A method, device, medium and product for delineating prospective mineral resources areas

Through the extraction of background samples in isolated forests and the construction of sparse autoencoder network, combined with the self-similarity fractal and convolutional structure, the problem of relying on marker samples and overfitting in the convergence of mineral resource prospects is solved, the recognition accuracy and efficiency are improved, and the recognition accuracy and efficiency are enhanced. The ability to identify geological information.

CN117994632BActive Publication Date: 2024-12-03INST OF MINERAL RESOURCES CHINESE ACAD OF GEOLOGICAL SCI
View PDF 9 Cites 0 Cited by

Patent Information

Application Number
CN202410193039.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-02-21
Publication Date
2024-12-03
Estimated Expiration
2044-02-21

AI Technical Summary

Technical Problem

The existing technology has limitations and overfitting problems of relying on labeled samples in the confinement of mineral resource vision areas. Especially in the case of insufficient deposits, the generalization ability of deep learning models is affected, and is ignored by traditional autoencoder networks. Multi-source spatial coupling information is obtained, resulting in poor recognition effect.

Method used

Isolated forests are used to extract background samples and build a sparse autoencoder network. By combining the self-similarity fractal method and the convolutional structure, sparse constraints and geological constraints are added to reduce feature redundancy and improve the generalization ability of the algorithm.

Benefits of technology

It improves the accuracy and efficiency of the mineral resource vision area, overcomes the dependence of traditional autoencoder networks, enhances the ability to identify geological information, and makes the prediction results more in line with the geological significance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117994632B_ABST
    Figure CN117994632B_ABST
Patent Text Reader

Abstract

The present invention discloses a method, device, medium and product for delineating a prospective mineral resource area, and relates to the field of data analysis technology. The present invention uses an isolation forest to extract background samples of a mineral deposit to be detected, without relying on prior knowledge, and can overcome the limitation that a traditional autoencoder network relies on labeled non-mineralized samples. In addition, the present invention uses a constructed sparse autoencoder network to obtain detection results based on background samples to achieve the delineation of a prospective mineral resource area, which can improve the accuracy and efficiency of the delineation, and make the delineation results more consistent with geological significance.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of data analysis, and in particular to a method, equipment, medium and product for delineating prospective mineral resource areas. Background Art

[0002] Mineral prospecting is a comprehensive work in mineral resource exploration, and its purpose is to obtain important information related to mineralization from multi-source heterogeneous geoscience data to obtain mineralization potential areas. Machine learning algorithms make full use of multi-source related geographic information to deeply explore the complex nonlinear relationship between multi-scale prospecting indicators and ore deposit locations, and are widely used in mineral prospecting prediction to achieve intelligent delineation of prospecting areas.

[0003] As a cutting-edge method in the field of data analysis, shallow machine learning, such as random forests and support vector machines, does not rely on data distribution, and can extract internal logic and relationships from complex geological data, mine hidden geological information, and combine multi-source fusion data to achieve efficient target identification. The limitation of shallow machine learning algorithms is that they have limited ability to express complex functions under limited samples, which restricts their generalization ability for complex data distribution problems.

[0004] Compared with the shallow machine learning method that relies on manual experience to extract sample features, deep learning relies on simulating the neuron structure of the human brain, transforming the original signal layer by layer, establishing a mapping from low-level data to high-level features, and constructing an automatic extraction model from raw data to effective features, so as to greatly improve the efficiency and quality of anomaly recognition, and has the characteristics of high efficiency and intelligence. At the same time, deep learning can fully extract the spatial characteristics of fused data. Deep learning has advantages in analyzing nonlinear, complex and massive high-dimensional data structures. It has broad prospects in deep geological information mining and integration. Combining the latest deep learning technology to achieve rapid and automatic identification of mine-induced anomalies has great application potential. However, deep learning usually requires a large amount of labeled data, and often requires a lot of manpower and material resources to label. If the training data is insufficient, the training is prone to overfitting, resulting in deviations in the prediction results. In the field of mineral resource evaluation, data with mineral points are usually very scarce, and the number of mineral deposits in the study area is often insufficient. You can try to use unsupervised classification to identify mineralization anomalies, but unsupervised classification usually requires manual input of background (normal) samples for training, but background samples are also subjective due to the large number of human selection factors. In addition, data-driven machine learning methods are highly dependent on training data, which affects the generalization ability of the algorithm.

