriptive statistics of the original and synthetic datasets

[0055] S202. Construct a dataset based on the real historical slope samples and the synthetic samples; each sample in the constructed dataset includes characteristic parameters other than the uniaxial compressive strength parameter, as well as RMR value and rock mass quality grade label.

[0056] After synthesizing multiple synthetic samples, a dataset containing multiple real historical slope samples and synthetic samples is obtained. In order to train a prediction model that does not rely on uniaxial compressive strength (UCS) as the output, in this embodiment, the uniaxial compressive strength parameter can be removed from the real historical slope samples and synthetic samples to construct a dataset, and the constructed dataset can be used for training the subsequent rock mass quality grade prediction model.

[0057] The cleaned and processed real historical slope samples and synthetic samples are constructed into a standard format suitable for training classification models, namely a "sample-feature-label" structure. Each row represents a rock mass sample, each column corresponds to an evaluation parameter, and the output label is the corresponding rock mass quality grade. Through the above processing, a basic dataset that can be used for subsequent model training, validation, and testing is formed.

[0058] S203. Based on the constructed dataset, the improved NRBO optimization algorithm is used to train the rock mass quality grade prediction model, output the optimal hyperparameter combination, and obtain the trained rock mass quality grade prediction model.

[0059] S2031. Generate the initial population distribution using the Sobol sequence; Generating a low-difference initial point set within a unit hypercube using Sobol sequences:

[0060] Among them, the Sobol sequence generates N in the [0,1]D-dimensional hypercube. p There are N initial points with extremely uniform distribution. p Let d be the population size, and d be the dimension of the decision variable. This represents the value of the nth Sobol point in the j-th dimension; Mapping points in the Sobol sequence to the actual range of values ​​for each hyperparameter:

[0061] Among them, lb j and ub j Let these represent the lower and upper bounds of the j-th hyperparameter, respectively. This represents the actual parameter values ​​after mapping; it maps points in the Sobol sequence to the actual value range of each hyperparameter. For example, if the mapping value for a certain dimension is 0.5 and the max_depth range is [3,10], then the initial value is 3 + (10...). 3) × 0.5 = 6.5 (rounded down).

[0062] And the initial population is obtained: ; S2032. In each iteration, the first candidate solution is generated based on the Newton-Raphson search rule of the standard Newton-Raphson optimization algorithm, and the second candidate solution is generated based on the golden sine update formula. The first candidate solution is generated according to the Newton-Raphson search rule of NRBO. Calculate the objective function value for each individual in the initial population to determine the optimal and worst positions; for the current individual Two different individuals are randomly selected to construct differential guiding terms, and candidate solutions are generated based on the Newton-Raphson search rule, combined with neighborhood displacement and random perturbation. After position combination, boundary constraints and trap avoidance processing, the individuals are updated according to a greedy strategy.

[0063] Furthermore, employing the golden sine guidance strategy (global exploration and escaping local optima), the golden sine update formula for the i-th individual in generation t is:

[0064] in, This represents the updated second candidate solution. Indicates the current individual's position. This represents the current globally optimal solution. and Let be a random variable; in the formula Used to introduce periodic perturbations to enhance global exploration capabilities; Used to control the search direction and gradually reduce the search radius, causing individuals to converge towards the global optimum; where: ; here, c1 and c2 are the golden ratio constants; c1 and c2 are the golden section coefficients; c1 and c2 together determine the shrinkage strength and search direction of the candidate solution.

[0065] S2033. Calculate the fitness values ​​of the first candidate solution and the second candidate solution respectively using the fitness function based on hierarchical ten-fold cross-validation; ; ; Among them, f N f is the fitness of the first candidate solution; G The fitness of the second candidate solution; S2034. Select the candidate solution with the smaller fitness value as the winning candidate solution, and update the current individual with the winning candidate solution when the fitness value of the winning candidate solution is less than the fitness value of the current individual.

[0066]

[0067] in, Indicates the winning candidate solution; For the current individual; For the updated individual; S2035. When the convergence condition is met or the maximum number of iterations is reached, the hyperparameter combination with the optimal fitness value is taken as the optimal hyperparameter combination.

[0068] For the same original individual The objective function values ​​corresponding to candidate solutions generated by Newton's method and updated by the golden sine method are calculated separately. The candidate solution with the smaller error between the two objective function values ​​is selected as the individual solution. The next generation of positions Specifically, an improved NRBO optimization algorithm is used to train the rock mass quality grade prediction model, and the flowchart of the optimal hyperparameter combination is shown below. Figure 5 As shown.

