Waste rock aggregate parent ore tracing optimization method based on multi-source mine geological data mining

Abstract

The application discloses a waste rock aggregate parent ore tracing optimization method based on multi-source mine geological data mining, relates to the technical field of mine solid waste resource utilization, and comprises the following steps: multi-source mine geological data acquisition and spatial positioning; parent rock source label construction based on geological feature fingerprints; aggregate quality sensitive feature screening and multi-modal data fusion; parent rock aggregate quality potential prediction model construction based on extreme gradient boosting XGBoost; unknown source waste rock parent rock tracing and aggregate quality potential prediction; and waste rock aggregate parent ore optimization decision based on multi-objective optimization. The application realizes front-end prediction and source optimization of mine waste rock aggregate quality by using the above waste rock aggregate parent ore tracing optimization method based on multi-source mine geological data mining, and provides a full-chain intelligent decision scheme for mine solid waste resource utilization.

Description

Technical Field

[0001] This invention relates to the field of resource utilization technology of mine solid waste, and in particular to a method for source tracing and optimization of waste rock aggregate parent ore based on multi-source mine geological data mining. Background Technology

[0002] As my country's infrastructure construction continues to expand, natural sand and gravel aggregate resources are becoming increasingly depleted. At the same time, mining generates a large amount of waste rock, with a cumulative stockpile of tens of billions of tons. Processing mine waste rock into manufactured aggregate has become an important way to solve the dual problems of natural aggregate shortage and mine solid waste disposal.

[0003] However, the quality of mine waste rock is highly dependent on the geological properties of its parent rock. Waste rock from different mining areas, strata, and lithologies varies significantly in mineral composition, chemical composition, and physical and mechanical properties, directly determining its suitability as aggregate. Currently, the core technological bottleneck in waste rock aggregate utilization lies in the fact that large quantities of waste rock generated in mines are simply mixed and piled up, resulting in large fluctuations and poor stability in overall quality, making it difficult to meet the requirements for high-quality aggregate production. Traditional waste rock evaluation methods typically involve sampling and testing after waste rock is generated, which is a "post-evaluation" model. This approach suffers from drawbacks such as lag (waste rock is already mixed and piled up, making it impossible to separate it by quality), low efficiency (laboratory analysis cycles can take several days to weeks), high costs, and a lack of source control capabilities.

[0004] In recent years, data mining and machine learning technologies have been increasingly widely applied in the field of geology and mineral resources, demonstrating superior performance compared to traditional statistical methods in geochemical data processing, lithology identification, and mineral resource prediction. Interpretability analysis methods such as SHAP have been used to reveal the decision-making mechanisms of machine learning models, improving the reliability of prediction results. However, the application of these technologies in the field of geology and mineral resources is mainly concentrated in mineral exploration and prospecting prediction; a systematic application to the tracing and quality selection of mine waste rock aggregates is still lacking. The following gaps remain in existing technologies: First, there is a lack of traceability methods for waste rock aggregate utilization scenarios. The traceability results are not linked to the aggregate quality potential of waste rock, and cannot directly guide classification and utilization decisions. Second, there is a lack of effective dimensionality reduction methods for high-dimensional geological features. Direct modeling leads to high model complexity and a high risk of overfitting. Third, there is a lack of interpretable traceability prediction mechanisms. The "black box" nature of machine learning models makes them lack geological basis and difficult to gain the trust of mining engineers. Fourth, there is a lack of dynamic update mechanisms. Static models gradually become ineffective as mining geological conditions change. Fifth, existing decisions usually only rank targets based on quality indicators, without taking recoverable reserves into consideration, making it difficult to guarantee long-term stable supply. Summary of the Invention

[0005] The purpose of this invention is to provide a method for tracing and optimizing the parent mineral of waste rock aggregate based on multi-source mine geological data mining. In view of the shortcomings of existing technologies, which cannot predict aggregate quality and trace and optimize the parent mineral at the source before or during waste rock generation, resulting in mixed waste rock stockpiling and uncontrollable aggregate quality, the invention aims to establish a full-chain technical solution of "data collection - feature extraction - traceability matching - optimization rating".

[0006] To achieve the above objectives, this invention provides a method for source tracing and optimization of waste rock aggregate parent ore based on multi-source mine geological data mining, comprising the following steps: Step S1: Multi-source mine geological data acquisition and spatial positioning. Obtain multi-source geological data of the mine to be analyzed, assign a unified spatial coordinate benchmark to the multi-source geological data, establish the correlation mapping between each data point and the three-dimensional spatial location of the mine, and form a spatial database. Step S2: Constructing source tags for parent rocks based on geological fingerprints. Based on the multi-source geological data collected in Step S1, extract the geological fingerprint vectors of parent rocks from each mining area and each stratum. The K-means clustering algorithm was used to cluster parent rock samples with known origins, the number of clusters K was determined, a mapping relationship database of geological feature fingerprints and parent rock origins was established, and each category was assigned a category label. ; Step S3: Screening of aggregate quality sensitive features and fusion of multimodal data. Quality testing data of aggregate samples prepared from waste rock from different parent rock sources are obtained. The SHAP interpretability analysis method is used to calculate the contribution of each geological fingerprint to aggregate quality indicators, and a subset of sensitive features that have a significant impact on aggregate quality is screened out. ; Step S4: Construct a potential prediction model for parent rock aggregate quality based on extreme gradient boosting XGBoost, using the sensitive feature subset selected in step S3. Using the aggregate quality detection level obtained in step S3 as the input feature, a parent rock aggregate quality potential prediction model based on XGBoost is constructed and trained, and the model with the best evaluation index is selected as the final prediction model. Step S5: Tracing the source of waste rock parent rock and predicting aggregate quality potential. For the waste rock sample to be traced, the same standard as in steps S1 and S2 is used to collect its multi-source geological data and extract geological feature fingerprints. The extracted features are input into the prediction model trained in step S4, and the parent rock source category of the waste rock sample and its corresponding aggregate quality potential level are output. Step S6: Optimization decision-making for waste rock aggregate mother ore based on multi-objective optimization. Obtain recoverable reserve data of waste rock from different parent rock sources in each mining area of ​​the mine. Use the aggregate quality potential level and recoverable reserves predicted in step S5 as optimization objectives to construct a multi-objective optimization decision-making model. A multi-objective optimization algorithm is used to solve for the Pareto optimal solution set or to calculate the weighted comprehensive score, and the parent ore preferred ranking list is output.

[0007] Preferably, step S1, which establishes the association mapping between each data point and the three-dimensional spatial location of the mine, specifically includes: Using the mine survey coordinate system or the geodetic coordinate system as a unified spatial reference system, coordinate transformation is performed on the borehole coordinates and borehole depth data in the geological exploration data, RTK real-time dynamic measurement and positioning is performed on the sampling points in the geochemical analysis data, and the mining face in the mining production data is spatially calibrated using the mining progress coordinates in the digital mining system. For historical data for which spatial coordinates cannot be directly obtained, geostatistical methods are used to perform spatial interpolation estimation based on adjacent data points, and the estimation confidence interval is marked.