[0005] Specifically, as a hierarchical machine learning algorithm with multi-level nonlinear transformations, deep learning can achieve complex function approximation by learning a deep-level representation of nonlinear network structure, and combine low-level features to form more abstract deep-level representation attributes or features, thereby achieving high-precision classification and prediction. Different from general neural networks, deep learning emphasizes learning and extracting sample features through deeper network models, and uses layer-by-layer training and back propagation to solve the global optimal solution problem. It can learn the complex spatiotemporal coupling relationship between multi-source prospecting information and ore deposits, and can characterize anomalies and patterns that cannot be discovered by conventional methods. It has been applied to the identification of mineral-induced anomalies and the evaluation of mineral resource potential. Chinese scholars used deep learning models to identify mineral-induced anomalies under complex geological conditions earlier, such as using continuous restricted Boltzmann machines to effectively identify multivariate geochemical anomalies. Based on this model, a multivariate geochemical anomaly identification model based on a deep autoencoder network was constructed. Based on the powerful feature extraction capability of deep learning, a hybrid model combining a deep belief network and a single-class support vector machine was further constructed. The deep feature information extracted by the deep learning model was used as the input of the single-class support vector machine anomaly detection algorithm, effectively extracting multivariate geochemical anomaly information. The deep autoencoder network was combined with a density-based clustering algorithm, and feature extraction was achieved using the deep autoencoder network. These studies found that the deep learning model does not rely on the distribution assumptions of exploration geochemical data, can handle complex and nonlinear spatial patterns, and can identify anomalies that traditional methods cannot identify. At the same time, attempts were made to use all geochemical variables, combining big data thinking with deep learning methods to fully consider the complexity and diversity of element combinations, providing a new way to characterize geochemical spatial patterns with nonlinear characteristics and extract implicit anomalies, which can be better applied to geochemical anomaly identification.

[0006] Depending on whether it relies on labeled data, machine learning for mineral prospect prediction can be divided into supervised, unsupervised, semi-supervised and hybrid methods. Supervised learning methods rely on labeled data related to known mineral deposits to establish classification models with and without prospects. The regional mineralization potential is evaluated by the obtained mineralization probability. The advantage of these supervised methods is that they can handle the nonlinear relationship between indicator factors and mineral deposits well without assuming the distribution of multi-source geoscience data, and delineate potential areas where mineralization characteristics are highly similar to known deposits. However, the training of supervised methods relies on labeled samples and requires sufficient known deposits. In some areas with low exploration levels, known mineral points are very limited, which may lead to poor recognition results.

[0007] Unsupervised learning methods do not rely on labeled samples, but extract mineralized anomalies based on the idea of ​​anomaly detection. They separate mineralized anomalies from the complex background of multi-source variables based on the special properties of mineralized anomalies. The higher the degree of abnormal mineralization, the higher the possibility of delineating mineralization, which is suitable for study areas with low exploration or undeveloped areas. For example, a class of support vector machines is used to extract multivariate anomaly areas, and a hypersphere is established so that background samples fall inside the hypersphere and anomaly samples fall outside the hypersphere. An autoencoder network is used to detect mineralized anomalies, and mineralized features are learned without labeled data through reconstruction errors. For example, deep regression is used in combination with multi-source geoscience to improve the prediction performance of neural networks. Geophysical anomaly analysis is extracted using wavelet transform based on geological data and geochemistry, and wavelet neural networks are applied to enhance the ability to capture weak and hidden information related to hidden minerals. Long Short Term Memory Networks (LSTM) are used to extract mineralization and weighted relationships based on the evidence layer generated by the joint singular weighting method. The gated recurrent unit LSTM network is used to carry out mineral potential prediction to better capture the nonlinear relationship between known deposits and indicator factors. Use graph deep learning to effectively capture spatial mineralization features; retain the anisotropy of the obtained features. Use Siamese networks to automatically extract generalized features to obtain new sample samples by measuring the similarity of the input, thereby realizing mineral prospecting prediction. According to the small reconstruction error of the autoencoder network for background samples and the large reconstruction error for abnormal samples, high-dimensional geochemical or mineralization anomaly recognition is realized. Background samples refer to areas that have no obvious anomalies compared to the surrounding background. Anomalies and background samples can be regarded as positive samples and negative samples, respectively. Using geochemical variables of different dimensions to realize geochemical anomaly recognition in the same study area shows that the autoencoder network has better anomaly recognition of high-dimensional data.

[0008] Due to the large reconstruction error of mineralized anomaly samples, traditional autoencoder networks have been widely used in the identification of mineralized anomalies. However, when identifying mineralized anomalies, traditional autoencoder networks tend to ignore multi-source spatial coupling information or indicator factors, have low generalization ability due to excessive feature redundancy, and rely on known non-mineralized samples, which has limitations.

[0009] In addition, the traditional autoencoder method uses one-dimensional input data of mineralization-related factors, ignoring the important spatial position information between mineralization factors, rather than being conducive to the establishment of mineralization position relationships. This may even lead to overfitting problems caused by excessive feature redundancy caused by one-dimensional vectors.

[0010] In the field of mineral resource evaluation, data on mineral deposits is usually very scarce. When the number of mineral deposits in the study area is often insufficient to define the target area for deep learning, unsupervised classification can be used to identify mineralization anomalies. However, unsupervised classification usually requires manual input of background (normal) samples for training. However, background samples are also subjective due to the large number of human selection factors. In addition, purely data-driven deep learning methods are highly dependent on training data, which affects the generalization ability of the algorithm. Summary of the invention

[0011] In order to solve the above problems existing in the prior art, the present invention provides a method, equipment, medium and product for delineating prospective areas of mineral resources.