[0069] In this embodiment, the improved mNRBO-XGBoost not only performs hyperparameter optimization but, more importantly, maintains the ability to continuously search for and anchor the most robust classification boundary in the complex hyperspace caused by the missing UCS, thereby avoiding interference with the classification results due to local optima. Its optimization goal is not merely to improve the mathematical fitness, but to restore the model's stable constraints on rock mass quality grade classification as much as possible when key variables are missing, making the classification results more consistent with the actual grading rules of engineering rock masses. Finally, when the algorithm meets the convergence condition or reaches the maximum number of iterations, it outputs the optimal hyperparameter combination, which is used for subsequent model training and prediction. Furthermore, the computational efficiency of the two models was recorded: mNRBO took 928.7 seconds to complete 100 iterations, while NRBO took 1,516.8 seconds, meaning that the computational cost was reduced by approximately 40%.

[0070] In its implementation, mNRBO is used to optimize key hyperparameters of the XGBoost classification model, including learning rate, maximum tree depth, subsampling ratio, column sampling ratio, and regularization parameters, enabling the model to achieve better generalization performance and class discrimination ability during training. Subsequently, the model performance is comprehensively evaluated using metrics such as accuracy (ACC), precision, recall, F1-score, and confusion matrix. ACC measures the overall classification accuracy, F1-score balances the accuracy and coverage of class predictions, and the confusion matrix analyzes the misclassification relationships and classification boundary stability among different rock mass quality grades. These evaluations verify the effectiveness of mNRBO in optimizing XGBoost model parameters and demonstrate the stability, accuracy, and engineering applicability of this method in rock mass quality grade prediction.

[0071] like Figure 6 As shown, to analyze the impact of different improvement mechanisms on the hyperparameter optimization process, under the same dataset, fitness function, hyperparameter search space, population size and maximum number of iterations, comparative experiments were conducted using standard NRBO, SNRBO with Sobol sequence initialization, GNRBO with golden sine update mechanism, and mNRBO with both Sobol sequence initialization and golden sine update mechanism. The fitness values ​​of each optimization algorithm as the number of iterations changed were recorded.

[0072] As shown by the convergence curves, the fitness value of standard NRBO decreases slowly in the early stages of iteration and enters a relatively stable state earlier in the later stages. Introducing Sobol sequence initialization generates more evenly distributed initial candidate solutions within the predefined hyperparameter search space, reducing the concentration of candidate solutions in local regions or blind spots caused by random initialization, thus improving the initial coverage and convergence stability of the optimization process. Compared to standard NRBO, GNRBO exhibits a more pronounced fitness decrease trend during iteration, especially retaining some search capability in the mid-to-late stages. This indicates that the golden sine update mechanism can increase the search direction and position changes of candidate solutions, improve the global search capability during the optimization process, and reduce the possibility of prematurely stopping in local optima. mNRBO simultaneously employs Sobol sequence initialization, NRBO search rules, and the golden sine candidate solution update mechanism. mNRBO coordinates global and local searches by using the same fitness evaluation conditions to competitively update two types of candidate solutions, reducing the model's dependence on single random initialization and a single search path. This demonstrates that Sobol sequence initialization and the golden sine candidate solution mechanism can work synergistically. The combination of the two is beneficial to improving the convergence efficiency and quality finding of XGBoost hyperparameter optimization, thereby enabling the constructed mNRBO-XGBoost model to exhibit better prediction stability.

[0073] like Figure 7 As shown, the graphs display the effects of the XGBoost model using default parameters, the NRBO-XGBoost model, the Optuna-XGBoost model, the PSO-XGBoost model, the Random Forest (RF) model, the CatBoost model, the ANN model, and the mNRBO-XGBoost model of this invention. Specifically: Figure 7 Each column represents the actual rock mass grade IV; each row represents the predicted rock mass grade IV; the value represents the number of samples for the corresponding category. For example... Figure 7As shown, the baseline model XGBoost has generally low classification accuracy, with extremely high misclassification rates in a few classes (Class I and Class V rock masses). Specifically, three Class I samples were misclassified as Class II, and three Class V samples were misclassified as Class IV; it cannot meet the high-precision requirements for engineering safety. Under the same data partitioning and evaluation conditions, the mNRBO-XGBoost model of this embodiment is compared with the standard NRBO-XGBoost, default parameter XGBoost, Optuna-XGBoost, PSO-XGBoost, Random Forest (RF), CatBoost, and ANN. When UCS parameters are unavailable, the improved mNRBO-XGBoost model proposed in this invention achieves an accuracy of 88.04%, precision of 88.42%, recall of 88.04%, and F1-score of 88.02% on the test set; the NRBO-XGBoost model achieves an accuracy of 83.69%, precision of 84.33%, recall of 83.69%, and F1-score of 83.36%; and the default parameter XGBoost model achieves an accuracy of 78.80%, precision of 80.00%, recall of 78.80%, and F1-score of 79.40%. Compared to the NRBO-XGBoost model, the accuracy, precision, recall, and F1-score are improved by 4.35, 4.09, 4.35, and 4.66 percentage points, respectively. Compared to the XGBoost model with default parameters, the accuracy, precision, recall, and F1-score are improved by 9.24, 8.42, 9.24, and 8.62 percentage points, respectively. The mNRBO-XGBoost model achieves the best performance among other machine learning models.

