Multi-source data fusion hidden ore body intelligent prediction model training and prediction method

Abstract

The application discloses a kind of multi-source data fusion's concealed ore body intelligent prediction model training and prediction method, after the corresponding multiple grid units of ore-bearing area are divided into positive sample or unmarked sample respectively, the initial characteristics corresponding to each are constructed;Using the feature gate layer of the measurement network to be trained, the attention weight of each column in the prototype matrix is calculated according to the initial characteristics, the attention weight and each column are weighted summation, obtain adaptive gate vector after the initial characteristics are modulated, the target characteristics after modulation are converted using multiple cascaded fully connected layers, obtain embedding vector;After the weight vector of output layer is normalized, the discriminant direction for distinguishing positive sample and unmarked sample is obtained, the projection square value of positive sample is calculated combined with embedding vector, and the noise subspace energy of unmarked sample, the measurement network is trained through the loss function with competition mechanism, obtain the trained measurement network for predicting concealed ore body.

Description

Technical Field

[0001] This invention relates to the field of geophysical exploration and artificial intelligence, and relates to, but is not limited to, training and prediction methods for intelligent prediction models of concealed ore bodies based on multi-source data fusion. Background Technology

[0002] In the southern part of a certain mining area, thick loess is widely distributed with few outcrops. Surface geological surveys are insufficient to directly identify deep mineralization information. It is necessary to use integrated geophysical exploration methods to obtain the material response of deep geological bodies in order to discover hidden ore bodies.

[0003] In related technologies, on the one hand, traditional geophysical inversion and multivariate statistical methods (such as least squares fitting, principal component analysis, or support vector machines), while computationally inexpensive, rely on linear or shallow nonlinear assumptions, making it difficult to characterize high-order nonlinear coupling relationships in multi-source heterogeneous data. Furthermore, multi-source data fusion requires independent anomaly extraction and binarized thresholding, easily resulting in the loss of weak anomaly information. On the other hand, deep learning methods can handle nonlinear relationships, but face the problem of extremely scarce known mining sites (i.e., positive samples) in the training data, and the overwhelming majority of unlabeled samples. This leads to overfitting, insufficient accuracy, and high false alarm rates in traditional prediction models due to sample imbalance.

[0004] Therefore, the question is how to efficiently and accurately solve the problem of predicting hidden ore bodies while effectively fusing multi-source data under the condition of sparse positive samples. Summary of the Invention

[0005] In view of this, embodiments of the present invention provide a training method for an intelligent prediction model of concealed ore bodies based on multi-source data fusion, which at least solves the problem that related technologies cannot efficiently and accurately solve the problem of effectively fusing multi-source data to predict concealed ore bodies under the condition of sparse positive samples.

[0006] According to a first aspect of the present invention, a method for training an intelligent prediction model for concealed ore bodies based on multi-source data fusion is provided, comprising: The ore-bearing area is divided into multiple raster units, and each raster unit is further divided into positive samples or unlabeled samples. Initial features are constructed based on the positive samples or unlabeled samples. Both positive samples and unlabeled samples include magnetic data, induced polarization data, CSAMT data, and geological data. Using the feature gating layer of the metric network to be trained, the attention weights corresponding to each column in the prototype matrix are calculated based on the initial features. The attention weights and each column are weighted and summed to obtain the adaptive gating vector. The initial features are then multiplied element-wise based on the adaptive gating vector to obtain the target features. Each column of the prototype matrix represents a combination of features for a mineralization mode. The target features are transformed using multiple cascaded fully connected layers in the metric network to be trained, resulting in an embedding vector. The weight vector of the output layer in the metric network to be trained is normalized to obtain the discriminative direction used to distinguish between positive samples and unlabeled samples. The squared projection value of the positive sample along the discrimination direction and the energy of the unlabeled sample in the noise subspace orthogonal to the discrimination direction are calculated based on the discrimination direction and the embedding vector. The loss is calculated using the squared projection value, the noise subspace energy, and a loss function with a competition mechanism. The loss is then used to train the metric network to be trained, resulting in a trained metric network for predicting concealed ore bodies.

[0007] Optionally, the step of calculating the attention weights corresponding to each column in the prototype matrix based on the initial features includes: Calculate the cosine similarity between the initial features and each column of the prototype matrix; The cosine similarity is normalized to obtain the attention weights corresponding to each column in the prototype matrix.

[0008] Optionally, the step of calculating the squared projection value of the positive sample along the discrimination direction and the energy of the unlabeled sample in the noise subspace orthogonal to the discrimination direction based on the discrimination direction and the embedding vector includes: Transpose the discrimination direction, substitute the transposed discrimination direction and the embedding vector corresponding to the positive sample into Equation 1 to obtain the projection scalar value; and square the projection scalar value to obtain the projection square value. Substituting the calculated squared projection values ​​of the unlabeled samples and the corresponding embedding vectors into Equation 2, we obtain the noise subspace energy; Formula 1 is as follows:

[0009] In the above formula, Let be the projected scalar value corresponding to the i-th positive sample. To determine direction, T stands for transpose. This is the embedding vector corresponding to the i-th positive sample; Formula 2 is as follows:

[0010] In the above formula, Let be the noise subspace energy corresponding to the j-th unlabeled sample. Let j be the embedding vector corresponding to the j-th unlabeled sample. Let be the squared projection value corresponding to the j-th unlabeled sample.

[0011] Optionally, the step of calculating the loss from the squared projection value, the noise subspace energy, and the loss function with a competition mechanism includes: Substituting the squared projection value and the noise subspace energy into the loss function with a competition mechanism yields the total loss; the loss function with a competition mechanism is as follows:

[0012] In the above formula, S is the total loss, m is the number of positive samples, and P is the set of positive samples. Here, n is the number of unlabeled samples, and U is the set of unlabeled samples. and This is a hyperparameter.

