Mineral resource classification prediction method and system based on multi-source small sample joint learning

Abstract

The present invention discloses a method and system for mineral resource classification prediction based on multi-source small sample joint learning. The method includes: pre-processing to construct a meta-training task; establishing an abnormal sample mining module for collaborative adversarial training at the latent space feature level and the sample level; training the abnormal sample mining module and generating samples to obtain highly abnormal samples; establishing a feature joint embedding module to obtain an initial class prototype, and randomly selecting highly abnormal samples for fine-tuning to obtain a final class prototype; and predicting the target task through iterative training of the network through the meta-training task. The present invention can effectively improve the quality of generated samples and the performance of the sample generator, so that the model can fully mine the deep features of multivariate geophysical and chemical data and enhance the ability to fit the true distribution of the data; the introduction of geological constraints can learn the metallogenic laws within the multi-source samples to a greater extent, and fully improve the robustness and accuracy of the mineral resource classification prediction model.

Description

Technical Field

[0001] The present invention relates to a mineral resource classification prediction method and system based on multi-source small sample joint learning, belonging to the fields of geoinformation science and computer vision. Background Art

[0002] Mineral prediction involves analyzing and evaluating various geological, geophysical, and geochemical data, combined with historical mining data, to predict the presence, size, and distribution of mineral deposits, providing a scientific basis for mineral exploration and mining. The development and utilization of mineral resources is crucial to national economic and social development, and therefore, mineral prediction has long been a hot topic in geological exploration. Traditional mineral prediction methods rely primarily on the experience and expertise of geological exploration engineers, which are subjective and limited, requiring significant time, manpower, and material resources. In recent years, with the continuous development and application of artificial intelligence (AI) technology, an increasing number of researchers have begun exploring its application to mineral prediction to improve accuracy and efficiency. Through automated learning and data mining, AI can unlock the deep insights of mineral exploration data, build predictive models, and achieve more accurate and rapid mineral predictions.

[0003] In recent years, the application of artificial intelligence technology in mineral prediction has gradually achieved certain results. However, due to the small sample problem in mineral resource prediction: the samples in the ore deposit area are very scarce compared to the background area samples, the imbalance of the data greatly limits the prediction performance of the model, resulting in the model being unable to fully explore the channels and spatial characteristics within the mineral data, and the generalization ability is weak, making it impossible to accurately predict the mineralization of the study area. Summary of the Invention

[0004] Purpose of the invention: In view of the shortcomings of the existing technology, the purpose of the present invention is to propose a mineral resource classification prediction method and system based on multi-source small sample joint learning, so as to more comprehensively and objectively find the laws between geophysical and chemical data, geological data and mineralization, so as to predict the probability of mineralization in new tasks.

[0005] Technical solution: To achieve the above-mentioned purpose of the invention, the present invention provides a method for mineral resource classification prediction based on multi-source small sample joint learning, comprising the following steps:

[0006] Step 1: Divide the study area into grid data, mark the mineralization area and background area, and construct multiple meta-training tasks;

[0007] Step 2: Establish an abnormal sample mining module for collaborative adversarial training at the latent space feature level and the sample level, which includes a latent space generation adversarial submodule and a sample generation adversarial submodule; the latent space generation adversarial submodule includes a latent space adversarial generator and a latent space adversarial discriminator, and the sample generation adversarial module includes a sample adversarial generator and a sample adversarial discriminator; the latent space features generated by the latent space adversarial generator are used as input to the sample adversarial generator to generate samples;

[0008] Step 3: Train the abnormal sample mining module until the network converges, and generate a corresponding false sample for each real sample in the grid data of the study area;

[0009] Step 4: Take the weighted average of the sample reconstruction error and feature reconstruction error between the generated sample and the real sample as the anomaly score of each generated sample, and sort the generated samples in descending order according to the anomaly score, retaining the first part of the samples as high anomaly samples;

[0010] Step 5: Establish a joint feature embedding module for the graph convolutional network and the convolutional neural network. Input the geochemical data of the same sample into the convolutional neural network for geochemical feature extraction and generate a geochemical feature map. Input the geological data into the graph convolutional network for geological feature extraction and generate a geological feature map. The geochemical feature map and the geological feature map are spliced ​​in the channel dimension and then flattened as the embedded representation of the sample in the metric space.

[0011] Step 6: Use the feature joint embedding module to calculate the initial class prototype of each category in the support set;

[0012] Step 7: Randomly select some high-abnormal samples and map them into the metric space to fine-tune the initial class prototype to obtain the final class prototype for each category;

[0013] Step 8: Calculate the distance between the query set sample embedding and the final class prototype of each category, complete the category division, calculate the loss, and backpropagate to update the network weights;

[0014] Step 9: Complete the iteration of the meta-training task and generate a mineral resource prediction map for the target prediction task.