[0008] Preferably, the K-means clustering algorithm in step S2 is performed according to the following steps: Step S21: Feature matrix construction, which involves extracting the geological feature fingerprint vectors of each parent rock sample. Arrange them in rows to construct a feature matrix X of size N×n, where N is the total number of samples and n is the total number of geological features; Z-score standardization is applied to each column of the feature matrix X to ensure that the mean of each feature is 0 and the standard deviation is 1, thus eliminating the influence of dimensional differences on the clustering results. The standardization formula is as follows: ; in, For the first The first sample The original values ​​of each feature, For the first The mean of the column features, For the first Standard deviation of the column features These are the standardized eigenvalues; Step S22: Determining the optimal number of clusters K. A joint decision strategy combining the silhouette coefficient method and the elbow rule is used to determine the optimal number of clusters K: Perform K-means clustering and calculate the silhouette coefficients of the clustering results. ; in, For the first The average distance from one sample to other samples in the same cluster For the first The minimum average distance from a sample to all samples in other clusters; Calculate the sum of squared intra-cluster errors of the clustering results: ; in For the k-th cluster, Let K be the center of the k-th cluster; plot the profile coefficient curve and elbow curve with K as the x-axis and the average profile coefficient and average SSE of R runs as the y-axis, and select the optimal number of clusters K. Step S23: Iterative calculation of K-means clustering; Step S24: Clustering results annotation and fingerprint database construction. After iterative convergence, output the cluster labels of each sample. Each cluster represents a type of parent rock source; For each cluster Calculate the cluster center eigenvector of this cluster. This serves as a standard geological fingerprint of the parent rock source; At the same time, the distribution of the actual source mining area and stratigraphic position of each sample in each cluster is statistically analyzed, and the mining area-stratigraphic position combination with the highest frequency is taken as the geological attribution label of the source of the parent rock of this type, thus completing the geological significance labeling of the clustering results; The cluster center feature vector, geological attribution label, number of samples within the cluster, and variance of sample features within the cluster are stored in the parent rock source fingerprint database.

[0009] Preferably, step S23 is as follows: Using the optimal number of clusters K determined in step S22 and the normalized feature matrix X' constructed in step S21 as input, the initial cluster centers are determined using the K-means++ initialization strategy: The first sample is randomly selected from all samples as the first initial cluster center. ; For each sample Calculate the shortest distance between it and the selected cluster centers. proportional to The probability of selecting the next cluster center is given, and this process continues until K initial cluster centers are selected. Starting with the selected K initial cluster centers, perform the following iterative calculations: Allocation steps: Calculate each sample To K cluster centers Euclidean distance Each sample is assigned to the nearest cluster center, forming K clusters; Update steps: For each cluster Recalculate its cluster centers: ; That is, the mean vector of all samples within the cluster; Alternately execute the assignment and update steps until one of the following convergence conditions is met: Change in cluster center location ,Right now ;in The preset convergence accuracy threshold has a range of values. ; Reaching the preset maximum number of iterations , The range of values ​​is .

[0010] Preferably, in step S3, the SHAP interpretability analysis method performs sensitive feature screening according to the following steps: Step S31: Baseline model training. Using the feature matrix X constructed in step S2 as input and the aggregate quality grade obtained in step S3 as the prediction target, a baseline extreme gradient boosting XGBoost model is trained as the basic model for SHAP value calculation. Step S32: SHAP value matrix calculation. For the trained baseline model, the TreeExplainer interpreter is used to calculate the SHAP value matrix for each sample. Each geological feature SHAP value of the prediction result ; Model predictions Compared with the benchmark expected value The difference between them is equal to the sum of the SHAP values ​​of all features: ; in, For the first The model prediction output for each sample. The mean of the predicted values ​​for all samples. For the first The first sample The SHAP value of each feature is calculated; the maximum absolute value of each category's SHAP value is taken as the overall SHAP value of that feature. Step S33: Ranking the global importance of features for each geological feature. Calculate the global mean of the absolute values ​​of its SHAP. : ; According to each feature The values ​​are sorted in descending order to obtain a list of global importance ranking of features. Step S34: Sensitive feature subset selection. Starting from the feature importance ranking list, add up the features one by one, starting with the first ranked feature, and calculate the cumulative SHAP contribution of the top m features. : ; When cumulative contribution First time reaching or exceeding the preset threshold Stop accumulating and select the previous value. These features constitute a subset of sensitive features for initial screening; For each feature selected into the sensitive feature subset, the Spearman rank correlation coefficient between it and the aggregate quality index is further calculated. Output sensitive feature subset: ; in This represents the actual number of features after correlation verification, and .

[0011] Preferably, in step S4, the XGBoost model is constructed and trained according to the following steps: Step S41: Training data preparation, using a selected subset of sensitive features. Corresponding feature matrix The aggregate quality grade labels obtained in step S3 are used as the model input and the prediction target. Construct a labeled dataset: ; in {Class I, Class II, Class III} are the aggregate quality grade labels; Step S42: Model objective function construction. The XGBoost model makes predictions by ensemble T classification and regression trees. The model output format is as follows: ; Where F is the function space consisting of all classification and regression trees. Let t be the function corresponding to the t-th tree; The objective function of the model It consists of two parts: a loss function term and a regularization term. ; in, The loss function is the multi-class cross-entropy loss function, calculated using the following formula: ; in, The number of aggregate quality grades ( ), For indicator functions, Let be the predicted probability that the i-th sample belongs to the k-th class. ; For regularization terms: ; in, For the first The number of leaf nodes in a tree. For the first The weight values ​​of each leaf node. The penalty coefficient for the number of leaf nodes. The L2 regularization coefficient; Step S43, Hyperparameter Definition and Search Space Setting: Define the core hyperparameters that need to be optimized in the XGBoost model and their search space as follows: Learning rate Range of values Step size 0.01; maximum tree depth Value range [3, 10], step size 1; number of estimators Value range [50, 400], step size 50; minimum subsample weight. Value range [1, 10], step size 1; column sampling ratio The value range is [0.5, 1.0], with a step size of 0.1; The L2 regularization coefficient lambda ranges from [0, 10] with a step size of 1; the L1 regularization coefficient alpha ranges from [0, 10] with a step size of 1. Step S44: Hyperparameter tuning based on Bayesian optimization. The hyperparameter search is performed using the Bayesian optimization method. The iterative process of Bayesian optimization is as follows: Ten sets of hyperparameter combinations were randomly sampled within the hyperparameter search space, and XGBoost models were trained and the validation set accuracy was calculated to serve as the initial observation data for the Gaussian process. Based on the current observation data, the posterior distribution of the Gaussian process surrogate model is updated to obtain the predicted mean and predicted variance of each hyperparameter combination. Maximize the expected improvement of the acquisition function and calculate the next set of hyperparameter combinations to be evaluated; The model was trained using the new hyperparameter combination and the accuracy on the validation set was evaluated. The results were then added to the observation data. Continue until the maximum number of iterations, 50, is reached; The set of hyperparameters that achieves the highest accuracy on the output validation set is taken as the optimal hyperparameter combination. Step S45, Model Training, Validation and Performance Evaluation: Configure the XGBoost model with the optimal hyperparameter combination obtained in step S44, train the final model on the entire training set, and evaluate its performance on the test set.