[0013] Optionally, constructing the corresponding initial features based on positive samples or unlabeled samples includes: The magnetic data were subjected to polarization, horizontal differentiation, vertical first-order differentiation, and upward continuation to obtain the corresponding magnetic anomaly derived field; the induced polarization data were measured to obtain the apparent polarizability and apparent resistivity; and the apparent polarizability and apparent resistivity were interpolated to obtain the metallicity factor. Based on CSAMT data, the resistivity values ​​corresponding to multiple preset depth layers are calculated; the lithology field in the geological data is thermally encoded to obtain the corresponding encoding vector; and the Euclidean distance between the corresponding raster cell and the known fault structure in the geological data is calculated. The magnetic anomaly-derived field, apparent polarizability, apparent resistivity, metallicity factor, resistivity value, encoding vector, and Euclidean distance are concatenated to obtain the initial features corresponding to each grid cell.

[0014] Optionally, before constructing the corresponding initial features based on positive samples or unlabeled samples, the method further includes: Obtain the coordinates of known mineral points in the ore-bearing area; calculate the Euclidean distance between the coordinates of the known mineral points and the center coordinates of each grid cell; When the Euclidean distance is not greater than the distance threshold, the grid cell is regarded as a positive sample; When the Euclidean distance is greater than the distance threshold, the grid cell is treated as an unlabeled sample.

[0015] According to a second aspect of the present invention, a method for intelligent prediction of concealed ore bodies based on multi-source data fusion is provided, comprising: The study area to be measured is divided into multiple grid cells to be measured; each grid cell contains corresponding magnetic data, induced polarization data, CSAMT data and geological data; The trained metric network is used to process each grid cell to be tested to obtain the target embedding vector; the trained metric network is trained using the concealed ore body intelligent prediction model training method described in any of the foregoing embodiments; The weight vector of the output layer of the trained metric network is normalized to obtain the optimal discrimination direction; and the target projection scalar value and the target noise subspace energy are calculated using the optimal discrimination direction and the target embedding vector. Substituting the target projection scalar value and the target noise subspace energy into the mineralization probability formula, the mineralization probability of each grid cell is obtained. By stitching together all mineralization probabilities according to their spatial location, a mineralization probability distribution map corresponding to the study area to be tested is obtained; The mineralization probability formula is as follows:

[0016] In the above formula, This is a mineralization probability distribution map. For the target projection scalar value, The energy of the target noise subspace. and These are two hyperparameters.

[0017] According to a third aspect of the present invention, an electronic device is provided, comprising: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other via the communication bus; the memory is used to store at least one executable instruction, wherein the executable instruction causes the processor to perform an operation corresponding to the method described in the first and second aspects.

[0018] According to the scheme provided in the embodiments of the present invention, the ore-bearing area is divided into multiple grid cells, and the multiple grid cells are respectively divided into positive samples or unlabeled samples; and initial features corresponding to each positive sample or unlabeled sample are constructed; both positive samples and unlabeled samples include magnetic data, induced polarization data, CSAMT data and geological data; using the feature gating layer of the metric network to be trained, the attention weights corresponding to each column in the prototype matrix are calculated according to the initial features, and the attention weights and each column are weighted and summed to obtain an adaptive gating vector; the initial features are then multiplied element-wise according to the adaptive gating vector to obtain the target features; wherein, each column of the prototype matrix Each represents a combination of features for a mineralization mode. Multiple cascaded fully connected layers in the metric network are used to transform the target features, resulting in embedding vectors. The weight vectors of the output layer in the metric network are normalized to obtain the discrimination direction used to distinguish between positive and unlabeled samples. The squared projection value of the positive sample along the discrimination direction and the energy of the unlabeled sample in the noise subspace orthogonal to the discrimination direction are calculated based on the discrimination direction and the embedding vectors. The loss is calculated using the squared projection value, the noise subspace energy, and a loss function with a competition mechanism. The loss is then used to train the metric network to be trained, resulting in a trained metric network for predicting concealed ore bodies. In this process, for multi-source data covering mineral-bearing areas, to achieve more efficient screening, a dedicated metric network designed for the mineral exploration scenario is first trained. Through the learnable feature gating layer of the metric network, an adaptive gating vector is generated using the prototype matrix of prior knowledge of mineral exploration to screen the constructed initial features, highlighting effective information and suppressing interference, providing high-quality input for subsequent direction search, and improving efficiency and prediction reliability. By normalizing the weights of the output layer of the metric network to be trained and using them as the direction of discrimination, iterative optimization maximizes the projection scalar value of positive samples, effectively avoiding overfitting due to scarce samples. Based on the projection scalar value of positive samples and the energy of the noise subspace of unlabeled samples, a competition mechanism is used to calculate the loss and stably optimize the direction of discrimination. The loss is calculated using the squared projection value, noise subspace energy, and a loss function with a competitive mechanism. This loss is then used to train the metric network to be trained. Its positive sample attraction term and unlabeled sample noise suppression term form an intrinsically competitive gradient by sharing a discriminant direction. Without external class prior estimation or manual weight setting, it can autonomously converge to the optimal discriminant direction, ensuring training stability and prediction accuracy. The final trained metric network is then obtained for predicting concealed ore bodies. Using this method, the trained metric network can effectively fuse multi-source data and achieve high-precision prediction of concealed ore bodies even under conditions of sparse positive samples. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein: Figure 1 This is a flowchart illustrating a training method for an intelligent prediction model of concealed ore bodies based on multi-source data fusion, provided in an embodiment of the present invention. Figure 2 A flowchart illustrating a method for intelligent prediction of concealed ore bodies based on multi-source data fusion, provided in an embodiment of the present invention. Figure 3 This is a mineralization probability distribution map of the study area to be tested, provided in an embodiment of the present invention. Figure 4 This is a schematic diagram illustrating the effect of the AUC metric provided in the embodiments of the present invention; Figure 5 This is a schematic diagram illustrating the effect of target area hit rate provided in an embodiment of the present invention; Figure 6 A schematic diagram illustrating the correlation between mineralization probability and drilling ore grade provided in an embodiment of the present invention; Figure 7 This is a schematic diagram illustrating the effect of single-instance prediction time provided in an embodiment of the present invention; Figure 8 This is a schematic diagram illustrating the effect corresponding to the number of training convergence iterations provided in an embodiment of the present invention; Figure 9 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. The following embodiments are used to illustrate the present invention, but are not intended to limit the scope of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0022] It should be noted that the terms "first, second, and third" used in the embodiments of the present invention are only used to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, and third" can be interchanged in a specific order or sequence where permitted, so that the embodiments of the present invention described herein can be implemented in an order other than that illustrated or described herein.

