Deposit type prediction method and system based on multiple algorithms

Abstract

The application discloses a deposit type prediction method based on multiple algorithms, comprising the following steps: collecting trace element data of a mineral and corresponding deposit genetic type data, and performing normalization processing on the trace element data; analyzing the correlation between each trace element and performing element recombination; dividing the element data into a training set and a test set according to a certain proportion, inputting the training set into multiple algorithm models for training, and selecting the best model through cross-validation; performing element feature selection on the model by using a recursive feature elimination method; retraining the model to obtain an optimal random forest model for predicting the deposit genetic type based on the trace element; selecting trace element features of a to-be-tested mineral according to the optimal element feature set, and inputting the optimal random forest model to obtain the deposit genetic type of the to-be-tested mineral. The application greatly improves the precision and efficiency of deposit genetic type identification.

Description

Technical Field

[0001] This invention relates to the field of mineral resource prediction and evaluation, and in particular to a method and system for predicting mineral deposit types based on multiple algorithms. Background Technology

[0002] Mineral resources are a vital material foundation for economic and social development, and their exploration and development are crucial to national welfare and security. Currently, the number of outcrops and shallow deposits available for exploration and development is drastically decreasing, making further exploration and mining of mineralization information in overburdened areas and deep mineral deposits particularly important. However, uncovering the metallogenic patterns and clues hidden within the vast amount of high-quality, multi-scale geoscientific data accumulated by predecessors remains a significant challenge.

[0003] With the continuous development of computer science and technology, the era of artificial intelligence and big data has begun. Machine learning methods are being applied to the field of geology. Their ability to classify, extract, and mine massive amounts of geological data based on certain algorithms has shown strong advantages in mineralization prediction, opening up a new perspective for geological exploration combined with intelligent technologies. Machine learning algorithms can extract subtle information from geoscientific data that is not easily detected directly. This information may provide valuable guidance on the type, type, and geological background of mineral resources, which is of great significance for mineral resource prediction and evaluation.

[0004] The current method of identifying different types of mineral deposits mainly relies on comprehensive research and analysis of mineralization characteristics, mineralization conditions and mineralization-controlling factors. This requires a lot of field mapping, logging, experimental testing and expert guidance to arrive at a result, which is time-consuming and labor-intensive, and the estimated results may also have significant discrepancies.

[0005] In the era of big data, we can easily summarize and reuse the same mineral chemical data collected by our predecessors in order to obtain more reliable patterns and new information. Summary of the Invention

[0006] The main objective of this invention is to provide a method and system for predicting mineral deposit types based on multiple algorithms, which can improve the accuracy and efficiency of identifying the genetic types of mineral deposits.

[0007] The technical solution adopted in this invention is:

[0008] A method for predicting mineral deposit types based on multiple algorithms is provided, including the following steps:

[0009] S1. Collect trace element data of a certain mineral and corresponding ore deposit genetic type data from a large amount of literature, and standardize the trace element data to obtain experimental data tables.

[0010] S2. Analyze the correlation between the trace elements in the experimental data table, and recombine the elements to obtain the characteristics of elements or combinations of elements that are not correlated, and update the experimental data table.

[0011] S3. Divide the element data in the experimental data table into training set and test set according to a certain ratio. Input the training set into multiple algorithm models for training and select the best model as the random forest model through cross-validation.

[0012] S4. Based on the random forest model, the recursive feature elimination method is used to filter the element features in the updated experimental data table, select the optimal element feature set, and update the random forest model.

[0013] S5. Retrain the updated random forest model to obtain the optimal random forest model based on the prediction of mineral deposit genetic types using trace elements.

[0014] S6. Select the trace element characteristics of the mineral to be tested based on the optimal element feature set, and input them into the optimal random forest model to obtain the ore deposit genetic type of the mineral to be tested.

[0015] Following the above technical solution, step S2 specifically involves:

[0016] S21. Calculate the correlation between each trace element in the experimental data table, screen out the elements with significant linear correlation, add them in pairs to obtain the combined elements, replace the original two elements, and update the experimental data table.

[0017] S22. Based on the box plots of the new elements in different types of mineral deposit genesis, divide the elements with overlapping values ​​by the elements that show opposite indication patterns and significant differences in numerical range in different types of mineral deposit genesis to obtain new combined elements, or accumulate the elements that show consistent indication patterns and significant differences in numerical range to obtain new combined features, and update the experimental data table.

[0018] S23. Delete elements with consistent indication patterns and overlapping combination elements from the box plots of different types of mineral deposit genesis, and then remove combination features with significant linear correlation to obtain the final experimental data table of elements or combination elements that are not correlated.

[0019] Following the above technical solution, in step S3, before model training, the data in the training set is mathematically transformed to make the data distribution closer to a normal distribution.

[0020] Following the above technical solution, in step S1, the standardization processing of trace element data includes replacing the element data with identifiers, uniformly converting the element data type to floating point type, calculating the proportion of missing values ​​for each element data, and removing elements with data volume below a certain value and duplicate sample data.

[0021] Following the above technical solution, in step S1, the normalization processing of trace element data includes filling missing values ​​in the element data, specifically including the following three methods: (1) filling the values ​​containing missing columns with the mean and median respectively; (2) filling by linear interpolation; (3) filling by KNN algorithm model.

[0022] Following the above technical solution, after filling in missing values, the reliability of the data is verified by using quantile-quantile plots.

[0023] Following the above technical solution, step S5 further includes evaluating the accuracy of the trained random forest model by using the remaining untrained samples, the overall dataset, and the external dataset as three test sets, and predicting the classification performance based on the accuracy evaluation report and confusion matrix of the test sets.

[0024] Following the above technical solution, based on the importance of each element feature in the optimal element feature set of the optimal random forest model, the features that contribute the most to the model are selected in turn, and they are paired up to serve as the x and y axes respectively, resulting in multiple spatial cross-sections. The decision boundaries on the spatial cross-sections are then visualized by adjusting certain parameters.