[0012] To achieve the above object, the present invention provides the following solutions:

[0013] A method for delineating a prospective mineral resource area, the method comprising:

[0014] Isolation forests were used to extract background samples of the mineral deposits to be tested;

[0015] Build a sparse autoencoder network;

[0016] Inputting the background sample into the sparse autoencoder network to obtain a detection result;

[0017] Based on the detection results, the prospective mineral resource areas are identified.

[0018] Optionally, an isolation forest is used to extract background samples of the mineral deposit to be detected, including:

[0019] Constructing a sample subset based on the multi-channel image of the mineral deposit to be detected; the sample subset has V pixels at the root node of the tree;

[0020] A dimension is randomly selected from all channels, and a split value is randomly generated; the split value is a random value from the minimum to the maximum value of the pixel point of the current node in the selected channel;

[0021] Splitting the split value to divide the pixel points of the current node into sub-nodes in the selected channel;

[0022] Recursively perform pixel division on the child nodes until there is only one pixel in the child node or the formed isolated tree reaches the set height limit to generate multiple isolated trees;

[0023] The generated multiple isolated trees are used to determine the normalized anomaly score;

[0024] The pixels whose normalized anomaly scores are less than a set value are taken as background pixels to generate background samples.

[0025] Optionally, the set value is 0.5.

[0026] Optionally, dividing the split value to divide the pixel points of the current node into sub-nodes in the selected channel specifically includes:

[0027] When the value of a pixel in the selected channel is less than the split value after segmentation, this pixel is used as the left child node of the current node;

[0028] When the value of a pixel in the selected channel is greater than or equal to the split value after segmentation, this pixel is used as the right child node of the current node.

[0029] Optionally, construct a sparse autoencoder network, specifically including:

[0030] The spatial distribution of known mineral deposits is quantified using self-similar fractal methods to obtain prior knowledge;

[0031] Adding sparse constraints to the convolutional autoencoder network, and using the prior knowledge as geological constraints to obtain an initial sparse autoencoder network;

[0032] Use Lone Tree to separate the background information of known mineral deposits to obtain background samples;

[0033] Using the background samples as training samples, and using the training samples to train the initial sparse autoencoder network to obtain a trained initial sparse autoencoder network;

[0034] The trained initial sparse autoencoder network is used as the sparse autoencoder network.

[0035] Optionally, the self-similarity fractal method adopted is a box counting method.

[0036] Optionally, the loss function of the sparse autoencoder network is:

[0037]

[0038] In the formula, Loss SCAE Represents the loss function value of the sparse autoencoder network, Loss MSE ReLU represents the activation function value, β represents the weight coefficient for controlling the sparse constraint, Represents the average activation value of the background samples and the constant ρ, λ represents the penalty factor, and c_loss represents the geological constraint.

[0039] A computer device comprises: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of any of the above-mentioned methods for delineating prospective mineral resources areas.

[0040] A computer-readable storage medium having a computer program stored thereon, characterized in that when the computer program is executed by a processor, the steps of any of the above-mentioned methods for delineating prospective mineral resources areas are implemented.

[0041] A computer program product, comprising a computer program, characterized in that when the computer program is executed by a processor, the steps of any of the above-mentioned methods for delineating prospective mineral resources areas are implemented.

[0042] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0043] The present invention uses isolation forest to extract background samples of the mineral deposit to be detected, without relying on prior knowledge, and can overcome the limitation of the traditional autoencoder network relying on labeled non-mineralized samples. In addition, the present invention uses a constructed sparse autoencoder network to obtain detection results based on background samples to achieve the delineation of mineral resource prospecting areas, which can improve the accuracy and efficiency of delineation and make the delineation results more consistent with geological significance. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0045] Figure 1 A flow chart of a method for delineating a prospective mineral resource area provided in Example 1 of the present invention;

[0046] Figure 2 A diagram of the preselection process of background samples provided in Example 1 of the present invention;

[0047] Figure 3 A schematic diagram of the method for identifying mineralization anomalies based on a convolutional sparse autoencoder network provided in Example 1 of the present invention;

[0048] Figure 4 A network structure diagram of a convolutional sparse autoencoder provided in Example 1 of the present invention;

[0049] Figure 5 A specific workflow diagram for mineralization anomaly identification based on a convolutional sparse autoencoder network and an isolation forest provided in Example 1 of the present invention;

[0050] Figure 6 The entire implementation architecture diagram of constructing a sparse autoencoder network provided in Example 1 of the present invention;

[0051] Figure 7 A diagram showing a specific construction process of an input grid sample provided in Embodiment 1 of the present invention;

[0052] Figure 8 A frequency histogram of anomaly scores provided in Example 1 of the present invention;

[0053] Fig. 9 A logarithmic graph of the number N(ε) of mineral deposits and the number of boxes provided in Example 1 of the present invention;

[0054] Fig.10 A sensitivity curve diagram of mineralization anomaly provided in Example 1 of the present invention;

[0055] Fig.11 This is a prediction success rate curve diagram provided in Example 1 of the present invention. DETAILED DESCRIPTION

[0056] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0057] The purpose of the present invention is to provide a method, equipment, medium and product for delineating prospective mineral resources areas, aiming to improve the generalization ability of the algorithm, overcome the limitation of traditional autoencoder networks that rely on labeled non-mineralized samples, and thus improve the accuracy and efficiency of delineation, so that the delineation results are more in line with geological significance.