[0015] Furthermore, each meta-training task in step 1 contains N mineral categories, and K samples are randomly sampled from each category to form a support set S = {(x1, y1), (x2, y2), ..., (x N×K ,y N×K )}, randomly sample T samples from the remaining samples to form a query set For the training of subsequent meta-learning models. s ,y s ), xs and y s Represents the s-th support set sample and its label (s=1,2,…,N×K); for the sample pair in the query set Q and They represent the qth query set sample and its label (q=1, 2,…, T) respectively.

[0016] Furthermore, the step 2 includes the following steps:

[0017] Step 21: Establish a latent space generation adversarial submodule, which consists of a convolutional autoencoder CAE and a latent space adversarial generator G L and latent space adversarial discriminator D L Composition; Convolutional Autoencoder CAE consists of Encoder and Decoder, which serves as the real branch of the latent space generation adversarial submodule; the preprocessed sample x is passed through the Encoder to obtain the real sample latent space features; the real sample latent space features are used as the input of the Decoder to obtain the reconstructed sample. The encoding and decoding process is optimized by the reconstruction error of the real sample and the reconstructed sample, which is expressed by the L2 norm of the difference between the real sample and the reconstructed sample, as shown below:

[0018]

[0019] Among them, E is the expected value, x~p data (x) represents x from the data distribution p data (x) is a randomly drawn sample.

[0020] Latent space adversarial generator G L and latent space adversarial discriminator D L As a false branch of the latent space generation adversarial submodule; randomly sample noise Z from the Gaussian space as the latent space adversarial generator G L Input, G L Generate false latent space features, and input the generated false latent space features and the real sample latent space features Encoder(x) into the latent space adversarial discriminator D L The adversarial loss is optimized in

[15] , and the loss function of the latent space generation adversarial submodule is as follows:

[0021]

[0022] Among them, p z (x) represents the noise data distribution.

[0023] Step 22: Establish a sample generation adversarial submodule, which is composed of the sample adversarial generator G I and sample adversarial discriminator D IComposition; use the false latent space features generated in step 21 as samples to counter the generator G I Input, generate the corresponding false sample G I (G L (z)), and send it together with the real sample x into the sample adversarial discriminator D I The gradient penalty loss is optimized in

[15] , and the loss function of the sample generation adversarial module is as follows:

[0024]

[0025] Among them, gradient_penalty is the gradient penalty term, ω and λ are constants.

[0026] Furthermore, the step 5 includes the following steps:

[0027] Step 51: Assume that the sample x in the study area i The corresponding geochemical data are The corresponding geological data are Converting geological data into graph-structured data To facilitate graph convolutional network processing; specifically, each pixel in the geological data image is converted into a node, and the data of different channels represent the attribute information of each node. The nodes with a distance less than or equal to a preset threshold are set as their neighbor nodes, and the edge information is stored;

[0028] Step 52: Let the feature extractor of the convolutional neural network be f, and the feature extractor of the graph convolutional network be g. f and g output feature maps of the same size. In order to improve the convolutional neural network's ability to mine channel features and spatial features, add a channel attention module and a spatial attention module to f.

[0029] Step 53: Geochemical data Input the feature extractor f of the convolutional neural network to generate a geochemical exploration feature map The corresponding geological map structure data Input the feature extractor g of the graph convolutional network to generate a geological feature map Concatenate the two feature maps in the channel dimension and then flatten them as the embedding representation emb of the sample in the metric space i , as shown below:

[0030]

[0031] Among them, concat is the splicing operation of the channel dimension, and flatten is the flattening operation to convert the feature map into a vector.

[0032] Furthermore, in step 6, the initial class prototype of each category is calculated on the support set, and the prototype proto of category c is c The calculation process is as follows:

[0033]

[0034] in, y i is the label corresponding to the i-th support set sample, y i =c when w i,c =1, otherwise w i,c =0,emb i is the embedding representation of the i-th support set sample.

[0035] Furthermore, step 7 includes the following steps:

[0036] Step 71: Randomly extract k unlabeled samples {x1 * ,x2 * ,…,x k *}, after forming the embedding in the metric space through the feature joint embedding module, the unlabeled sample x is calculated based on the Euclidean distance between the unlabeled sample and each initial class prototype. i * The probability p(c|x i * ,{proto c}), the category with the highest probability is used as its label information, p(c|x i * ,{proto c The calculation process of}) is as follows:

[0037]

[0038] Among them, emb represents the feature joint embedding module, emb(x i * ) represents the embedding representation of the i-th unlabeled sample, i = 1, 2, ..., k, For category c * The initial class prototype of

[0039] Step 72: After assigning the category to the unlabeled high anomaly samples, fine-tune all class prototypes. The final class prototype proto of category c is c * The calculation process is as follows:

[0040]

[0041] in,

[0042] emb(x i ) is the embedding representation of the i-th labeled sample, emb(x j * ) is the embedding representation of the j-th unlabeled sample.