[0025] This invention also provides a system for predicting mineral deposit types based on multiple algorithms, including:

[0026] The data collection module is used to collect trace element data of a certain mineral and corresponding ore deposit genetic type data from a large amount of literature, and to standardize the trace element data to obtain experimental data tables.

[0027] The element combination selection module is used to analyze the correlation between various trace elements in the experimental data table, and to recombine the elements to obtain the characteristics of elements or combinations of elements that are not correlated, and update the experimental data table.

[0028] The model selection and validation module is used to divide the element data in the experimental data table into training set and test set according to a certain ratio, input the training set into various algorithm models for training, and select the best model as the random forest model through cross-validation.

[0029] The element feature selection module is used to filter the element features in the updated experimental data table based on the random forest model using the recursive feature elimination method, select the optimal set of element features, and update the random forest model.

[0030] The model training module is used to retrain the updated random forest model to obtain the optimal random forest model based on the prediction of the genetic type of mineral deposits using trace elements.

[0031] The prediction module is used to select the trace element characteristics of the mineral to be tested based on the optimal element feature set, and input them into the optimal random forest model to obtain the ore deposit genetic type of the mineral to be tested.

[0032] The present invention also provides a computer storage medium storing a computer program executable by a processor, the computer program executing the mineral deposit type prediction method based on multiple algorithms described in the above technical solution.

[0033] The beneficial effects of this invention are as follows: By collecting, analyzing, and integrating trace element data of a specific mineral from a vast amount of literature, this invention calculates the relationships between various trace elements and recombines them, making full use of each retained element to find the simplest element combination index—that is, expressing the most information with the fewest indicators. Using the recombined trace element data table, the best model is selected from multiple algorithm models, and the optimal element feature set is further selected. The model is then updated and trained to obtain the optimal random forest model, which is used to identify the genetic type of the mineral deposit under test. This invention significantly improves the accuracy and efficiency of mineral deposit genetic type identification.

[0034] Furthermore, this invention fills in missing values ​​using three different methods and verifies the distribution of the data before and after filling, which greatly enhances the reliability of data filling and reduces data waste to a certain extent.

[0035] Furthermore, after a series of preprocessing steps, the collected data inevitably loses one or more trace elements. By calculating the correlation coefficients between these elements, a preliminary exploration of their relationships is conducted, screening out elements with significant linear correlations. These two types of elements are then added together to create a cumulative element combination, which serves as a new feature replacing the two correlated elements. A two-dimensional box plot is then used to visually observe the discriminative power of the elements, identifying the elements with the most significant numerical differences across different deposit types. To maximize the retention of each element and ensure that each element combination feature has clear indicative power, elements with overlapping values ​​are divided by elements that exhibit opposite indicative patterns across different deposit types to obtain new combination elements. Finally, the combined elements are displayed again in the form of box plots, retaining combinations with significantly different and completely non-overlapping numerical ranges, while excluding features with consistent indicative power and similar numerical ranges, leaving the optimal element combination. This trace element combination method proposed in this invention simplifies feature indicators from complex variables without losing information about the sample itself, and amplifies the weak information of a single element through combination.

[0036] Furthermore, the classification performance of the data can be intuitively seen through the decision boundaries of each aspect of the optimal random forest model, realizing visualization of the ore deposit type discrimination, thereby revealing the discrimination rules of new features step by step. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This is a flowchart of a mineral deposit type prediction method based on multiple algorithms according to an embodiment of the present invention;

[0039] Figure 2 This is a trace element correlation diagram of turquoise according to an embodiment of the present invention;

[0040] Figure 3 This is a box plot of trace elements in chlorite after incorporating elements according to an embodiment of the present invention;

[0041] Figure 4 This is a box plot of feature elements after further element combination in an embodiment of the present invention;

[0042] Figure 5 This is a correlation diagram of feature elements in an embodiment of the present invention;

[0043] Figure 6 This is a feature selection diagram of an embodiment of the present invention;

[0044] Figure 7 This is a learning curve graph of an embodiment of the present invention;

[0045] Figure 8 This is the confusion matrix diagram of test set 1 in this embodiment of the invention;

[0046] Figure 9 This is the confusion matrix diagram of test set 2 in this embodiment of the invention;

[0047] Figure 10 This is a ranking of feature importance in the random forest model of this invention.

[0048] Figure 11 This is a classification decision surface diagram according to an embodiment of the present invention;

[0049] Figure 12 This is a two-dimensional scatter plot of an embodiment of the present invention;

[0050] Figure 13 This is a three-dimensional classification scatter plot of the feature indicator in an embodiment of the present invention;

[0051] Figure 14 This is a flowchart of the mineral deposit type prediction method based on multiple algorithms in Embodiment 3 of the present invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0053] Example 1

[0054] like Figure 1 As shown, the mineral deposit type prediction method based on multiple algorithms in this embodiment of the invention includes the following steps:

[0055] S1. Collect trace element data of a certain mineral and corresponding ore deposit genetic type data from a large amount of literature, and standardize the trace element data to obtain experimental data tables.

[0056] S2. Analyze the correlation between the trace elements in the experimental data table, and recombine the elements to obtain the characteristics of elements or combinations of elements that are not correlated, and update the experimental data table.

[0057] S3. Divide the element data in the experimental data table into training set and test set according to a certain ratio. Input the training set into multiple algorithm models for training and select the best model as the random forest model through cross-validation.

[0058] S4. Based on the random forest model, the recursive feature elimination method is used to filter the element features in the updated experimental data table, select the optimal element feature set, and update the random forest model.

[0059] S5. Retrain the updated random forest model to obtain the optimal random forest model based on the prediction of mineral deposit genetic types using trace elements.

[0060] S6. Select the trace element characteristics of the mineral to be tested based on the optimal element feature set, and input them into the optimal random forest model to obtain the ore deposit genetic type of the mineral to be tested.

[0061] The above embodiments analyze and integrate trace element data for a specific mineral collected from a vast amount of literature, calculate the relationships between various trace elements, and recombine them to fully utilize each element as much as possible, finding the simplest element combination index—that is, expressing the most information with the fewest indicators. Using the recombined trace element data table, the best model is selected from multiple algorithm models, and the optimal element feature set is further chosen. The model is then updated and trained to obtain the optimal random forest model, which is used to identify the genetic type of the tested mineral deposit. This invention significantly improves the accuracy and efficiency of mineral deposit genetic type identification.

