A semi-supervised mineral resource prediction method, device, equipment and storage medium
By employing a semi-supervised learning method, combining principal component analysis and cosine distance to select representative unlabeled samples, supervised and semi-supervised models are constructed, solving the problem of insufficient labeled data in mineral resource prediction and achieving more comprehensive mineral resource prediction and evaluation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA UNIV OF GEOSCIENCES (WUHAN)
- Filing Date
- 2023-09-26
- Publication Date
- 2026-06-09
AI Technical Summary
In mineral resource forecasting, limited labeled data leads to poor performance and generalization of supervised learning methods, while failing to effectively utilize abundant unlabeled data.
A semi-supervised learning method is adopted to evaluate the similarity of unlabeled samples through principal component analysis and cosine distance, select representative samples, and construct supervised and semi-supervised machine learning models. These models are trained using both labeled and unlabeled data, and self-training and generative adversarial networks are used to improve the generalization of the models.
It improves the accuracy and generalization ability of mineral resource prediction, effectively utilizes unlabeled data, and enhances the model's prediction performance and geological interpretation capabilities.
Smart Images
Figure CN117313932B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of mineral resource prediction and evaluation, specifically to a semi-supervised mineral resource prediction method, apparatus, equipment, and storage medium. Background Technology
[0002] Mineral resource prediction aims to delineate metallogenic potential areas and determine priorities for mineral exploration by analyzing and synthesizing relevant geological exploration data, playing a crucial guiding role in mineral exploration. Against the backdrop of surging demand for mineral resources, mineral resource prediction and evaluation play a vital role in further exploration investigations of previously explored and nearly unexplored areas. In recent years, data-driven mineral resource prediction based on artificial intelligence (AI) algorithms has sparked a new wave of quantitative mineral resource assessment. As the core of AI algorithms, machine learning (ML) can automatically analyze and uncover complex relationships between known mineral deposits. To date, numerous ML methods have been used for data-driven mineral potential evaluation. The most commonly used machine learning methods include logistic regression, support vector machines, random forests, artificial neural networks, and self-organizing map networks. Recently, deep learning (DL) methods, due to their powerful feature representation capabilities and ability to learn sample data features from high-dimensional space, have been successfully applied to the field of mineral resource prediction. These deep learning network models have proven to be very powerful in extracting high-dimensional relationships and identifying complex spatial patterns associated with mineralization anomalies. Common deep learning methods include convolutional neural networks, deep neural networks, deep autoencoders, generative adversarial networks, and graph neural networks.
[0003] In the field of mineral resource prediction and evaluation, machine learning methods have been extensively researched and explored, with an increasing number of advanced algorithms being practically validated and applied. Generally speaking, based on whether the training samples for machine learning need to be labeled, the use of machine learning methods in mineral resource prediction can be broadly divided into unsupervised learning and supervised learning. The goal of unsupervised learning is to achieve data clustering by analyzing and mining the underlying structure or distribution of a large amount of unlabeled data in the data space, or to capture geological anomalies by reconstructing the error between the data and the real data. However, the main drawback of unsupervised learning methods is that the results usually require geological interpretation and calibration, and the accuracy of prediction results is generally lower than that of supervised learning methods. As an alternative, supervised learning methods can automatically mine the complex mapping relationship between exploration data and mineral deposits and automatically interpret the prediction results. However, the performance of supervised learning methods largely depends on the amount of labeled data. Supervised learning methods place more emphasis on increasing labeled data but neglect the auxiliary role of a large amount of unlabeled data in classification. In the field of mineral resource prediction, the rarity of mineralization leads to insufficient available labeled data. With limited labeled data, supervised learning methods typically perform poorly in terms of model performance and generalization, while also failing to utilize the abundant unlabeled data within the study area. To overcome the aforementioned challenges, semi-supervised learning (SSL) combines limited labeled data and a large amount of unlabeled data within the study area, enabling it to better learn the data distribution. This improves supervised learning methods and enhances model performance and geological generalization ability. Therefore, in the field of mineral resource prediction, the SSL method, by combining limited labeled data from mineralized areas with a large amount of unlabeled data within the study area, possesses unique advantages and is a crucial technical problem that urgently needs to be solved: providing a mineral resource prediction method based on semi-supervised machine learning. Summary of the Invention
[0004] To address the aforementioned technical problems, this invention proposes a semi-supervised mineral resource prediction method, apparatus, equipment, and storage medium. This method, while making reasonable use of limited labeled data, considers a large amount of unlabeled data in the study area, breaks away from the conventional supervised learning model, and researches and develops a semi-supervised mineral prediction modeling algorithm with stronger geological generalization, thereby achieving more comprehensive mineral resource prediction and evaluation.