[0058] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0059] Example 1

[0060] like Figure 1 As shown, the method for delineating a prospective mineral resource area in this embodiment includes:

[0061] Step 100: Use isolation forest to extract background samples of the mineral deposit to be detected.

[0062] In actual applications, when using a convolutional sparse autoencoder network to identify mineralization anomalies, background samples must be input. Background samples are usually non-mineral point grids, that is, areas where the anomalies are not significant compared to the surrounding environment. For unlabeled non-mineral samples, the convolutional sparse autoencoder network cannot automatically learn the characteristics of background samples. In order to address this limitation, the present invention uses isolation forests to autonomously calculate the anomaly scores of all multi-channel image samples without relying on prior knowledge, and selects areas without obvious anomalies as background samples. According to its characteristics, it is believed that normal sample points generally have longer separation paths, so samples with lower anomaly scores can be taken as background samples. Based on this, in the pre-selection process of background samples, background samples are mainly isolated forests, such as Figure 2 As shown, it includes two parts: one part is to establish an isolated tree (ie, an isolated tree), and the other part is to make a fusion decision of all isolated trees.

[0063] 1) Establishment of an isolated tree. Assume that the training data set X contains multi-channel image samples of all indicator mineral factors. X belongs to R M×C , M is all the pixels in the single-channel image, and C is the number of multiple channels. Specifically, the process of establishing an isolation tree is:

[0064] Step 1: Create a sample subset for training. Randomly select V pixels as training points, subset X', belonging to R M×C , the sample subset has V pixels at the root node of the tree.

[0065] Step 2: Randomly select a dimension from all channels and randomly generate a split value p, which is a random value from the minimum to the maximum value of the current node pixel in the selected channel.

[0066] Step 3: The split value p divides the pixel of the current node into two child nodes in the selected channel. If the value of a pixel in the channel is less than the value p, the pixel will be isolated to the left of the current node, that is, as the left child node. Otherwise, the pixel will be isolated to the right of the current node, that is, as the right child node.

[0067] Step 4: Recursively perform steps 2 and 3 on the child node. This achieves the continuous growth of the isolated tree until there is only one pixel in the child node or the entire tree reaches the height limit. In this way, all pixels of the sample subset X are placed on the child node.

[0068] 2) Fusion decision of all isolated trees. The normalized anomaly score is calculated using the generated t isolated trees [i.e., formula (1)]. r (x) is the isolation path length of sample x in the nth isolation tree, r = {1, 2, ..., t}, [Formula (2)] is the average path length of x under all isolation trees, reflecting the integrated fusion decision. c(V) [Formula (3)] is a value for adjusting the added path length, which is determined by the number of samples in the sample subset.

[0069]

[0070] Because background points are more likely to be isolated in the depth of the isolation tree and farther from the root, it is possible to determine whether a sample is a background sample by judging its path length during the isolation process. Therefore, the anomaly score of the isolation forest is used to obtain the background sample. When the anomaly score of a sample is much smaller than the set value (for example, 0.5), it can be regarded as a background sample. Isolation forests can rely on prior knowledge to separate background samples from the distribution of the data itself, overcoming the subjectivity of human selection.

[0071] Step 101: Build a sparse autoencoder network.

[0072] In practical applications, the construction process of the sparse autoencoder network can be:

[0073] 1) Use the self-similar fractal method to quantify the spatial distribution of known mineral deposits and obtain prior knowledge.

[0074] Specifically, the box counting method is used to quantify the spatial distribution of mineral deposits. The entire study area is divided into 10 square grids according to the size of ε. According to the different sizes of ε in the grids, the number of grids with mineral deposits N(ε) is counted. N(ε) is obviously related to ε. When ε is small enough, the minimum number of grids N(ε) required is proportional to ε. -D is proportional, that is:

[0075] N(ε)=Cε -D (4)

[0076] Where C is a constant and D is the fractal dimension. The slope of the graph of N(ε) and the logarithm of ε is the D value in formula (4). Different fractal dimensions D represent different degrees of self-similarity at different scales. The smaller D is, the more concentrated the mineral deposits are in space. The minimum value d at a certain distance from D is used to establish the radius-density function, which is P(x):

[0077] P(x)=Cx -D (x <d) (5)

[0078] The radius-density function can characterize the spatial pattern of a certain type of mineral deposit distribution. The radius-density function can roughly calculate the density scale of the function P(x) at a given distance x. The function P(x) represents the mineralization probability Y through the normalization process of formula (6). Among them, P(x) min and P(x) max are the minimum and maximum values ​​of P.