[0062] Example 2

[0063] Chlorite is a common hydrous layered silicate mineral found in many different types of magmatic-hydrothermal deposits. The varied chemical composition of chlorite is of great significance for the genesis and mineral exploration of different hydrothermal deposits. Therefore, this invention also establishes a method to process publicly available trace element data of chlorite minerals worldwide, aiming to build an algorithmic model for identifying and predicting different genetic deposit types based on chlorite minerals, and to discover new chemical element discrimination indicators. Furthermore, the decision-making behavior of the model is visualized, revealing the discrimination rules of new characteristics step by step.

[0064] The following examples use chlorite as an example to implement the deposit type prediction method based on multiple algorithms described in the above examples.

[0065] (1) Data collection

[0066] Collect the published chlorite trace element data from large literature databases such as Web of Science and CNKI, and merge them according to the element names. For example, the specific trace elements collected are Li, B, Na, Mg, Al, Si, K, Ca, Ti, V, Cr, Mn, Fe, Co, Ni, Cu, Zn, As, Sr, Y, Zr, Nb, Ag, Sn, Sb, Ba, La, Ce, Eu, Gd, Yb, Lu, Hf, Ta, Au, Tl, Pb, Bi, Th, U, a total of 40 kinds. Since the trace element data are all from different literatures, there are certain differences in the identification of sample data with values below the detection limit and some unmeasured results in different literatures. In order to uniformly replace the identification of such data, such as: LOD, <LOD, No_measures, 0 value, etc., all are replaced with blank values (NaN); the data containing "<" a certain value is removed the less-than sign and replaced with half of its value; since the trace element data belongs to quantitative data, the element data type is uniformly converted to floating point type. Finally, calculate the proportion of missing values in each element data, and剔除 the elements with data volume less than 30% of the total and the duplicate samples to obtain the experimental data table.

[0067] (2) Filling missing values

[0068] For huge geological data, data missing is a common problem when processing geological data. In this regard, predecessors usually choose to directly delete it simply or use traditional mean filling and other methods to obtain a complete data table. These processing methods often cause waste of data, or due to the lack of further verification of the interpolated data, the data deviates from its actual distribution.

[0069] Therefore, in order to more fully consider the data reliability issue, the present invention combines three different methods including traditional processing methods to fill the missing values, and verifies the distribution of the data before and after filling: ① Fill the values of the columns containing missing values with the mean and median respectively; ② Fill with the simple linear interpolation method; ③ Fill through the KNN algorithm model.

[0070] A quantile-quantile (QQ) plot was used to visualize and compare the distribution of the three imputed data with the original data. For each group of elements with missing values, the method with the highest consistency with the original data distribution was selected. Ultimately, the KNN algorithm was found to be the most effective at imputing missing values, and therefore, KNN was chosen for this purpose. Furthermore, elements that deviated significantly from the original data distribution were discarded to obtain a complete data table. Traditional KNN missing value imputation simply involves interpolation followed by direct application. However, this method, by incorporating QQ plot reliability verification, significantly enhanced the reliability of the data imputation and reduced data waste to some extent.

[0071] KNN missing value imputation principle: The KNN algorithm is a method of imputing missing values ​​by using the average value of the k nearest neighboring samples of the sample to be estimated. First, we need to calculate the distance between the sample point containing the missing value and the other sample points. In the case of missing coordinates, the Euclidean distance between sample points is calculated by ignoring the missing values ​​and increasing the weight (w) of the non-missing coordinates.

[0072] Weight w = Total number of coordinate dimensions / Number of non-missing coordinate dimensions

[0073] The formula for calculating the Euclidean distance between two points is:

[0074]

[0075] Where i represents the coordinate dimension, i = 1, 2, 3...n.

[0076] Secondly, specify a K value, which is the number of nearest neighbor points, usually 5 by default. Based on the calculated Euclidean distance between the unknown sample and the known sample, the coordinate values ​​of the missing dimension corresponding to the K sample points with the smallest distance are added together and averaged to obtain the estimated value of each item corresponding to the missing coordinate point.

[0077] The QQ plot, short for quantile-quantile plot, is used to test whether two sets of data follow the same distribution. Besides checking if a single set of data follows a normal distribution, it can also check if two sets of data follow the same distribution. By comparing the quantiles of a sample set of data with the quantiles of a set of data with a known distribution, the distribution of the data can be examined. A percentile is the value of a data set arranged in ascending order, with the cumulative percentile calculated. It can be represented as: a set of n observations arranged numerically. For example, the value at the p% position is called the p-th percentile.

[0078] (3) Trace element combination

[0079] After a series of preprocessing steps, one or more trace elements will inevitably be lost. Therefore, in order to make full use of each element that is retained, we try to find the simplest combination of elements as much as possible without losing any element, that is, to express the most information with the fewest indicators.

[0080] Since the distribution of trace elements often does not conform to a normal distribution, we choose to calculate the Spearman correlation coefficient between each element to preliminarily explore the correlation between them, screening out elements with |R|>0.5 indicating a significant linear correlation. Then, we perform hypothesis testing to observe whether the p-value is less than 0.05 (confidence level α is set to 0.05). If p<0.05, it indicates that the two elements are indeed correlated. Finally, we add these two types of elements together, creating a cumulative element combination as a new feature to replace the two correlated elements.

[0081] Secondly, using two-dimensional box plots, the discriminative power of elements is visually observed beforehand to identify the elements with the most significant numerical differences across different deposit types. To maximize the retention of each element and ensure that each element combination has clear indicative characteristics, elements with significantly overlapping numerical ranges across different deposit types (i.e., those with poor indicative performance) are compared with elements whose numerical trends are opposite and whose numerical ranges differ significantly, yielding new features. Elements with consistent indicative trends and significantly different numerical ranges are then accumulated to obtain new features. Finally, all the combined elements are displayed again in the form of box plots, retaining combinations with large, non-overlapping numerical ranges and excluding those with consistent indicative power and similar numerical ranges, leaving only the optimal element combinations.