[0043] Furthermore, in step 8, the samples in the query set Q are embedded by the feature joint embedding module emb, and the distance between the sample and the final prototype of each category is calculated using the distance metric, and the category with the smallest distance is selected as the predicted label of the query set sample; for the samples in the query set Q Calculate the loss and back-propagate to update the network weights, the loss L p The definition is as follows:

[0044]

[0045] Among them, T is the number of query set samples, d is the Euclidean distance metric, is the embedding representation of the t-th query set sample, is the true category of the t-th query set sample The corresponding final class prototype, For category c * The final class prototype.

[0046] Based on the same inventive concept, the present invention provides a mineral resource classification prediction system based on multi-source small sample joint learning, comprising:

[0047] The data preprocessing module is used to divide the study area into grid data, mark the mineralization area and background area, and construct multiple meta-training tasks;

[0048] An unsupervised anomaly mining module is used to establish an abnormal sample mining module for collaborative adversarial training at the latent space feature level and the sample level, which includes a latent space generation adversarial submodule and a sample generation adversarial submodule; the latent space generation adversarial submodule includes a latent space adversarial generator and a latent space adversarial discriminator, and the sample generation adversarial module includes a sample adversarial generator and a sample adversarial discriminator; the latent space features generated by the latent space adversarial generator are used as input to the sample adversarial generator to generate samples; the abnormal sample mining module is trained until the network converges, and a corresponding false sample is generated for each real sample in the grid data of the study area; and the weighted average of the sample reconstruction error and the feature reconstruction error between the generated sample and the real sample is used as the abnormal score of each generated sample, and the generated samples are sorted in descending order according to the high and low abnormal scores, and the first part of the samples are retained as high abnormal samples;

[0049] The prototype representation module is used to establish a feature joint embedding module for the graph convolutional network and the convolutional neural network. The geochemical data of the same sample is input into the convolutional neural network for geochemical feature extraction to generate a geochemical feature map. The geological data is input into the graph convolutional network for geological feature extraction to generate a geological feature map. The geochemical feature map and the geological feature map are spliced ​​in the channel dimension and then flattened as the embedded representation of the sample in the metric space. The feature joint embedding module is used to calculate the initial class prototype of each category in the support set. A portion of highly abnormal samples are randomly selected and mapped into the metric space to fine-tune the initial class prototype to obtain the final class prototype of each category.

[0050] The network training module is used to determine the category of the query set sample by calculating the distance between the embedding and the final class prototype, calculate the loss, and backpropagate to update the network weights until all meta-training tasks are iterated and the final model is saved;

[0051] The target task prediction and visualization module is used to use the trained network model to perform mineralization prediction on the target task and generate a mineral resource prediction map.

[0052] Based on the same inventive concept, the present invention provides a computer system comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded into the processor, the method for mineral resource classification and prediction based on multi-source small sample joint learning is implemented.

[0053] Based on the same inventive concept, the present invention also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the mineral resource classification prediction method based on multi-source small sample joint learning.

[0054] Beneficial effect: The present invention provides a mineral resource classification and prediction method and system based on multi-source small sample joint learning, which has the following advantages: in order to more accurately and objectively find the relationship between geophysical and chemical data, geological data and mineralization laws, the present invention first establishes a sample generation adversarial submodule and a latent space generation adversarial submodule to mine the characteristics of abnormal samples in the study area. Compared with the traditional use of random noise as the input of the generator, the generator's ability to fit the true distribution of the data is greatly enhanced, and the quality of the generated samples is continuously improved; secondly, the present invention introduces geological knowledge constraints and, according to the characteristics of multi-source data, uses convolutional neural networks and graph convolutional networks with channel attention and spatial attention to extract features of geochemical data and geological data respectively and integrate geochemical knowledge and geological knowledge, so that the characterization ability of category prototypes is greatly enhanced; finally, the present invention fine-tunes the category prototype by introducing highly abnormal unlabeled samples, which to a certain extent alleviates the prototype selection sensitivity problem of the prototype network, effectively improves the generalization ability and robustness of the network model, and can greatly improve the model accuracy under limited mineral deposit samples, making the final result more accurate and reliable. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 is an overall flow chart of an embodiment of the present invention;

[0056] Figure 2 This is a network structure diagram of the abnormal sample mining module in the present invention;

[0057] Figure 3 This is a network structure diagram of the feature joint embedding module in the present invention. DETAILED DESCRIPTION

[0058] The present invention is further illustrated below with reference to specific examples. It should be understood that these examples are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, modifications of various equivalent forms of the present invention made by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0059] like Figure 1 As shown, the embodiment of the present invention discloses a mineral resource classification prediction method based on multi-source small sample joint learning, which mainly includes the following steps:

[0060] Step 1: Divide the study area into grid data according to a certain superpixel size. The grid data contains geochemical and geological information, which is mainly reflected in different channels within the data. If there is a mineral deposit in the grid, it is set as a mineralization area, otherwise it is set as a background area. After preprocessing the mineral deposit samples and background area samples, a spatiotemporal database containing geophysical, chemical and geological data is established. On this basis, M meta-training tasks are constructed.