[0074] In practical engineering scenarios, misclassifying inferior rock masses as superior ones leads to higher engineering risks; conversely, misclassifying superior rock masses as inferior ones often results in overly conservative support measures, increased material and labor inputs, reduced construction efficiency, and higher engineering costs. The mNRBO-XGBoost model constructed in this invention maintains high classification accuracy in minority class identification, with only one misclassification at level I and only two at level V. Furthermore, the classification performance at intermediate levels is further improved, with 16 correctly identified at level II, 69 at level III, and 56 at level IV—all optimal values ​​among the eight models. Moreover, the model's misclassified samples are primarily concentrated in adjacent levels, with no severe misclassifications spanning two or more levels, indicating that this method effectively reduces the probability of high-risk misclassifications. These results demonstrate that the improved mNRBO-XGBoost model, through its optimization strategy, overcomes the constraints of local extrema in the parameter space, significantly improving the model's generalization ability, prediction stability, and engineering applicability compared to the XGBoost model.

[0075] The Receiver Operating Characteristic Curve (ROC) and its Area Under the Curve (AUC) are also commonly used metrics for evaluating the predictive performance of classifiers. The ROC curve reflects the relationship between the true positive rate (TPR, i.e., recall) and the false positive rate (FPR). The closer the AUC value is to 1, the better the classifier's performance. The ROC curves and corresponding AUC values ​​of each classifier are shown in the figure. For multi-class tasks, both micro-average ROC curves and macro-average ROC curves were used. The micro-average curve reflects the overall performance of the classifier, while the macro-average curve describes the performance of each rock mass category. The figure shows that among all individual classifiers, the mNRBO-XGBoost classifier exhibits the best predictive performance. The predictive performance of the XGBoost and RF classifiers is relatively low, with both having a micro-average AUC value of 0.962 and macro-average AUC values ​​of 0.955 and 0.961, respectively. Figure 8 This diagram illustrates the comparison of the ROC curves and AUC data of the mNRBO-XGBoost model of this invention with those of standard NRBO-XGBoost, default parameter XGBoost, Optuna-XGBoost, PSO-XGBoost, Random Forest (RF), CatBoost, and ANN.

[0076] In one embodiment, the method further includes: after the rock mass quality grade prediction model outputs the rock mass quality grade prediction result, using the SHAP attribution module to analyze the marginal contribution of each input feature parameter to the output of the rock mass quality grade prediction model under different combination conditions, calculating the corresponding attribution value, and identifying the key parameters that play a dominant role in the classification result in the current sample, and outputting the contribution ranking of each feature parameter.

[0077] The SHAP attribution module outputs the contribution ranking of each feature parameter, single-sample contribution decomposition, and category response direction, which characterize the decisive weight of each geomechanical indicator in the current evaluation. By presenting the classification results output by the model simultaneously with the corresponding attribution results, field engineers can intuitively obtain the main basis for determining the quality grade of the rock mass, thereby avoiding the risk of misjudgment caused by the lack of interpretability of purely data-driven models.

[0078] S204. Collect the original image dataset of the slope to be measured, and perform three-dimensional reconstruction of the slope to be measured based on the original image dataset to obtain the point cloud dataset of the slope to be measured.

[0079] Non-contact spatial information of the slope under test is acquired, and three-dimensional spatial data is generated for interpreting structural surface parameters. The non-contact spatial information includes multi-view images obtained by UAVs or other equivalent image acquisition devices; the three-dimensional spatial data includes point cloud data, three-dimensional reality models, or other data that can characterize the spatial morphology of the slope surface.

[0080] In this embodiment, an appropriate UAV model and corresponding photographic equipment can be selected for oblique photogrammetry. Specifically, based on the extent, slope undulation, and occlusion of the slope to be measured, a UAV can be used to acquire multi-view images covering the target area. The acquired images should reflect the spatial morphology of the slope surface, local rock outcrops, and identifiable structural surfaces, and meet the basic requirements for subsequent 3D spatial data construction. In this embodiment, a suitable multi-rotor UAV and photographic payload are selected to carry out oblique photogrammetry operations according to the survey requirements. A UAV equipped with a high-resolution image sensor is used to take multi-angle photos of the slope to be measured along a preset trajectory, with the acquisition perspective covering vertical to the ground and multiple oblique orientations, ensuring that the top surface, slope surface, slope toe, and sidewall areas of the slope are completely captured. Regarding aerial photography parameter configuration, reasonable heading and lateral overlap indices are set to ensure sufficient overlap between adjacent images, supporting feature matching and spatial calculation in the subsequent 3D reconstruction stage.

