Machine Learning-Based Methods and Systems for Solid Mineral Prediction

Abstract

This invention relates to the field of artificial intelligence technology, specifically to a method and system for predicting solid mineral resources based on machine learning. The method includes the following steps: acquiring fault line data and mineralization anomaly data of the target area; constructing a raster prediction unit and dividing the fault segments; calculating the fault line saliency, low mineralization response value of the fault zone, and misalignment recurrence value of mineralization anomalies to generate late-stage mineralization structural marker values; generating uncorrected mineralization probabilities using a machine learning model; using the fault segments as reinforcement learning decision objects, outputting the fault zone probability suppression coefficient and misalignment anomaly recurrence coefficient, and correcting the uncorrected mineralization probabilities in logarithmic probability space to obtain the final mineralization probability. This invention can reduce prediction bias caused by late-stage mineralization faults being misidentified as ore-controlling structures, and improve the reliability of solid mineralization target area delineation in multi-stage tectonic superposition areas.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and specifically to a method and system for predicting solid mineral resources based on machine learning. Background Technology

[0002] Solid mineral prediction is a crucial step in mineral resource exploration. It typically requires a comprehensive assessment of the mineralization potential of a target area, taking into account factors such as fault structures, lithology, geochemical anomalies, geophysical anomalies, remote sensing interpretation results, and the distribution of known mineral deposits. With the development of artificial intelligence technology, machine learning-based solid mineral prediction methods are increasingly being applied to the evaluation of prospective mineralization areas and the delineation of target areas. These methods usually divide the target area into multiple prediction units and convert data such as fault distance, fault density, and mineralization anomaly intensity into model inputs. The machine learning model then learns the mineralization probability distribution based on known mineral deposit samples and non-mineralized samples.

[0003] However, in regions with multiple superimposed tectonic phases, not all fault structures have the same mineralization significance. Early faults may act as ore-guiding, ore-hosting, or ore-controlling structures, playing a positive role in the formation of ore bodies. Later faults, on the other hand, may cause dislocation, displacement, or destruction of existing ore bodies, alteration zones, or mineralization anomaly zones. Since later-stage ore-breaking faults may also exhibit strong continuity, obvious linear response, and large extension length in planar structural data, conventional machine learning models are prone to mistakenly treating these faults as ore-controlling structures for positive learning, leading to an erroneous increase in the probability of mineralization near later-stage ore-breaking fault zones. Furthermore, mineralization anomalies dislocated by later-stage ore-breaking faults may still have a staggered recurrence relationship on both sides of the fault, but conventional models find it difficult to use this staggered recurrence relationship to restore the potential mineralization continuity, thus causing misjudgment of the target area direction for solid mineral prediction. Summary of the Invention

[0004] The purpose of this invention is to address the problems in the background technology where mid-to-late stage mineralization faults in multi-stage tectonic superposition regions have high linear significance on the plane, are easily mistaken by machine learning models as ore-controlling structures for positive learning, resulting in high erroneous mineralization probability near the fault zone and the disruption of the continuous relationship of misaligned mineralization anomalies on both sides of the fault. The invention proposes a solid mineral prediction method and system based on machine learning.

[0005] The technical solution of this invention: a solid mineral prediction method based on machine learning, comprising: Acquire fault line data and mineralization anomaly data of the target area, divide the target area into multiple raster prediction units, and divide the fault line data into multiple fault segments; For each fracture segment, the saliency of the fracture line is calculated based on the continuous length and linear response intensity of the fracture segment, and the low response value of mineralization of the fracture zone is calculated based on the mineralization anomaly intensity in the corresponding fracture zone grid set. For each fault segment, the displacement correlation calculation is performed on the mineralization anomaly profiles on both sides of the fault along the extension direction of the fault segment to determine the mineralization anomaly misalignment recurrence value. Based on the significance of the fracture line, the low response value of mineralization in the fracture zone, and the recurrence value of mineralization anomalies, the later mineralization structural marker values ​​of the fracture segment are generated. Using the mineralization anomaly intensity and basic tectonic response values ​​of the raster prediction unit as input, the uncorrected mineralization probability of each raster prediction unit is generated through a machine learning model. Using the fault segment as the decision object of reinforcement learning, the state including the later mineralization structure marker value is input into the reinforcement learning policy network to obtain the fault zone probability suppression coefficient and the misalignment anomaly reconnection coefficient. Based on the later-stage mineralization structural marker values, fault zone probability suppression coefficients, and misalignment anomaly reconnection coefficients, the uncorrected mineralization probabilities are corrected to generate the final mineralization probabilities for each grid prediction unit. The target area for solid mineralization prediction is then determined based on the final mineralization probabilities.

[0006] Preferably, the mineralization anomaly intensity is obtained by normalizing the mineralization anomaly data; when the mineralization anomaly data includes multiple mineralization anomaly indicators, the multiple mineralization anomaly indicators are normalized respectively, and the fusion weight is determined according to the response consistency between each mineralization anomaly indicator and the known mineral point sample label, so as to obtain the comprehensive mineralization anomaly intensity of each grid prediction unit.

[0007] Preferably, the significance of the fracture line is obtained by normalizing the product of the continuous length of the fracture segment and the linear response intensity of the fracture segment; the low response value of mineralization in the fracture zone is obtained by normalizing and reversing the mean value of mineralization anomaly intensity in the corresponding fracture zone grid set for each fracture segment.

[0008] Preferably, for each fault segment, the mineralization anomaly profiles on both sides of the fault are subjected to displacement correlation calculations along the extension direction of the fault segment to determine the mineralization anomaly recurrence value, including: Using the extension direction of the target fracture segment as the tangential direction and the direction perpendicular to the target fracture segment as the normal direction, the mineralization anomaly intensity of adjacent grids on both sides of the target fracture segment is extracted to form a first side anomaly profile and a second side anomaly profile. The correlation calculations under non-zero displacement are performed on the first side anomaly profile and the second side anomaly profile along the tangential direction to obtain a set of non-zero displacement correlation values; The effective misalignment reproduction conditions are determined based on the distribution statistics of the non-zero misalignment correlation value set, and the non-zero misalignment amount that satisfies the effective misalignment reproduction conditions and has the largest correlation value is determined as the optimal misalignment reproduction displacement. Based on the optimal staggered reproduction displacement, establish the correspondence between the grid prediction units on both sides of the fracture. The mineralization anomaly misalignment value is determined based on the correlation value corresponding to the optimal misalignment recurrence displacement and the ratio of the optimal misalignment recurrence displacement to the continuous length of the fracture segment.

[0009] Preferably, the valid misalignment reproduction conditions are determined based on the distribution statistics of the set of non-zero misalignment correlation values, including: Calculate the central tendency and dispersion values ​​of the set of non-zero shift correlation values; Non-zero shift correlation values ​​that are higher than the central tendency value and exceed the dispersion value constraint are determined as valid correlation values. The non-zero shift quantity with the aforementioned effective correlation value is determined as the non-zero shift quantity that satisfies the effective misalignment reproduction condition.

[0010] Preferably, the late-stage mineralization structural marker value is obtained by multiplying the fracture line significance, the low mineralization response value of the fracture zone, and the mineralization anomaly misalignment recurrence value.