[0061] Specifically, the grid data of the study area in step 1 includes the longitude and latitude coordinates of the known mining area, the contents of various geochemical elements in the known mining area, various geophysical data information of the known mining area, and various geological data information of the known mining area, which are obtained by dividing the geological image formed by the inverse distance interpolation method in ArcGIS software; on this basis, combined with the presence or absence of mineral deposits in the grid, each grid data is set as a mineral deposit sample or background area sample. After preprocessing, a spatiotemporal database containing geophysical and chemical data and geological data is established, and M meta-training tasks are constructed. Specifically, each meta-training task contains N mineral categories, and K samples are randomly sampled from each category to form a support set S = {(x1, y1), (x2, y2),…, (x N×K ,y N×K )}, randomly sample T samples from the remaining samples to form a query set For the training of subsequent meta-learning models. s ,y s ), x s and y s Represents the s-th support set sample and its label (s=1,2,…,N×K); for the sample pair in the query set Q and They represent the qth query set sample and its label (q=1, 2,…, T) respectively.

[0062] Step 2: Establish an abnormal sample mining module that performs collaborative adversarial training at the latent space feature level and the sample level. This module includes two submodules: a latent space generative adversarial submodule (including a latent space adversarial generator and a latent space adversarial discriminator) and a sample generative adversarial submodule (including a sample adversarial generator and a sample adversarial discriminator). The latent space features generated by the latent space adversarial generator replace the traditional random noise and serve as the input of the sample adversarial generator to generate samples.

[0063] Step 2 includes the following steps:

[0064] Step 21: Establish a latent space generation adversarial submodule, which consists of a convolutional autoencoder CAE and a latent space adversarial generator G L and latent space adversarial discriminator D L Composition; Convolutional Autoencoder CAE consists of Encoder and Decoder, which serves as the real branch of the latent space generation adversarial submodule; the preprocessed sample x is passed through the Encoder to obtain the real sample latent space features; the real sample latent space features are used as the input of the Decoder to obtain the reconstructed sample. The encoding and decoding process is optimized by the reconstruction error of the real sample and the reconstructed sample, which is expressed by the L2 norm of the difference between the real sample and the reconstructed sample, as shown below:

[0065]

[0066] Among them, E is the expected value, x~p data (x) represents x from the data distribution p data (x) is a randomly drawn sample.

[0067] Latent space adversarial generator G L and latent space adversarial discriminator D L As a false branch of the latent space generation adversarial submodule; randomly sample noise Z from the Gaussian space as the latent space adversarial generator G L Input, G L Generate false latent space features, and input the generated false latent space features and the real sample latent space features Encoder(x) into the latent space adversarial discriminator D L The adversarial loss is optimized in

[15] , and the loss function of the latent space generation adversarial submodule is as follows:

[0068]

[0069] Among them, p z (z) represents the noise data distribution.

[0070] Step 22: Establish a sample generation adversarial submodule, which is composed of the sample adversarial generator G I and sample adversarial discriminator D I Composition; use the false latent space features generated in step 21 as samples to counter the generator G I Input, generate the corresponding false sample G I (G L (z)), and send it together with the real sample x into the sample adversarial discriminator D I The gradient penalty loss is optimized in

[15] , and the loss function of the sample generation adversarial module is as follows:

[0071]

[0072] Among them, gradient_penalty is the gradient penalty term, ω and λ are constants.

[0073] Step 3: Use all the geochemical data of the study area to train the anomaly sample mining module until the network converges. Suppose the study area is divided into s grid samples. After reaching Nash equilibrium, each grid sample x in the study area is i (i=1,2,…,s) is input into the abnormal sample mining module, and the sample adversarial generator G I Generate corresponding false samples x′ i .

[0074] Step 4: Use the pre-trained anomaly sample mining module to generate samples, and take the weighted average of the sample reconstruction error and feature reconstruction error between the generated sample and the real sample as the anomaly score of each generated sample. Then sort the generated samples in descending order according to the anomaly score, and retain the top n highly anomaly samples.