[0081] Before data collection, a data collection plan should be developed based on the site terrain, operational safety conditions, and the visibility of the target area. For engineering scenarios requiring spatial positioning or precision control, the three-dimensional spatial data can be constrained by combining positioning information, control point information, or existing measurement data. Specific data collection equipment, flight path settings, and positioning methods can be determined according to engineering conditions.

[0082] The acquired multi-view images are subjected to 3D reconstruction processing to obtain 3D spatial data of the slope surface. The 3D reconstruction can be achieved using existing photogrammetry, dense matching, or other equivalent methods. This invention does not limit the specific software platform, reconstruction algorithm, or processing order. In one embodiment, the method further includes preprocessing the acquired raw point cloud data.

[0083] Specifically, voxel filtering can be used to downsample the point cloud. This involves dividing the 3D space into several regular voxel units, using the geometric center of the point set within each voxel as the representative point of that voxel. This significantly reduces the number of points while preserving as much of the original point cloud spatial morphology and local geometric features as possible. It effectively removes outliers and points far from the target while successfully preserving key feature points. Further processing of the 3D spatial data, including denoising, cropping, density adjustment, and outlier removal, reduces the impact of irrelevant data and reconstruction noise on the interpretation of structural surface parameters.

[0084] S205. Extract the feature point set of each structural surface from the point cloud dataset.

[0085] After point cloud preprocessing, a spatial neighborhood index needs to be established for each point to support subsequent normal vector estimation. A spatial index structure is used to organize the point cloud data to quickly retrieve the neighboring point sets of each point, providing data support for subsequent local plane fitting and normal vector calculation. The point cloud data obtained through 3D reconstruction consists of a large number of discrete points, each recording the 3D coordinates of a specific location on the rock mass surface and the normal vector of that surface. Since the point cloud covering the same structural plane exhibits consistency in the normal vector direction, while the normal vectors of point clouds from different structural planes show distinguishable deviations, the normal vector is used as the criterion to classify and aggregate discrete points with similar geometric properties and spatial adjacency in the point cloud. This allows for the separation of feature point sets corresponding to each structural plane from the point cloud data.

[0086] Specifically, a point is selected from the point cloud dataset as an initial seed point and included in set Q; Calculate the angle difference between the normal of the initial seed point and the normal of the adjacent points of the initial seed point. When the angle difference is less than a preset threshold, include the adjacent points of the initial seed point in the set Q. When a set Q contains three points that define a plane, new neighboring points are added to the two planar regions before and after set Q. and The normal vectors of the two planar regions are derived respectively. and The mean square error between the two normal vectors is less than a preset threshold:

[0087]

[0088] In the formula, k is the number of points in the plane. It is a plane normal vector. Let m be a point within the point set, and m be the centroid of the plane. This is the set error threshold; When the mean square error between the two corresponding normal vectors of a new neighboring point is less than a preset threshold, and the new neighboring point is on the plane When the vertical distance is less than the set distance threshold, the new adjacent points are divided into a feature point set of the same structural surface; Set the minimum number of plane points to extract. and maximum value When the number of planar points generated by clustering is within the range of minimum and maximum values, the final extracted feature point set of the structural surface is output.

[0089] S206. Extract feature parameters of the structural surfaces based at least on the feature point set of each structural surface; the feature parameters do not include uniaxial compressive strength parameters.

[0090] Specifically, based on the feature point set of each structural surface, the attitude, trace length, and spacing of the structural surfaces are extracted; the attitude of the structural surfaces includes dip angle and dip direction; rock quality indicators are estimated based at least on the attitude, trace length, and spacing of the structural surfaces; and the roughness, opening, weathering degree, and groundwater condition of each structural surface are determined by combining the feature point set of each structural surface, the original image data, and the field survey records.

[0091] The calculation of the orientation of structural surfaces involves treating structural surfaces as planes in three-dimensional space. For the point cloud data of each acquired structural surface, the point cloud data of the same plane are fitted using the least squares method to obtain the plane equation. The feature point sets of each structural surface are considered as a group of planar points in three-dimensional space. The least squares method is used to fit the point cloud data belonging to the same structural surface to the spatial plane, and the equation expression of the plane is solved.

[0092] Its plane normal vector is:

[0093] Subsequently, the system automatically calculates the dip angle of the structural surface based on the normal vector using the inverse trigonometric function transformation formula. ( ) and tendencies ( ).

[0094] Structural surface trace length calculation: The system automatically identifies the two points farthest from both ends of the structural surface as the endpoints of the trace, denoted as _____. and The distance between the farthest points in the structural surface point set is determined, and the calculation result serves as the basis for scoring the trace length sub-item of the structural surface condition (R4) in the RMR system. Based on the Euclidean distance formula in three-dimensional space, the formula for calculating the trace length L is as follows: .