[0012] Preferably, step S5 outputs the parent rock source category of the waste rock sample and its corresponding aggregate quality potential level, specifically including: By inputting the sensitive feature vector of the sample to be tested into the prediction model, the probability distribution of the sample belonging to each parent rock source category is obtained. ,in Indicates belonging to a category The probability, and ; The category with the highest probability is taken as the source of the parent rock for the sample, and the confidence level of the sample belonging to that category is output. The confidence level is equal to the maximum probability value. If the maximum probability value is lower than the preset confidence threshold, it is judged as unclassified and a manual review process is triggered. Explanatory evidence is generated in the following ways: For the prediction model trained in step S4, the SHAP method is used again to calculate the SHAP value of each sensitive feature of the test sample on the prediction result. Sort the SHAP values ​​of each sensitive feature by absolute value, and select the top 3 to 5 features with the largest absolute values ​​as the key features that affect the prediction result. Generate an interpretable report containing key feature names, SHAP values, contribution directions, and current measured values ​​of the features, serving as the geological basis for the source tracing results.

[0013] Preferably, in step S6, the first optimization objective of the multi-objective optimization decision model is to maximize the recoverable reserves Q, and the second optimization objective is to maximize the aggregate quality comprehensive score S. The aggregate quality comprehensive score S is calculated based on the weighted sum of quality grades and quality indicators. ; in, This represents the total number of aggregate quality evaluation indicators. For the first The dimensionless score value of the aggregate quality index after positive normalization is in the range of [0,1]. For the first Overall weight of each indicator: ; in, This is the preference coefficient, with a value range of [0,1]. The analytic hierarchy process (AHP) is used to determine the matrix through pairwise comparisons; The entropy weight method is used to determine the information entropy of each indicator; The higher the final overall quality score S value, the better the quality of the aggregate prepared from the waste rock sourced from that parent rock.

[0014] Preferably, the sorting, stripping, and differentiation processing in step S6 includes: Waste rock from parent rock that is assessed as having high-quality aggregate potential is stripped, transported, crushed, and screened separately, and the products are used in high-strength concrete and high-grade highway pavement applications. Waste rock from parent rock that is assessed as having general aggregate potential will be used in ordinary concrete and general fill works. Waste rock from parent rock that is assessed as having low aggregate potential or alkali reactivity risk will be used for non-load-bearing structures, base course filling, or goaf filling.

[0015] Therefore, the above-mentioned method for source tracing and optimization of waste rock aggregate parent ore based on multi-source mine geological data mining has the following beneficial effects: (1) The present invention can quickly determine the source of the parent rock and predict the quality grade of aggregate by means of geological feature fingerprint when the waste rock is generated, providing a real-time decision-making basis for the "stripping, sorting and stacking and differentiated processing" of waste rock.

[0016] (2) This invention achieves interpretable tracing by using SHAP in two stages: the first stage is used for pre-screening of sensitive features, and the second stage is used for post-interpretation of prediction results (outputting key features and their contribution directions).

[0017] (3) Construct a dual-objective optimization model of "quality-reserves", use NSGA-II to solve the Pareto front and TOPSIS to select the optimal compromise solution, which takes into account both the minimum requirements for aggregate quality and the long-term supply stability, and is more in line with the actual production of mines.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating an embodiment of the waste rock aggregate parent ore tracing and optimization method based on multi-source mine geological data mining of the present invention; Figure 2 This is a schematic diagram showing the accuracy comparison of model performance under different feature sets in an embodiment of the present invention. Figure 3 This is a schematic diagram showing the experimental time comparison of model performance under different feature sets in an embodiment of the present invention. Detailed Implementation

[0020] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0021] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0022] Example Please see Figures 1-3 This invention provides a method for source tracing and optimization of waste rock aggregate parent ore based on multi-source mine geological data mining, including the following: Step S1: Multi-source mine geological data acquisition and spatial positioning. Obtain multi-source geological data of the mine to be analyzed. The multi-source geological data should include at least the following three types: Geological exploration data: borehole core logging data, stratigraphic columnar section, geological profile, recording the rock type, mineral assemblage, and structural features of each stratum.

[0023] Geochemical analysis data: Analysis data on the content of major elements, trace elements and rare earth elements in representative rock samples from each mining area.

[0024] Mining production data: blasting records for each mining face, coordinates of ore extraction locations, and waste rock output records.

[0025] A unified spatial coordinate benchmark is assigned to multi-source geological data, and a correlation mapping between each data point and the three-dimensional spatial location of the mine is established to form a spatial database.

[0026] This includes establishing a mapping relationship between each data point and its three-dimensional spatial location in the mine, specifically: Using the mine survey coordinate system or the geodetic coordinate system as a unified spatial reference system, coordinate transformation is performed on the borehole coordinates and borehole depth data in the geological exploration data, RTK real-time dynamic measurement and positioning is performed on the sampling points in the geochemical analysis data, and the mining face in the mining production data is spatially calibrated using the mining progress coordinates in the digital mining system. For historical data for which spatial coordinates cannot be directly obtained, geostatistical methods are used to perform spatial interpolation estimation based on adjacent data points, and the estimation confidence interval is marked.

[0027] Step S2: Constructing source tags for parent rocks based on geological fingerprints. Based on the multi-source geological data collected in Step S1, extract the geological fingerprint vectors of parent rocks from each mining area and each stratum. The feature vector includes rock and mineral assemblage features, geochemical features, and structural features; Based on the differences in geological fingerprints, the K-means clustering algorithm was used to cluster parent rock samples with known origins, determining the number of clusters K, and dividing the parent rock samples into K parent rock origin categories. A mapping relationship database of "geological fingerprint - parent rock origin" was established, and each category was assigned a category label. .