[0011] Preferably, the basic structural response value is obtained by normalizing and back-calculating based on the distance from the grid prediction unit to the nearest fault segment, and the machine learning model includes at least one of the following: random forest model, gradient boosting tree model, support vector machine model, or neural network model that can output the probability of mineralization.

[0012] Preferably, the state input to the reinforcement learning policy network includes: the late-stage mineralization structure flag value; the ratio of the optimal misaligned recurrence displacement to the continuous length of the fault segment; the average uncorrected mineralization probability within the fault zone grid set; and the average difference in uncorrected mineralization probability between corresponding grid prediction units on both sides of the fault.

[0013] Preferably, correcting the uncorrected mineralization probability includes: When the target grid prediction unit is located within the fault zone grid set corresponding to the target fault segment, the uncorrected mineralization probability of the target grid prediction unit is suppressed according to the later mineralization structure indicator value of the target fault segment and the fault zone probability suppression coefficient. When the target grid prediction unit belongs to the grid prediction unit with a misaligned recurrence correspondence on both sides of the target fault segment, the uncorrected mineralization probability of the target grid prediction unit is adjusted according to the later mineralization structure indicator value of the target fault segment, the misaligned anomaly reconnection coefficient, and the difference in uncorrected mineralization probability between the corresponding grid prediction unit and the target grid prediction unit. When correcting the uncorrected mineralization probability, the correction amount of one or more fault segments to the same grid prediction unit is superimposed in the logarithmic probability space, and the final mineralization probability is obtained by transformation through the Sigmoid function.

[0014] Compared with the prior art, the above-mentioned technical solution of the present invention has the following beneficial technical effects: This invention generates late-stage mineralization structural marker values ​​by combining fracture line saliency, low mineralization response value of fracture zone, and mineralization anomaly misalignment recurrence value. It can identify late-stage fractures that are linearly significant on the plane but whose mineralization significance is dislocation, displacement, or destruction from ordinary fracture structural evidence, avoiding machine learning models from learning all significant fractures as favorable factors for mineralization. By using the mineralization anomaly misalignment recurrence relationship on both sides of the fracture to establish the correspondence between grid prediction units on both sides of the fracture, the continuity of mineralization anomalies after being dislocated by late-stage fractures can participate in the mineralization probability correction, reducing the problem of erroneous extension of the prediction target area along the direction of late-stage mineralization fracture. This invention uses fault segments as the decision-making objects in reinforcement learning. The reinforcement learning strategy network adaptively outputs the fault zone probability suppression coefficient and the misalignment anomaly reconnection coefficient based on the later mineralization structural marker values, normalized misalignment recurrence displacement, the mean probability of mineralization without correction of the fault zone, and the probability difference between the corresponding grids on both sides. This avoids insufficient or excessive correction caused by manually fixing the correction intensity. The correction amount of different fault segments on the same grid prediction unit is superimposed in the log probability space, and the final mineralization probability is obtained by transforming it through the Sigmoid function. This can improve the stability of solid mineral prediction results in multi-stage tectonic superposition areas while maintaining the validity of probability values. Attached Figure Description

[0015] Figure 1 This is a flowchart of the machine learning-based solid mineral prediction method proposed in this invention; Figure 2 This is a block diagram of the machine learning-based solid mineral prediction system proposed in this invention. Figure 3 This is a schematic diagram of the late-stage ore-breaking structure markers proposed in this invention; Figure 4 This is a schematic diagram illustrating the calculation of the mineralization anomaly staggered recurrence relationship proposed in this invention; Figure 5 This is a schematic diagram of the final mineralization probability correction result proposed in this invention. Detailed Implementation

[0016] Example 1, as Figure 1 As shown, the solid mineral prediction method based on machine learning proposed in this invention is used to predict the mineralization probability of solid minerals in areas with superimposed tectonic phases. This method can identify late-stage mineralization structures with significant fault lines on the plane but low mineralization response of the fault zone itself and a staggered recurrence relationship of mineralization anomalies on both sides of the fault, and correct the machine learning prediction results based on the late-stage mineralization structures. In this embodiment, the target area is a multi-phase tectonic superposition mineralization area; the input data includes fault line data, mineralization anomaly data, known mineral occurrence samples and non-mineralization samples of the target area; wherein, the fault line data can be derived from geological structure interpretation map, remote sensing linear structure interpretation map, geophysical linear anomaly interpretation map or fault interpretation confidence map; the mineralization anomaly data can be single element anomaly, combined element anomaly, alteration anomaly or comprehensive anomaly related to the target mineral type; In this embodiment, the non-mineralized samples can be composed of grid prediction units corresponding to boreholes, trenches, or engineering verification locations where mineralization has not been found, as confirmed by existing exploration data. When engineering verification locations are lacking, non-mineralized samples can be selected from unverified areas using a data-adaptive approach. Specifically, the distance from each candidate grid prediction unit to the nearest known mineral point is calculated, and the distance distribution of all candidate grid prediction units is calculated. Grid prediction units whose distance is higher than the median level of this distance distribution, whose mineralization anomaly intensity is lower than the median level of the mineralization anomaly intensity distribution, and which do not intersect with known ore-controlling structural zones are selected as candidate non-mineralized samples. For candidate non-mineralized samples, a spatially uniform sampling method is used to select them, ensuring that the selected non-mineralized samples are distributed dispersedly within the target area, avoiding the concentration of non-mineralized samples in a single geological background area. This avoids directly using areas that have not yet been found to be mineralized but may still have mineralization potential as non-mineralized samples.

[0017] Establish a unified spatial grid; divide the target area into multiple raster prediction units according to a spatial scale adapted to the spatial accuracy of the fault line data and mineralization anomaly data, denoted as: ; in, Represents the set of raster prediction cells. The j-th raster prediction unit is represented by n, which represents the number of raster prediction units. In this embodiment, the raster prediction unit refers to the smallest prediction and evaluation unit formed after dividing the target area according to a uniform spatial scale. Each raster prediction unit corresponds to a mineralization anomaly intensity, basic structural response value, uncorrected mineralization probability, and final mineralization probability. The fracture lines within the target area are divided into multiple fracture segments based on geometric continuity, consistency of extension direction, or consistency of interpretation properties, denoted as: ; in, Represents a set of fracture segments. Let m represent the i-th fracture segment and m represent the number of fracture segments. In this embodiment, the fracture segment refers to a linear structural unit obtained by segmenting the fracture line within the target area according to the geometric continuity, consistency of extension direction, or consistency of interpretation properties of the fracture line.