[0023] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which these embodiments of the invention pertain. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0024] Figure 1 This is a flowchart illustrating a training method for an intelligent prediction model of concealed ore bodies based on multi-source data fusion, provided in an embodiment of the present invention. This method can be executed by an electronic device, such as a computer or server.

[0025] like Figure 1 As shown, the training method for an intelligent prediction model of concealed ore bodies based on multi-source data fusion includes: S101. Divide the ore-bearing area into multiple raster units, and divide the multiple raster units into positive samples or unlabeled samples respectively; and construct the corresponding initial features based on the positive samples or unlabeled samples; the positive samples or unlabeled samples include magnetic data, induced polarization data, CSAMT data and geological data.

[0026] In an embodiment of the present invention, magnetic data, induced polarization data, geological data, and controlled-source audio-frequency magnetotellurics (CSAMT) data of known ore-bearing areas are first collected across the entire region. Then, the spatial extent of the ore-bearing area can be plotted in appropriate software, such as geological mapping software, by presetting the grid spacing, such as... The spatial extent of the ore-bearing area is divided into N raster cells. Each raster cell can be further divided into positive samples and unlabeled samples based on known mineralization points. Positive samples are raster cells within the ore-bearing study area with known mineralization points and clear mineralization labels. Unlabeled samples are raster cells within the ore-bearing study area where mineralization is unknown, including background and potential mineralization. Each raster cell carries coordinate information. The coordinates of the collected magnetic, induced polarization, CSAMT, and geological data are matched and located with the coordinates of the raster cells. Spatial interpolation methods are then used to estimate the corresponding magnetic, induced polarization, CSAMT, and geological data for each raster cell.

[0027] S102. Using the feature gating layer of the metric network to be trained, the attention weights corresponding to each column in the prototype matrix are calculated based on the initial features. The attention weights and each column are weighted and summed to obtain the adaptive gating vector. The initial features are multiplied element-wise based on the adaptive gating vector to obtain the target features. Each column of the prototype matrix represents a combination of features of a mineralization mode.

[0028] In an embodiment of the present invention, a dedicated metric network is designed for mineral exploration scenarios. Its structure is to add a learnable feature gating layer before the input layer of a traditional linear classifier (containing three fully connected layers and one output layer). That is, the metric network contains a feature gating layer, three fully connected layers and one output layer.

[0029] Specifically, the initial features are input into the feature gating layer of the metric network to be trained. The initial features and each column of the prototype matrix are calculated to obtain the corresponding attention weight for each column. The attention weights and each column are then weighted and summed, and finally processed by the Sigmoid activation function to obtain the corresponding adaptive gating vector. Each component of this adaptive gating vector represents the importance of a certain dimension of the initial features to the mineralization prediction of the current grid cell. Finally, the adaptive gating vector and the initial features are multiplied element-wise to obtain the target features. When the adaptive gating vector is close to 1, the initial features are almost completely preserved through element-wise multiplication. When it is suppressed to near 0, the initial features are essentially masked. This operation achieves adaptive feature selection, ensuring that high-importance features are retained and fed into subsequent layers of the metric network, while low-importance or interfering features are suppressed.

[0030] In this prototype matrix, each column represents a typical combination of mineralization pattern features, and the dimension of the combination of mineralization pattern features is the same as the dimension of the initial features. The combination of mineralization pattern features is randomly initialized in the initial stage, and the trained prototype matrix is ​​obtained after the metric network is trained.

[0031] S103. The target features are transformed using multiple cascaded fully connected layers in the metric network to be trained to obtain the embedding vector; and the weight vector of the output layer in the metric network to be trained is normalized to obtain the discrimination direction used to distinguish between positive samples and unlabeled samples.

[0032] In an embodiment of the present invention, three cascaded fully connected layers sequentially complete feature encoding, feature abstraction, and low-dimensional embedding mapping. After the target features are input into the first fully connected layer for feature encoding, the output high-order features related to mineralization are input into the second fully connected layer for further compression and abstraction of feature representation, enhancing mineralization signals and suppressing noise. The output high-dimensional features are input into the last fully connected layer for mapping to a low-dimensional embedding space, and an embedding vector is output. This embedding vector condenses the fingerprint features most relevant to mineralization in the original multi-source geoscience information.

[0033] Simultaneously, substituting the weights of the output layer into the normalization formula yields the discrimination direction. The normalization formula is as follows:

[0034] In the above formula, w represents the direction of discrimination. This is the weight vector of the output layer.

[0035] The metric network adopts a linear classifier structure, and the output layer weights are the decision hyperplane normal vectors. The direction of these normal vectors is the classification discriminant direction. To stabilize the training process, the weights are normalized so that the magnitude of the weight vector is fixed at 1, thereby decoupling the magnitude and direction. This allows the training to optimize only the direction. During this optimization process, the discriminant direction gradually points to the clustering center of known points in the embedding space under the drive of the competition mechanism, and is then used to distinguish between positive samples and unlabeled samples.