[0079]

[0080] 2) Add sparse constraints to the convolutional autoencoder network, and use prior knowledge as geological constraints to obtain the initial sparse autoencoder network.

[0081] In actual applications, through convolution and pooling operations, the encoder maps the multi-channel input image to the latent space to express it and extract higher-level representation features. Through deconvolution and depooling operations, the decoder maps the latent features to generate a reconstructed image layer by layer, and the reconstruction error is as follows (7). i is the background sample, y i To reconstruct data. The present invention uses the ReLU activation function to introduce sparse constraints through convolution and deconvolution layers. When the output value of the feature unit passed by the ReLU function is less than zero, this feature unit is forced to be replaced by zero. The use of the ReLU method ensures that some potential features are suppressed.

[0082]

[0083] In addition, in order to ensure that most of the latent layer neuron units are in an inhibited state, the present invention adds another sparse constraint [i.e., reconstruction error, formula (7)] as a penalty factor to the mean square error of the input and output. The loss function of the autoencoder network with convolutional sparse constraints is shown in formula (8). When the input background sample x i (i={1,2,...,n}) The activation value of feature unit k is A k (x i ), where k = {1, 2, ..., m}. Here the average activation value of the feature unit over all input background samples It can be obtained by formula (9). Let constant ρ be the desired sparse parameter, and the constant value close to zero can be manually set. The relative entropy of the distance between [Formula (10)] introduces the loss function. β is the weight coefficient for controlling the sparsity constraint. During the training process, as long as As ρ is approached, more feature units can be suppressed.

[0084]

[0085] The prior knowledge obtained from fractal analysis is used as the geological constraint c_loss, and the spatial distribution probability (i.e., mineralization probability Y) is used as the additional geological term constraint model. Let Y_error be Loss MSE The reconstruction error is based on the L2 distance in formula (7) as a non-negative value. The Tanh activation function is added to convert Y_error into probability Pano The Tanh activation function is given by:

[0086]

[0087] The mean square error (MSE) function (also known as L2 loss) is used to measure the predicted probability P ano The difference between and the weight within distance d. The c_loss function based on MSE is expressed as:

[0088]

[0089] P ano = {P i |i=1, 2,...,N} and Y={Y i |i=1,2,...,N} is the probability value, where N is the number of pixels. Under this geological constraint, within a certain range d, the area around the known ore deposit (x≤d) should have a high probability of mineralization anomaly. The additional loss term c_loss can calculate the model output of the difference between the predefined anomaly probability value and the error. The constraint term ensures the consistency of the geological model output during the process of adjusting the gradient descent direction during training. The addition of c_loss can improve the training process, effectively promote the training of the network, and strengthen the consistency of the model output with the geological law. Among them, λ is the penalty coefficient factor (0-1), which can be adjusted to balance the model geological loss.

[0090] Through back propagation, the weights of the convolutional sparse autoencoder network are continuously adjusted to minimize the loss function [Formula (11)], which not only realizes the continuous reconstruction of the image to approximate the distribution of the input image, but also suppresses most of the feature units to improve the generalization ability of the mineralization anomaly recognition model. The principle of identifying mineralization anomalies based on the convolutional sparse autoencoder network is as follows: Figure 3 shown.

[0091] 3) Using the lonely tree to separate the background information of the known mineral deposit to obtain the background sample. The specific implementation process of this step can be referred to the above step 100.

[0092] 4) Using the background samples as training samples, and using the training samples to train the initial sparse autoencoder network, to obtain a trained initial sparse autoencoder network.

[0093] In practical applications, the adjustment of the weight of extracted features depends on the distribution of background samples of the data. The trained network can approximately generate background samples with smaller reconstruction errors. Since mineralized anomaly samples are rare events, the distribution and potential of the data are different from the characteristics of abnormal samples and background samples. When the input abnormal samples are input into the training network, the network cannot correctly reconstruct the abnormal samples. Therefore, samples with larger reconstruction errors are considered to be mineralized anomaly samples. The larger the reconstruction error, the greater the possibility of discovering potential mineral resources.

[0094] The convolutional sparse autoencoder network structure used this time is as follows Figure 4 As shown in Figure 2. Through the hierarchical calculation of convolutional layers and pooling layers, the predicted samples are gradually mapped to high-level feature representations of potential samples, and then the mapping gradually becomes a reconstructed image.

[0095] Among them, Figure 4 A brief description of what the different layers mean is: “Con3×3, 10, s2, p1, ReLU, @40×40” means it is a convolutional layer of 10 filters with a stride value of 2 and a padding value of 1, after which the ReLU function is used for this convolutional layer, and the output feature size is 40×40. Max pooling 2×2, s2, p0, index 1, @20×20, means it is a max pooling layer with a stride value of 2, a padding value of 0, a pooling index of 1, and the output feature size is 20×20. Deconv3×3, 40, s1, p1, op0, ReLU, @10×10, means it is a deconvolutional layer of 40 filters with a stride value of 1, a padding value of 1, and an output padding value of 0, after which the ReLU function is used for this deconvolutional layer, and the output feature size is 10×10. MaxUnpool2×2,s2,p0,indices3,@10×10, means it is a maximum unpooling layer with a step size of 2, a padding value of 0, a pooling index of 3, and an output size characteristic of 10×10. It is worth noting that the index layers of the pooling symmetric pooling and non-pooling must be consistent during programming, and the ReLU activation function used effectively avoids the disappearance of the gradient.