[0082] Finally, the correlation between all selected features is checked again. If the correlation between features is close to 0, it means that each feature is independent of the others and there are no collinear features. All selected features can participate in the modeling.

[0083] Considering that some minerals (such as chlorite) have data at ppm or lower, and the information of a single element is relatively weak, the trace element combination method proposed in this invention can simplify the characteristic indicators from complex variables without losing the information of the sample itself, and amplify the weak information of a single element through combination.

[0084] (4) Model selection

[0085] The data is divided into training and test sets in an 8:2 ratio (the data ratio can be adjusted according to actual needs). The training set is then input into various algorithm models for training (in this embodiment, nine algorithm models are selected, including logistic regression, support vector machine, K-nearest neighbors, Naive Bayes, decision tree, random forest, gradient boosting tree, perceptron, and artificial neural network). The model with the best classification performance is selected through cross-validation scoring.

[0086] In one embodiment of the present invention, a 5-fold cross-validation method is used. The 5-fold cross-validation score is obtained by dividing the training set into 5 equal parts, selecting 4 of them to input into the model for training, and using the remaining 1 / 5 to test the trained model. This process is repeated to select different 4 equal parts to train 5 models. The 5-fold cross-validation score is obtained by summing the validation scores of the 5 models and averaging them.

[0087] (5) Feature selection

[0088] A recursive feature elimination method based on a random forest model is employed for feature selection. The basic idea is to recursively build the model, repeatedly selecting features from an increasingly smaller feature set. For n feature variables, the first random forest model is trained on the initial feature set, and the importance ranking of each feature in the model is obtained after training. Then, the least important features are removed from the n features, and the model is trained again based on the n-1 feature set. This recursive training is repeated n times, and the optimal feature set size can be selected using the score curve. In this way, we can eliminate features that have little or no effect on the model, ensuring that the remaining features all positively contribute to the model's improvement.

[0089] (6) Accuracy assessment

[0090] After the model is trained, three test sets are used: the remaining untrained samples (e.g., 20% of the remaining samples, which can be adjusted according to actual needs), the overall dataset, and the external dataset. The classification prediction effect of the model is tested by the accuracy evaluation report of the test sets, confusion matrix, and other indicators.

[0091] Based on model predictions, there are generally four possible outcomes:

[0092] TP (True Positive): Successfully predicts a positive sample as positive. TN (True Negative): Successfully predicts a negative sample as negative. FP (False Positive): Incorrectly predicts a negative sample as positive. FN (False Negative): Incorrectly predicts a positive sample as negative.

[0093] Meaning of each evaluation indicator in the accuracy evaluation report (precision, recall, f1-score, support):

[0094] 1) Accuracy: The proportion of the total number of correctly predicted samples out of the total number of samples. It is applicable to balanced datasets.

[0095]

[0096] 2) Precision: Represents the proportion of samples of a certain class that are actually predicted in the predicted results, i.e., the probability that a certain class was predicted correctly.

[0097]

[0098] 3) Recall: Represents the proportion of samples of a certain class that are predicted from the actual samples of that class, that is, the probability of a certain class being predicted.

[0099]

[0100] 4) F1 score: It is a statistical metric used to measure the precision of a binary classification model. It is used to measure the precision of imbalanced data. The F1 score represents a weighted average of the model's precision and recall, and its value ranges from [1,0].

[0101]

[0102] Macro F1 average (avg): The average of the F1 scores for all labels, generally used in multi-class classification. Weighted F1 average (avg): The weighted average of the F1 scores for all labels, generally used in multi-class classification where imbalance exists.

[0103] 5) Support: The number of samples in each category.

[0104] Confusion matrix: mainly used to compare classification results with actual measured values, intuitively showing which part the classification model will confuse when making predictions, the degree of confusion and the number of confusions.

[0105]

[0106] (7) Decision boundary of the model

[0107] The internal process of how machine learning models classify and predict targets based on multi-dimensional input features has always been a black box problem. Therefore, we attempt to find a method to extract certain aspects of the model and combine them with two-dimensional and three-dimensional diagrams to further understand how our multi-dimensional random forest model makes decisions. This allows us to see the so-called "black box" more intuitively and understand more clearly how the model distinguishes different types based on different features, thus enhancing the model's reliability. This embodiment of the invention mainly calls the `plot_decision_regions` API interface in the `mlxtend` (machine learning extensions) library in Python to draw the model's decision regions. It mainly obtains the drawing results by passing in a feature array, a prediction target array, a classifier model, and some control parameters (value, width). By continuously adjusting certain parameters, the function can be made to use any two features as a planar coordinate system and output the model's classification decision projection surface on the specified two feature axes. This allows for the display of the classifier's decision cross-section with multi-dimensional input features, giving us a more intuitive understanding of the model's classification performance.

[0108] Chlorite is a common hydrous layered silicate mineral found in many different types of magmatic-hydrothermal deposits. The varied chemical composition of chlorite is of great significance for the genesis and mineral exploration of different hydrothermal deposits. Therefore, this invention establishes a method for processing publicly available trace element data of chlorite minerals worldwide, aiming to develop an algorithmic model for identifying and predicting different genetic deposit types based on chlorite minerals, and to discover new chemical element discrimination indicators. Furthermore, the decision-making behavior of the model is visualized, revealing the discrimination rules of new characteristics step by step.

[0109] Example 3

[0110] The following section further uses chlorite as an example to determine specific mineralization types, such as... Figure 14 As shown, the method specifically includes the following steps:

[0111] 1. Data Collection

[0112] We collected published trace element data for chlorite from large databases such as Web of Science and CNKI, representing both hydrothermal and sedimentary deposit genetic types. We calculated the percentage of missing values ​​for each element, removed elements with less than 30% of the total data and duplicate samples, resulting in 2761 records containing 21 elements (Mn, Na, Ca, K, Ti, Li, B, V, Cr, Co, Ni, Cu, Zn, As, Sr, Y, Zr, Sn, Ba, Pb, Mg) and their corresponding deposit genetic types.