[0036] During training, the parameters and discrimination direction of the metric network are jointly updated.

[0037] S104. Calculate the squared projection value of the positive sample along the discrimination direction and the energy of the unlabeled sample in the noise subspace orthogonal to the discrimination direction based on the discrimination direction and the embedding vector.

[0038] S105. The loss is calculated using the squared projection value, the energy of the noise subspace, and the loss function with a competition mechanism. The loss is then used to train the metric network to be trained, resulting in a trained metric network for predicting concealed ore bodies.

[0039] In embodiments of the present invention, the projection scalar value is the projection length of the embedding vector along the discrimination direction. A larger value indicates a higher probability of mineralization in the grid cell. Squaring the projection scalar value eliminates symbolic ambiguity. The noise subspace energy refers to the portion of the feature energy in the embedding vector of the grid cell after mapping by the metric network that cannot be explained by typical mineralization patterns, reflecting the intensity of noise or background changes unrelated to mineralization. Based on the discrimination direction and the embedding vector, the squared projection value of the embedding vector along the discrimination direction and the energy in the noise subspace orthogonal to the discrimination direction are calculated. The loss is then calculated using the squared projection value and the noise subspace energy to train the metric network. The trained metric network can then be used to predict concealed ore bodies.

[0040] In summary, specifically, this invention chooses the PU learning algorithm to address the problem of sparse positive samples because it only requires a small number of known mineral points as positive samples, treating the remaining large amount of data as unlabeled samples, eliminating the need for negative sample labels. Its core advantage lies in its ability to extract effective information from sparse positive samples even when facing imbalanced data. While traditional PU learning algorithms are theoretically applicable to scenarios with sparse positive samples, their underlying assumptions fundamentally contradict prediction of deep, concealed mineral deposits. Traditional PU learning assumes that positive samples form compact clusters in the feature space, thereby identifying similar samples from unlabeled samples. However, in deep mineral deposit prediction, positive samples are extremely scarce, insufficient to form statistically significant clusters. Furthermore, the geophysical responses of deep, concealed mineral deposits, such as low-resistivity, high-polarization anomalies, are often submerged in background noise, and positive samples lack a clear clustering structure, leading to insufficient prediction accuracy. To address the limitation of the traditional PU learning algorithm's positive sample clustering assumption failing under sparse conditions, a novel learning paradigm centered on discriminative direction search is proposed. By constructing a loss function with an inherent competitive mechanism (explained in detail later), the metric network autonomously seeks the optimal discriminative direction that maximizes the projected energy of positive samples while suppressing the noise energy of unlabeled samples, thus solving the underfitting problem caused by positive sample sparsity. This paradigm is based on a metric network specifically designed for mineral exploration scenarios.

[0041] It is understood that, in the embodiments of the present invention, the ore-bearing area is divided into multiple grid cells, and each grid cell is further divided into positive samples or unlabeled samples; and initial features are constructed based on the positive samples or unlabeled samples; both positive samples and unlabeled samples include magnetic data, induced polarization data, CSAMT data, and geological data; using the feature gating layer of the metric network to be trained, the attention weights corresponding to each column in the prototype matrix are calculated based on the initial features, and the attention weights and each column are weighted and summed to obtain an adaptive gating vector; the initial features are then multiplied element-wise based on the adaptive gating vector to obtain the target features; wherein, each column in the prototype matrix... Each column represents a combination of features for a mineralization mode. The target features are transformed using multiple cascaded fully connected layers in the metric network to obtain embedding vectors. The weight vectors of the output layer in the metric network are normalized to obtain the discrimination direction used to distinguish between positive and unlabeled samples. The squared projection value of the positive sample along the discrimination direction and the energy of the unlabeled sample in the noise subspace orthogonal to the discrimination direction are calculated based on the discrimination direction and the embedding vector. The loss is calculated using the squared projection value, the noise subspace energy, and a loss function with a competition mechanism. The loss is then used to train the metric network to be trained, resulting in a trained metric network for predicting concealed ore bodies. In this process, for multi-source data in the coverage area, to achieve more efficient screening, a dedicated metric network designed for mineral exploration scenarios is first used. Through the learnable feature gating layer of the metric network, an adaptive gating vector is generated using the prototype matrix of prior mineral exploration knowledge. This vector is used to screen the initial features based on mineralization pattern matching, highlighting effective information and suppressing interference. This provides high-quality input for subsequent direction discrimination search, improving efficiency and prediction reliability. Furthermore, by normalizing the weights of the output layer of the metric network to be trained and using them as the direction discrimination, iterative optimization maximizes the positive sample projection scalar value, effectively avoiding the problems caused by scarce samples. Overfitting is addressed by employing a competitive mechanism to calculate the loss based on the projected scalar values ​​of positive samples and the energy of the noise subspace of unlabeled samples. This mechanism stabilizes and optimizes the discrimination direction. The loss is calculated using the squared projection value, the energy of the noise subspace, and a loss function with a competitive mechanism. This loss is then used to train the metric network. The positive sample attraction term and the unlabeled sample noise suppression term form an intrinsically competitive gradient by sharing the discrimination direction. This eliminates the need for external class prior estimation or manual weight setting, allowing the network to autonomously converge to the optimal discrimination direction, ensuring training stability and prediction accuracy. The resulting trained metric network is used to predict concealed ore bodies. Through this method, the trained metric network can effectively fuse multi-source data and achieve high-precision prediction of concealed ore bodies even under conditions of sparse positive samples.

[0042] In some embodiments of the present invention, the calculation of the attention weights corresponding to each column in the prototype matrix based on the initial features in S102 can be achieved by the following steps: calculating the cosine similarity between the initial features and each column in the prototype matrix; normalizing the cosine similarity to obtain the attention weights corresponding to each column in the prototype matrix.