[0096] Based on the above content, by adding two sparsity constraints, feature redundancy can be reduced, and geological prior knowledge is added to constrain the loss function. The extracted variables of the geological and geochemical spatial coupling correlation extracted by convolution operations are retained, which can accelerate network convergence and improve the generalization ability of the algorithm, making the prediction results more consistent with geological significance and have a certain degree of interpretability.

[0097] 5) Use the trained initial sparse autoencoder network as the sparse autoencoder network.

[0098] Based on the above description, the entire implementation architecture of building a sparse autoencoder network is as follows Figure 6The specific workflow of mineralization anomaly recognition based on convolutional sparse autoencoder network and isolation forest mainly includes three stages (such as Figure 5 to identify mineralized anomalies without relying on prior knowledge.

[0099] In the first stage, the isolation forest is used to extract background samples. Samples with lower anomaly scores are used as background samples. Based on the image samples composed of multi-source mineral-containing factors, an isolation forest is established, the anomaly score of each sample is calculated, and a suitable threshold is selected. When the anomaly score of a sample is less than this threshold, the sample is classified as a background sample, otherwise it is classified as an anomaly sample. In this process, training samples are obtained through a convolutional sparse autoencoder network.

[0100] In the second stage, the background samples obtained from the isolation forest are uniformly input into the convolutional sparse autoencoder network for training. The weights of the network are continuously adjusted through back propagation until the overall loss function converges to a smaller value. In this way, a reconstruction model of background samples with smaller reconstruction errors is established. Then, the average reconstruction error of these background samples is calculated to calculate the E value as the standard for judging mineralized abnormal samples.

[0101] In the third stage, a sample is input into the trained convolutional sparse autoencoder network to obtain the reconstruction error E' of the sample. When E' is greater than E, the sample is considered to be a mineralized anomaly. The farther E' is from E, the higher the possibility of finding a deposit in the area.

[0102] Step 102: Input the background sample into the sparse autoencoder network to obtain the detection result.

[0103] Step 103: Based on the detection results, the prospective mineral resource areas are identified.

[0104] An example is provided below to illustrate the construction process of the sparse autoencoder network provided above in the present invention and the detection effect it has.

[0105] In order to establish a unified image format for multi-source geographic information, this example converts geological data into a buffer distance map and geochemical data into a geochemical concentration map by interpolation. Each color predictor image is converted into a grayscale single-channel image so that each grayscale predictor image has the same image size of 5760 pixels × 2160 pixels. According to the prediction results, a suitable grid size is selected to divide each 80-pixel grayscale prediction image into grids to obtain 72×27 sample grids. For a certain grid position, the 11 single-channel prediction images at the same position of the grid are formed into input samples of several multi-channel grids and input into the convolutional sparse autoencoder network with isolation forest. Then, 58 grids with mineral deposits are selected as positive samples. All grid samples have the same image size of 80 pixels × 80 pixels × 11 channels. The specific construction process of the input grid sample is as follows. Figure 7 shown.

[0106] Set the number of isolated trees to 100 and the number of samples in the input subset to 256. Input all grid samples into the isolation forest for training to obtain the anomaly score. Figure 8 is the anomaly score of the frequency histogram, where the abscissa is the anomaly score and the ordinate is the corresponding density value calculated by dividing the frequency by the interval width.

[0107] The fractal dimension of the ore deposit was analyzed and the radius density power law function was established. The fractal dimension D corresponding to the number N(ε) of ore deposits in the study area with different box sizes was fitted with the slope of the straight line in the logarithmic graph of N(ε) and the number of boxes, as shown in Fig. 9 As shown. For a range of less than or equal to 16 km, D = 0.46, representing the local scale. The power law function is N(ε) = 515.19ε -0.46 (ε<16 km). The λ value ranges from 0 to 1 to represent the density of mineral deposits under the guidance of the probability fractal theory of the distance to known mineral deposits.

[0108] The lower the outlier value, the higher the probability that the sample is a background sample. The higher the outlier value, the higher the probability that the sample is an abnormal sample. Based on this, appropriate thresholds were selected and different background sample sets were constructed, with thresholds of 0.36, 0.37, 0.38, 0.39, and 0.40, respectively. The background sample set of the mineralization anomaly recognition model based on the convolutional sparse autoencoder network was trained using these different features. Then, the ability to predict positive samples (values) and the sensitivity of these mineralization anomalies (such as Fig.10The model trained based on background samples is shown in Figure 1. When the threshold is 0.37, it has a better performance. Therefore, the anomaly score of 0.37 can be selected as the upper limit of the background sample. In other words, the samples with anomaly scores less than 0.37 are selected as background samples. 58 background samples are randomly selected as test samples, and the remaining background samples are used to participate in the training of the convolutional sparse autoencoder network.