[0075] Step 4 includes the following steps:

[0076] Step 41: For each true-generated sample pair (x i ,x′ i ), use the weighted average of the sample reconstruction error and feature reconstruction error of the real sample and the generated sample as the anomaly score AnomalyScore of each generated sample i , AnomalyScore i The calculation formula is as follows:

[0077] AnomalyScore i =α1*‖x i -x i ′‖2+α2*‖Encoder(x i )-Encoder(x i ′)‖2

[0078] Where: Encoder is the encoder pre-trained in step 3 in the latent space generative adversarial network module, and α1 and α2 are constants.

[0079] Step 42: Sort the anomaly scores of all generated samples in descending order from high to low, and retain the first k generated samples as high anomaly samples.

[0080] Step 5: Establish a feature joint embedding module of graph convolutional network and convolutional neural network, input the geochemical data of the same sample into the convolutional neural network for geochemical feature extraction, generate a geochemical feature map, input the geological data into the graph convolutional network for geological feature extraction, generate a geological feature map, splice the geochemical feature map and the geological feature map in the channel dimension, and then flatten them as the embedded representation of the sample in the metric space.

[0081] Step 5 includes the following steps:

[0082] Step 51: Assume that the sample x in the study area i The corresponding geochemical data are The corresponding geological data are Converting geological data into graph-structured data To facilitate graph convolutional network processing; specifically, each pixel in the geological data image is converted into a node, and the data of different channels represent the attribute information of each node. The nodes with a distance less than or equal to a preset threshold are set as their neighbor nodes, and the edge information is stored;

[0083] Step 52: Let the feature extractor of the convolutional neural network be f, and the feature extractor of the graph convolutional network be g. f and g output feature maps of the same size. To improve the convolutional neural network's ability to mine channel features and spatial features, a channel attention module and a spatial attention module are added to f. In the channel attention module, the shallow feature map is input into the channel attention module to learn the channel-based attention weights and multiply them with the shallow feature map to obtain a feature map fused with channel attention. Specifically, the input feature map I (H×W×C) is first subjected to global maximum pooling and global average pooling based on sample width W and height H, respectively; then, the features aggregated by global maximum pooling and global average pooling are input into a two-layer fully connected neural network, with the number of neurons in the first layer being C / R (C is the number of channels in feature map I, R is the decay rate), and the number of neurons in the second layer being C; then, the output feature maps are element-wise added and nonlinearly mapped using an activation function to generate a channel attention weight map; finally, the channel attention weight map is element-wise multiplied with the original feature map to obtain the channel attention feature map map. c , map c The calculation formula is as follows:

[0084] map c =σ(MLP(MaxPool(I))+MLP(AvgPool(I)))

[0085] Among them, MLP is the fully connected mapping function, MaxPool is the global maximum pooling, AvgPool is the global average pooling, and σ is the activation function.

[0086] The spatial attention module is added after the channel attention module to map the channel attention feature map c Input into the spatial attention module to learn the spatial attention weights and map c Multiply them together to get the feature map of fusion channel and spatial attention. Specifically, first map the channel attention feature map c(H×W×C) is used to compress the channel dimension, and the global maximum pooling and global average pooling based on the channel dimension are performed respectively; then the features aggregated by the global maximum pooling and the features aggregated by the global average pooling are spliced, and the spliced ​​feature map is compressed in the channel dimension with a 7×7 convolution kernel to generate a spatial attention weight map; finally, the spatial attention weight map is element-wise multiplied with the channel attention feature map to obtain the feature map of the fusion channel and spatial attention. s , drawing on the idea of ​​residual, the deep feature map of the fusion attention s Added to the initial shallow feature map I, as the final feature map output, map s The calculation formula is as follows:

[0087] map s =σ(Conv 7×7 ([Avgpool(map c );Avgpool(map c )]))

[0088] Among them, Conv 7×7 It is a 7×7 convolution operation.

[0089] Step 53: Geochemical data Input the feature extractor f of the convolutional neural network to generate a geochemical exploration feature map The corresponding geological map structure data Input the feature extractor g of the graph convolutional network to generate a geological feature map Concatenate the two feature maps in the channel dimension and then flatten them as the embedding representation emb of the sample in the metric space i , as shown below:

[0090]

[0091] Among them, concat is the splicing operation of the channel dimension, and flatten is the flattening operation to convert the feature map into a vector.

[0092] Step 6: Calculate the initial class prototype of each category on the support set, the prototype proto of category c c The calculation process is as follows:

[0093]

[0094] in, y i is the label corresponding to the i-th support set sample, emb i is the embedding representation of the i-th support set sample.

[0095] Step 7: Randomly select k high-abnormal samples and map them into the metric space to fine-tune the initial class prototype to obtain the final class prototype for each category.