[0018] For each raster prediction unit Obtain its original mineralization anomaly value The mineralization anomaly intensity was obtained by normalization. : ; in, This represents the mineralization anomaly intensity of the j-th raster prediction cell. This represents the original mineralization anomaly value of the j-th raster prediction cell. This represents the minimum value of the original mineralization anomaly within the target area. Indicates the maximum value of the original mineralization anomaly within the target area; when When the corresponding normalization result is set to 0, it indicates that the mineralization anomaly index does not participate in the differential evaluation within the target area. When mineralization anomaly data includes multiple mineralization anomaly indicators, each indicator is first normalized to obtain the normalized value of the q-th mineralization anomaly indicator in the j-th raster prediction cell. Then, the fusion weights are determined based on the consistency of the responses between each mineralization anomaly index and the known mineral deposit sample labels. : ; in, Indicates the number of mineralization anomaly indicators. Let represent the value sequence of the q-th mineralization anomaly index in the training samples. Represents the training sample label sequence. This represents the correlation between the q-th mineralization anomaly index and the training sample label. This represents the fusion weight of the q-th mineralization anomaly index. When the sum of the absolute values ​​of the correlations between each mineralization anomaly index and the training sample label is 0, the fusion weights of each mineralization anomaly index are set to the same value, so that they are only used as a component for balancing the overall mineralization anomaly intensity. In the case of multiple indicators, the mineralization anomaly intensity of the j-th raster prediction unit Represented as: ; Obtain the mineralization anomaly intensity for each raster prediction cell. ; For each fracture segment Extract the corresponding fault zone raster set. In this embodiment, the fault zone grid set refers to a set of grid prediction units that have a spatial proximity to a certain fault segment and are used to characterize the mineralization response near the fault segment; the set is composed of grid prediction units that intersect with the fault segment, or grid prediction units located within the neighborhood of the fault segment, and the neighborhood range is determined according to the grid scale and the spatial accuracy of the fault line data. In one implementation, it will be related to the fracture segment. Intersecting raster prediction cells form a set of fault zone raster cells. : ; in, This represents the set of fault zones corresponding to the i-th fault segment; when a certain fault segment Corresponding fault zone grid set If the value is empty, the fault zone grid set is redefined after expanding the fault segment by one grid width according to the grid scale; if it is still empty after expansion, the fault segment will not participate in the subsequent calculation of the later mineralization structure marker value. Calculate the significance of the fracture line For the i-th fracture segment Get its continuous length and linear response strength In this embodiment, the linear response intensity refers to the degree of linear response exhibited by the fault segment in structural interpretation data, remote sensing linear structural data, geophysical linear anomaly data, or fault interpretation confidence data; the fault line significance refers to the comprehensive quantitative result of the continuous length of the fault segment and the linear response intensity, used to characterize the significance of the fault segment in the planar structural data. When fracture line data has fields for fracture interpretation confidence, linear tectonic response value, or geophysical linear anomaly intensity, the fracture segment... The average value of the corresponding field at each sampling point is used as the linear response strength. When the fault line data does not have the above fields, but the target area has a remote sensing linear enhancement map, a geophysical linear anomaly map, or a constructed linear response raster map, along the fault segment... Extract the corresponding raster values ​​and use the average of the extracted values ​​as the linear response intensity. When only vector fracture line data is available and no response intensity field or response raster is available, the linear response intensity of each fracture segment will be used. The significance of the break line is uniformly set to 1. Continuous length of the fractured segment The dominant position is determined; make: ; The significance of the fracture line Represented as: ; in, Represents the continuous length of the i-th fracture segment With linear response strength The product of Indicates all fracture segments corresponding to The minimum value in, Indicates all fracture segments corresponding to The maximum value in; when At that time, the salience of the fracture line of each fracture segment was determined. Set to 0 to indicate that the significance of the fracture line does not participate in the differential evaluation within the target area; Furthermore, the low response value of mineralization in the fault zone was calculated. In this embodiment, the low mineralization response value of the fault zone refers to the quantitative result of the degree of low mineralization anomaly intensity within the fault zone grid set corresponding to the fault segment; the higher the low mineralization response value of the fault zone, the weaker the mineralization anomaly response of the fault zone itself corresponding to the fault segment. Calculate the fault zone grid set Mean intensity of internal mineralization anomaly: ; in, This represents the set of fault zones corresponding to the i-th fault segment. The mean intensity of internal mineralization anomaly, Represents a set of fault zone grids The number of raster prediction cells in the data; Normalize and reverse-process the mean mineralization anomaly intensity of each fault segment to obtain the low response value of the fault zone mineralization. : ; in, Indicates all fracture segments corresponding to The minimum value in, Indicates all fracture segments corresponding to The maximum value in; when At that time, the low response value of the fracture zone mineralization of each fracture segment was determined. Set to 0 to indicate that low response values ​​of fault zone mineralization do not participate in differential evaluation within the target area.

[0019] Calculate the correlation of mineralization anomalies on both sides of the fault; in this embodiment, the correlation of mineralization anomalies refers to the phenomenon that the mineralization anomaly profiles on both sides of the fault segment exhibit a spatial correspondence after non-zero displacement along the fault extension direction; this phenomenon is used to characterize that the mineralization anomalies on both sides of the fault may originally belong to a continuous mineralization anomaly zone, but later formed a misaligned distribution due to fault displacement. For the fracture segment The direction of extension is taken as the tangential direction, and the direction perpendicular to the fracture segment is taken as the normal direction; multiple sampling positions s are set along the tangential direction at sampling intervals adapted to the grid size; at each sampling position s, sampling is performed on the fracture segment. The first and second sides are selected from the grid prediction cells adjacent to the fracture segment to form the anomaly profile on the first side. Second side abnormal profile ; in: ; In the formula, This represents the intensity of the mineralization anomaly on the first side of the i-th fracture segment at the tangential position s. This represents the intensity of the mineralization anomaly on the second side of the i-th fracture segment at the tangential position s. This indicates the grid number corresponding to the first lateral tangential position s. This indicates the grid number corresponding to the second lateral tangential position s; Multiple non-zero displacements are set along the fracture extension direction. For the first side abnormal profile Second side abnormal profile Perform shift-related calculations; shift amount The search range is based on the fracture segment The continuous length and the effective overlap length after the misalignment of the two abnormal profiles are determined; for each misalignment amount Only the first side anomalous profile is used. Second side abnormal profile Data with overlapping sampling positions after the shift are used in relevant calculations; The shift correlation calculation uses the following formula: ; in, Indicates the displacement of the i-th fracture segment. Correlation values ​​of mineralization anomaly profiles on both sides below. Indicates the amount of current shift involved. The mean value of the first side anomaly profile was calculated. Indicates the amount of current shift involved. The mean value of the second-side anomaly profile was calculated; if it is included in the current displacement. If the number of overlapping sampling locations calculated is insufficient to form an effective correlation calculation, or if the variance of the anomaly profile on either side is 0, then the shift amount... The corresponding values ​​are not involved in determining the optimal staggered recurrence displacement; To avoid forcibly establishing a misalignment relationship when there is no actual correspondence between the anomalies on both sides, this embodiment determines the effective misalignment reproduction conditions based on the distribution statistics of the non-zero misalignment correlation value set; specifically, for the same fracture segment... The set consists of all valid non-zero shift correlation values. Calculate the mean of the set. and standard deviation : ; in, This represents the set of non-zero shift correlation values ​​corresponding to the i-th fracture segment. This represents the mean of the set. This represents the standard deviation of the set; when the non-zero shifted correlation value set... When the number of valid correlation values ​​is less than two, the fracture segment is considered to be... There is no effective mineralization anomaly misalignment replication relationship, and the mineralization anomaly misalignment replication value is not included. Set to 0; In the set of nonzero shift correlation values When a distribution statistic can be formed, the non-zero shift quantity that meets the following conditions is determined as the non-zero shift quantity that meets the effective misalignment reproduction condition: ; and: ; in, Used to exclude negatively correlated shift results. Used to screen for shift results that are significantly higher than the overall correlation level from the distribution of non-zero shift correlation values ​​in the same fracture segment; when When the value is less than or equal to 0, the positive correlation condition is used as the basis for effective stagger reproduction, and the correlation value corresponding to the selected non-zero shift is still required to be the maximum correlation value among all positively correlated non-zero shifts in the fault segment; the effective stagger reproduction condition is adaptively determined by the distribution of non-zero shift correlation values ​​in the same fault segment. When a non-zero shift exists that satisfies the effective staggered reproduction condition, the non-zero shift with the largest correlation value is determined as the optimal staggered reproduction displacement. : ; in, This represents the optimal staggered recurrence displacement corresponding to the i-th fracture segment; if there is no displacement that simultaneously satisfies the positive correlation condition and If the non-zero displacement is considered, then the fracture segment is considered to be... There is no effective mineralization anomaly staggered relationship on both sides of this fault segment. Mineralization anomaly misalignment recurrence value Set it to 0, and define the fracture segment. The normalized staggered recurrence displacement is set to 0, and no correspondence is established between the grid prediction units on both sides of the fracture. Determining the optimal staggered reproduction displacement Then, the displacement is reproduced based on the optimal staggered arrangement. Establish the correspondence between the grid prediction units on both sides of the fracture; specifically, for the fracture segment... Any raster prediction unit on either side Take the raster prediction unit center point , center point Projected onto the fracture segment In the local coordinate system, the tangential coordinates along the fracture extension direction are obtained. Normal coordinates perpendicular to the fracture direction ; If the raster prediction unit Located in the fracture segment The first side, then according to and Determine the target location on the second side of the fracture; if the raster prediction unit Located in the fracture segment The second side, then according to and Determine the target location on the first side of the fracture; define the raster prediction cell containing the target location as... : ; in, This indicates the displacement of the i-th fracture segment in the optimal staggered reproduction. Under the action, the grid prediction cell number on the other side of the fracture corresponding to the j-th grid prediction cell is... This indicates the raster prediction cell number corresponding to the other side of the fracture. When the target's corresponding position does not fall into any valid raster prediction cell, it is considered that there is no correspondence between the raster prediction cells and they are not included in the set of raster prediction cells with corresponding relationships on both sides of the fracture. .