[0113] 2. Handling Missing Data Values

[0114] The KNN algorithm model was used for filling, and the distribution of the filled data was visualized and compared with the original data using a QQ graph. Elements that deviated significantly from the original data distribution were removed, and finally 12 relatively reliable elements were obtained: Mn, Na, K, Li, B, V, Co, Zn, Sr, Y, Sn, and Mg.

[0115] 3. Feature Combination

[0116] To test the correlation between elements, such as Figure 2 As shown, the correlation coefficient between Mn and Zn is 0.66, indicating a positive correlation. Therefore, the two elements are added together to create a new feature, Mn+Zn, which replaces the Mn and Zn elements.

[0117] like Figure 3 As shown, by visually observing the discriminative properties of elements in advance through two-dimensional box plots, it was found that there are 6 elements, namely Li, V, Co, Sn, Mg, and Zn+Mn, whose two types of box plots indicate the same element with almost no overlap, and 5 elements, namely Na, K, B, Sr, and Y, whose box plots have overlap.

[0118] The indicative patterns of Li, V, Co, Mg, Zn+Mn, and B are generally lower for sedimentary rock types and higher for hydrothermal types; the indicative patterns of Na and K are generally higher for sedimentary rock types and lower for hydrothermal types; Sr and Y have no obvious indicative patterns due to partial overlap.

[0119] To preserve each element to the greatest extent and ensure that each feature has obvious indicative power, elements with poor indicative performance (i.e., overlapping) are divided pairwise with elements whose general indicative patterns are opposite and whose numerical ranges differ significantly, resulting in new feature combinations such as Na / Li, Na / V, Na / Co, Na / Mg, Na / Zn+Mn, K / Li, K / V, K / Co, K / Mg, and K / Zn+Mn. Elements without obvious indicative patterns are accumulated with elements whose numerical ranges differ significantly, resulting in Sr+Zn+Mn, Sr+Mg, Y+Zn+Mn, Y+V, Y+Co, Y+Mg, and Y+Zn+Mn.

[0120] After comparing and selecting using box plots, features with completely identical indications and overlapping combinations of features were removed. Then, the correlation of each feature was examined, and combinations of features with a correlation coefficient |R|>=5, indicating significant linear correlation, were eliminated. Figure 4 As shown, six features were ultimately obtained: Sn, Mn+Zn, Na / Co, Y+V, Sr+Mg, and Li / K. The correlation among these six features was found to be close to zero, indicating the absence of collinearity. Figure 5 As shown.

[0121] 4. Model Selection

[0122] The data was divided into training and test sets in an 8:2 ratio. The training set was then fed into nine different algorithm models for training (logistic regression, support vector machine, K-nearest neighbors, Naive Bayes, decision tree, random forest, gradient boosting tree, perceptron, and artificial neural network). The best-performing model was selected as the random forest model based on the F1 score of the 5-fold cross-validation set (see Table 1 below). Furthermore, before training, some data transformations, such as logarithmic transformation, can be performed to make the data distribution closer to a normal distribution. This allows models sensitive to data distribution to achieve better scores, enabling a more objective evaluation and selection of models. The model with the highest final score was random forest, and the order of the impact of data transformation on the model, from highest to lowest, was: logistic regression, perceptron, artificial neural network, nearest neighbor, Naive Bayes, support vector machine, and decision tree. Moreover, it was found that different data distributions had no impact on the random forest and gradient boosting tree models.

[0123] Table 1: Cross-validation scores (F1) of the models

[0124] Model No transformation Logarithmic transformation (log) Random Forest 0.994 0.994 Gradient boosting classification tree 0.992 0.992 Decision Tree 0.990 0.991 nearest neighbor 0.985 0.991 Artificial Neural Networks 0.755 0.983 Support Vector Machine 0.983 0.984 Logistic Regression 0.950 0.981 Naive Bayes 0.985 0.981 Perceptron 0.619 0.981

[0125] 5. Feature Selection

[0126] On the divided training set, a recursive feature elimination method based on a random forest model is used to perform minimum optimal feature selection. By continuously changing the number of features, the model is repeatedly built to obtain the highest cross-validation score. After training and selection, no redundant features are found. Figure 6 As shown, the six input features are the optimal features. Adding any one of these features will increase the model score. Therefore, it can be ensured that the six features Sn, Mn+Zn, Na / Co, Y+V, Sr+Mg, and Li / K all contribute to improving the model.

[0127] 6. Accuracy Assessment

[0128] After updating the features, retrain the model and examine the learning curves of its training score and 4-fold cross-validation score to determine if the model has a fitting problem. The learning curves plot the changes in accuracy of the training set and cross-validation under different training set sizes, thereby determining whether the model is underfitting or overfitting due to excessively high or low variance or bias, and whether increasing the training set size can reduce the fitting problem.

[0129] like Figure 7 As shown in the diagram, the learning curve reveals that the model's training score remains consistently at 1, indicating a perfect fit to the training set. The cross-validation score gradually increases and tends towards 1 as the number of training samples increases, demonstrating an increasingly better fit and continuously improving generalization performance to its optimal level. The last two points almost coincide, indicating extremely high model accuracy and no fitting issues.

[0130] The remaining 20% ​​of test set samples that were not used in the entire training process, the overall dataset samples, and the samples from the new external dataset were used as test sets, and the classification prediction performance was evaluated using metrics such as accuracy evaluation reports and confusion matrices.

[0131] Test set 1: Original data table.

[0132] Using test set 1 as the test set, the generated accuracy report is shown in Table 2 below, including precision, recall, F1 score, and support.