[0096] Step 7 includes the following steps:

[0097] Step 71: Randomly extract k unlabeled samples {x1 * ,x2 * ,…,x k *}, after forming the embedding in the metric space through the feature joint embedding module, the unlabeled sample x is calculated based on the Euclidean distance between the unlabeled sample and each initial class prototype. i * The probability p(c|x i * ,{proto c}), the category with the highest probability is used as its label information, p(c|x i * ,{proto c The calculation process of}) is as follows:

[0098]

[0099] Among them, emb represents the feature joint embedding module, emb(x i * ) represents the embedding representation of the i-th unlabeled sample, i = 1, 2, ..., k, For category c * The initial class prototype of

[0100] Step 72: After assigning the category to the unlabeled high anomaly samples, fine-tune all class prototypes. The final class prototype proto of category c is c * The calculation process is as follows:

[0101]

[0102] in,

[0103] emb(x i ) is the embedding representation of the i-th labeled sample, emb(x j * ) is the embedding representation of the j-th unlabeled sample.

[0104] Step 8: The samples in the query set Q are embedded by the feature joint embedding module emb, and the distance between them and the final prototype of each category is calculated using the distance metric. The category with the smallest distance is selected as the predicted label of the query set sample; for the samples in the query set Q Calculate the loss and back-propagate to update the network weights, the loss L p The definition is as follows:

[0105]

[0106] Among them, T is the number of query set samples, d is the Euclidean distance metric, is the embedding representation of the t-th query set sample, is the true category of the t-th query set sample The corresponding final class prototype, For category c * The final class prototype.

[0107] Step 9: Complete the iterative process of the meta-training task, save the final model parameters, input the geophysical and chemical data and geological data in the target prediction task into the system, and use the sliding window algorithm to input the samples in the window into the model in turn to obtain the mineralization probability of each window, and finally form a mineral resource prediction map for the target task.

[0108] Based on the same inventive concept, a mineral resource classification and prediction system based on multi-source small sample joint learning includes: a data preprocessing module for dividing the study area into grid data, and marking it into mining areas and background areas, and constructing multiple meta-training tasks; an unsupervised anomaly mining module for establishing an abnormal sample mining module for collaborative adversarial training at the latent space feature level and the sample level, which includes a latent space generation adversarial submodule and a sample generation adversarial submodule; the latent space generation adversarial submodule includes a latent space adversarial generator and a latent space adversarial discriminator, and the sample generation adversarial module includes a sample adversarial generator and a sample adversarial discriminator; the latent space features generated by the latent space adversarial generator are used as input to the sample adversarial generator for generating samples; the abnormal sample mining module is trained until the network converges, and a corresponding false sample is generated for each real sample in the grid data of the study area; and the weighted average of the sample reconstruction error and the feature reconstruction error of the generated sample and the real sample is used as the abnormal score of each generated sample, and the generated samples are ranked according to the high or low abnormal score. Sort in descending order and retain the first part of the samples as high-anomaly samples; the prototype representation module is used to establish a feature joint embedding module of the graph convolutional network and the convolutional neural network, input the geochemical data in the same sample into the convolutional neural network for geochemical feature extraction, generate a geochemical feature map, input the geological data into the graph convolutional network for geological feature extraction, generate a geological feature map, splice the geochemical feature map and the geological feature map in the channel dimension, and then flatten them as the embedding representation of the sample in the metric space; use the feature joint embedding module to calculate the initial class prototype of each category in the support set; and randomly select some high-anomaly samples to map to the metric space to fine-tune the initial class prototype to obtain the final class prototype of each category; the network training module is used to determine the category to which the query set sample belongs by calculating the distance between the embedding and the final class prototype, calculate the loss, and backpropagate to update the network weights until all meta-training tasks are iterated and the final model is saved; the target task prediction and visualization module is used to use the trained network model to perform mineralization prediction on the target task and generate a mineral resource prediction map.

[0109] The specific working process of each module described above can refer to the corresponding process in the aforementioned method embodiment and will not be repeated here. The division of the modules is only a logical function division. In actual implementation, there may be other division methods, such as combining multiple modules or integrating them into another system.

[0110] Based on the same inventive concept, an embodiment of the present invention discloses a computer system, including a memory, a processor, and a computer program stored in the memory and runnable on the processor. When the computer program is loaded into the processor, it implements the mineral resource classification prediction method based on multi-source small sample joint learning.

[0111] Based on the same inventive concept, an embodiment of the present invention discloses a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the mineral resource classification prediction method based on multi-source small sample joint learning.

[0112] Those skilled in the art will appreciate that the technical solution of the present invention, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes instructions for causing a computer system (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in the embodiments of the present invention. Storage media include various media capable of storing computer programs, such as USB flash drives, mobile hard drives, read-only memories (ROMs), random-access memories (RAMs), magnetic disks, or optical disks.