[0043] Specifically, the cosine similarity of each column in the prototype matrix corresponding to each grid cell is calculated and normalized to obtain the corresponding exclusive attention weight.

[0044] In some embodiments of the present invention, the calculation of the projected square value of the positive sample along the discrimination direction and the energy of the unlabeled sample in the noise subspace orthogonal to the discrimination direction in S104 can be achieved by the following steps: transposing the discrimination direction, substituting the transposed discrimination direction and the embedding vector corresponding to the positive sample into Formula 1 to obtain the projected scalar value; squaring the projected scalar value to obtain the projected square value; substituting the calculated projected square value of the unlabeled sample and the corresponding embedding vector into Formula 2 to obtain the noise subspace energy; Formula 1 is as follows:

[0045] In the above formula, Let be the projected scalar value corresponding to the i-th positive sample. To determine direction, T stands for transpose. This is the embedding vector corresponding to the i-th positive sample; Formula 2 is as follows:

[0046] In the above formula, Let be the noise subspace energy corresponding to the j-th unlabeled sample. Let j be the embedding vector corresponding to the j-th unlabeled sample. Let be the squared projection value corresponding to the j-th unlabeled sample.

[0047] Specifically, when calculating the projected square value corresponding to the j-th unlabeled sample, the discrimination direction and the embedding vector corresponding to the j-th unlabeled sample are substituted into Formula 1, where i is replaced by j, and the corresponding projected scalar value is calculated. Then, the projected square value is obtained by squaring the projected scalar value.

[0048] In Formula 1, if the discrimination direction is trained to point towards the center of the positive sample cluster, then the embedding vector of a highly mineralized sample will be highly aligned with the discrimination direction. Larger unlabeled samples show weak alignment. Smaller. In Equation 2, when a sample contains strong mineralization information, its energy should be mainly concentrated in the signal subspace. The energy level approaches 0; when the sample is pure background or interference, the energy proportion of the signal subspace is small. Relatively large.

[0049] In some embodiments of the present invention, the calculation of the loss in S105 using the squared projection value, the noise subspace energy, and the loss function with a competition mechanism can be achieved through the following steps: substituting the squared projection value and the noise subspace energy into the loss function with a competition mechanism to obtain the loss; the loss function with a competition mechanism is as follows:

[0050] In the above formula, S is the total loss, m is the number of positive samples, and P is the set of positive samples. Here, n is the number of unlabeled samples, and U is the set of unlabeled samples. and This is a hyperparameter.

[0051] The first term of the loss function The aim is to ensure that the embedding vectors of positive samples have a large projected square value in the discriminative direction, so a negative log-likelihood form is used. When the value is very small, the output of the sigmoid activation function approaches zero, and then a strong penalty is imposed. When the value is large enough, the sigmoid activation function saturates at 1, the loss approaches zero, and it no longer over-penalizes. To preset hyperparameters and control the saturation threshold. Because the number of unlabeled samples is large, if squared loss or linear penalty is used, the model may tolerate a certain degree of noise energy during the prediction process, so the second term... The exponential function approach aims to constrain the energy of the noise subspace of unlabeled samples. To minimize this, an exponential penalty is imposed on the noise subspace energy of the unlabeled samples using an exponential function. When the noise subspace energy increases, leading to a rise in loss, the coupling of the two terms in the loss function is reduced. The coupling is not an independent weighted sum, but rather through a shared discrimination direction. The coupling mechanism involves the second term of the loss function competing with the first term. Positive samples force the discrimination direction to point towards their aggregation direction, while unlabeled samples constrain the discrimination direction to minimize noise energy. This means maximizing the projected energy of positive samples and suppressing the noise energy of unlabeled samples. The result of this balance is that the discrimination direction points towards a unique direction where positive samples deviate from the noise subspace energy of unlabeled samples—the optimal discrimination direction—the direction of the mineralization feature.

[0052] The Adam optimizer is used to iteratively optimize the parameters of the metric network to be trained, with each batch containing both positive and unlabeled samples. During training, training ends when the relative transformation rate of the first term of the loss function (positive sample attraction term), the decrease in total loss on the validation set, and the change in the Euclidean distance of the discriminant direction all tend to stabilize. Training also ends when, for five consecutive epochs, the transformation rate of the first term of the loss function is less than a preset transformation threshold, the decrease rate of the total loss on the validation set is less than a preset decrease threshold, and the Euclidean distance between the discriminant direction of the current epoch and the discriminant direction of the previous epoch is less than a preset distance threshold. The preset transformation threshold, preset decrease threshold, and preset distance threshold can all be manually set by those skilled in the art based on experimental results or actual application scenarios, and are not limited here. Preferably, the preset transformation threshold can be 10. -4 The preset threshold for the decrease can be 10. -3 The preset distance threshold can be 10. -5 .in, and All are greater than 0.

[0053] In some embodiments of the present invention, the construction of corresponding initial features based on positive or unlabeled samples in S101 can be achieved through the following steps: polarization, horizontal differentiation, vertical first-order differentiation, and upward extension of the magnetic data to obtain the corresponding magnetic anomaly derived field; measurement of induced polarization data to obtain apparent polarizability and apparent resistivity; interpolation of apparent polarizability and apparent resistivity to obtain the metal factor; and calculation of resistivity values ​​corresponding to multiple preset depth layers based on CSAMT data. The lithology field in the geological data is uniquely encoded to obtain the corresponding encoding vector; the Euclidean distance between the corresponding raster cell and the known fracture structure in the geological data is calculated; the magnetic anomaly derived field, apparent polarizability, apparent resistivity, metallicity factor, resistivity value, encoding vector and Euclidean distance are concatenated to obtain the initial feature corresponding to each raster cell.