[0005] According to a first aspect of the present invention, a semi-supervised mineral resource prediction method includes the following steps:
[0006] The mineralization evidence layer is trimmed and segmented to establish standardized labeled data and candidate unlabeled data;
[0007] The similarity between candidate unlabeled sample data and known mineral deposit or mineralized area labeled data is evaluated using a similarity strategy based on principal component analysis and cosine distance. All candidate unlabeled samples in the study area are ranked according to the similarity score, and representative unlabeled samples are selected.
[0008] Supervised machine learning and semi-supervised machine learning mineral prediction models are constructed. Labeled data is input into the supervised machine learning model for training, and labeled data and representative unlabeled data are input into the semi-supervised machine learning model for training. After training, the optimal mineral prediction model is obtained.
[0009] The optimal mineral prediction model is used to predict and evaluate the mineral prospects of the study area.
[0010] Furthermore, the step of trimming and segmenting the mineralization evidence layer to establish standardized labeled data and candidate unlabeled data includes:
[0011] Obtain raw point data within the study area, including: geochemical data, regional geological data, and geophysical data;
[0012] The raw point data is used as the mineralization evidence layer for machine learning based on point data input to prepare the dataset;
[0013] Inverse distance weighted interpolation is performed on the original point data to obtain raster data. All raster data are converted into grayscale images as a mineralization evidence layer based on image data input.
[0014] The mineralization evidence layer based on image data input is cropped according to the locations of representative mineralization points and non-mineralization points to determine the labeled training data. Candidate unlabeled data are obtained by dividing the mineralization evidence layer into blocks.
[0015] Further, the step of evaluating the similarity between candidate unlabeled sample data and known deposit or mineralized area labeled data based on a similarity strategy using principal component analysis and cosine distance, ranking all candidate unlabeled samples within the study area according to the similarity score, and selecting representative unlabeled samples includes:
[0016] Principal component analysis was used to reduce the dimensionality of the mineralization evidence layer data, resulting in redundancy-free evidence layer data.
[0017] Cosine distance is used to evaluate the similarity between unlabeled samples and samples from known deposits or mineralized areas. Two sample vectors A = (x... A1 x A2 , ..., x An ) and B = (x B1 x B2 , ..., x Bn The cosine distance of ) is calculated as follows:
[0018]
[0019] Where θ represents the spatial angle between vectors A and B, and A and B are n-dimensional vectors;
[0020] Use X d = (x1, x2, ..., x n () represents the set of known mineral deposits or mineralized areas, x u Let x represent an unlabeled sample. u and X d The average similarity of all known deposits or mineralized areas in the region is:
[0021]
[0022] All candidate unlabeled samples within the study area are ranked according to similarity scores, and representative unlabeled samples are selected.
[0023] Furthermore, the steps for constructing a semi-supervised machine learning mineral prediction model include:
[0024] The model is trained using a self-trained semi-supervised approach by iteratively updating the training set. Each iteration consists of two steps: first, a base classifier is selected and trained on the initial labeled dataset to obtain a teacher model; second, the teacher model is used to predict unlabeled data u. m Generate pseudo tags for it Select a subset of pseudo-labeled data with a confidence level higher than the set value. subset Finally, the data is merged into the labeled dataset for the next iteration of training.
[0025] The iteration stops when the maximum specified number of iterations is reached or no new pseudo-label dataset is added to the label dataset, resulting in a semi-supervised machine learning mineral prediction model.
[0026] Furthermore, the base classifiers include: random forest model and support vector machine model.
[0027] Furthermore, the steps for constructing a semi-supervised machine learning mineral prediction model also include:
[0028] A semi-supervised generative adversarial network is constructed, consisting of a generator and a discriminator. The discriminator of the semi-supervised generative adversarial network is transformed into a three-class classifier, where the first two classes correspond to the categories of the classification task, and the last class indicates whether the input sample comes from real exploration data in the study area or fake data generated by the generator.