Claims

1. A mineral resource classification prediction method based on multi-source small sample joint learning, characterized in that: The steps include: Step 1: Divide the study area into grid data, mark the mineralization area and background area, and construct multiple meta-training tasks; Step 2: Establish an abnormal sample mining module for collaborative adversarial training at the latent space feature level and the sample level, which includes a latent space generation adversarial submodule and a sample generation adversarial submodule; the latent space generation adversarial submodule includes a latent space adversarial generator and a latent space adversarial discriminator, and the sample generation adversarial module includes a sample adversarial generator and a sample adversarial discriminator; the latent space features generated by the latent space adversarial generator are used as input to the sample adversarial generator to generate samples; Step 3: Train the abnormal sample mining module until the network converges, and generate a corresponding false sample for each real sample in the grid data of the study area; Step 4: Take the weighted average of the sample reconstruction error and feature reconstruction error between the generated sample and the real sample as the anomaly score of each generated sample, and sort the generated samples in descending order according to the anomaly score, retaining the first part of the samples as high anomaly samples; Step 5: Establish a joint feature embedding module for the graph convolutional network and the convolutional neural network. Input the geochemical data of the same sample into the convolutional neural network for geochemical feature extraction and generate a geochemical feature map. Input the geological data into the graph convolutional network for geological feature extraction and generate a geological feature map. The geochemical feature map and the geological feature map are spliced ​​in the channel dimension and then flattened as the embedded representation of the sample in the metric space. Step 6: Use the feature joint embedding module to calculate the initial class prototype of each category in the support set; Step 7: Randomly select some high-abnormal samples and map them into the metric space to fine-tune the initial class prototype to obtain the final class prototype for each category; Step 8: Calculate the distance between the query set sample embedding and the final class prototype of each category, complete the category division, calculate the loss, and backpropagate to update the network weights; Step 9: Complete the iteration of the meta-training task and generate a mineral resource prediction map for the target prediction task.

2. The mineral resource classification prediction method based on multi-source small sample joint learning according to claim 1 is characterized in that: Each meta-training task in step 1 contains N mineral categories, and K samples are randomly sampled from each category to form a support set S = {(x1, y1), (x2, y2), ..., (x N×K ,y N×K )}, randomly sample T samples from the remaining samples to form a query set For the training of subsequent meta-learning models; for the sample pairs (x s ,y s ), x s and y s Represents the s-th support set sample and its label, s=1,2,…,N×K; for the sample pair in the query set Q and They represent the qth query set sample and its label, q = 1, 2,…, T.

3. The mineral resource classification prediction method based on multi-source small sample joint learning according to claim 1 is characterized in that: The step 2 comprises the following steps: Step 21: Establish a latent space generation adversarial submodule, which consists of a convolutional autoencoder CAE and a latent space adversarial generator G L and latent space adversarial discriminator D L Composition: The convolutional autoencoder CAE consists of an encoder and a decoder, which serves as the real branch of the latent space generation adversarial submodule; The preprocessed sample x is passed through the encoder to obtain the real sample latent space features; The latent space features of the real samples are used as the input of the decoder to obtain the reconstructed samples. The encoding and decoding process is optimized by the reconstruction error of the real samples and the reconstructed samples, which is expressed by the L2 norm of the difference between the real samples and the reconstructed samples, as shown below: Among them, E represents the expected value, x~p data (x) represents x from the data distribution p data (x) is a randomly selected sample; Latent space adversarial generator G L and latent space adversarial discriminator D L As a false branch of the latent space generation adversarial submodule; randomly sample noise Z from the Gaussian space as the latent space adversarial generator G L Input, G L Generate false latent space features, and input the generated false latent space features and the real sample latent space features Encoder(x) into the latent space adversarial discriminator D L The adversarial loss is optimized in [15], and the loss function of the latent space generation adversarial submodule is as follows: Among them, p z (x) represents the noise data distribution; Step 22: Establish a sample generation adversarial submodule, which is composed of the sample adversarial generator G I and sample adversarial discriminator D I Composition; use the false latent space features generated in step 21 as samples to counter the generator G I Input, generate the corresponding false sample G I (G L (z)), and send it together with the real sample x into the sample adversarial discriminator D I The gradient penalty loss is optimized in [15], and the loss function of the sample generation adversarial module is as follows: Among them, gradient_penalty is the gradient penalty term, ω and λ are constants.

4. The mineral resource classification prediction method based on multi-source small sample joint learning according to claim 1 is characterized in that: The step 5 comprises the following steps: Step 51: Assume that the sample x in the study area i The corresponding geochemical data are The corresponding geological data are Converting geological data into graph-structured data To facilitate graph convolutional network processing; specifically, each pixel in the geological data image is converted into a node, and the data of different channels represent the attribute information of each node. The nodes with a distance less than or equal to a preset threshold are set as their neighbor nodes, and the edge information is stored; Step 52: Let the feature extractor of the convolutional neural network be f, and the feature extractor of the graph convolutional network be g. f and g output feature maps of the same size. In order to improve the convolutional neural network's ability to mine channel features and spatial features, add a channel attention module and a spatial attention module to f. Step 53: Geochemical data Input the feature extractor f of the convolutional neural network to generate a geochemical exploration feature map The corresponding geological map structure data Input the feature extractor g of the graph convolutional network to generate a geological feature map Concatenate the two feature maps in the channel dimension and then flatten them as the embedding representation emb of the sample in the metric space i , as shown below: Among them, concat is the splicing operation of the channel dimension, and flatten is the flattening operation to convert the feature map into a vector.