[0133] Table 2: Accuracy Report for Test Set 1

[0134] Accuracy Recall rate F1 score Support Sedimentary rock type 0.99 0.99 0.99 354 hydrothermal type 1.00 1.00 1.00 2407 accuracy 1.00 2761 Macro F1 1.00 1.00 1.00 2761 Weighted F1 1.00 1.00 1.00 2761

[0135] like Figure 8As shown, the 2761 samples from test set 1 were input into the trained random forest model for prediction and discrimination. The results showed that approximately 99% of the sedimentary rock types in the 354 sedimentary rock samples were successfully predicted, with only 2 misidentified. Similarly, only 2 out of the 2407 hydrothermal rock samples were confused. All accuracy metrics scored 1, indicating that the model achieved almost completely accurate predictions for the samples in test set 1. However, since the training data was derived from the original dataset, the high prediction results are to be expected; therefore, these results should only be considered as a reference.

[0136] Test set 2: To further verify the model's performance, the remaining 20% ​​of the original dataset that was not used in the model training was selected as the test set.

[0137] Using test set 2 as the test set, the generated accuracy report is shown in Table 3 below, including precision, recall, F1 score, and support.

[0138] Table 3: Accuracy Report for Test Set 2

[0139] Accuracy Recall rate F1 score Support Sedimentary rock type 0.97 0.97 0.97 71 hydrothermal type 1.00 1.00 1.00 482 accuracy 0.99 553 Macro F1 0.98 0.98 0.98 553 Weighted F1 0.99 0.99 0.99 553

[0140] like Figure 9 As shown, the 553 samples from test set 2 were input into the trained random forest model for prediction and discrimination. The results showed that approximately 97% of the 71 sedimentary rock samples were successfully predicted, with only 2 misidentified. Similarly, only 2 out of 482 hydrothermal samples were incorrectly identified. Due to the imbalance between positive and negative samples, we can focus on the F1 score among the various indicators. The F1 score for sedimentary rock types is approximately 0.97, while the score for hydrothermal types is approximately 1. The overall weighted F1 score is 0.99, indicating that the model can provide relatively accurate predictions and discriminations for the input samples.

[0141] Test set 3: New external dataset.

[0142] Using test set 3 as the test set, the generated accuracy report is shown in Table 4 below, including precision, recall, F1 score, and support.

[0143] Table 4: Accuracy Report for Test Set 3

[0144] Accuracy Recall rate F1 score Support hydrothermal type 1.00 1.00 1.00 288 accuracy 1.00 288 Macro F1 1.00 1.00 1.00 288 Weighted F1 1.00 1.00 1.00 288

[0145] Twelve trace element data from 288 newly collected chlorite samples from the Luobugangmu lead-zinc mine and the Cima Banshuo copper mine in Tibet were combined to obtain six features: Sn, Mn+Zn, Na / Co, Y+V, Sr+Mg, and Li / K. These features were then input into a trained random forest model for prediction. The results predicted that the deposits in both regions were hydrothermal. Field investigations confirmed that the Luobugangmu deposit was an epithermal lead-zinc deposit and the Cima Banshuo deposit was a porphyry copper deposit. The fact that both deposits are hydrothermal demonstrates the reliability of the prediction results.

[0146] 7. Visualization of RF Models

[0147] a. Visualization of 6D classification decision boundary

[0148] The random forest model can make comprehensive decisions and give prediction results based on six input features. However, the specific decision-making process cannot be fully demonstrated due to the complexity of the model itself and the high dimensionality of the input features.

[0149] like Figure 10 As shown, based on the importance of each feature in the random forest model, the features with the largest contribution weights to the model—Sr+Mg, Mn+Zn, Y+V, Sn, Na / Co, and Li / K—are selected in order. These features are then paired and used as the x and y axes to obtain 15 spatial cross-sections. The decision boundaries on these cross-sections are then visualized by adjusting certain parameters. Figure 11 As shown.

[0150] Since a 6-dimensional random forest model cannot accurately represent the overall high-dimensional space, projecting the decision boundary on the cross-section onto a two-dimensional plane is only an angle. In addition, we can combine three-dimensional and two-dimensional low-dimensional scatter plots to further understand and explore the model.

[0151] b. Scatter plots of 3D and 2D raw data

[0152] Using Python, we can directly visualize the best combinations of two-dimensional and three-dimensional feature elements for differentiating mineral deposit types from the original dataset. Based on the six optimal feature elements selected—Sr+Mg, Mn+Zn, Y+V, Sn, Na / Co, and Li / K—15 unique two-dimensional scatter plots were created (e.g., Figure 12 As shown in the figure, it corresponds exactly to the model cross-sectional decision diagram mentioned above. By comparing the actual projections of all the data, we can further verify that the model's classification decision is indeed based on the data itself.

[0153] like Figure 13As shown, by projecting points in space using three unique feature indicators as three coordinate axes, we can observe and find the three-dimensional feature indicators with optimal discrimination performance. This also serves as an important and intuitive reference tool, mutually validating the selected feature elements and the classifier's classification decision process in the established high-dimensional classifier model, thereby enhancing our deeper understanding of the model's classification decision process. Although high-dimensional machine learning models cannot yet be visualized, we can explore certain classification aspects from multiple perspectives, thereby improving our understanding of the model and the intrinsic relationship between the model and the input data.

[0154] This invention applies machine learning methods to the geological field. Based on specific algorithms, its ability to classify, extract, and mine massive amounts of geological data demonstrates a powerful advantage in mineralization prediction, opening up a new perspective for geological exploration combined with intelligent technologies. Furthermore, machine learning algorithms can extract subtle information from geoscientific data that is not easily detected directly. This information may provide valuable guidance on the type, type, and geological background of mineral resources, which is of great significance for mineral resource prediction and evaluation.

[0155] In the above-described method embodiments, Python can be used for data mining, which can integrate and reuse the massive amounts of geoscientific data accumulated by predecessors, uncovering new and useful information and subtle information that is not easily discovered directly. Furthermore, the model created using the method of this invention can predict new target information using unknown data, providing valuable guidance for the rapid identification of mineral deposit types and holding significant importance for mineral resource prediction and evaluation.

[0156] Example 4

[0157] The system for predicting ore deposit types based on multiple algorithms in this invention is mainly used to implement the above-described method embodiments. The system includes:

[0158] The data collection module is used to collect trace element data of a certain mineral and corresponding ore deposit genetic type data from a large amount of literature, and to standardize the trace element data to obtain experimental data tables.