[0020] Calculate the recurrence value of mineralization anomaly misalignment based on the maximum correlation value and displacement ratio. : ; in, This represents the recurrence value of mineralization anomalies in the i-th fault segment. Represents the optimal staggered recurrence displacement The absolute value, Indicates the fracture segment The continuous length; The larger the value, the more likely the mineralization anomalies on both sides of the fault are to re-correspond after displacement, indicating that the fault segment is more likely to exhibit later-stage displacement mineralization anomalies.

[0021] Calculate the structural indices of late-stage ore breaking In this embodiment, the late-stage mineralization structural marker value refers to a comprehensive index determined by the significance of the fracture line, the low mineralization response value of the fracture zone, and the misalignment and recurrence value of mineralization anomalies. This index is used to identify fracture segments with significant fracture lines on the plane, low mineralization response of the fracture zone itself, and misalignment and recurrence relationships of mineralization anomalies on both sides of the fracture. Later ore-breaking structural marker values Represented as: ; in, This represents the later-stage mineralization structural indicator value of the i-th fault segment. This indicates the significance of the fracture line in the i-th fracture segment. This indicates the low response value of the fracture zone mineralization in the i-th fracture segment. This represents the mineralization anomaly recurrence value of the i-th fault segment; when there is no non-zero shift that satisfies the valid recurrence condition, due to =0, The corresponding value is 0, thus avoiding the fault segment from being identified as a later ore-breaking structure due to accidental correlation.

[0022] Establish an initial machine learning mineral prediction model; for each grid prediction unit Constructing basic mineralization evidence vectors : ; in, This represents the basic mineralization evidence vector for the j-th raster prediction unit. This represents the mineralization anomaly intensity of the j-th raster prediction cell. This represents the basic structural response value of the j-th grid prediction unit. In this embodiment, the basic structural response value refers to the structural influence obtained based on the spatial relationship between the grid prediction unit and the fault segment when ore-controlling faults and later ore-breaking faults are not distinguished. Basic structural response value According to raster prediction unit Distance to the nearest fracture segment Perform normalized inverse calculation: ; in, This represents the distance from the j-th raster prediction cell to the nearest break segment. This represents the minimum distance among all raster prediction cells to the nearest break segment. Represents the maximum value among all raster prediction cells' distances to the nearest break segment; when At that time, the basic structural response value will be... Set to 0 to indicate that the basic structural response value does not participate in the differential evaluation within the target area; Train a machine learning model using known mineral deposit samples and unmineralized samples. Output each raster prediction unit Uncorrected mineralization probability : ; in, This represents the uncorrected mineralization probability of the j-th raster prediction cell. Represents a machine learning model. The parameters of the machine learning model are represented here. In this embodiment, the uncorrected mineralization probability refers to the mineralization probability output by the machine learning model based on the mineralization anomaly intensity and the basic tectonic response value, which has not yet been corrected according to the later mineralization breaking tectonic indicators. Machine learning models Random forest, gradient boosting tree, support vector machine, neural network, or other models capable of outputting mineralization probabilities can be used. The training objective employs cross-entropy loss to make the predicted probability of known mineralized samples approach 1 and the predicted probability of non-mineralized samples approach 0. ; in, Represents the training sample set, This represents the label of the j-th training sample. Indicates a mineralized sample. This indicates a non-mineralized sample.

[0023] Furthermore, a reinforcement learning policy network is used to determine the correction intensity of different fracture segments. In this embodiment, the reinforcement learning process is a single-step reinforcement learning decision-making process with the fracture segment as the decision object. Each fracture segment in the target area corresponds to an environmental state. The reinforcement learning policy network outputs probability suppression actions and misalignment reconnection actions for the environmental state. After the actions are executed, the reward value is obtained based on the loss predicted by the verification samples and the high probability penalty of the fracture zone error. The parameters of the reinforcement learning policy network are updated with the reward value. With each fracture segment As the decision-making object in reinforcement learning, constructing reinforcement learning states. : ; in, This represents the reinforcement learning state corresponding to the i-th fracture segment. This represents the later-stage mineralization structural indicator value of the i-th fault segment. This represents the normalized staggered recurrence displacement of the i-th fracture segment. This represents the set of fault zones corresponding to the i-th fault segment. Mean probability of mineralization without internal correction. This represents the mean difference in uncorrected mineralization probability between raster prediction cells that have a corresponding relationship on both sides of the i-th fault segment; when there is no effective optimal misaligned recurrence displacement. hour, Set to 0; Specifically, Represented as: ; Represented as: ; in, Indicates the i-th fracture segment A set of raster prediction cells with corresponding relationships on both sides. This represents the uncorrected mineralization probability of the raster prediction unit on the other side of the fault corresponding to the j-th raster prediction unit; when there is no correspondence between the raster prediction units on both sides of the fault... It is empty at this time. Set to 0; Reinforcement learning strategy network According to the status Output Action : ; in, This represents the action corresponding to the i-th fracture segment. This represents the probability suppression coefficient of the fault zone. The term "reconnection coefficient" refers to the coefficient used to control the reduction in mineralization probability within the corresponding fault zone of the later mineralization structure. The term "reconnection coefficient" refers to the coefficient used to control the adjustment of mineralization probability between raster prediction units with misalignment relationships on both sides of the fault. The term "reconnection" refers to the process of introducing the uncorrected mineralization probability difference between corresponding raster prediction units into the probability correction process of the target raster prediction unit based on the misalignment relationship on both sides of the fault, in order to reflect the potential mineralization continuity after the fault displacement. Reinforcement learning strategy network For a feedforward neural network, the number of nodes in its input layer and its state... The dimensions are consistent, and the output layer includes two output nodes, which correspond to the probability suppression coefficient of the fault zone, respectively. and misaligned exception return coefficient To ensure the stability of the correction direction, the output layer uses a non-negative output function. and Non-negative values; fault zone probability suppression coefficient and misaligned exception return coefficient The effective range is determined by the process of maximizing the reward of the validation samples; Reinforcement learning strategy network The training process includes: setting the state corresponding to each fracture segment. Input the current policy network to obtain the action. According to the action Generate temporary correction probabilities for raster prediction cells within the neighborhood of each fracture segment; calculate reward values ​​based on temporary correction probabilities and validation sample labels. Based on reward value Update strategy network parameters Repeat the above process until the changes in reward values ​​in each round satisfy the convergence condition, thus obtaining the trained reinforcement learning policy network. .