[0029] Furthermore, in the semi-supervised generative adversarial network, the discriminator's loss function includes unsupervised loss and additional supervised loss, the specific form of which is as follows:
[0030] L D =L supervised +L unsupervised
[0031]
[0032]
[0033] Among them, L supervised L represents the negative log probability of a true labeled sample. unsupervised Equivalent to the loss function of the original GAN, log p model (y|x, y<K+1) represents the log probability that the discriminator predicts the true sample as the corresponding class, log(1-p model (y=K+1|x)) represents the log probability that the discriminator predicts the true sample as the true sample class, log(p model (y=K+1|x)) represents the log probability that the discriminator will predict a generated sample as a false sample, x~p data (x, y) and x ~ p data (x) represents the data distribution of the real samples, x~G represents the data distribution of the samples generated by the generator, and K represents the class of the real samples; during the training phase, L unsupervised Used to guide the discriminator to distinguish real samples p data (x) and the fake samples p generated by the generator G (x), L supervised This is used to help the discriminator correctly predict the real samples as their corresponding categories; in semi-supervised generative adversarial networks, the generator loss function uses feature matching to minimize the statistical difference between features of real and fake samples from the generator, typically using the L2 distance of the sample feature vectors.
[0034]
[0035] Where f(x) represents the feature vector of the real sample, f(G(z)) represents the feature vector of the fake sample generated by the generator, and x ~ p data (x) represents the data distribution of the true sample, z ~ p z (z) represents the data distribution of the fake samples generated by the generator.
[0036] According to a second aspect of the present invention, a semi-supervised mineral resource prediction device includes the following modules:
[0037] The dataset construction module is used to trim and segment the mineralization evidence layer, and to establish standardized labeled data and candidate unlabeled data.
[0038] The similarity screening module is used to evaluate the similarity between candidate unlabeled sample data and known mineral deposit or mineralized area labeled data based on a similarity strategy based on principal component analysis and cosine distance. Based on the similarity score, all candidate unlabeled samples in the study area are sorted and representative unlabeled samples are selected.
[0039] The modeling and training module is used to build supervised machine learning and semi-supervised machine learning mineral prediction models. Labeled data is input into the supervised machine learning model for training, and labeled data and representative unlabeled data are input into the semi-supervised machine learning model for training. After training, the optimal mineral prediction model is obtained.
[0040] The prediction and evaluation module is used to predict and evaluate the mineral prospects of the study area using the optimal mineral prediction model.
[0041] According to a third aspect of the present invention, an electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the steps of the semi-supervised mineral resource prediction method.
[0042] According to a fourth aspect of the invention, a storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the semi-supervised mineral resource prediction method.
[0043] The technical solution provided by this invention has the following beneficial effects:
[0044] This invention first establishes standardized labeled data and candidate unlabeled data by trimming and segmenting the mineralization evidence layer. Then, it evaluates the similarity between the candidate unlabeled sample data and the labeled data of known mineral deposits or mineralized areas using a similarity strategy based on principal component analysis and cosine distance. Based on the similarity scores, all candidate unlabeled samples within the study area are ranked, and representative unlabeled samples are selected. Next, supervised machine learning and semi-supervised machine learning mineral prediction models are constructed. Labeled data is input into the supervised machine learning model for training, while labeled data and representative unlabeled data are input into the semi-supervised machine learning model for training. The optimal mineral prediction model is obtained after training. Finally, the optimal mineral prediction model is used to predict and evaluate the mineral prospects of the study area. This invention breaks with the conventional supervised learning mineral resource prediction modeling mechanism. Addressing the problem of the limited number of known mineral deposits in mineral exploration practice, it can better learn the spatial patterns of mineralization anomalies and has better generalization ability than supervised learning. This invention utilizes a similarity strategy based on principal component analysis and cosine distance to select representative unlabeled samples, enabling the identification of mineralization anomalies, excluding high-risk unlabeled samples, and better assisting semi-supervised learning. Attached Figure Description
[0045] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings:
[0046] Figure 1 This is a flowchart illustrating the overall process of a semi-supervised mineral resource prediction method in an embodiment of the present invention.