[0159] The element combination selection module is used to analyze the correlation between various trace elements in the experimental data table, and to recombine the elements to obtain the characteristics of elements or combinations of elements that are not correlated, and update the experimental data table.

[0160] The model selection and validation module is used to divide the element data in the experimental data table into training set and test set according to a certain ratio, input the training set into various algorithm models for training, and select the best model as the random forest model through cross-validation.

[0161] The element feature selection module is used to filter the element features in the updated experimental data table based on the random forest model using the recursive feature elimination method, select the optimal set of element features, and update the random forest model.

[0162] The model training module is used to retrain the updated random forest model to obtain the optimal random forest model based on the prediction of the genetic type of mineral deposits using trace elements.

[0163] The prediction module is used to select the trace element characteristics of the mineral to be tested based on the optimal element feature set, and input them into the optimal random forest model to obtain the ore deposit genetic type of the mineral to be tested.

[0164] Furthermore, the element combination selection module specifically includes the following sub-modules:

[0165] Submodule 1: Used to calculate the correlation between various trace elements in the experimental data table, filter out elements with significant linear correlation, add them in pairs to obtain combined elements, replace the original two elements, and update the experimental data table;

[0166] Submodule 2: Based on the box plots of the new elements in different types of mineral deposit genesis, divide the elements with overlapping values ​​by the elements that show opposite indications and significant differences in numerical ranges in different types of mineral deposit genesis to obtain new combined elements, or accumulate the elements that show consistent indications and significant differences in numerical ranges to obtain new combined features, and update the experimental data table.

[0167] Submodule 3: This module is used to remove elements and overlapping combinations that show consistent patterns in box plots of different types of mineral deposit genesis, and further eliminate combinations with significant linear correlations to obtain the final experimental data table of elements or combinations that are not correlated.

[0168] The specific implementation details of each submodule are shown in the method implementation examples above, and will not be repeated here.

[0169] Specifically, before model training, data can be transformed, such as by logarithmic transformation (log), to make the data distribution closer to a normal distribution. This allows models sensitive to data distribution to achieve better scores, enabling more objective evaluation and selection of models. See the method implementation examples above for details, which will not be repeated here.

[0170] Furthermore, the standardization process for trace element data includes replacing element data identifiers, uniformly converting element data types to floating-point types, calculating the percentage of missing values ​​for each element, and removing elements with data volumes below a certain value and duplicate sample data. See the method implementation examples above for details, which will not be repeated here.

[0171] The standardization of trace element data includes imputation of missing values, specifically in three ways: (1) imputing missing values ​​in columns with the mean and median respectively; (2) imputing using linear interpolation; and (3) imputing using the KNN algorithm model. After imputation, the reliability of the data can be verified by using quantile-quantile plots. See the above method implementation examples for details, which will not be repeated here.

[0172] Furthermore, the system for predicting ore deposit types based on multiple algorithms also includes an accuracy evaluation module, used to evaluate the accuracy of the trained random forest model. Three test sets are used: the remaining untrained samples, the overall dataset, and the external dataset. Classification performance is predicted based on the accuracy evaluation report and confusion matrix of the test sets. See the method implementation examples above for details, which will not be repeated here.

[0173] Furthermore, the system for predicting ore deposit types based on multiple algorithms also includes a decision boundary visualization module. This module selects the features that contribute the most to the model based on the importance of each element feature in the optimal feature set of the optimal random forest model, and pairs them together as the x and y axes to obtain multiple spatial cross-sections. By adjusting certain parameters, the decision boundary on the spatial cross-sections is visualized. See the method implementation examples above for details, which will not be repeated here.

[0174] Example 5

[0175] This invention also provides a computer-readable storage medium, such as flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, disk, optical disk, server, app store, etc., which stores a computer program that performs corresponding functions when executed by a processor. The computer-readable storage medium of this embodiment is used to implement the mineral deposit type prediction method based on multiple algorithms of the method embodiment when executed by a processor.

[0176] In summary, this invention, through data mining, integrates and reuses massive amounts of accumulated geoscientific data to uncover new, useful information and subtle information that is not easily discovered directly. The random forest model of this invention can predict new target information using unknown data, providing valuable guidance for the rapid identification of mineral deposit types and holding significant importance for mineral resource prediction and evaluation. Furthermore, by visualizing the decision boundaries of various aspects of the model, we can intuitively see the classification performance of the data, developing a new method for revealing new information based on mineral data.

[0177] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for predicting ore deposit types based on multiple algorithms, characterized in that, Includes the following steps: S1. Collect trace element data of chlorite and corresponding ore deposit genetic types from a large amount of literature, and standardize the trace element data to obtain experimental data tables; specifically, it includes 21 elements, Mn, Na, Ca, K, Ti, Li, B, V, Cr, Co, Ni, Cu, Zn, As, Sr, Y, Zr, Sn, Ba, Pb, and Mg, and their corresponding ore deposit genetic types; the ore deposit genetic types include hydrothermal and sedimentary rock types; S2. Analyze the correlation between the trace elements in the experimental data table, and recombine the elements to obtain the characteristics of elements or combinations of elements that are not correlated, and update the experimental data table. S3. Divide the element data in the experimental data table into training set and test set according to a certain ratio. Input the training set into various algorithm models for training. Select the model with the best classification effect through cross-validation scoring. This model is the random forest model. S4. Based on the random forest model, the recursive feature elimination method is used to filter the element features in the updated experimental data table, select the optimal element feature set, and update the random forest model. S5. Retrain the updated random forest model to obtain the optimal random forest model based on the prediction of mineral deposit genetic types using trace elements. S6. Select the trace element characteristics of the chlorite to be tested based on the optimal element feature set, and input them into the optimal random forest model to obtain the ore deposit genetic type of the chlorite to be tested. Specifically, step S2 is as follows: S21. Calculate the correlation between each trace element in the experimental data table, screen out the elements Mn and Zn with significant linear correlation, add the two elements together, and create a new feature Mn+Zn to replace the Mn and Zn elements. S22. Based on the box plots of the new elements in different types of ore deposit genesis, the non-overlapping elements include 6 elements or combinations: Li, V, Co, Sn, Mg, and Zn+Mn. The overlapping elements include 5 elements: Na, K, B, Sr, and Y. The overlapping elements are divided pairwise by the elements whose general indicative patterns are opposite and whose numerical ranges differ significantly, resulting in new characteristic combinations: Na / Li, Na / V, Na / Co, Na / Mg, Na / Zn+Mn, K / Li, K / V, K / Co, K / Mg, and K / Zn+Mn. The overlapping elements are summed with the elements whose numerical ranges differ significantly, resulting in Sr+Zn+Mn, Sr+Mg, Y+Zn+Mn, Y+V, Y+Co, Y+Mg, and Y+Zn+Mn. S23. Delete elements with consistent indication patterns and overlapping combinations from the box plots of different types of mineral deposit genesis, and then remove combinations with significant linear correlation to obtain the final experimental data table of elements or combinations without correlation. Specifically, remove combinations with significant linear correlations that have a correlation coefficient greater than or equal to 5, and finally obtain 6 features: Sn, Mn+Zn, Na / Co, Y+V, Sr+Mg, and Li / K. Test that the correlation between the 6 features is close to 0, then there are no collinear features.