[0028] The K-means clustering algorithm is executed according to the following steps: Step S21: Feature matrix construction, which involves extracting the geological feature fingerprint vectors of each parent rock sample. Arrange them in rows to construct a feature matrix X of size N×n, where N is the total number of samples and n is the total number of geological features; Z-score standardization is applied to each column of the feature matrix X to ensure that the mean of each feature is 0 and the standard deviation is 1, thus eliminating the influence of dimensional differences on the clustering results. The standardization formula is as follows: ; in, For the first The first sample The original values ​​of each feature, For the first The mean of the column features, For the first Standard deviation of the column features These are the standardized eigenvalues; Step S22: Determining the optimal number of clusters K. The optimal number of clusters K is determined using a joint decision strategy combining the silhouette coefficient method and the elbow rule: the range of values ​​for K is set as follows: For each value of K, repeat the following operation R times ( Each time, cluster centers are randomly initialized: Perform K-means clustering and calculate the silhouette coefficients of the clustering results: ; in Let be the average distance from the i-th sample to the other samples in the same cluster. SC represents the minimum average distance from the i-th sample to all samples in other clusters, and its value ranges from 1 to 2. A larger value indicates a better clustering effect; Calculate the sum of squared intra-cluster errors of the clustering results: ; in For the k-th cluster, The center of the kth cluster is K; plot the profile coefficient curve and elbow curve with K as the x-axis and the average profile coefficient and average SSE of R runs as the y-axis, respectively. Select the K value corresponding to the local maximum value in the silhouette coefficient curve, and at the same time select the K value corresponding to the inflection point where the slope of the elbow curve changes significantly. If the two are the same, then the K value is the optimal number of clusters; if they are not the same, the K value corresponding to the larger silhouette coefficient is selected first, but it must be ensured that the average silhouette coefficient corresponding to the final K value is greater than the preset threshold of 0.6. Step S23: K-means clustering iterative calculation. Using the optimal number of clusters K determined in step S22 and the normalized feature matrix X' constructed in step S21 as input, the initial cluster centers are determined using the K-means++ initialization strategy. Step 1: Randomly select the first sample from all samples as the first initial cluster center. ; Step 2: For each sample Calculate the shortest distance between it and the selected cluster centers. proportional to The probability of selecting the next cluster center; Step 3: Repeat step 2 until K initial cluster centers are selected; starting from the selected K initial cluster centers, perform the following iterative calculations: Allocation steps: Calculate each sample To K cluster centers Euclidean distance Each sample is assigned to the nearest cluster center, forming K clusters; Update steps: For each cluster Recalculate its cluster centers That is, the mean vector of all samples within the cluster; Alternately execute the assignment and update steps until one of the following convergence conditions is met: (a) the change in the location of the cluster centers. ,Right now ;in The preset convergence accuracy threshold has a range of values. (b) Reaching the preset maximum number of iterations. , The range of values ​​is ; Step S24: Clustering results annotation and fingerprint database construction. After iterative convergence, output the cluster labels of each sample. Each cluster represents a type of parent rock source; For each cluster Calculate the cluster center eigenvector of this cluster. This serves as a standard geological fingerprint of the parent rock source; At the same time, the distribution of the actual source mining area and stratigraphic position of each sample in each cluster is statistically analyzed, and the mining area-stratigraphic position combination with the highest frequency is taken as the geological attribution label of the source of the parent rock of this type, thus completing the geological significance labeling of the clustering results; The cluster center feature vector, geological attribution label, number of samples within the cluster, and variance of sample features within the cluster are stored in the parent rock source fingerprint database.

[0029] Step S3: Screening of aggregate quality-sensitive features and fusion of multimodal data to obtain quality test data of aggregate samples prepared from waste rock from different parent rock sources. The quality test data includes at least crushing index, needle-like and flaky content, soundness, Los Angeles abrasion loss and alkali reactivity index. Interpretability analysis was used to calculate the contribution of geological feature fingerprints from various regions to aggregate quality indicators, and a subset of sensitive features that have a significant impact on aggregate quality was selected. This enables a dimensionality reduction mapping from a high-dimensional geological feature space to a aggregate quality-sensitive feature space. The sensitive feature subset contains fewer sensitive features than the total number of features in the geological feature fingerprint vector F, and the absolute value of the correlation coefficient between each sensitive feature and the aggregate quality index is greater than the preset threshold. The screening of aggregate quality-sensitive characteristics and the fusion of multimodal data were used to obtain quality test data of aggregate samples prepared from waste rock from different parent rock sources. The quality test data included at least the crushing index, needle-like and flaky content, soundness, Los Angeles abrasion loss and alkali reactivity index. The SHAP interpretability analysis method was used to calculate the contribution of various geological feature fingerprints to aggregate quality indicators, and to screen out the sensitive feature subsets that have a significant impact on aggregate quality. This enables a dimensionality reduction mapping from a high-dimensional geological feature space to a aggregate quality-sensitive feature space. The sensitive feature subset contains fewer sensitive features than the total number of features in the geological feature fingerprint vector F, and the absolute value of the correlation coefficient between each sensitive feature and the aggregate quality index is greater than the preset threshold.

[0030] The SHAP interpretability analysis method performs sensitive feature screening according to the following steps: Step S31: Baseline model training, using the feature matrix X (of size X) constructed in step S2. , The total number of samples, Using the total number of geological features as input and the aggregate quality grade obtained in step S3 as the prediction target, a benchmark extreme gradient boosting (XGBoost) model is trained as the basic model for calculating SHAP value. The dataset was randomly divided into training and test sets in a 7:3 ratio. Five-fold cross-validation was used, with classification accuracy as the optimization objective. The basic hyperparameters of the XGBoost model were determined by grid search: the learning rate ranged from [0.01, 0.3], the maximum tree depth ranged from [3, 10], the number of estimators ranged from [50, 300], and the L2 regularization coefficient ranged from [0, 10]. Step S32: SHAP value matrix calculation. For the trained benchmark model, the TreeExplainer interpreter is used to calculate each geological feature of each sample i. SHAP value of the prediction result ; The SHAP value, based on the game theory principle of Shapley values, represents the value of the first... The marginal contribution of each feature to the prediction result satisfies the additivity: the model prediction value Compared with the benchmark expected value The difference between them is equal to the sum of the SHAP values ​​of all features, that is: in, For the first The model prediction output for each sample. The mean of the predicted values ​​for all samples. For the first The first sample The SHAP values ​​of each feature; for each sample, the SHAP values ​​satisfy the assignment property, that is, the dimension of the SHAP value matrix of the prediction results for each category is . ,in The number of aggregate quality grades is used as the total SAP value for the feature, and the maximum absolute value of each SAP is taken as the overall SAP value for that feature. Step S33: Ranking the global importance of features for each geological feature. Calculate the global mean of the absolute values ​​of its SHAP. : ; According to each feature The values ​​are sorted in descending order to obtain a list of global importance ranking of features. Simultaneously, a SHAP summary diagram is generated, with the distribution of SHAP values ​​of features and the level of feature values ​​as the vertical and horizontal axes, respectively, to visually demonstrate the direction and intensity of the influence of each feature on the prediction results. Step S34: Sensitive feature subset selection. Starting from the feature importance ranking list, add up the features one by one, starting with the first ranked feature, and calculate the cumulative SHAP contribution of the top m features. : ; When cumulative contribution First time reaching or exceeding the preset threshold ( When the value reaches 85%, stop accumulating and select the previous value. These features constitute a subset of the initial sensitive features; if the number of selected sensitive features... If there are fewer than 3, the top 3 most important features will be forcibly selected and included in the sensitive feature subset. For each feature selected into the sensitive feature subset, the Spearman rank correlation coefficient between it and the aggregate quality index is further calculated. Remove Weakly correlated features with a correlation value less than 0.1 are ultimately output as a subset of sensitive features. ,in This represents the actual number of features after correlation verification, and .

[0031] In step S3, the aggregate quality inspection grade is divided into three levels according to the national standard "Construction Gravel and Crushed Stone": Class I aggregate: crushing index not greater than 10%, needle-like and flaky content not greater than 5%, and soundness loss not greater than 5%; Class II aggregate: crushing index not greater than 20%, needle-like and flaky content not greater than 15%, and soundness loss not greater than 10%; Class III aggregate: crushing index not greater than 30%, needle-like and flaky content not greater than 25%, and soundness loss not greater than 12%. For aggregates with alkali reactivity, regardless of other indicators, they are automatically downgraded or marked as requiring special treatment.

[0032] Step S4: Construct a parent rock aggregate quality potential prediction model based on Extreme Gradient Boosting (XGBoost), using the sensitive feature subset selected in Step S3. Using the aggregate quality detection level obtained in step S3 as the input feature, a parent rock aggregate quality potential prediction model based on XGBoost is constructed and trained. The dataset is divided into training, validation and test sets. The Bayesian optimization method is used to optimize the model hyperparameters. The classification accuracy and macro-average F1 score are used as model evaluation metrics. The model with the best evaluation metrics is selected as the final prediction model.