[0047] Figure 2 These are some of the stream sedimentary geochemical data in this embodiment of the invention, where (a) corresponds to the W element concentration map; (b) corresponds to the Sn element concentration map; (c) corresponds to the Bi element concentration map; (d) corresponds to the Be element concentration map; (e) is the fracture buffer boundary; and (f) is the granite body buffer boundary.
[0048] Figure 3 This is the workflow of the similarity strategy based on principal component analysis and cosine distance in this embodiment of the invention, where (a) represents data dimensionality reduction through principal component analysis, (b) represents calculating the average similarity between candidate unlabeled samples and all known deposits or mineralized area samples, and (c) is a similarity map.
[0049] Figure 4 The present invention relates to a semi-supervised machine learning network structure for mineral prediction, wherein (a) is a semi-supervised generative adversarial network structure and (b) is a self-trained semi-supervised learning network structure.
[0050] Figure 5 This is the training curve of the semi-supervised generative adversarial network in an embodiment of the present invention.
[0051] Figure 6 These are mineral prospect maps obtained by various methods in the embodiments of the present invention, wherein (a) corresponds to SemiGAN; (b) corresponds to SSRF; (c) corresponds to S3VM; (d) corresponds to RF; and (e) corresponds to SVM.
[0052] Figure 7 These are the prediction success rate curves of various methods in the embodiments of the present invention, where (a) corresponds to the comparison results of various semi-supervised learning methods; (b) corresponds to the comparison results of semi-supervised learning and supervised learning methods; (c) corresponds to the comparison results of SSRF and RF; and (d) corresponds to the comparison results of S3VM and SVM.
[0053] Figure 8 This is a schematic diagram of the structure of a semi-supervised mineral resource prediction device in an embodiment of the present invention.
[0054] Figure 9 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention. Detailed Implementation
[0055] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0056] In the field of mineral resource prediction and evaluation, the rarity of mineralization leads to a shortage of available labeled data. With limited labeled data, supervised learning methods typically perform poorly in terms of model performance and generalization, while also failing to utilize the abundant unlabeled data within the study area. To overcome these challenges, this invention provides a semi-supervised machine learning-based mineral prediction modeling algorithm. This algorithm rationally utilizes limited labeled data while considering the large amount of unlabeled data in the study area, breaking away from conventional supervised learning models. It researches and develops a semi-supervised mineral prediction modeling algorithm with stronger geological generalization capabilities, achieving more comprehensive mineral resource prediction and evaluation.
[0057] Please refer to Figure 1 The semi-supervised mineral resource prediction method of this invention includes the following steps:
[0058] S1: Trim and segment the mineralization evidence layer to establish standardized labeled data and candidate unlabeled data;
[0059] S2: Evaluate the similarity between candidate unlabeled sample data and known deposit or mineralized area labeled data based on a similarity strategy using principal component analysis and cosine distance. Sort all candidate unlabeled samples in the study area according to the similarity score and select representative unlabeled samples.
[0060] S3: Construct supervised machine learning and semi-supervised machine learning mineral prediction models. Input labeled data into the supervised machine learning model for training, and input labeled data and representative unlabeled data into the semi-supervised machine learning model for training. After training, obtain the optimal mineral prediction model.
[0061] S4: Utilize the optimal mineral prediction model to predict and evaluate the mineral prospects of the study area.
[0062] Based on, but not limited to, the above methods, the specific implementation process of step S1 is as follows:
[0063] The dataset used in this embodiment of the invention originates from the Nanling region. Based on the mineralization process of tungsten-tin polymetallic deposits in the Nanling region, this embodiment selects geochemical, regional geological, and geophysical data from the study area for modeling the prospective tungsten-tin deposits in the Nanling region. Specifically, it includes geochemical data of eight stream sediments closely related to W-Sn mineralization (W, Sn, Bi, Be, Pb, Ag, Mo, Zn), local gravity anomalies, fault buffer zones, and granite boundary buffer zones (such as...). Figure 2 The original data (shown in the image) were selected as indicator factors for mineral prospective modeling. These raw data points, being point data, were directly used as the mineralization evidence layer for point-based machine learning (Random Forest (RF) and Support Vector Machine (SVM)). To suit deep learning image data input, the original data was interpolated using Inverse Distance Weights (IDW) to create 501×834 pixel raster data, with each pixel (grid) measuring 0.6km×0.6km. All raster images were converted to grayscale images for the mineralization evidence layer based on image input methods. Preparing the training dataset is a crucial step in using machine learning for mineral prospective prediction. Ultimately, 29 mineralized locations and 31 non-mineralized locations, along with their corresponding mineralization evidence layers, were selected to construct the dataset (including training and testing data). All unlabeled samples and prediction data were derived from the original evidence layer through block partitioning or sliding cropping.