[0024] Training reinforcement learning strategy network At that time, according to the action Generate temporary correction probabilities; based on the action Before calculating the temporary correction probability, first divide the fracture segment. The temporary correction probability of each grid prediction cell within the correction range is initialized to its uncorrected mineralization probability, i.e.: ; in, This indicates that the j-th raster prediction unit is in the current action. The temporary correction probability is obtained from the grid prediction cells and the fault segment. The relationship is that the fault zone probability suppression correction amount and the misalignment anomaly reconnection correction amount are superimposed in sequence; For a grid set located in a fault zone Within the raster prediction unit Perform probability suppression in the log-odds space: ; in, ; For raster prediction units that belong to the correspondence of staggered recurrence on both sides of the fracture Perform a pullback adjustment in logarithmic odds space: ; in, This represents the uncorrected mineralization probability of the raster prediction unit on the other side of the fracture corresponding to the j-th raster prediction unit; to avoid abnormal logarithmic probability calculation caused by the probability value being equal to 0 or 1, the probability is restricted to an open interval before calculation so that it is not equal to 0 or 1. The reinforcement learning reward function is used to evaluate the current action. The effect; in one implementation, the reward function simultaneously considers the accuracy of the verification sample prediction and the suppression effect of the high probability of errors in the later mineralization fracture zone; for the i-th fracture segment ,set up Given the set of validation samples in the neighborhood of the fracture segment, the validation loss is... for: ; in, This represents the verification loss within the neighborhood of the i-th fracture segment. This represents the set of verification samples in the neighborhood of the i-th fracture segment. Indicates the number of validation samples. This represents the label of the j-th verification sample. This indicates that the j-th raster prediction unit is in the current action. The temporary correction probability obtained below; High probability penalty for fault zone error for: ; in, This represents the high-probability penalty term for fault zone errors corresponding to the i-th fault segment. This represents the set of fault zone grids corresponding to the i-th fault segment; This embodiment will verify the loss. High probability penalty item for fault zone errors Perform in-batch normalization separately to obtain the normalized validation loss. High probability penalty term for normalized fault zone errors Intra-batch normalization refers to normalizing the data of each fracture segment within the set of fracture segments updated in the same round of training. and Perform maximum-minimum normalization; when the maximum value of a certain item in the same training round is equal to the minimum value, set the normalization result corresponding to that item to 0, so that the item does not participate in the differentiated reward evaluation in the current training round. Reward Value Represented as: ; in, Indicates the action corresponding to the i-th fracture segment. The reward value, This represents the normalized verification loss. The normalized fault zone error high probability penalty term is represented; the reward function simultaneously constrains the accuracy of the verification sample prediction and the effect of suppressing the high probability of fault zone errors in the later mining process.

[0025] The reward value increases when the current action reduces verification losses and decreases the probability of temporary mineralization in fault zone areas with high mineralization structural marker values ​​in the later stages. Enlarge; Reinforcement Learning Strategy Network The strategy network parameters are updated by maximizing the sum of the reward values ​​of each fracture segment: ; in, This represents the parameters of the policy network after training is complete; specifically, in each round of training, the fragmentation segments... status Input reinforcement learning policy network Output action Calculate the reward value after performing this action. , and then As the optimization loss of the policy network, the policy network parameters are updated through backpropagation. This allows the strategy network to gradually learn the probability suppression actions of fracture zones and the reconnection actions of misaligned anomalies under different late-stage mineralization tectonic states. After training, the reinforcement learning policy network outputs the fault zone probability suppression coefficient and misalignment anomaly reconnection coefficient for each fault segment: .

[0026] The final mineralization probability is generated; in this embodiment, the correction amount refers to the probability correction amount calculated based on the later mineralization structural marker value, the fault zone probability suppression coefficient, and the misalignment anomaly reconnection coefficient; the logarithmic probability space refers to the probability correction amount obtained through... The computational space formed by the transformation of probability values ​​by the function can be superimposed with the correction amount in the log probability space, which can avoid the problem of probability values ​​being less than 0 or greater than 1 caused by directly superimposing in the probability space; For any raster prediction unit To determine whether it is located in a certain fault segment Corresponding fault zone grid set Within, and whether it belongs to that fault segment. The raster prediction cells in the corresponding relationship are reproduced by staggered arrays on both sides; when Located in the fault zone grid set Within the time frame, the i-th fracture segment right The generated fracture zone probability suppresses the correction amount Represented as: ; in, Indicates the i-th fracture segment For the j-th raster prediction unit The generated fracture zone probability suppresses the correction amount; when When the raster prediction cell belongs to the category of raster cells with staggered recurrence correspondences on both sides of the fracture, the i-th fracture segment right The resulting misalignment anomaly correction amount Represented as: ; in, Indicates the i-th fracture segment For the j-th raster prediction unit The generated misalignment anomaly correction amount; When the same raster prediction unit Relative to the same fracture segment Simultaneously satisfying the condition of being located in the fault zone grid set When considering conditions that include internal and non-interventional correspondences, calculate the fracture segment separately. For this raster prediction unit The generated fracture zone probability suppresses the correction amount And misalignment error correction amount The two types of correction amounts are added together to form the fracture segment. For this raster prediction unit Total correction amount : ; When the same raster prediction unit If only one of the conditions is met, the corresponding correction amount is taken as the total correction amount. When the same raster prediction unit Relative to the fracture segment Not located in the fault zone grid set If the segment does not belong to a misaligned recurrence correspondence, then the fracture segment... For this raster prediction unit Total correction amount Set to 0; When the same raster prediction unit When affected by multiple fracture segments simultaneously, the total correction amount corresponding to each fracture segment will be... Unified superposition; final mineralization probability Represented as: ; in, This represents the final mineralization probability of the j-th raster prediction cell. , Indicates the i-th fracture segment The range of corrective action, This represents the sum of the correction amounts of all broken segments that affect the j-th raster prediction unit; Obtain the final mineralization probability of all raster prediction cells. Then, according to the final mineralization probability The target areas are sorted from highest to lowest probability, and candidate target units are determined based on the final mineralization probability distribution of all units within the target area; specifically, the final mineralization probability is... Raster prediction units with values ​​higher than the upper quantile level of the final mineralization probability distribution of the entire target area are identified as candidate target area units; spatially adjacent candidate target area units are merged into connected components to form candidate target area patches; the mean final mineralization probability, the highest final mineralization probability, and the patch area within each candidate target area patch are calculated, and the candidate target area patches are sorted according to the mean final mineralization probability, the highest final mineralization probability, and the patch area, and the ranking results of the solid mineralization prediction target areas are output.