[0109] In the mineralization anomaly recognition, the performance of the convolutional sparse autoencoder network (CSAE) and the traditional convolutional autoencoder network (CAE model) were compared. The number of applied iterations was set to 1000, the learning rate was set to 0.01, the expected sparsity constraint was set to 0.01, and the sparsity weight was set to 0.01. After training, the average reconstruction error of the background samples of the CAE and CSAE models was: 0.007085 for the CAE model. 0.029146 for the CSAE model. The mineralization anomaly recognition model was tested after training using 58 background samples (negative samples) and 58 known samples with combined storage samples (positive samples).

[0110] The average value of the reconstruction error of the background sample is taken as the metric of the mineralization anomaly sample. When the reconstruction error of the test sample is greater than the average reconstruction error of the background sample, the sample is considered to be a mineralization anomaly sample. Otherwise, it will be identified as a background sample. By comparing with the true label of the test sample, in order to evaluate the effectiveness of the mineralization anomaly recognition model, the prediction success rate curve (such as Fig.11 As shown in Figure 2, it is used to evaluate the prediction results. Fig.11 It can be seen that the success rate curve of the CSAE model on the upper left is steeper. In a smaller prospect area (16.77%), the CSAE model can accurately identify more known deposits (68.96%). It can be seen that the CSAE model established for mineralization anomaly identification has higher accuracy than the CAE model. With the addition of sparse constraints and geological constraints, the CSAE model has a higher correct recognition rate for both positive and negative samples.

[0111] Based on the above description, the present invention has the following advantages over the prior art:

[0112] 1. The present invention adopts a convolutional sparse autoencoder network to realize the identification of mineralization anomalies. This method retains the spatial coupling correlation of geological and geochemical variables extracted by adding convolution operations, and learns the spatial coupling relationship between mineralization characteristics according to the mineralization location, which is beneficial to the results of geological structure analysis. The sparse terms are established using the ReLU activation function, and sparse constraints are added to the target loss function to improve the generalization ability of the algorithm. The entire network reduces redundant features by suppressing multiple feature units. In addition, the isolation forest is used as an autonomous extractor of background multi-channel image samples, which overcomes the limitation of traditional autoencoder networks that rely on labeled non-mineralized samples. Experiments have shown that the combined application of convolutional sparse autoencoders and isolation forests can improve the accuracy and efficiency of predictions.

[0113] 2. To solve the overfitting problem caused by excessive feature redundancy caused by one-dimensional vectors, the present invention proposes to use a convolutional sparse autoencoder network to identify mineralization anomalies, which can directly input multi-channel indicator factor images and retain spatial feature extraction between multi-source geographic information. Through the convolution structure, in order to reduce redundant features, the present invention introduces two sparse methods, ReLU function and sparse constraint factors in the loss function, which can suppress multiple feature units and improve generalization ability. In addition, the present invention lacks non-mineral sample areas on the label.

[0114] 3. Manual selection of background samples will be greatly affected by subjectivity. In order to solve this limitation, the present invention uses unsupervised isolation forest as an automatic extractor of the background, which does not require prior knowledge.

[0115] 4. The present invention introduces geological prior knowledge as a constraint into the training process, which can improve the interpretability of machine learning and accelerate network convergence, making the prediction results more consistent with geological significance. Data-driven + geological knowledge-driven mineral deposit prediction is a major trend in future deep learning.

[0116] Example 2

[0117] A computer device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method for delineating prospective mineral resource areas in Example 1.

[0118] Example 3

[0119] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the method for delineating a prospective mineral resource area in Example 1.

[0120] Example 4

[0121] A computer program product includes a computer program, which, when executed by a processor, implements the steps of the method for delineating prospective mineral resource areas in Example 1.

[0122] Example 5

[0123] A computer device, which may be a database. The computer device includes a processor, a memory, an input / output interface (I / O for short) and a communication interface. The processor, the memory and the input / output interface are connected via a system bus, and the communication interface is connected to the system bus via the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store pending transactions. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, the method for delineating a prospective area of ​​mineral resources in Example 1 is implemented.

[0124] It should be noted that the object information (including but not limited to object device information, object personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in the present invention are all information and data authorized by the object or fully authorized by all parties, and the collection, use and processing of relevant data need to comply with relevant laws, regulations and standards of relevant countries and regions.

[0125] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided by the present invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided by the present invention may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited thereto. The processor involved in each embodiment provided by the present invention may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but are not limited thereto.