[0033] The Extreme Gradient Boosting (XGBoost) model is constructed and trained according to the following steps: Step S41: Training data preparation, using the sensitive feature subset selected in steps S3-4. Corresponding feature matrix (N×m') is the model input, and the aggregate quality grade label obtained in step S3 is the prediction target. Build a labeled dataset ,in {Class I, Class II, Class III} are the aggregate quality grade labels; A stratified sampling method was adopted, and the dataset was divided into a training set (60%), a validation set (20%), and a test set (20%) according to the proportion of each aggregate quality grade in the original dataset, to ensure that the distribution ratio of each category of samples in each dataset is consistent with that of the original dataset. Step S42: Model objective function construction. The XGBoost model makes predictions by ensemble T classification and regression trees. The model output format is as follows: ; Where F is the function space consisting of all possible classification and regression trees, f t Let t be the function corresponding to the t-th tree; The objective function L(θ) of the model consists of two parts: a loss function term and a regularization term. ; in, The loss function is the multi-class cross-entropy loss function, calculated using the following formula: ; in, The number of aggregate quality grades ( ), For indicator functions, Let be the predicted probability that the i-th sample belongs to the k-th class. ; This is a regularization term used to control model complexity and prevent overfitting. ;in, For the first The number of leaf nodes in a tree. For the first The weight values ​​of each leaf node. The penalty coefficient for the number of leaf nodes (range of values) ), L2 regularization coefficient (range of values) ); Step S43, Hyperparameter Definition and Search Space Setting: Define the core hyperparameters that need to be optimized in the XGBoost model and their search space as follows: Learning rate : Controls the contribution weight of each tree to the final result, with a range of values. Step size 0.01; Maximum depth of tree : Controls the complexity of each tree, with a value range of [3, 10] and a step size of 1; Number of estimators : That is, the total number of trees T, with a value range of [50, 400] and a step size of 50; Minimum Subsample Weight : The minimum sum of sample weights required for a leaf node, with a value range of [1,10] and a step size of 1; Column sampling ratio The proportion of features randomly sampled from each tree, with a value range of [0.5, 1.0] and a step size of 0.1; L2 regularization coefficient lambda: range [0, 10], step size 1; L1 regularization coefficient alpha: ranges from [0, 10], with a step size of 1; Step S44: Hyperparameter tuning based on Bayesian optimization. A Bayesian optimization method is used for hyperparameter search, with classification accuracy on the validation set as the optimization objective. The maximum number of iterations is set to 50. A Gaussian process is used as the surrogate model, and the acquisition function adopts an expected improvement strategy. The iterative process of Bayesian optimization is as follows: Step 1: Randomly sample 10 sets of hyperparameter combinations within the hyperparameter search space, train the XGBoost model for each combination, and calculate the validation set accuracy as the initial observation data for the Gaussian process; Step 2: Based on the current observation data, update the posterior distribution of the Gaussian process surrogate model to obtain the predicted mean and predicted variance of each hyperparameter combination; Step 3: Maximize the expected improvement of the acquisition function and calculate the next set of hyperparameter combinations to be evaluated; Step 4: Train the model using the new hyperparameter combination and evaluate the accuracy on the validation set, then add the results to the observation data; Repeat steps two through four until the maximum number of iterations of 50 is reached; The set of hyperparameters with the highest accuracy on the output validation set is taken as the optimal hyperparameter combination. At the same time, an early stopping mechanism is introduced in the Bayesian optimization process. When the accuracy on the validation set no longer improves after 10 consecutive iterations, the training of the set of hyperparameters is terminated in advance, and the number of trees is rolled back to the number of trees when the performance on the validation set is optimal. Step S45: Model Training, Validation, and Performance Evaluation. The XGBoost model is configured with the optimal hyperparameter combination obtained in Step S4-4. The final model is trained on the entire training set, and its performance is evaluated on the test set. Model performance evaluation metrics include classification accuracy, macro-average F1 score, precision and recall for each category, and the confusion matrix. The macro-average F1 score is calculated using the following formula: ; For multi-class classification problems, a one-to-many strategy is used to calculate the value of each class. and ; When the model's macro-average F1 score on the test set falls below a preset threshold of 0.80, one or more of the following remedial measures are triggered: (a) Expand the training sample size and collect additional rock samples from different mining areas and strata; (b) Return to step S3 to re-execute the SHAP sensitive feature screening and adjust the cumulative contribution threshold. (c) Expand the search space range of Bayesian optimization and increase the maximum number of iterations to 80; (d) Use a hierarchical cross-validation strategy instead of ordinary validation set partitioning, and use the average accuracy of 3-layer hierarchical cross-validation as the hyperparameter optimization target.

[0034] Step S5: Tracing the source of waste rock parent rock and predicting aggregate quality potential. For the waste rock sample to be traced, the same standard as in steps S1 and S2 is used to collect its multi-source geological data and extract geological feature fingerprints. The extracted features are input into the prediction model trained in step S4, and the parent rock source category of the waste rock sample and its corresponding aggregate quality potential level are output. Geological data collection for waste rock samples to be traced was completed at the mining site. Portable X-ray fluorescence spectrometers were used for rapid major element analysis. The acquired data was uploaded in real time to a cloud server or mine data center via wireless network. The prediction model deployed in the cloud returned the traceability results and quality prediction results in real time.

[0035] Simultaneously, using the same SHAP interpretability analysis method as in step S3, the top N key geological features affecting the prediction result and their contribution ranking are output, providing an interpretable basis for the source tracing results.

[0036] Step S5 outputs the parent rock source category of the waste rock sample and its corresponding aggregate quality potential level, specifically including: By inputting the sensitive feature vector of the sample to be tested into the prediction model, the probability distribution of the sample belonging to each parent rock source category is obtained. ,in Indicates belonging to a category The probability, and ; The category with the highest probability is taken as the source of the parent rock for the sample, and the confidence level of the sample belonging to that category is output. The confidence level is equal to the maximum probability value. If the maximum probability value is lower than the preset confidence threshold, it is judged as "unclassified" and a manual review process is triggered. Explanatory evidence is generated in the following ways: For the prediction model trained in step S4, the SHAP method is used again to calculate the SHAP value of each sensitive feature of the test sample on the prediction result. The SHAP values ​​of each sensitive feature are sorted by absolute value, and the top 3 to 5 features with the largest absolute values ​​are selected as the key features affecting the prediction result. An interpretable report containing the key feature name, SHAP value, contribution direction and current measured value of the feature is generated as the geological basis for the source tracing result.