[0027] Through the above steps, fault zones with high mineralization indices in later stages will not be mistakenly elevated to high mineralization probability areas simply because the fault lines are significant; mineralization anomaly areas that still have a staggered recurrence relationship after being dissected on both sides of the fault can have their potential continuity restored by the staggered anomaly reconnection coefficient, thereby improving the reliability of solid mineral prediction results in multi-stage tectonic superposition areas.

[0028] Example 2, based on Example 1, further illustrates the machine learning model. The specific construction, training, and prediction process; In this embodiment, the machine learning model Used for raster prediction units mineralization anomaly intensity and basic structural response values Generate the raster prediction unit Uncorrected mineralization probability The uncorrected mineralization probability This refers to the fact that the ore-breaking structural indicator values ​​have not yet been determined. Initial ore-forming probability before probability suppression and misalignment reconnection; In this embodiment, the machine learning model A gradient boosting tree classification model is adopted. This model is formed by the sequential stacking of multiple regression trees. Each regression tree is used to fit the residual direction between the previous prediction result and the sample label, thus gradually improving the model's ability to distinguish between known mineralized samples and non-mineralized samples. The reason for adopting the gradient boosting tree classification model is that it reflects the intensity of mineralization anomalies. and basic structural response values There may be a non-linear relationship between them. The gradient boosting tree classification model can learn the correspondence between the two and the mineralization probability without pre-setting a linear discrimination boundary. For any training sample in the training sample set T Construct machine learning input vectors: ; in, This represents the machine learning input vector for the j-th training sample. This represents the mineralization anomaly intensity of the j-th raster prediction cell. The basic construction response value of the j-th raster prediction unit is denoted as ; the training sample label is denoted as . ,in, This indicates that the training sample is a known mineral deposit sample. This indicates that the training sample is a non-mineralized sample; Training machine learning models Previously, the training sample set T was spatially grouped. Specifically, based on the spatial location of training samples within the target region, adjacent training samples were grouped into the same spatial group, and training subsets and validation subsets were divided using spatial groups as units. This prevented adjacent grid prediction units from simultaneously entering both the training and validation subsets, which could lead to inflated validation results and affect the machine learning model. It can more realistically reflect the mineralization prediction capabilities at different spatial locations within the target area; Machine learning models The initial predicted score is denoted as In one implementation, the initial prediction score is determined based on the proportion of positive samples in the training sample set T: ; in, This represents the initial predicted score of the j-th training sample. This represents the proportion of known mining sites in the training sample set T; if or If so, it means that the training sample set T does not meet the conditions for binary classification training. The training sample set T should be reconstructed so that the training sample set T includes both known mineral point samples and non-mineralized samples. For the t-th iteration, the machine learning model uses the predicted score from the previous iteration. Calculate the current predicted probability of the j-th training sample: ; in, This represents the predicted probability of the j-th training sample after round t-1. Represents the Sigmoid function, and: ; Based on training sample labels and current predicted probability Calculate the residual term in the t-th iteration: ; in, Let represent the residual term of the j-th training sample in the t-th iteration; this residual term is used to represent the direction of deviation between the current prediction result of the machine learning model and the true sample label. With training sample input vector As input, with residual terms To fit the target, train the t-th regression tree. After training is complete, update the predicted scores: ; in, This represents the predicted score of the j-th training sample after the t-th iteration. This represents the t-th regression tree for the input vector. Output residual fitted values, Indicates the learning step size; learning step size This is used to control the update magnitude of each regression tree to the model's predicted score, and its value is determined by the decrease in cross-entropy loss on the validation subset. After multiple iterations, a gradient boosting tree classification model consisting of multiple regression trees is obtained: ; in, This indicates that the machine learning model responds to the input vector. The output probability of mineralization. Represents the set of parameters for a machine learning model. This indicates the number of regression trees involved in the integration; Based on the change in validation subset loss, when adding more regression trees no longer reduces the cross-entropy loss of the validation subset, stop adding regression trees to avoid overfitting the model to the training samples. Machine learning models The training objective is to minimize the cross-entropy loss on the training sample set T: ; in, This represents the probability of mineralization output by the machine learning model for the j-th training sample; to avoid abnormal logarithmic calculations caused by probability values ​​being equal to 0 or 1, when calculating the cross-entropy loss, Restrict it to an open interval, making it not equal to 0 or 1; In the training samples, the number of known mineralized samples is usually less than the number of unmineralized samples; to avoid machine learning models... This implementation avoids over-biasing towards unmineralized samples. Instead, it sets sample weights based on the number of samples with different labels in the training sample set T. Weights are then set separately for known mineralized samples and unmineralized samples. ; in, This represents the weight of a known mineral deposit sample. Indicates the weight of unmineralized samples. This represents the total number of samples in the training sample set T. This represents the number of known mining site samples in the training sample set T. This represents the number of non-mineralized samples in the training sample set T; after applying the above weights, the training objective is expressed as: ; When the number of known mineral deposit samples is small, the contribution of these samples to model training is increased, avoiding the need for machine learning models to... Only the distribution characteristics of the majority of non-mineralized samples were learned; After the model is trained, predict each grid cell within the target area. machine learning input vector Input the trained machine learning model The probability of mineralization without correction is obtained as follows: ; in, This represents the uncorrected mineralization probability of the j-th raster prediction cell; this uncorrected mineralization probability Based solely on the intensity of mineralization anomalies and basic structural response values The calculated values ​​do not include later ore-breaking structural markers. Fault zone probability suppression coefficient and misaligned exception return coefficient Uncorrected mineralization probability It can serve as the basis for subsequent reinforcement learning correction, enabling subsequent steps to identify and correct high probability errors and misalignment / abnormal segmentation problems caused by later ore breaking and fracture. To ensure the probability of mineralization without correction This embodiment can be used for subsequent log-odds space correction of machine learning models. The output probability undergoes boundary handling; when When it equals 0, replace it with the minimum predicted probability greater than 0 within the target region; when When the value is equal to 1, it is replaced with the maximum predicted probability less than 1 within the target region; the probability after boundary processing is still denoted as... And used for subsequent calculations. ; During the model validation phase, each sample from the validation subset is input into the machine learning model. The validation prediction probabilities are obtained, and the validation subset cross-entropy loss is calculated. If the validation subset cross-entropy loss continuously decreases with the increase of the number of regression trees, training continues. If the validation subset cross-entropy loss no longer decreases, training stops, and the model parameters corresponding to the lowest validation subset cross-entropy loss are retained. This method is used to determine the machine learning model. The final parameters can reduce the overfitting of machine learning models to training samples, thus reducing the probability of uncorrected mineralization. It is more suitable as the basic input for subsequent correction of ore-breaking structures; In another implementation, the machine learning model Alternatively, a random forest classification model can be used; in this case, with As input to each decision tree, multiple decision trees are trained. Each decision tree outputs the probability that the j-th grid prediction cell belongs to the mineralization category. (Machine learning model) Output the average of all decision tree probability results as the uncorrected mineralization probability. ; In another implementation, the machine learning model Alternatively, a neural network classification model can be used; in this case, with As input to the neural network, intermediate representations are calculated through hidden layers, and the uncorrected mineralization probability is obtained through the Sigmoid output layer. The training objective of the neural network classification model remains cross-entropy loss, and the output after training is the uncorrected mineralization probability. It also serves as input for subsequent reinforcement learning correction steps.