[0064] For semi-supervised learning, unlabeled samples should have a similar data distribution (distribution consistency) to labeled samples. Therefore, this invention proposes a similarity strategy of principal component analysis + cosine distance to select representative unlabeled samples.
[0065] Based on, but not limited to, the above methods, the specific implementation process of step S2 is as follows:
[0066] First, principal component analysis (PCA) is used to reduce the dimensionality of the mineral exploration evidence layer, retaining the most important information, reducing data dimensionality, and eliminating redundant features. This helps capture potential structures and patterns in a more compact representation. Next, cosine distance is used to evaluate the similarity between unlabeled samples and samples from known deposits or mineralized areas. Cosine distance emphasizes the directional difference between two vectors, is unaffected by vector scale, and can effectively capture the directional similarity between vectors. Taking two sample vectors A = (x... A1 x A2 , ..., x An ) and B = (x B1 x B2 , ..., x Bn For example, the cosine distance between two vectors can be calculated as follows:
[0067]
[0068] In practical mineral resource forecasting tasks, assume X d = (x1, x2, ..., x n () represents the set of known mineral deposits or mineralized areas, x u Let x represent an unlabeled sample. u and X d The average similarity of all known deposits or mineralized areas in the region is:
[0069]
[0070] Finally, all unlabeled samples within the study area are ranked according to similarity scores, and representative unlabeled samples are selected. The similarity measurement method proposed in this invention, based on principal component analysis and cosine distance (e.g., Figure 3 As shown in the figure, the aim is to select representative unlabeled samples to facilitate the effective use of the abundant unlabeled samples within the study area. By incorporating these diverse and information-rich unlabeled samples into the model's learning process, the model's performance can be further improved.
[0071] Based on, but not limited to, the above methods, the specific implementation process of step S3 is as follows:
[0072] Self-training is an iterative method widely used in semi-supervised learning (SSL). It utilizes a classification model trained on labeled data to predict pseudo-labels on unlabeled data, thus training an optimal classification model by combining labeled and unlabeled data. Training involves iteratively updating the training set to train the model. Each iteration consists of two steps: First, a base classifier is selected and trained on the labeled dataset to obtain a teacher model. Second, the teacher model predicts the pseudo-labels on the unlabeled data u. mGenerate pseudo tags for it Select a subset of pseudo-labeled data with relatively high confidence levels. subset Finally merged into the labeled dataset, i.e., The next iteration of training is then performed. The stopping condition for this iteration is reaching the maximum specified number of iterations or no new pseudo-label dataset is added to the training set. This invention selects the most commonly used random forest and support vector machine as base classifiers. Using the above method, a self-trained semi-supervised random forest and semi-supervised support vector machine method can be implemented for mineral resource prediction. Figure 4 b). This method enables the model to simultaneously utilize limited labeled data and a large amount of unlabeled data, better learn the data distribution, and thus improve the prediction accuracy and geological generalization of the mineral prediction model.
[0073] Meanwhile, semi-supervised generative adversarial networks (GANs) overcome the limitations of limited labeled data by effectively utilizing both labeled and unlabeled data, providing a powerful learning model for fully leveraging unlabeled data. Semi-supervised GANs aim to enhance the model's ability to capture the underlying data distribution, thereby improving model performance, and represent a state-of-the-art semi-supervised learning method. Compared to unsupervised GANs, the discriminator in a semi-supervised GAN is a multi-classifier, playing multiple roles in label prediction and distinguishing between real and generated fake samples. The model's final prediction is achieved through the multi-classifier. Specifically, the discriminator divides input samples into K+1 classes. Among these classes, the first K classes correspond to the categories of the classification task, while the K+1 classes are used to distinguish between real and generated samples. In the field of mineral resource prediction, the discriminator needs to be transformed into a three-classifier, where the first two classes correspond to the various categories of the classification task (mineral deposits or non-mineral deposits), and the last class indicates whether the input sample comes from real exploration data within the study area or fake data generated by the generator.