[0095] Calculation of spacing between structural surfaces: Randomly select one point on each of the two structural surfaces to be measured. and The direction vectors originating from these two points are calculated, and their direction cosine values ​​are also calculated. A reference point, denoted as P1 and P2, is selected from each of the two sets of structural surface feature points for the distance to be measured. The line connecting these two points is used as the survey line, and the coordinate components of the direction vector of this survey line are calculated.

[0096] Then, substituting the values ​​into the inverse trigonometric function formula, we can obtain the inclination angle of the survey line. With tendency Combined with the known dip angle of the structural plane With tendency The actual spacing between structural surfaces is automatically calculated based on the spatial geometric projection formula. The dip angle and dip direction of the survey line are calculated from the components of the survey line direction. Then, combined with the known attitude parameters of the structural planes, the true normal distance between the two structural planes is automatically obtained through spatial projection calculation. This distance value will correspond to the score of the structural plane spacing term (R3) in the RMR system.

[0097] RQD Calculation: When the measured RQD value of borehole cores cannot be directly obtained, the volumetric joint number of the rock mass can be calculated using the spacing method based on known information such as the attitude and spacing of structural planes. :

[0098] in, To determine the number of structural surface groups within the study area, For the first The number of structural surfaces per meter of length along the normal direction in a group of structural surfaces. The number of non-grouped joints per cubic meter of rock mass; based on the empirical correlation proposed by Palmström. Estimate the RQD.

[0099] Structural surface roughness calculation: Extracting cross-sectional coordinates of the structural surface under test along multiple directions. Calculate the horizontal projection length Length of profile curve Relative fluctuation and elongation and according to and The combined characteristics of the joint surface are used to classify the macroscopic morphology of the structure into three categories: straight, wavy, and sawtooth. By observing whether the undulation of the joint surface is obvious, whether the surface is rough, whether it is wavy, whether it is smooth, and whether there are scratches or polishing phenomena, the roughness of the structure surface is judged. Combined with the joint conditions in the RMR system to classify the roughness, the qualitative and quantitative comprehensive evaluation of the roughness of the structure surface is realized.

[0100] Structural plane opening calculation: Based on 3D point cloud interpretation technology, the system extracts and fits the point cloud data of both sides (upper and lower walls) of the open structural plane. The least squares method is used to obtain the best-fit plane equation for the point clouds of both structural planes. The initial opening value is determined by calculating the visible normal vertical distance between two parallel (or nearly parallel) planes. To improve accuracy, sampling points are evenly distributed along the grid within the structural plane area (sampling interval controlled at 1~2mm to ensure sufficient resolution). The vertical Euclidean distance between the corresponding points on both structural planes is measured, and the average value is taken as the final opening. According to the rock mass engineering quality grading standard, the automatically calculated opening values ​​are classified: less than 0.2mm is defined as closed, 0.2~1.0mm as slightly open, 1.0~5.0mm as open, and greater than 5.0mm as wide open. Furthermore, for structural planes identified as being affected by infill material, the system will provide additional annotations for differentiation.

[0101] Weathering Degree Evaluation: Based on the 3D point cloud interpretation results, combined with the surface color changes of structural surfaces, the degree of mineral particle loosening, the integrity of the original rock structure, and the development characteristics of the weathering crust, the weathering degree of structural surfaces is qualitatively identified and graded. Preferably, a structural surface that retains the original unweathered characteristics, has minimal color changes, and an intact particle structure is classified as unweathered; a surface showing only slight color changes, minor local mineral decomposition, or slight surface coarsening is classified as slightly weathered; a surface with significant color changes, local structural damage, and relatively significant mineral decomposition is classified as moderately weathered; a surface with extensive structural damage, significantly weakened original rock tissue characteristics, and a thick weathering crust is classified as strongly weathered; and a surface with essentially lost original rock structure, surface soilification, or extreme loosening is classified as completely weathered. The identification results can be further combined with rock mass quality grading standards to assign corresponding level labels to the weathering degree.

[0102] Groundwater condition assessment: Based on images, 3D point clouds, and field survey information, combined with the surface wetness of structural surfaces, seepage traces, and dripping or linear water outflow phenomena, the groundwater condition is qualitatively identified and classified. A surface with no obvious wetness traces and no seepage is classified as dry; a surface with localized dampness but no visible water outflow is classified as slightly damp; a surface that is generally damp but without dripping is classified as damp; a surface with obvious dripping is classified as dripping; and a surface with continuous or linear water outflow is classified as flowing water. Furthermore, the degree of groundwater influence can be comprehensively corrected by combining groundwater recharge conditions, structural surface connectivity, and hydrogeological survey results to improve the accuracy of rock mass quality classification.

[0103] S207. Input the feature parameters into the trained rock mass quality grade prediction model to predict the rock mass quality grade of the slope to be tested.