[0029] Example 3, as Figure 2 As shown, the machine learning-based solid mineral prediction system proposed in this invention is used to execute the machine learning-based solid mineral prediction method described in Embodiment 1, including: The data grid building module is used to acquire fault line data and mineralization anomaly data of the target area, divide the target area into multiple grid prediction units, and divide the fault line data into multiple fault segments. The fracture response module is used to calculate the fracture line saliency for each fracture segment based on the continuous length and linear response intensity of the fracture segment, and to calculate the low mineralization response value of the fracture zone based on the mineralization anomaly intensity within the corresponding fracture zone grid set of the fracture segment. The misalignment reproduction module is used to perform misalignment correlation calculations on the mineralization anomaly profiles on both sides of each fault segment along the extension direction of the fault segment, and determine the mineralization anomaly misalignment reproduction value. The mineralization breaking indicator module is used to generate the later mineralization breaking structural indicator value of the fault segment based on the significance of the fault line, the low response value of mineralization in the fault zone, and the recurrence value of mineralization anomaly misalignment. The initial prediction module is used to generate the uncorrected mineralization probability of each grid prediction unit by taking the mineralization anomaly intensity and basic tectonic response value of the grid prediction unit as input and through a machine learning model. The reinforcement decision module is used to take the fault segment as the reinforcement learning decision object. It inputs the state, including the later mineralization structure marker value, into the reinforcement learning policy network to obtain the fault zone probability suppression coefficient and the misalignment anomaly reconnection coefficient. The probability correction module is used to correct the uncorrected mineralization probability based on the later mineralization structural marker value, the fault zone probability suppression coefficient, and the misalignment anomaly reconnection coefficient, to generate the final mineralization probability of each grid prediction unit, and to determine the solid mineralization prediction target area based on the final mineralization probability.

[0030] like Figure 3 As shown, this figure is used to represent the spatial identification relationship of late-stage mineralization faults in a multi-stage tectonic superposition region. The target area is divided into multiple raster prediction units. The fault segment passes through the raster area in a continuous linear form, indicating that the fault segment has high fault line significance in the planar tectonic data, with strong continuity and obvious linear response. Light-colored areas near the fault segment represent low-response areas of mineralization in the fault zone, indicating that the mineralization anomaly intensity corresponding to the fault segment itself is low and does not show a mineralization-controlling response directly enhanced along the fault zone. On both sides of the fault segment, mineralization anomaly patches are shown with dashed boxes. The two mineralization anomaly patches are relatively misaligned along the fault extension direction, but after applying the corresponding misalignment amount, they can form a spatial correspondence, indicating that there is a misaligned recurrence of mineralization anomalies on both sides of the fault. The figure is composed of significant fault lines, low mineralization response in the fault zone, and misaligned recurrence of mineralization anomalies on both sides, which constitute the late-stage mineralization tectonic markers. This indicates that the fault segment is more likely to be a structure that causes misalignment, displacement, or destruction of existing mineralization anomalies after mineralization, and should not be directly used as a positive mineralization-controlling structure to enhance the probability of mineralization in the machine learning prediction process.

[0031] like Figure 4 As shown in the figure, this diagram illustrates the calculation process of displacement correlation between mineralization anomaly profiles on both sides of a fault segment. The extension direction of the fault segment is taken as the tangential direction, and the direction perpendicular to the fault segment is taken as the normal direction. The mineralization anomaly intensities of adjacent raster prediction units on both sides of the fault segment are extracted and arranged sequentially along the tangential direction to form the first and second side anomaly profiles. During the calculation, multiple non-zero displacements are traversed along the tangential direction, allowing the first and second side anomaly profiles to be correlated under different displacement values. The non-zero displacement with the largest correlation value that satisfies the effective misalignment reproduction condition is selected as the optimal misalignment reproduction displacement. After determining the optimal misalignment reproduction displacement, the raster prediction units on both sides of the fault are spatially mapped according to this displacement to establish a correspondence between the two raster units. This correspondence is used to characterize the spatial relationship where mineralization anomalies on both sides of the fault can still be reproduced after displacement. Further calculations of mineralization anomaly misalignment reproduction values ​​can be performed, and these values, along with the fault line saliency and the low mineralization response value of the fault zone, are used to generate subsequent mineralization breaking structure marker values.

[0032] like Figure 5As shown, this figure illustrates the target area variation after correcting the uncorrected mineralization probability based on later ore-breaking structural indicators and reinforcement learning correction coefficients. The left side represents the uncorrected mineralization probability map, where areas near the fault segment exhibit erroneously high mineralization probabilities due to the significant fault lines. These high-probability areas extend along the fault zone, easily leading machine learning models to mistakenly identify later ore-breaking faults as ore-controlling structures. The right side represents the corrected mineralization probability map. After correction based on later ore-breaking structural indicators, fault zone probability suppression coefficients, and misalignment anomaly reconnection coefficients, the erroneously high mineralization probability near the fault zone decreases. The probability of mineralization is reduced, and the mineralization anomaly areas with staggered recurrence relationships on both sides of the fault are reconnected to form a more continuous mineralization probability response. The dashed box in the figure represents the main probability concentration area participating in the target area evaluation after correction, and the elliptical area represents the solid mineral prediction target area determined based on the final mineralization probability. By suppressing the probability of later mineralization fault zones and reconnecting the staggered anomaly areas on both sides of the fault, the problem of erroneous extension of prediction results along the direction of later mineralization faults can be reduced, and the rationality of the delineation of solid mineral prediction target areas in multi-stage tectonic superposition areas can be improved.