5. The mineral resource classification prediction method based on multi-source small sample joint learning according to claim 1 is characterized in that: In step 6, the initial class prototype of each category is calculated on the support set, and the prototype proto of category c is c The calculation process is as follows: in, y i is the label corresponding to the i-th support set sample, y i =c when w i,c =1, otherwise w i,c =0,emb i is the embedding representation of the i-th support set sample.

6. The mineral resource classification prediction method based on multi-source small sample joint learning according to claim 1 is characterized in that: The step 7 comprises the following steps: Step 71: Randomly extract k unlabeled samples {x1 * ,x2 * ,…,x k * }, after forming the embedding in the metric space through the feature joint embedding module, the unlabeled sample x is calculated based on the Euclidean distance between the unlabeled sample and each initial class prototype. i * The probability p(c|x i * ,{proto c }), the category with the highest probability is used as its label information, p(c|x i * ,{proto c The calculation process of}) is as follows: Among them, emb represents the feature joint embedding module, emb(x i * ) represents the embedding representation of the i-th unlabeled sample, i = 1, 2, ..., k, For category c * The initial class prototype of Step 72: After assigning the category to the unlabeled high anomaly samples, fine-tune all class prototypes. The final class prototype proto of category c is c * The calculation process is as follows: in, emb(x i ) is the embedding representation of the i-th labeled sample, emb(x j * ) is the embedding representation of the j-th unlabeled sample.

7. The method for mineral resource classification prediction based on multi-source small sample joint learning according to claim 1 is characterized in that: In step 8, the samples in the query set Q are embedded in the query set sample by the feature joint embedding module emb, and the distance between the query set sample and the final class prototype of each category is calculated using the distance metric, and the category with the smallest distance is selected as the predicted label of the query set sample; For the samples in the query set Q Calculate the loss and back-propagate to update the network weights, the loss L p The definition is as follows: Among them, T is the number of query set samples, d is the Euclidean distance metric, is the embedding representation of the t-th query set sample, is the true category of the t-th query set sample The corresponding final class prototype, For category c * The final class prototype.

8. A mineral resource classification prediction system based on multi-source small sample joint learning, characterized by: include: The data preprocessing module is used to divide the study area into grid data, mark the mineralization area and background area, and construct multiple meta-training tasks; An unsupervised anomaly mining module is used to establish an anomaly sample mining module for collaborative adversarial training at the latent space feature level and the sample level, which includes a latent space generation adversarial submodule and a sample generation adversarial submodule; the latent space generation adversarial submodule includes a latent space adversarial generator and a latent space adversarial discriminator, and the sample generation adversarial module includes a sample adversarial generator and a sample adversarial discriminator; the latent space features generated by the latent space adversarial generator are used as input to the sample adversarial generator to generate samples; The abnormal sample mining module is trained until the network converges, and a corresponding false sample is generated for each real sample in the grid data of the study area. The weighted average of the sample reconstruction error and feature reconstruction error between the generated sample and the real sample is used as the abnormal score of each generated sample. The generated samples are sorted in descending order according to the abnormal score, and the first samples are retained as high abnormal samples. The prototype representation module is used to establish a joint feature embedding module for the graph convolutional network and the convolutional neural network. The geochemical data of the same sample is input into the convolutional neural network for geochemical feature extraction and the generation of a geochemical feature map. The geological data is input into the graph convolutional network for geological feature extraction and the generation of a geological feature map. The geochemical feature map and the geological feature map are spliced ​​in the channel dimension and then flattened as the embedded representation of the sample in the metric space. The feature joint embedding module is used to calculate the initial class prototype of each category in the support set; And randomly select some high-abnormal samples and map them into the metric space to fine-tune the initial class prototype to obtain the final class prototype of each category; The network training module is used to determine the category of the query set sample by calculating the distance between the embedding and the final class prototype, calculate the loss, and backpropagate to update the network weights until all meta-training tasks are iterated and the final model is saved; The target task prediction and visualization module is used to use the trained network model to perform mineralization prediction on the target task and generate a mineral resource prediction map.

9. A computer system comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the computer program is loaded into a processor, a mineral resource classification prediction method based on multi-source small sample joint learning according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, a mineral resource classification prediction method based on multi-source small sample joint learning according to any one of claims 1 to 7 is implemented.

Images