2. The mineral deposit type prediction method based on multiple algorithms according to claim 1, characterized in that, Before model training in step S3, the data in the training set is mathematically transformed to make the data distribution closer to a normal distribution.

3. The mineral deposit type prediction method based on multiple algorithms according to claim 1, characterized in that, In step S1, the standardization process of trace element data includes replacing the element data with identifiers, uniformly converting the element data type to floating point type, calculating the proportion of missing values ​​for each element data, and removing elements with data volume below a certain value and duplicate sample data.

4. The mineral deposit type prediction method based on multiple algorithms according to claim 1, characterized in that, In step S1, the normalization of trace element data includes filling missing values ​​in the element data, specifically including the following three methods: (1) filling the values ​​with missing columns with the mean and median respectively; (2) filling by linear interpolation; (3) filling by KNN algorithm model.

5. The mineral deposit type prediction method based on multiple algorithms according to claim 4, characterized in that, After imputing missing values, the reliability of the data is verified by using quantile-quantile plots.

6. The mineral deposit type prediction method based on multiple algorithms according to claim 1, characterized in that, Step S5 also includes evaluating the accuracy of the trained random forest model by using the remaining untrained samples, the overall dataset, and the external dataset as three test sets, and predicting the classification performance based on the accuracy evaluation report and confusion matrix of the test sets.

7. The mineral deposit type prediction method based on multiple algorithms according to claim 1, characterized in that, Based on the importance of each element feature in the optimal element feature set of the optimal random forest model, the features that contribute the most to the model are selected in turn, and they are paired up to serve as the x and y axes respectively, resulting in multiple spatial cross-sections. The decision boundaries on the spatial cross-sections are then visualized by adjusting certain parameters.

8. A system for predicting ore deposit types based on multiple algorithms, characterized in that, include: The data collection module is used to collect trace element data of chlorite and corresponding ore deposit genetic types from a large amount of literature, and to standardize the trace element data to obtain experimental data tables. Specifically, it includes data on 21 elements, Mn, Na, Ca, K, Ti, Li, B, V, Cr, Co, Ni, Cu, Zn, As, Sr, Y, Zr, Sn, Ba, Pb, and Mg, and their corresponding ore deposit genetic types, including hydrothermal and sedimentary types. The element combination selection module is used to analyze the correlation between various trace elements in the experimental data table, and to recombine the elements to obtain the characteristics of elements or combinations of elements that are not correlated, and update the experimental data table. The model selection and validation module is used to divide the element data in the experimental data table into training set and test set according to a certain ratio, input the training set into various algorithm models for training, and select the model with the best classification effect through cross-validation scoring. This model is the random forest model. The element feature selection module is used to filter the element features in the updated experimental data table based on the random forest model using the recursive feature elimination method, select the optimal set of element features, and update the random forest model. The model training module is used to retrain the updated random forest model to obtain the optimal random forest model based on the prediction of the genetic type of mineral deposits using trace elements. The prediction module is used to select the trace element characteristics of the chlorite to be tested based on the optimal element feature set, and input them into the optimal random forest model to obtain the ore deposit genetic type of the chlorite to be tested. The element combination selection module is specifically used for: calculating the correlation between various trace elements in the experimental data table, screening out elements Mn and Zn with significant linear correlation, adding the two elements to create a new feature Mn+Zn to replace Mn and Zn elements; based on the box plots of the new elements in different types of mineral deposit genesis, obtaining non-overlapping elements including Li, V, Co, Sn, Mg, and Zn+Mn (6 elements or elements), and overlapping elements including Na, K, B, Sr, and Y (5 elements); dividing the overlapping elements pairwise with elements whose approximate indicative patterns are opposite and whose numerical ranges differ significantly to obtain new feature combinations Na / Li, Na / V, Na / Co, Na / Mg, Na / Zn+Mn, K / Li, K / V, K / Co, etc. K / Mg and K / Zn+Mn are used to sum elements with overlapping values ​​and those with significantly different numerical ranges to obtain Sr+Zn+Mn, Sr+Mg, Y+Zn+Mn, Y+V, Y+Co, Y+Mg, and Y+Zn+Mn. Elements with consistent indication patterns and overlapping combinations in box plots of different ore deposit genetic types are removed, and combinations with significant linear correlations are further eliminated to obtain the final experimental data table of elements or combinations without correlation. Specifically, combinations with significant linear correlations and correlation coefficients greater than or equal to 5 are removed, resulting in six features: Sn, Mn+Zn, Na / Co, Y+V, Sr+Mg, and Li / K. The correlation among these six features is close to 0, indicating the absence of collinearity.

9. A computer storage medium, characterized in that, It contains a computer program that can be executed by a processor, which performs the mineral deposit type prediction method based on multiple algorithms as described in any one of claims 1-7.

Images