[0126] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0127] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core ideas of the present invention. The same and similar parts between the embodiments can be referred to each other. At the same time, for those skilled in the art, according to the ideas of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A method for delineating prospective mineral resource areas. It is characterized in that The method comprises: Isolation forests were used to extract background samples of the mineral deposits to be tested; Build a sparse autoencoder network; Inputting the background sample into the sparse autoencoder network to obtain a detection result; Based on the detection results, the prospective area of ​​mineral resources is identified and delineated; Isolation forest is used to extract background samples of the mineral deposits to be tested, including: Constructing a sample subset based on the multi-channel image of the mineral deposit to be detected; the sample subset has V pixels at the root node of the tree; A dimension is randomly selected from all channels, and a split value is randomly generated; the split value is a random value from the minimum to the maximum value of the pixel point of the current node in the selected channel; Splitting the split value to divide the pixel points of the current node into sub-nodes in the selected channel; Recursively perform pixel division on the child nodes until there is only one pixel in the child node or the formed isolated tree reaches the set height limit to generate multiple isolated trees; The generated multiple isolated trees are used to determine the normalized anomaly score; The pixels whose normalized anomaly score is less than a set value are taken as background pixels to generate background samples; Construct a sparse autoencoder network, including: The spatial distribution of known mineral deposits is quantified using self-similar fractal methods to obtain prior knowledge; Adding sparse constraints to the convolutional autoencoder network, and using the prior knowledge as geological constraints to obtain an initial sparse autoencoder network; Use Lone Tree to separate the background information of known mineral deposits to obtain background samples; Separating the background information of the known mineral deposit to obtain background samples as training samples, and using the training samples to train the initial sparse autoencoder network to obtain a trained initial sparse autoencoder network; Using the trained initial sparse autoencoder network as the sparse autoencoder network; The sparse constraint is added to the convolutional autoencoder network, and the prior knowledge is used as a geological constraint to obtain an initial sparse autoencoder network, including: The ReLU activation function is used to introduce sparse constraints through convolution and deconvolution layers; when the output value of the feature unit passing the ReLU activation function is less than zero, the feature unit is replaced by zero; the ReLU activation function is expressed as: In the formula, x i represents the background sample, y i Represents reconstructed data; Add another sparse constraint as a penalty factor for the mean squared error between input and output; another sparse constraint is the reconstruction error; When the background sample x is input i (i={1,2,…,n}) The activation value of feature unit k is A k (x i ), the average activation value of this feature unit over all input background samples is calculated using the formula Assume that the constant ρ is the desired sparse parameter and the average activation value of the measured background samples is The relative entropy of the distance between Introduce the loss function; where the average activation value of the background sample The formula for determining the relative entropy of the distance between and the constant ρ is: The loss function of the sparse autoencoder network is: In the formula, Loss SCAE Represents the loss function value of the sparse autoencoder network, Loss MSE represents the ReLU activation function value, β represents the weight coefficient controlling the sparse constraint, λ represents the penalty coefficient factor, c_loss represents the geological constraint, i = {1,2,…,n}; The prior knowledge obtained from fractal analysis is used as the geological constraint c_loss, and the spatial distribution probability is used as the additional geological term constraint model; the Tanh activation function is added to convert the reconstruction error Y_error of the ReLU activation function into the probability P ano ; The Tanh activation function is expressed as: The mean square error function is used to measure the probability P ano The difference between and the weight within the distance d; the geological constraints based on the mean square error function are expressed as: In the formula, x represents the area around the deposit, P ano = {P i |i=1,2,…,N} and Y={Y i |i=1,2,…,N} is the probability value, Y is the mineralization probability, and N is the number of pixels; The geological constraint c_loss calculates the model output as the difference between a predefined anomaly probability value and the error.

2. The method for delineating a prospective mineral resource area according to claim 1, It is characterized in that The set value is 0.

5.

3. The method for delineating a prospective mineral resource area according to claim 1, It is characterized in that Splitting the split value to divide the pixel points of the current node into sub-nodes in the selected channel specifically includes: When the value of a pixel in the selected channel is less than the split value after segmentation, this pixel is used as the left child node of the current node; When the value of a pixel in the selected channel is greater than or equal to the split value after segmentation, this pixel is used as the right child node of the current node.

4. The method for delineating a prospective mineral resource area according to claim 1, It is characterized in that The self-similar fractal method adopted is the box counting method.

5. A computer device, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method for delineating prospective mineral resource areas as described in any one of claims 1 to 4.

6. A computer-readable storage medium having a computer program stored thereon, It is characterized in that When the computer program is executed by a processor, the steps of the method for delineating a prospective mineral resource area described in any one of claims 1 to 4 are implemented.

7. A computer program product comprising a computer program, It is characterized in that When the computer program is executed by a processor, the steps of the method for delineating a prospective mineral resource area described in any one of claims 1 to 4 are implemented.

Citation Information

Patent Citations

  • Power operation and maintenance data cleaning method based on isolation forest algorithm and neural network

    CN108776683A

  • Coal mine underground human body posture monitoring method

    CN110163347A

  • Submarine sulfide mineralization quantitative prediction method based on machine learning

    CN114330841A

  • Method and system for training positive sample by utilizing mineralization prediction of information amount expansion

    CN115272798A

  • Mineral prediction method and system based on multi-scale sample unevenness

    CN115511214A