[0054] Specifically, magnetic data, induced polarization data, CSAMT data, and geological data are collected from the entire ore-bearing area. Magnetic data can be collected using a proton magnetometer, induced polarization data using an induced polarization meter, CSAMT data using a ground-based electromagnetic detection system, and geological data can be collected using a combination of GPS, geological compass, and field logbook. Each grid cell contains coordinate information. The collected data is matched and located according to the coordinates carried and the coordinates corresponding to the grid cell. Spatial interpolation methods are used to estimate the four types of data corresponding to each grid cell.

[0055] Furthermore, the magnetic data were subjected to polarization, horizontal derivative, vertical first-order derivative, and upward continuation to obtain the polarization anomaly field, horizontal derivative field, vertical derivative field, and upward continuation field, resulting in a total of four magnetic anomaly derived fields. The induced polarization data were measured, and the ratio of the measured apparent polarizability to apparent resistivity was calculated to obtain the metallicity factor. Impedance calculations were performed on the electric and magnetic fields in the CSAMT data to obtain the apparent resistivity, followed by two-dimensional inversion to obtain the subsurface resistivity profile. Interpolation of the subsurface resistivity profile yielded a three-dimensional resistivity volume. The subsurface space of the study area was divided into multiple depth layers along the depth direction, such as five depth layers (e.g., 50m, 100m, 200m and 400m, 800m). The resistivity value at each depth layer was inferred using the three-dimensional resistivity volume. The lithology field in the geological data is uniquely encoded to obtain the corresponding encoding vector. Then, known fault structures in the geological data are obtained. These known fault structures are rock strata or stratigraphic structures that have fractured, i.e., crustal rocks have fractured and displaced under tectonic stress. If there are multiple known fault structures, the Euclidean distance between the nearest known fault structures is calculated for each grid cell. Multiple depth levels can be manually set.

[0056] In some embodiments of the present invention, before constructing the corresponding initial features based on positive samples or unlabeled samples in S101, the following steps are also included: obtaining the coordinates of known mineral points in the ore-bearing area; calculating the Euclidean distance between the coordinates of the known mineral points and the center coordinates of each grid cell; when the Euclidean distance is not greater than a distance threshold, the grid cell is used as a positive sample; when the Euclidean distance is greater than the distance threshold, the grid cell is used as an unlabeled sample.

[0057] Specifically, the distance threshold can be set manually, and can be set according to the size of the ore body and the density of the exploration network in the ore-bearing area, such as 50m.

[0058] Figure 2 This is a flowchart illustrating a method for intelligent prediction of concealed ore bodies based on multi-source data fusion, provided in an embodiment of the present invention. The specific steps include: S201. Divide the study area to be measured into multiple grid cells to be measured; each grid cell to be measured contains corresponding magnetic data, induced polarization data, CSAMT data and geological data.

[0059] S202. Call the trained metric network to process each grid cell to be tested to obtain the target embedding vector.

[0060] S203. Normalize the weight vector of the output layer of the trained metric network to obtain the optimal discrimination direction; and calculate the target projection scalar value and the target noise subspace energy through the optimal discrimination direction and the target embedding vector.

[0061] S204. Substitute the target projection scalar value and the target noise subspace energy into the mineralization probability formula to obtain the mineralization probability of each grid cell.

[0062] S205. All mineralization probabilities are spliced ​​together according to their spatial location to obtain the mineralization probability distribution map corresponding to the area to be studied.

[0063] The mineralization probability formula is as follows:

[0064] In the above formula, This is a mineralization probability distribution map. For the target projection scalar value, The energy of the target noise subspace. and These are two hyperparameters.

[0065] Specifically, after training, the optimal discrimination direction is obtained along with the trained metric network. This is achieved by normalizing the weight vector of the output layer of the trained metric network. and These are two hyperparameters: the signal gain coefficient and the noise suppression coefficient. These can be manually set and are both greater than 0. The study area is processed through steps S201 to S205 to obtain the corresponding mineralization probability distribution map.

[0066] Figure 3 This is a mineralization probability distribution map of the study area to be tested, provided by an embodiment of the present invention. The horizontal and vertical axes represent the grid cell numbers in different directions within the study area. Known mineral points (green triangles) are ore bodies discovered on the surface or in shallow areas. All known mineral points fall within the high-probability zone (yellow-orange), verifying the accuracy of the prediction. The high-probability zone extends into the depths and surrounding areas, indicating the spatial distribution of concealed mineralization. Local high-probability points represent the specific locations of the inferred concealed ore bodies, which can be verified by subsequent drilling. The above results show that the present invention can effectively integrate multi-source data, accurately locate known mineralization locations, and infer the spatial distribution of concealed ore bodies, achieving highly accurate prediction of concealed ore bodies.

[0067] To verify the effectiveness of the present invention, the following comparative experiment was conducted: One set is T2 (traditional PU learning algorithm): It targets a weighted classification loss between positive and unlabeled samples, relying on class prior estimation and positive sample clustering assumptions, without employing discriminative direction search and coupling loss design. The other set is T4 (the scheme of this invention). Both sets of experiments are based on multi-source data collected from the same study area, totaling 20,000 raster cells, of which 5 are known mining sites with positive samples. Stratified five-fold cross-validation is used. The data is randomly divided into 5 folds, each containing approximately 4,000 samples. Stratified sampling ensures that each fold contains one positive sample. Four folds (approximately 16,000 samples, including 4 positive samples) are used as the training set, and the remaining fold (approximately 4,000 samples, including 1 positive sample) is used as the test set. This process is repeated 5 times. Evaluation metrics include: Area Under the Curve (AUC): the model's ability to rank predicted mineralization probabilities, reflecting the degree to which positive samples rank higher than unlabeled samples; Target Hit Rate (%): the proportion of known mineralization points or subsequent drilling mineralization points actually contained in the top 5% of grids with the highest predicted probabilities; Correlation between Mineralization Probability and Drilling Grade: the Pearson correlation coefficient between the model's output probability and the actual gold grade (g / t) at the validation borehole location; Prediction Time per Run (s): the average time from input features to output mineralization probability and target map (excluding data preprocessing); Number of Training Convergence Iterations (epochs): the number of iterations required from the start of training to meet the preset convergence criteria, reflecting the model's training efficiency.