[0037] Step S6: Optimization of waste rock aggregate parent ore based on multi-objective optimization. Obtain recoverable reserves data for waste rock from different parent rock sources in each mining area. Using the aggregate quality potential level and recoverable reserves predicted in Step S5 as optimization objectives, construct a multi-objective optimization decision model. The first optimization objective of the multi-objective optimization decision model is to maximize recoverable reserves Q, and the second optimization objective is to maximize the comprehensive aggregate quality score S. The comprehensive aggregate quality score S is calculated based on the weighted sum of quality level and quality indicators. ; in, This represents the total number of aggregate quality evaluation indicators. For the first The dimensionless score value of the aggregate quality index after positive normalization is in the range of [0,1]. Determined according to indicator type as follows: For indicators such as crushing index, needle-like and flaky content, soundness loss, Los Angeles abrasion loss, and alkali activity expansion rate, the lower the value, the better. ; For indicators such as apparent density and bulk density, where higher is always better: ; For the first The overall weight of each indicator is determined by a combination of subjective and objective weights: ; in, This is the preference coefficient, with a value range of [0,1]. The analytic hierarchy process (AHP) was used, with domain experts determining the matrix through pairwise comparisons. The entropy weight method is used to determine the information entropy of each indicator; the larger the final comprehensive quality score S value, the better the quality of the aggregate prepared from the waste rock from that parent rock.

[0038] A multi-objective optimization algorithm is used to solve for the Pareto optimal solution set or to calculate a weighted comprehensive score, outputting a preferred ranking list of parent ore. This ranking list guides the classification, stripping, separate stockpiling, and differentiated processing of mine waste rock according to its parent rock origin. The classification, stripping, and differentiated processing include: Waste rock from parent rock that is assessed as having high-quality aggregate potential is stripped, transported, crushed, and screened separately, and the products are used in high-strength concrete and high-grade highway pavement applications. Waste rock from parent rock that is assessed as having general aggregate potential will be used in ordinary concrete and general fill works. Waste rock from parent rock that is assessed as having low aggregate potential or alkali reactivity risk will be used for non-load-bearing structures, base course filling, or goaf filling.

[0039] The multi-objective optimization algorithm is the Non-Dominated Sorting Genetic Algorithm (NSGA-II). The Pareto front solution set is obtained by solving the Pareto front using the NSGA-II algorithm, where each solution corresponds to an optimal scheme for a source of parent minerals. When the number of Pareto optimal solutions is greater than 1, the optimal compromise solution is selected from the Pareto front as the final recommended scheme using the approximation of ideal solution sorting method.

[0040] Example This embodiment takes a large limestone mine as an example to fully implement the method of the present invention and verify the feasibility and effectiveness of the technical solution.

[0041] Step S1: Multi-source mine geological data acquisition and spatial positioning.

[0042] In this embodiment, the mining area is approximately 2.5 square kilometers, divided into three mining areas: the East Mining Area, the West Mining Area, and the North Mining Area. The mining level ranges from +120m to +20m, with a total of six elevations.

[0043] Geological exploration data acquisition.

[0044] A total of 87 boreholes with a total drilling footage of approximately 8,500 meters were collected and core data were compiled. Lithological descriptions, mineral assemblages, and structural features of each stratum were extracted. Thin sections of representative core samples were analyzed using a polarizing microscope to identify mineral types and their relative abundances.

[0045] Geochemical analysis data acquisition.

[0046] Systematic sampling was conducted on representative core samples from 87 boreholes, resulting in 310 samples collected from different mining areas and strata. A handheld RTK device was used to precisely locate each sampling point (accuracy better than 5 cm), and the three-dimensional spatial coordinates were recorded using the CGCS2000 national geodetic coordinate system. The samples were processed into 200-mesh powder, and major elements were analyzed using X-ray fluorescence spectrometry, selecting 10 elements: SiO2, Al2O3, Fe2O3, CaO, MgO, Na2O, K2O, TiO2, P2O5, and MnO.

[0047] Mining production data collection.

[0048] Collect data such as blasting records, ore extraction location coordinates, and waste rock output from the mine over the past three years.

[0049] The above three types of data are entered into a spatial database, and each data point is associated with its three-dimensional spatial coordinates.

[0050] Step S2: Construction of parent rock origin tags based on geological feature fingerprints Geological feature fingerprint vector extraction.

[0051] Geological fingerprint vectors were extracted from each of the 310 samples, comprising 13 features: Mineral assemblage characteristics (3 items): calcite content, dolomite content, clay content Geochemical characteristics (8 items): SiO2, Al2O3, Fe2O3, CaO, MgO content, Na2O+K2O, SiO2 / Al2O3 ratio, CaO / MgO ratio Structural features (2 items): grain size index, fracture density Feature matrix construction and standardization.

[0052] Construct a 310×13 feature matrix and perform Z-score normalization; Determining the optimal number of clusters The silhouette coefficient method and the elbow rule are used together to determine the value of K. The range of K is set to [2,10], and K-means clustering is repeated 20 times for each K value.

[0053] Table 1 Clustering evaluation indicators under different K values

[0054] The profile coefficient is highest when K=4 (0.73>0.6), and the elbow curve inflection point also appears at K=4, thus determining the optimal number of clusters K=4.

[0055] K-means clustering iterative calculation and labeling.

[0056] Using the K-means++ initialization strategy, after 25 iterations and convergence, 310 samples were divided into 4 clusters. For each cluster, the actual source mining area and stratigraphic distribution were statistically analyzed to complete the geological significance labeling.

[0057] Step S3: Screening of aggregate quality-sensitive features and fusion of multimodal data.

[0058] Aggregate quality testing.

[0059] Representative samples were taken from the four types of parent rock, processed into aggregates according to GB / T 14685 standard, and their quality indicators were tested. Table 2 Quality test results of aggregates from various parent rock sources

[0060] Training the SHAP baseline model.

[0061] Using 13 geological features from 310 samples as input and aggregate quality grade as the prediction target, a benchmark XGBoost model (5-fold cross-validation) was trained, achieving an accuracy of 83.7% on the test set.

[0062] SHAP value calculation and feature importance ranking.

[0063] TreeExplainer is used to calculate the SHAP value of each feature. For feature j, the SHAP value is calculated. The sorting results are as follows: Table 3 Ranking of global importance of each feature SHAP

[0064] Cumulative contribution The process stops when 85% is reached, with 6 features initially screened. Further calculation of Spearman rank correlation coefficients is performed; all 6 features with |ρ| greater than 0.1 are retained. Final sensitive feature subset. ={CaO content, CaO / MgO ratio, calcite content, MgO content, SiO2 content, dolomite content}, a total of 6 items.

[0065] Step S4: Construction of XGBoost parent rock aggregate quality potential prediction model.

[0066] Using six selected sensitive features as input and quality level as the prediction target, stratified sampling was employed to divide the dataset into training, validation, and test sets at 60% / 20% / 20% of the total dataset.

[0067] Bayesian optimization is used for hyperparameter search, with a search space of η∈[0.01,0.30]. , , , lambda alpha The maximum number of iterations is 50. The acquisition function adopts an expected improvement strategy and introduces an early stopping mechanism.

[0068] Optimal hyperparameter: η = 0.12 =5, =180, =3, =0.8, lambda=2.0, alpha=0.5. Confusion matrix of the model on the test set: Table 4. Confusion Matrix of XGBoost Model Test Set

[0069] Step S5: Tracing the parent rock of waste rock from unknown sources and predicting the potential of aggregate quality.

[0070] Forty-five new waste rock samples were discovered at the +60m level in the eastern mining area. Six sensitive features were rapidly analyzed on-site using a portable XRF scanner, and the data was uploaded to the cloud via a 5G private network.