[0033] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. A machine learning-based method for predicting solid mineral resources, characterized in that, Includes the following steps: Acquire fault line data and mineralization anomaly data of the target area, divide the target area into multiple raster prediction units, and divide the fault line data into multiple fault segments; For each fracture segment, the saliency of the fracture line is calculated based on the continuous length and linear response intensity of the fracture segment, and the low response value of mineralization of the fracture zone is calculated based on the mineralization anomaly intensity in the corresponding fracture zone grid set. For each fault segment, the displacement correlation calculation of the mineralization anomaly profile on both sides of the fault is performed along the extension direction of the fault segment to determine the mineralization anomaly misalignment recurrence value. Based on the significance of the fracture line, the low response value of mineralization in the fracture zone, and the recurrence value of mineralization anomalies, the later mineralization structural marker values ​​of the fracture segment are generated. Using the mineralization anomaly intensity and basic tectonic response values ​​of the grid prediction unit as input, the uncorrected mineralization probability of each grid prediction unit is generated through a machine learning model. Using the fault segment as the decision object of reinforcement learning, the state including the later mineralization structure marker value is input into the reinforcement learning policy network to obtain the fault zone probability suppression coefficient and the misalignment anomaly reconnection coefficient. Based on the later-stage mineralization structural marker values, fault zone probability suppression coefficients, and misalignment anomaly reconnection coefficients, the uncorrected mineralization probabilities are corrected to generate the final mineralization probabilities for each grid prediction unit. The target area for solid mineralization prediction is then determined based on the final mineralization probabilities.

2. The solid mineral prediction method based on machine learning according to claim 1, characterized in that, The mineralization anomaly intensity is obtained by normalizing the mineralization anomaly data. When the mineralization anomaly data includes multiple mineralization anomaly indicators, the multiple mineralization anomaly indicators are normalized respectively, and the fusion weight is determined according to the response consistency between each mineralization anomaly indicator and the known mineral point sample label to obtain the comprehensive mineralization anomaly intensity of each grid prediction unit.

3. The solid mineral prediction method based on machine learning according to claim 1, characterized in that, The significance of the fracture line is obtained by normalizing the product of the continuous length of the fracture segment and the linear response intensity of the fracture segment; the low response value of mineralization in the fracture zone is obtained by normalizing and reversing the mean value of mineralization anomaly intensity in the corresponding fracture zone grid set for each fracture segment.

4. The solid mineral prediction method based on machine learning according to claim 1, characterized in that, For each fault segment, the displacement correlation of the mineralization anomaly profiles on both sides of the fault is calculated along the extension direction of the fault segment to determine the recurrence value of the mineralization anomaly misalignment, including: Using the extension direction of the target fracture segment as the tangential direction and the direction perpendicular to the target fracture segment as the normal direction, the mineralization anomaly intensity of adjacent grids on both sides of the target fracture segment is extracted to form a first side anomaly profile and a second side anomaly profile. The correlation calculations under non-zero displacement are performed on the first side anomaly profile and the second side anomaly profile along the tangential direction to obtain a set of non-zero displacement correlation values; The effective misalignment reproduction conditions are determined based on the distribution statistics of the non-zero misalignment correlation value set, and the non-zero misalignment amount that satisfies the effective misalignment reproduction conditions and has the largest correlation value is determined as the optimal misalignment reproduction displacement. Based on the optimal staggered reproduction displacement, establish the correspondence between the grid prediction units on both sides of the fracture. The mineralization anomaly misalignment value is determined based on the correlation value corresponding to the optimal misalignment recurrence displacement and the ratio of the optimal misalignment recurrence displacement to the continuous length of the fracture segment.

5. The solid mineral prediction method based on machine learning according to claim 4, characterized in that, The valid conditions for the reproduction of misaligned columns are determined based on the distribution statistics of the set of non-zero shift correlation values, including: Calculate the central tendency and dispersion values ​​of the set of non-zero shift correlation values; Non-zero shift correlation values ​​that are higher than the central tendency value and exceed the dispersion value constraint are determined as valid correlation values. The non-zero shift quantity with the aforementioned effective correlation value is determined as the non-zero shift quantity that satisfies the effective misalignment reproduction condition.

6. The solid mineral prediction method based on machine learning according to claim 1, characterized in that, The late-stage mineralization structural marker value is obtained by multiplying the fracture line significance, the low mineralization response value of the fracture zone, and the recurrence value of the mineralization anomaly misalignment.

7. The solid mineral prediction method based on machine learning according to claim 1, characterized in that, The basic structural response value is obtained by normalizing and back-calculating the distance from the grid prediction unit to the nearest fault segment. The machine learning model includes at least one of the following: random forest model, gradient boosting tree model, support vector machine model, or neural network model that can output the probability of mineralization.

8. The machine learning-based solid mineral prediction method according to claim 1, characterized in that, The input states of the reinforcement learning policy network include: the late-stage mineralization structure flag value; the ratio of the optimal misaligned recurrence displacement to the continuous length of the fault segment; the average uncorrected mineralization probability within the fault zone grid set; and the average difference in uncorrected mineralization probability between corresponding grid prediction units on both sides of the fault.

9. The solid mineral prediction method based on machine learning according to claim 1, characterized in that, Correcting the uncorrected mineralization probability includes: When the target grid prediction unit is located within the fault zone grid set corresponding to the target fault segment, the uncorrected mineralization probability of the target grid prediction unit is suppressed according to the later mineralization structure indicator value of the target fault segment and the fault zone probability suppression coefficient. When the target grid prediction unit belongs to the grid prediction unit with a misaligned recurrence correspondence on both sides of the target fault segment, the uncorrected mineralization probability of the target grid prediction unit is adjusted according to the later mineralization structure indicator value of the target fault segment, the misaligned anomaly reconnection coefficient, and the difference in uncorrected mineralization probability between the corresponding grid prediction unit and the target grid prediction unit. When correcting the uncorrected mineralization probability, the correction amount of one or more fault segments to the same grid prediction unit is superimposed in the logarithmic probability space, and the final mineralization probability is obtained by transformation through the Sigmoid function.

10. A machine learning-based solid mineral prediction system, used to execute the machine learning-based solid mineral prediction method according to any one of claims 1-9, characterized in that, Specifically, it includes: The data grid construction module is used to acquire fault line data and mineralization anomaly data of the target area, divide the target area into multiple grid prediction units, and divide the fault line data into multiple fault segments; The fracture response module is used to calculate the fracture line saliency for each fracture segment based on the continuous length and linear response intensity of the fracture segment, and to calculate the low mineralization response value of the fracture zone based on the mineralization anomaly intensity within the corresponding fracture zone grid set of the fracture segment. The misalignment reproduction module is used to perform misalignment correlation calculations on the mineralization anomaly profiles on both sides of each fault segment along the extension direction of the fault segment, and determine the mineralization anomaly misalignment reproduction value. The mineralization breaking indicator module is used to generate the later mineralization breaking structural indicator value of the fault segment based on the significance of the fault line, the low response value of mineralization in the fault zone, and the recurrence value of mineralization anomaly misalignment. The initial prediction module is used to generate the uncorrected mineralization probability of each grid prediction unit by taking the mineralization anomaly intensity and basic tectonic response value of the grid prediction unit as input and through a machine learning model. The reinforcement decision module is used to take the fault segment as the reinforcement learning decision object. It inputs the state, including the later mineralization structure marker value, into the reinforcement learning policy network to obtain the fault zone probability suppression coefficient and the misalignment anomaly reconnection coefficient. The probability correction module is used to correct the uncorrected mineralization probability based on the later mineralization structural marker value, the fault zone probability suppression coefficient, and the misalignment anomaly reconnection coefficient, to generate the final mineralization probability of each grid prediction unit, and to determine the solid mineralization prediction target area based on the final mineralization probability.

Images