[0104] It should be noted that the rock mass quality grade prediction model can use the XGBoost classification model, and the improved NRBO method can be used to optimize the parameters of the XGBoost classification model in order to improve the accuracy of the predicted rock mass quality grade when UCS data is missing in the input.

[0105] In a specific application example, a rock slope is used as a case study for verification. The slope in the study area has a horizontal length of approximately 80 m, a width of approximately 50 m, and a height of approximately 30 m. The structural planes are well-developed and exhibit a continuous, interconnected distribution. The rock slope has a distinct layered structure with well-developed wedge-shaped bedding, and the geological conditions are relatively complex, making it a suitable research object for the construction and verification of a three-dimensional network model of structural planes.

[0106] In practical engineering applications, this approach falls within a reasonable range. Using drones to collect multi-view original images of the slope under test, after 3D reconstruction, the generated point cloud dataset is input into the automated structural surface extraction algorithm for recognition processing. By automatically clustering and segmenting point sets with similar attitude characteristics, the structural surface is interpreted, and its recognition effect is as follows: Figure 9 As shown, the proposed method can not only completely interpret large-area main structural surfaces, but also accurately capture information on small and scattered local structural surfaces. To verify the accuracy and reliability of the automated extraction logic, the three-point method measurement value was used as a benchmark, and the interpretation results of the attitude obtained by the method of this invention were compared and verified. The dip angle deviation and dip direction deviation were both stably controlled within 5°.

[0107] Furthermore, the slope structural surfaces were divided into three groups, and the corresponding layers of each group were screened, with the main structural surfaces selected for detailed analysis. Based on the field investigation and geological survey data of the slope area, the rock strata exhibit a medium-thick layer structure. The uniaxial compressive strength of the rock was determined to be 40 MPa. The RQD needs to be calculated based on the volumetric joint number Jv of the rock mass. The calculated Jv of the slope rock mass was 7.56, and the RQD was 91.1. The spacing between these rock structural surfaces is approximately 20-60 cm; most of the internal rock structural surfaces are closed, filled with clay, and have slightly rough surfaces. The trace lengths are mainly concentrated between 1 and 3 m, and the groundwater conditions are moist.

[0108] Using the RMR (Real Rock Mass Basic Quality) grading standard, Table 4 shows that the overall quality levels of the three rock mass groups are between Grade III and IV, reflecting the representativeness and engineering complexity of the research objects. Combining the basic rock quality indicators and the modified basic rock quality indicators, it can be seen that, without relying on UCS (Unified Geometry System) data, the model algorithm proposed in this invention can achieve accurate non-contact deduction of rock mass quality grades, and the predicted results are consistent with actual engineering judgments. This result not only verifies the efficient utilization of remaining geological information under complex geological conditions, but also demonstrates its significant advantages in improving rock mass grading efficiency, enhancing discrimination consistency, and reducing on-site operational risks, thus providing reliable technical support for rapid rock mass quality assessment in complex, high-risk slopes and scenarios with limited samples.

[0109] Table 4 Results of Engineering Rock Mass Quality Classification Using the RMR Classification Method

[0110] This invention addresses the key technical challenge of extracting and identifying structural surface sets. Based on an optimized region growing method combined with digital processing techniques, it interprets structural surfaces from UAV 3D point cloud data to obtain structural surface attitude information. Simultaneously, it utilizes analytical geometry theory and a self-developed program to automatically identify the trace length and spacing of structural surfaces, thereby improving the automation and processing efficiency of structural surface geometric parameter extraction. Considering the high cost, limited sample size, and strong field dependence of RMR data acquisition in real-world engineering, where small sample sizes are relatively limited for machine learning modeling and may affect the model's generalization ability, this invention further employs a neighborhood feature resampling strategy to construct a synthetic dataset. This expands the effective sample size while preserving the original local correlation structure and distribution characteristics as much as possible, improving the reliability of model training and evaluation. Addressing the issues of performance degradation, unstable convergence, and getting trapped in local optima when machine learning models are improperly configured with hyperparameters, this invention further introduces an improved NRBO algorithm to globally optimize model hyperparameters, thereby improving model stability, classification performance, and generalization ability. This process framework utilizes structural surface geometry information and other residual parameters acquired by UAVs without relying on UCS data. It employs machine learning methods to extrapolate rock mass quality grades and further combines SHAP interpretability analysis to reveal the contribution mechanism of each input feature to the prediction results, thereby enhancing the transparency and engineering credibility of the model output. This enables non-contact, intelligent, and interpretable evaluation of rock mass quality grading based on RMR under complex geological conditions.