[0074] In semi-supervised generative adversarial networks (GANs), the discriminator has two main tasks. First, it aims to distinguish between real and generated samples to help the generator learn to produce more realistic samples. Second, it utilizes generated fake samples along with labeled and unlabeled real data to aid the model in learning classification tasks. Therefore, the discriminator's loss function includes unsupervised loss and an additional supervised loss. The specific form of the loss function is as follows:
[0075] L D =L supervised +L unsupervised (3)
[0076]
[0077]
[0078] Among them, L supervised L represents the negative log probability of a true labeled sample. unsupervised This is equivalent to the loss function of the original GAN. During the training phase, L... unsupervised Used to guide the discriminator to distinguish real samples p data (x) and the fake samples p generated by the generator G (x). The discriminator works by minimizing the negative log-probability log(1-p) of correctly predicting a true sample as the true sample class. model (y=K+1|x)) and the negative log probability log(p) of correctly predicting the generated sample as a false sample class. model (y=K+1|x)) is optimized. supervised This is used to help the discriminator correctly predict the true sample as its corresponding class. The discriminator works by minimizing the negative log-probability logp of correctly predicting the true sample as its corresponding class. model Optimization is achieved using (y|x, y<K+1). In the prediction phase, the discriminator predicts the top K classes to fulfill the specific classification task. In semi-supervised generative adversarial networks, the generator's task is to generate more realistic samples by learning the underlying data distribution. Simultaneously, the generated fake samples serve as boundary supplementary samples, resulting in a better classification (decision) boundary. The generator loss function utilizes feature matching to minimize the statistical difference between features from real and fake samples from the generator, typically using the L2 distance of the sample feature vectors.
[0079]
[0080] Here, f(x) represents the feature vector of the real sample, and f(G(z)) represents the feature vector of the fake sample generated by the generator. Minimizing the feature matching loss function can encourage the generator to produce more realistic samples that conform to the underlying data distribution.
[0081] This invention designs a semi-supervised generative adversarial network for the field of mineral resource prediction and evaluation (e.g., Figure 4 (As shown in a). The network structure mainly consists of two parts: a generator and a discriminator. The generator contains three upsampling layers that transform the input random noise into real-data size through transposed convolution operations. To adapt to the mineral resource prediction classification task with two categories, the discriminator is designed as a three-class classifier, with the input data being real samples (labeled and unlabeled samples) from the training set and fake samples generated by the generator. The discriminator consists of three convolutional layers and two fully connected layers. Considering the spatial size of the input samples in the mineral prospect prediction domain, the convolutional kernel size is set to 3×3. The ReLU nonlinear activation function is used to alleviate the problems of gradient vanishing and gradient saturation.
[0082] Based on, but not limited to, the above methods, the specific implementation process of step S4 is as follows:
[0083] In this embodiment, five machine learning models—SemiGAN, SSRF, S3VM, RF, and SVM—were selected for comparative experiments. The semi-supervised learning methods (SemiGAN, SSRF, and S3VM) were trained on unlabeled samples selected in schemes one and two, respectively. During the training of SemiGAN, the number of training iterations was set to 20, the learning rate to 0.0005, and the batch size to 64. The loss curve is shown below. Figure 5 As shown. To avoid overfitting due to overtraining, this embodiment chooses to stop training after approximately 15 epochs. During the training of the self-trained SSL methods (SSRF and S3VM), the core hyperparameters of the methods are a maximum of 20 iterations and a prediction probability value greater than 0.9 (confidence strategy). Subsequently, the trained model was used for comprehensive experiments and practical applications.