[0068] Furthermore, Table 1 compares the prediction performance of different methods for concealed ore bodies. Figure 4 This is a schematic diagram illustrating the effect of the AUC metric provided in this embodiment of the invention. Figure 5 This is a schematic diagram illustrating the effect of target area hit rate provided in an embodiment of the present invention. Figure 6 This is a schematic diagram illustrating the correlation between mineralization probability and drilling ore grade provided in an embodiment of the present invention. Figure 7 This is a schematic diagram illustrating the effect of single-instance prediction time provided in an embodiment of the present invention. Figure 8 The diagram illustrates the effect of the number of training convergence iterations provided in this embodiment of the invention. Table 1 is shown below: Table 1 compares the prediction performance of different methods for concealed ore bodies.

[0069] Table 1 shows the average results of five independent experiments on the test set. The data show that the proposed method outperforms the traditional PU learning algorithm in terms of AUC, target hit rate, correlation between mineralization probability and drilling grade, single-instance prediction time, and number of training convergence iterations.

[0070] Figure 4The horizontal axis represents the false positive rate, reflecting the degree to which background noise is mistakenly identified as ore bodies; the vertical axis represents the true positive rate, reflecting the degree to which known ore bodies are correctly identified; the larger the AUC area, the higher the detection rate can be obtained with a smaller false positive cost. Figure 8 The heatmap matrix is ​​2 rows x 5 columns, with rows corresponding to methods (T2 and T4) and columns corresponding to each fold of the five-fold cross-validation. Colors range from dark blue (low iteration count) to dark red (high iteration count), with T4 having a significantly lower iteration count per fold than T2 (which has a more bluish hue). Figures 4 to 8 These figures show that the AUC of T4 is 0.82, an increase of 0.11 from T2's 0.71, indicating that the invention significantly enhances the ability to distinguish between positive and unlabeled samples. The target area hit rate increased from 48.5% to 67.3%, an increase of 18.8 percentage points, meaning that with the same limited drilling verification workload, the probability of discovering concealed ore bodies increased significantly. The correlation between mineralization probability and drilling grade improved from 0.42 to 0.61, indicating that the probabilities output by the invention are not only accurately ranked, but also have stronger consistency with the actual mineralization intensity, providing a more reliable quantitative basis for subsequent resource estimation. The single-case prediction time increased slightly from 2.1 seconds to 2.3 seconds (an increase of 0.2 seconds), remaining essentially unchanged. This is because although the feature gating layer introduces a small amount of additional computation (Sigmoid and element-wise multiplication), the effective feature dimension is reduced after sparsification (about 50% of the feature values ​​after training the gating vector are below 0.2, which is essentially masked), and the floating-point operation of the subsequent fully connected layer is reduced accordingly, so the two cancel each other out. Reference Figure 9 The diagram shows a structural schematic of an electronic device according to an embodiment of the present invention. The specific embodiments of the present invention do not limit the specific implementation of the electronic device.

[0071] like Figure 9 As shown, the electronic device may include: a processor 502, a communications interface 504, a memory 506, and a communications bus 508.

[0072] in: The processor 502, communication interface 504, and memory 506 communicate with each other via communication bus 508.

[0073] Communication interface 504 is used to communicate with other electronic devices or servers.

[0074] The processor 502 is used to execute program 510, specifically the relevant steps in the above method embodiments.

[0075] Specifically, program 510 may include program code that includes computer operation instructions.

[0076] Processor 502 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention. The smart device may include one or more processors of the same type, such as one or more CPUs; or it may include processors of different types, such as one or more CPUs and one or more ASICs.

[0077] Memory 506 is used to store program 510. Memory 506 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0078] Specifically, program 510 can be used to cause processor 502 to perform the operations corresponding to the methods described in the above method embodiments.

[0079] The specific implementation of each step in program 510 can be found in the corresponding descriptions of the steps and units in the above method embodiments, and will not be repeated here. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the devices and modules described above can be referred to the corresponding process descriptions in the foregoing method embodiments, and will not be repeated here.

[0080] It should be noted that, depending on the implementation needs, the various components / steps described in the embodiments of the present invention can be broken down into more components / steps, or two or more components / steps or parts of the operation of components / steps can be combined into new components / steps to achieve the purpose of the embodiments of the present invention.

[0081] The methods described above according to embodiments of the present invention can be implemented in hardware, firmware, or as software or computer code that can be stored in a recording medium (such as a CD-ROM, RAM, floppy disk, hard disk, or magneto-optical disk), or as computer code originally stored on a remote recording medium or a non-transitory machine-readable medium and subsequently stored on a local recording medium, downloaded via a network. Thus, the methods described herein can be processed by software stored on a recording medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware (such as an ASIC or FPGA). It is understood that the computer, processor, microprocessor controller, or programmable hardware includes storage components (e.g., RAM, ROM, flash memory, etc.) capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods described herein. Furthermore, when a general-purpose computer accesses code used to implement the methods shown herein, the execution of the code transforms the general-purpose computer into a dedicated computer for executing the methods shown herein.

[0082] Those skilled in the art will recognize that the units and method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of the embodiments of the present invention.