[0111] Example 3 Figure 10 This is a schematic diagram of a rock mass quality grade estimation and evaluation system provided in Embodiment 3 of the present invention. Figure 10 As shown, the device includes: The sample acquisition unit 301 is used to acquire several real historical slope samples and generate synthetic samples based on the real historical slope samples using a neighborhood feature resampling strategy. Dataset construction unit 302 is used to construct a dataset based on the real historical slope samples and the synthetic samples; each sample data in the constructed dataset includes feature parameters other than the uniaxial compressive strength parameter, as well as RMR value and rock mass quality grade label; Model training unit 303 is used to train the rock mass quality grade prediction model based on the constructed dataset using an improved NRBO optimization algorithm, output the optimal hyperparameter combination, and obtain the trained rock mass quality grade prediction model. The parameter extraction unit 304 is used to collect the original image dataset of the slope to be measured, and extract the characteristic parameters of the slope to be measured based on the original image dataset and the field survey data; the characteristic parameters do not include the uniaxial compressive strength parameter; The prediction unit 305 is used to input the feature parameters into the trained rock mass quality grade prediction model to predict the rock mass quality grade of the slope to be tested.

[0112] Example 4 Figure 11 A schematic diagram of an electronic device 10, which can be used to implement embodiments of the present invention, is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (e.g., helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0113] like Figure 11 As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded from storage unit 18 into the RAM 13. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0114] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0115] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as a method for rock mass quality grade estimation and evaluation.

[0116] In some embodiments, a rock mass quality grade estimation and evaluation method may be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program may be loaded and / or installed on electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the rock mass quality grade estimation and evaluation method described above may be performed. Alternatively, in other embodiments, processor 11 may be configured to perform a rock mass quality grade estimation and evaluation method by any other suitable means (e.g., by means of firmware).

[0117] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0118] Computer programs used to implement the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0119] In the context of this invention, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0120] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0121] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0122] A computing system can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system to address the shortcomings of traditional physical hosts and VPS services, such as high management difficulty and weak business scalability.

[0123] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0124] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for deducing and evaluating the quality grade of rock mass, characterized in that, include: Several real historical slope samples are obtained, and synthetic samples are generated based on the real historical slope samples using a neighborhood feature resampling strategy. A dataset is constructed based on the real historical slope samples and the synthetic samples. Each sample in the constructed dataset includes characteristic parameters other than the uniaxial compressive strength parameter, as well as RMR value and rock mass quality grade label. Based on the constructed dataset, the improved NRBO optimization algorithm is used to train the rock mass quality grade prediction model, output the optimal hyperparameter combination, and obtain the trained rock mass quality grade prediction model. Collect the original image dataset of the slope to be measured, and extract the characteristic parameters of the slope to be measured based on the original image dataset and the field survey data; The characteristic parameters do not include uniaxial compressive strength parameters; The feature parameters are input into the trained rock mass quality grade prediction model to predict the rock mass quality grade of the slope to be tested.

2. The method for deducing and evaluating rock mass quality grades according to claim 1, characterized in that, The process of acquiring several real historical slope samples and generating synthetic samples based on these samples using a neighborhood feature resampling strategy includes: Several real historical slope samples are obtained, a portion of real historical slope samples are selected from the real historical slope samples, and the parameters in the real historical slope samples are normalized. In the normalized feature space, calculate the Manhattan distance between a current sample data and other sample data; Based on the Manhattan distance value, select the k samples with the smallest Manhattan distance value from the current sample data as the neighborhood sample set of the current sample data; Each parameter value is randomly extracted from the current sample data and the neighboring sample set, and the RMR value is calculated based on the extracted parameter values. The rock mass quality grade label corresponding to each sample is determined according to the RMR value to synthesize a new sample; the real historical slope sample includes uniaxial compressive strength parameters. Repeat the above steps until the required number of synthetic samples are synthesized.

3. The method for deducing and evaluating rock mass quality grades according to claim 2, characterized in that, The dataset constructed based on the real historical slope samples and the synthetic samples includes: The uniaxial compressive strength parameters are removed from the real historical slope samples and the synthetic samples. A dataset is constructed based on the real historical slope samples and the synthetic samples after removing the uniaxial compressive strength parameters.

4. The method for deducing and evaluating rock mass quality grades according to claim 1, characterized in that, The improved NRBO optimization algorithm is used to train the rock mass quality grade prediction model and output the optimal hyperparameter combination, including: The initial population distribution is generated using Sobol sequences; In each iteration, the first candidate solution is generated based on the Newton-Raphson search rule of the standard Newton-Raphson optimization algorithm, and the second candidate solution is generated based on the golden sine update formula. The fitness values ​​of the first candidate solution and the second candidate solution are calculated using a fitness function based on hierarchical ten-fold cross-validation, respectively. Select the candidate solution with the smaller fitness value as the winning candidate solution, and update the current individual with the winning candidate solution when the fitness value of the winning candidate solution is less than the fitness value of the current individual; When the convergence condition is met or the maximum number of iterations is reached, the hyperparameter combination with the optimal fitness value is taken as the optimal hyperparameter combination.