[0084] In this embodiment, a comparative analysis experiment was conducted between semi-supervised learning methods (SemiGAN, SSRF, and S3VM) and supervised learning methods (RF and SVM). Using the optimal semi-supervised learning models (SemiGAN, SSRF, and S3VM) and supervised learning models (RF and SVM) obtained through training, a prospective map of tungsten-tin deposits in the Nanling region was drawn based on the prediction set. Figure 6 In all the prospective maps, most of the known tungsten-tin deposits are located in high-potential areas. These high-potential areas are spatially closely correlated with known W-Sn mineralizations. To further quantify the generalization of the mineral prediction model, the prediction success rate curves of all known tungsten-tin deposits were used to further evaluate the model's performance. Figure 7 As can be observed from the figure, SemiGAN exhibits the best prediction performance in the 5% and 10% high-potential regions, containing 105 and 142 known mineral points respectively. This is because SemiGAN utilizes adversarial learning between the generator and discriminator, naturally possessing a strong ability to learn data distributions. Next, the SSRF model contains 105 and 131 known mineral points respectively. Compared to supervised learning methods, semi-supervised learning methods contain more known mineral points in the top 5% and 10% high-potential regions, resulting in significantly better prediction performance in these areas. This suggests that semi-supervised learning can effectively learn data distributions, improving the model's prediction performance and generalization ability. Overall, the semi-supervised prediction performance is superior to supervised methods, indicating that the addition of unlabeled data can alleviate the lack of labeled samples, improve the learning of data distributions, and further enhance the model's prediction results and generalization ability.
[0085] The following describes a semi-supervised mineral resource prediction device provided by the present invention. The semi-supervised mineral resource prediction device described below and the semi-supervised mineral resource prediction method described above can be referred to and correspond to each other.
[0086] like Figure 8 As shown, an example of a semi-supervised mineral resource prediction device is presented, comprising the following modules:
[0087] The dataset construction module 810 is used to trim and segment the mineralization evidence layer, and to establish standardized labeled data and candidate unlabeled data.
[0088] The similarity screening module 820 is used to evaluate the similarity between candidate unlabeled sample data and known mineral deposit or mineralized area labeled data according to the similarity strategy based on principal component analysis and cosine distance. It sorts all candidate unlabeled samples in the study area according to the similarity score and selects representative unlabeled samples.
[0089] The modeling and training module 830 is used to build supervised machine learning and semi-supervised machine learning mineral prediction models. Labeled data is input into the supervised machine learning model for training, and labeled data and representative unlabeled data are input into the semi-supervised machine learning model for training. After training, the optimal mineral prediction model is obtained.
[0090] The prediction and evaluation module 840 is used to predict and evaluate the mineral prospects of the study area using the optimal mineral prediction model.
[0091] like Figure 9As shown, a schematic diagram of the physical structure of an electronic device is illustrated. The electronic device may include: a processor 910, a communication interface 920, a memory 930, and a communication bus 940, wherein the processor 910, the communication interface 920, and the memory 930 communicate with each other through the communication bus 940. The processor 910 can call logic instructions in the memory 930 to execute the aforementioned semi-supervised mineral resource prediction method, including: trimming and segmenting the mineralization evidence layer to establish standardized labeled data and candidate unlabeled data; evaluating the similarity between candidate unlabeled sample data and labeled data of known deposits or mineralized areas based on a similarity strategy using principal component analysis and cosine distance; ranking all candidate unlabeled samples in the study area according to the similarity score and selecting representative unlabeled samples; constructing supervised machine learning and semi-supervised machine learning mineral prediction models, inputting labeled data into the supervised machine learning model for training, and inputting labeled data and representative unlabeled data into the semi-supervised machine learning model for training, obtaining the optimal mineral prediction model after training; and using the optimal mineral prediction model to predict and evaluate the mineral prospects of the study area.
[0092] Furthermore, the logical instructions in the aforementioned memory 930 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0093] In another aspect, embodiments of the present invention also provide a storage medium storing a computer program. When executed by a processor, this computer program implements the aforementioned semi-supervised mineral resource prediction method, including: trimming and segmenting the mineralization evidence layer to establish standardized labeled data and candidate unlabeled data; evaluating the similarity between candidate unlabeled sample data and labeled data of known deposits or mineralized areas based on a similarity strategy using principal component analysis and cosine distance; ranking all candidate unlabeled samples within the study area according to the similarity score and selecting representative unlabeled samples; constructing supervised machine learning and semi-supervised machine learning mineral prediction models; inputting labeled data into the supervised machine learning model for training; inputting labeled data and representative unlabeled data into the semi-supervised machine learning model for training; obtaining the optimal mineral prediction model after training; and using the optimal mineral prediction model to predict and evaluate the mineral prospects of the study area.