[0071] Taking sample X-023 as an example: CaO=49.23%, CaO / MgO=12.35, calcite=74.6%, MgO=3.98%, SiO2=3.45%, dolomite=16.8%. Model output probability: =0.12, =0.78, =0.08, p4=0.02, confidence level 0.78>0.70, classified as C2 (Class II aggregate potential).

[0072] The interpretability report was calculated again using SHAP: Table 5. SHAP Interpretability Analysis of Sample X-023

[0073] Prediction results for 45 samples: 15 samples of C1 category, 18 samples of C2 category, 8 samples of C3 category, 2 samples of C4 category, and 2 samples unclassified (confidence level <0.70), triggering manual review.

[0074] Step S6: Optimal decision-making for waste rock aggregate mother ore based on multi-objective optimization.

[0075] Recoverable reserves: C1 category 850,000 tons, C2 category 620,000 tons, C3 category 430,000 tons, C4 category 180,000 tons. Overall quality score. ,in For the smaller the better type of index, the positive normalized value is... Determined by a combination of the analytic hierarchy process and the entropy weight method ( =0.6). Calculation yields: =1.000, =0.664, =0.364, =0.000.

[0076] With maximizing both Q and S as the dual objectives, NSGA-II was used for the solution (population 100, iterations 200 generations). The Pareto front consisted of C1 and C2. The TOPSIS method was used to select the optimal compromise solution, with C1 having the highest closeness (1.000), making it the preferred source of parent minerals.

[0077] Recommendations: Prioritize mining of C1 type (upper middle part of the eastern mining area) and process it separately for use as aggregate for high-grade highways; use C2 type as a supplementary material source; use C3 type as an alternative for low-grade concrete; and use C4 type directly for backfilling.

[0078] The following experiment demonstrates the effectiveness of SHAP sensitive feature screening in step S3.

[0079] A comparative experiment was designed using the same dataset of 310 samples as in Example 1: Experimental group: Six sensitive features (CaO content, CaO / MgO ratio, calcite content, MgO content, SiO2 content, and dolomite content) were screened using the SHAP method in step S3 of this invention. Control group 1: All 13 original characteristics.

[0080] Control group 2: Six characteristics were randomly selected (CaO content, Na2O+K2O, Fe2O3 content, TiO2 content, particle size index, and fracture density). All three groups used the same XGBoost algorithm and the same hyperparameter settings. , , With the same training / test set split (70% / 30%), run each set 10 times and take the average result: Table 6. Comparison of model performance under different feature sets

[0081] Analysis of experimental results: The accuracy difference between the experimental group (SHAP screening) and the control group 1 (all 13 features) was only 0.6 percentage points, but the number of features was reduced by 53.8% (from 13 to 6), the training time was reduced by 52.1% (from 4.8 seconds to 2.3 seconds), the model complexity was significantly reduced, and the risk of overfitting was smaller.

[0082] The accuracy of control group 2 (randomly selected features) was only 71.5%, indicating that randomly selected features cannot obtain reliable prediction results, thus proving the necessity of sensitive feature screening.

[0083] This embodiment verifies the effectiveness and superiority of the SHAP sensitive feature screening method in step S3.

[0084] Therefore, this invention adopts the above-mentioned waste rock aggregate parent ore tracing and optimization method based on multi-source mine geological data mining, and achieves effective dimensionality reduction of high-dimensional geological features through interpretability analysis. While outputting tracing conclusions, it provides key criteria and contribution directions, thereby enhancing the credibility of the decision-making process.

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for source tracing and optimization of waste rock aggregate parent ore based on multi-source mine geological data mining, characterized in that, Includes the following steps: Step S1: Obtain multi-source geological data of the mine to be analyzed, assign a unified spatial coordinate benchmark to the multi-source geological data, establish a correlation mapping between each data point and the three-dimensional spatial location of the mine, and form a spatial database. Step S2: Based on the multi-source geological data collected in Step S1, extract the geological characteristic fingerprint vectors of the parent rock in each mining area and at each stratum. The K-means clustering algorithm was used to cluster parent rock samples with known origins, determine the number of clusters K, establish a mapping relationship database of geological feature fingerprints and parent rock origins, and assign a category label to each category. ; Step S3: Obtain quality testing data for aggregate samples prepared from waste rock from different parent rock sources. Use the SHAP interpretability analysis method to calculate the contribution of each geological feature fingerprint to the aggregate quality index, and screen out the sensitive feature subset that has a significant impact on aggregate quality. ; Step S4: The sensitive feature subset selected in step S3 Using the aggregate quality detection level obtained in step S3 as the input feature, a parent rock aggregate quality potential prediction model based on XGBoost is constructed and trained, and the model with the best evaluation index is selected as the final prediction model. Step S5: For the waste rock sample to be traced, collect its multi-source geological data and extract geological feature fingerprints using the same standards as in steps S1 and S2. Input the extracted features into the prediction model trained in step S4 and output the parent rock source category of the waste rock sample and its corresponding aggregate quality potential level. Step S6: Obtain the recoverable reserves of waste rock from different parent rock sources in each mining area of ​​the mine. Using the aggregate quality potential level and recoverable reserves predicted in Step S5 as optimization objectives, construct a multi-objective optimization decision model and output the parent ore optimization ranking list after solving it.

2. The method for source tracing and optimization of waste rock aggregate parent ore based on multi-source mine geological data mining according to claim 1, characterized in that, Step S1 establishes the association mapping between each data point and the three-dimensional spatial location of the mine, specifically including: Using the mine survey coordinate system or the geodetic coordinate system as a unified spatial reference system, coordinate transformation is performed on the borehole coordinates and borehole depth data in the geological exploration data, RTK real-time dynamic measurement and positioning is performed on the sampling points in the geochemical analysis data, and the mining face in the mining production data is spatially calibrated using the mining progress coordinates in the digital mining system. For historical data for which spatial coordinates cannot be directly obtained, geostatistical methods are used to perform spatial interpolation estimation based on adjacent data points, and the estimation confidence interval is marked.

3. The method for source tracing and optimization of waste rock aggregate parent ore based on multi-source mine geological data mining according to claim 2, characterized in that, In step S2, the K-means clustering algorithm is executed according to the following steps: Step S21: Feature matrix construction, which involves extracting the geological feature fingerprint vectors of each parent rock sample. Arrange them in rows to construct a feature matrix X of size N×n, where N is the total number of samples and n is the total number of geological features; Z-score standardization is applied to each column of the feature matrix X to ensure that the mean of each feature is 0 and the standard deviation is 1, thus eliminating the influence of dimensional differences on the clustering results. The standardization formula is as follows: ； in, For the first The first sample The original values ​​of each feature, For the first The mean of the column features, For the first Standard deviation of the column features These are the standardized eigenvalues; Step S22: Determining the optimal number of clusters K. A joint decision strategy combining the silhouette coefficient method and the elbow rule is used to determine the optimal number of clusters K: Perform K-means clustering and calculate the silhouette coefficients of the clustering results. ； in, For the first The average distance from one sample to other samples in the same cluster For the first The minimum average distance from a sample to all samples in other clusters; Calculate the sum of squared intra-cluster errors of the clustering results: ； in For the k-th cluster, Let K be the center of the k-th cluster; plot the profile coefficient curve and elbow curve with K as the x-axis and the average profile coefficient and average SSE of R runs as the y-axis, and select the optimal number of clusters K. Step S23: Iterative calculation of K-means clustering; Step S24: Clustering results annotation and fingerprint database construction. After iterative convergence, output the cluster labels of each sample. Each cluster represents a type of parent rock source; For each cluster Calculate the cluster center eigenvector of this cluster. This serves as a standard geological fingerprint of the parent rock source; The distribution of actual source mining areas and strata for each sample within each cluster is statistically analyzed. The mining area-strata combination with the highest frequency is taken as the geological attribution label of the source of the parent rock of this type, thus completing the geological significance labeling of the clustering results. The cluster center feature vector, geological attribution label, number of samples within the cluster, and variance information of sample features within the cluster are stored in the parent rock source fingerprint database.