5. The method for deducing and evaluating rock mass quality grades according to claim 4, characterized in that, The improved NRBO optimization algorithm is used to train the rock mass quality grade prediction model and output the optimal hyperparameter combination, including: Generating a low-difference initial point set within a unit hypercube using Sobol sequences: Where, N p Let d be the population size, and d be the dimension of the decision variable. This represents the value of the nth Sobol point in the j-th dimension; Mapping points in the Sobol sequence to the actual range of values ​​for each hyperparameter: Among them, lb j and ub j Let these represent the lower and upper bounds of the j-th hyperparameter, respectively. These are the actual parameter values ​​after mapping; the points in the Sobol sequence are mapped to the actual value range of each hyperparameter, and the initial population is obtained: ; The first candidate solution is generated based on the Newton-Raphson search rule of the standard Newton-Raphson optimization algorithm. ; A competitive update strategy guided by the golden sine wave is introduced during the iteration process: For the i-th individual in the t-th generation, its golden sine update formula is: in, This represents the updated second candidate solution. Indicates the current individual's position. This represents the current globally optimal solution. and Let be a random variable; in the formula Used to introduce periodic perturbations to enhance global exploration capabilities; Used to control the search direction and gradually reduce the search radius, causing individuals to converge towards the global optimum; where: ； here, c1 and c2 are the golden ratio constants; c1 and c2 are the golden ratio coefficients that control the search step size and boundary weights of the algorithm; c1 and c2 together determine the shrinkage strength and search direction of the candidate solution. The fitness of the first candidate solution and the second candidate solution is calculated using a fitness function based on hierarchical ten-fold cross-validation: ； ； Among them, f N f is the fitness of the first candidate solution; G The fitness of the second candidate solution; The fitness values ​​of the first candidate solution and the second candidate solution are calculated using a fitness function based on hierarchical ten-fold cross-validation, respectively. The candidate solution with the smaller fitness value is selected as the winning candidate solution: in, Indicates the winning candidate solution; And if the fitness value of the winning candidate solution is less than the fitness value of the current individual, update the current individual with the winning candidate solution: in, For the current individual; For the updated individual; When the convergence condition is met or the maximum number of iterations is reached, the hyperparameter combination with the optimal fitness value is taken as the optimal hyperparameter combination.

6. The method for deducing and evaluating rock mass quality grades according to claim 1, characterized in that, Also includes: After the rock mass quality grade prediction model outputs the rock mass quality grade prediction results, the SHAP attribution module is used to analyze the marginal contribution of each input feature parameter to the output of the rock mass quality grade prediction model under different combinations of conditions, calculate the corresponding attribution value, and identify the key parameters that play a dominant role in the classification results in the current sample, and output the contribution ranking of each feature parameter.

7. The method for deducing and evaluating rock mass quality grades according to claim 1, characterized in that, The extraction of feature parameters of the slope to be measured based on the original image dataset includes: Based on the original image dataset, a three-dimensional reconstruction of the slope under test is performed to obtain the point cloud dataset of the slope under test. The feature point set of each structural surface is extracted from the point cloud dataset; Feature parameters of the slope to be measured are extracted based on at least the feature point set of each structural surface; The extraction of feature parameters of the slope under test based at least on the feature point set of each structural surface includes: Based on the feature point set of each structural plane, the orientation, trace length, and spacing of the structural planes are extracted; the orientation of the structural planes includes dip angle and dip direction. Rock quality indicators should be estimated based at least on the orientation of structural planes, the length of structural plane traces, and the spacing between structural planes. By combining the feature point set of each structural surface, the original image data and the field survey record, the roughness, opening degree, weathering degree and groundwater condition of each structural surface are determined.

8. A rock mass quality grade estimation and evaluation system, characterized in that, include: The sample acquisition unit is used to acquire several real historical slope samples and generate synthetic samples based on the real historical slope samples using a neighborhood feature resampling strategy. The dataset construction unit is used to construct a dataset based on the real historical slope samples and the synthetic samples. Each sample in the constructed dataset includes feature parameters other than the uniaxial compressive strength parameter, as well as RMR value and rock mass quality grade label. The model training unit is used to train the rock mass quality grade prediction model based on the constructed dataset using the improved NRBO optimization algorithm, output the optimal hyperparameter combination, and obtain the trained rock mass quality grade prediction model. The parameter extraction unit is used to collect the original image dataset of the slope to be measured, and extract the characteristic parameters of the slope to be measured based on the original image dataset and the field survey data. The characteristic parameters do not include uniaxial compressive strength parameters; The prediction unit is used to input the feature parameters into the trained rock mass quality grade prediction model to predict the rock mass quality grade of the slope to be measured.

9. An electronic device, characterized in that, The electronic device includes: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the rock mass quality grade derivation and evaluation method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that are used to cause a processor to execute the rock mass quality grade derivation and evaluation method according to any one of claims 1-7.

Images