[0083] The above embodiments are only used to illustrate the embodiments of the present invention, and are not intended to limit the embodiments of the present invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the embodiments of the present invention. Therefore, all equivalent technical solutions also fall within the scope of the embodiments of the present invention, and the patent protection scope of the embodiments of the present invention should be defined by the claims.

Claims

1. A training method for an intelligent prediction model of concealed ore bodies based on multi-source data fusion, characterized in that, include: The ore-bearing area is divided into multiple grid cells, and each grid cell is further divided into positive samples or unlabeled samples; and initial features are constructed based on the positive samples or unlabeled samples. Positive or unlabeled samples include magnetic data, induced polarization data, CSAMT data, and geological data; Using the feature gating layer of the metric network to be trained, the attention weights corresponding to each column in the prototype matrix are calculated based on the initial features. The attention weights and each column are weighted and summed to obtain the adaptive gating vector. The initial features are then multiplied element-wise based on the adaptive gating vector to obtain the target features. Each column of the prototype matrix represents a combination of features for a mineralization mode. The target features are transformed using multiple cascaded fully connected layers in the metric network to be trained, resulting in an embedding vector. The weight vector of the output layer in the metric network to be trained is normalized to obtain the discriminative direction used to distinguish between positive samples and unlabeled samples. The squared projection value of the positive sample along the discrimination direction and the energy of the unlabeled sample in the noise subspace orthogonal to the discrimination direction are calculated based on the discrimination direction and the embedding vector. The loss is calculated using the squared projection value, the noise subspace energy, and a loss function with a competition mechanism. The loss is then used to train the metric network to be trained, resulting in a trained metric network for predicting concealed ore bodies.

2. The method according to claim 1, characterized in that, The step of calculating the attention weights corresponding to each column in the prototype matrix based on the initial features includes: Calculate the cosine similarity between the initial features and each column of the prototype matrix; The cosine similarity is normalized to obtain the attention weights corresponding to each column in the prototype matrix.

3. The method according to claim 1, characterized in that, The calculation of the squared projection value of the positive sample along the discrimination direction and the energy of the unlabeled sample in the noise subspace orthogonal to the discrimination direction based on the discrimination direction and the embedding vector includes: Transpose the discrimination direction, substitute the transposed discrimination direction and the embedding vector corresponding to the positive sample into Equation 1 to obtain the projection scalar value; and square the projection scalar value to obtain the projection square value. Substituting the calculated squared projection values ​​of the unlabeled samples and the corresponding embedding vectors into Equation 2, we obtain the noise subspace energy; Formula 1 is as follows: In the above formula, Let be the projected scalar value corresponding to the i-th positive sample. To determine direction, T stands for transpose. This is the embedding vector corresponding to the i-th positive sample; Formula 2 is as follows: In the above formula, Let be the noise subspace energy corresponding to the j-th unlabeled sample. Let j be the embedding vector corresponding to the j-th unlabeled sample. Let be the squared projection value corresponding to the j-th unlabeled sample.

4. The method according to claim 3, characterized in that, The loss is calculated using the squared projection value, the noise subspace energy, and a loss function with a competition mechanism, including: Substituting the squared projection value and the noise subspace energy into the loss function with a competition mechanism yields the total loss; the loss function with a competition mechanism is as follows: In the above formula, S is the total loss, m is the number of positive samples, and P is the set of positive samples. Here, n is the number of unlabeled samples, and U is the set of unlabeled samples. and This is a hyperparameter.

5. The method according to claim 1, characterized in that, The construction of initial features based on positive or unlabeled samples includes: The magnetic data were subjected to polarization, horizontal differentiation, vertical first-order differentiation, and upward continuation to obtain the corresponding magnetic anomaly derived field; the induced polarization data were measured to obtain the apparent polarizability and apparent resistivity; and the apparent polarizability and apparent resistivity were interpolated to obtain the metallicity factor. Based on CSAMT data, the resistivity values ​​corresponding to multiple preset depth layers are calculated; the lithology field in the geological data is thermally encoded to obtain the corresponding encoding vector; and the Euclidean distance between the corresponding raster cell and the known fault structure in the geological data is calculated. The magnetic anomaly-derived field, apparent polarizability, apparent resistivity, metallicity factor, resistivity value, encoding vector, and Euclidean distance are concatenated to obtain the initial features corresponding to each grid cell.

6. The method according to claim 1, characterized in that, Before constructing the corresponding initial features based on positive samples or unlabeled samples, the method further includes: Obtain the coordinates of known mineral points in the ore-bearing area; calculate the Euclidean distance between the coordinates of the known mineral points and the center coordinates of each grid cell; When the Euclidean distance is not greater than the distance threshold, the grid cell is regarded as a positive sample; When the Euclidean distance is greater than the distance threshold, the grid cell is treated as an unlabeled sample.

7. A method for intelligent prediction of concealed ore bodies through multi-source data fusion, characterized in that, include: The study area to be measured is divided into multiple grid cells to be measured; each grid cell contains corresponding magnetic data, induced polarization data, CSAMT data and geological data; The trained metric network is invoked to process each grid cell to be tested, and the target embedding vector is obtained. The trained metric network was obtained by training the concealed ore body intelligent prediction model training method according to any one of claims 16; The weight vector of the output layer of the trained metric network is normalized to obtain the optimal discrimination direction; and the target projection scalar value and the target noise subspace energy are calculated using the optimal discrimination direction and the target embedding vector. Substituting the target projection scalar value and the target noise subspace energy into the mineralization probability formula, the mineralization probability of each grid cell is obtained. By stitching together all mineralization probabilities according to their spatial location, a mineralization probability distribution map corresponding to the study area to be tested is obtained; The mineralization probability formula is as follows: In the above formula, This is a mineralization probability distribution map. For the target projection scalar value, The energy of the target noise subspace. and These are two hyperparameters.

Images