4. The method for source tracing and optimization of waste rock aggregate parent ore based on multi-source mine geological data mining according to claim 3, characterized in that, Step S23 is as follows: Using the optimal number of clusters K determined in step S22 and the standardized feature matrix X' constructed in step S21 as input, the initial cluster centers are determined using the K-means++ initialization strategy: The first sample is randomly selected from all samples as the first initial cluster center. ; For each sample Calculate the shortest distance between it and the selected cluster centers. proportional to The probability of selecting the next cluster center is given, and this process continues until K initial cluster centers are selected. Starting with the selected K initial cluster centers, perform the following iterative calculations: Assignment steps, calculate each sample To K cluster centers Euclidean distance: ； Each sample is assigned to the nearest cluster center, forming K clusters; Update steps, for each cluster Recalculate its cluster centers, which are also the mean vectors of all samples within the cluster: ； Alternately execute the assignment and update steps until one of the following convergence conditions is met: Change in cluster center location : ； in The preset convergence accuracy threshold has a range of values. ; Reaching the preset maximum number of iterations , The range of values ​​is .

5. The method for source tracing and optimization of waste rock aggregate parent ore based on multi-source mine geological data mining according to claim 4, characterized in that, In step S3, the SHAP interpretability analysis method performs sensitive feature screening according to the following steps: Step S31: Baseline model training. Using the feature matrix X constructed in step S2 as input and the aggregate quality grade obtained in step S3 as the prediction target, a baseline extreme gradient boosting XGBoost model is trained as the basic model for SHAP value calculation. Step S32: SHAP value matrix calculation. For the trained baseline model, the TreeExplainer interpreter is used to calculate the SHAP value matrix for each sample. Each geological feature SHAP value of the prediction result ; Model predictions Compared with the benchmark expected value The difference between them is equal to the sum of the SHAP values ​​of all features: ； in, For the first The model prediction output for each sample. The mean of the predicted values ​​for all samples. For the first The first sample The SHAP value of each feature is calculated; the maximum absolute value of each category's SHAP value is taken as the overall SHAP value of that feature. Step S33: Ranking the global importance of features for each geological feature. Calculate the global mean of the absolute values ​​of its SHAP. : ； According to each feature The values ​​are sorted in descending order to obtain a list of global importance ranking of features. Step S34: Sensitive feature subset selection. Starting from the feature importance ranking list, add up the features one by one, starting with the first ranked feature, and calculate the cumulative SHAP contribution of the top m features. : ； When cumulative contribution First time reaching or exceeding the preset threshold Stop accumulating and select the previous value. These features constitute a subset of sensitive features for initial screening; For each feature selected into the sensitive feature subset, the Spearman rank correlation coefficient between it and the aggregate quality index is further calculated. Output sensitive feature subset: ； in This represents the actual number of features after correlation verification, and .

6. The method for source tracing and optimization of waste rock aggregate parent ore based on multi-source mine geological data mining according to claim 5, characterized in that, In step S4, the XGBoost model is built and trained according to the following steps: Step S41: Using the selected sensitive feature subset Corresponding feature matrix The aggregate quality grade labels obtained in step S3 are used as the model input and the prediction target. Build a labeled dataset: ； in {Class I, Class II, Class III} are the aggregate quality grade labels; Step S42: The XGBoost model makes predictions by ensembles T classification and regression trees. The model output format is as follows: ； Where F is the function space consisting of all classification and regression trees. Let t be the function corresponding to the t-th tree; The objective function of the model It consists of two parts: a loss function term and a regularization term. ； in, The loss function is the multi-class cross-entropy loss function, calculated using the following formula: ； in, This represents the number of aggregate quality grade categories. For indicator functions, Let be the predicted probability that the i-th sample belongs to the k-th class. ; For regularization terms; Step S43: Define the core hyperparameters that the XGBoost model needs to optimize and their search space as follows: Key hyperparameters include the learning rate. Maximum depth of the tree Number of estimators Minimum subsample weight Column sampling ratio L2 regularization coefficient lambda, L1 regularization coefficient alpha; Step S44: Hyperparameter tuning based on Bayesian optimization. The Bayesian optimization method is used to search for hyperparameters and select the optimal combination of hyperparameters. Step S45, Model Training, Validation and Performance Evaluation: Configure the XGBoost model with the optimal hyperparameter combination obtained in step S44, train the final model on the entire training set, and evaluate its performance on the test set.

7. The method for source tracing and optimization of waste rock aggregate parent ore based on multi-source mine geological data mining according to claim 6, characterized in that, Step S5 outputs the parent rock source category of the waste rock sample and its corresponding aggregate quality potential level, specifically including: By inputting the sensitive feature vector of the sample to be tested into the prediction model, the probability distribution of the sample belonging to each parent rock source category is obtained. ,in Indicates belonging to a category The probability, and ; The category with the highest probability is taken as the source of the parent rock for the sample, and the confidence level of the sample belonging to that category is output. The confidence level is equal to the maximum probability value. If the maximum probability value is lower than the preset confidence threshold, it is judged as unclassified and a manual review process is triggered.

8. The method for source tracing and optimization of waste rock aggregate parent ore based on multi-source mine geological data mining according to claim 7, characterized in that, In step S6, the first optimization objective of the multi-objective optimization decision model is to maximize the recoverable reserves Q, and the second optimization objective is to maximize the aggregate quality comprehensive score S. The aggregate quality comprehensive score S is calculated based on the weighted sum of quality grades and quality indicators. ； in, This represents the total number of aggregate quality evaluation indicators. For the first The dimensionless score value of the aggregate quality index after positive normalization processing, with a value range of [0,1]; For the first Overall weight of each indicator: ； in, This is the preference coefficient, with a value range of [0,1]. The analytic hierarchy process (AHP) is used to determine the matrix through pairwise comparisons; The entropy weight method is used to determine the information entropy of each indicator; The higher the final overall quality score S value, the better the quality of the aggregate prepared from the waste rock sourced from that parent rock.

9. The method for source tracing and optimization of waste rock aggregate parent ore based on multi-source mine geological data mining according to claim 8, characterized in that... Step S6, the classification, stripping, and differentiated processing, includes: Waste rock from parent rock that is assessed as having high-quality aggregate potential is stripped, transported, crushed, and screened separately, and the products are used in high-strength concrete and high-grade highway pavement applications. Waste rock from parent rock that is assessed as having general aggregate potential will be used in ordinary concrete and general fill works. Waste rock from parent rock that is assessed as having low aggregate potential or alkali reactivity risk will be used for non-load-bearing structures, base course filling, or goaf filling.

Images