[0094] This invention proposes a semi-supervised mineral resource prediction method, apparatus, equipment, and storage medium. By pruning and segmenting the mineralization evidence layer, standardized labeled data and candidate unlabeled data are established. Representative unlabeled samples are selected based on a similarity strategy using principal component analysis and cosine distance. Supervised and semi-supervised machine learning mineral prediction models are constructed. Labeled data is input into the supervised machine learning model for training, while labeled data and representative unlabeled data are input into the semi-supervised machine learning model for training. The optimal mineral prediction model is obtained after training. This optimal model is then used to predict and evaluate the mineral prospects of the study area. This approach rationally utilizes limited labeled data while considering the large amount of unlabeled data in the study area, breaking away from conventional supervised learning models and developing a more geologically generalizable mineral prediction algorithm, thus achieving more comprehensive mineral resource prediction and evaluation.
[0095] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.
[0096] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments. In the unit claims listing several devices, several of these devices may be embodied by the same hardware item. The use of the terms first, second, and third, etc., does not indicate any order and can be interpreted as identifiers.
[0097] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.
Claims
1. A semi-supervised mineral resource prediction method, characterized in that, Includes the following steps: The mineralization evidence layer is trimmed and segmented to establish standardized labeled data and candidate unlabeled data; The similarity between candidate unlabeled sample data and known mineral deposit or mineralized area labeled data is evaluated using a similarity strategy based on principal component analysis and cosine distance. All candidate unlabeled samples in the study area are ranked according to the similarity score, and representative unlabeled samples are selected. Supervised machine learning and semi-supervised machine learning mineral prediction models are constructed. Labeled data is input into the supervised machine learning model for training, and labeled data and representative unlabeled data are input into the semi-supervised machine learning model for training. After training, the optimal mineral prediction model is obtained. The optimal mineral prediction model is used to predict and evaluate the mineral prospects of the study area. The steps for building a semi-supervised machine learning mineral prediction model include: The model is trained using a self-trained semi-supervised approach by iteratively updating the training set. Each iteration consists of two steps: first, a base classifier is selected and trained on the initial labeled dataset to obtain a teacher model; second, the teacher model is used to predict unlabeled data. Generate pseudo tags for it Select a subset of pseudo-labeled data with a confidence level higher than a set value. subset Finally, the data is merged into the labeled dataset for the next iteration of training. The iteration stops when the maximum specified number of iterations is reached or no new pseudo-label dataset is added to the label dataset, resulting in a semi-supervised machine learning mineral prediction model. The steps for building a semi-supervised machine learning mineral prediction model also include: Construct a semi-supervised generative adversarial network, including a generator and a discriminator; convert the discriminator of the semi-supervised generative adversarial network into a three-class classifier, where the first two classes correspond to the categories of the classification task, and the last class indicates whether the input sample comes from real exploration data in the study area or fake data generated by the generator; In the semi-supervised generative adversarial network, the discriminator's loss function includes unsupervised loss and an additional supervised loss, the specific form of which is as follows: in, This represents the negative logarithmic probability of the true sample with the label. Equivalent to the loss function of the original GAN, This represents the log probability that the discriminator predicts a true sample as belonging to the corresponding class. This represents the log probability that the discriminator predicts a real sample as belonging to the real sample class. This represents the log probability that the discriminator will predict a generated sample as a false sample. and Represents the data distribution of the real samples. This represents the data distribution of the samples generated by the generator, where K represents the class of the real samples; during the training phase, Used to guide the discriminator to distinguish real samples and fake samples generated by the generator , This is used to help the discriminator correctly predict the real samples as their corresponding categories; in semi-supervised generative adversarial networks, the generator loss function uses feature matching to minimize the statistical difference between features of real and fake samples from the generator, typically using the L2 distance of the sample feature vectors. in, The feature vector representing the real sample. This represents the feature vector of the fake samples generated by the generator. Represents the data distribution of the real samples. This represents the data distribution of the fake samples generated by the generator.
2. The semi-supervised mineral resource prediction method according to claim 1, characterized in that, The steps of trimming and segmenting the mineralization evidence layer to establish standardized labeled data and candidate unlabeled data include: Obtain raw point data within the study area, including: geochemical data, regional geological data, and geophysical data; The raw point data is used as the mineralization evidence layer for machine learning based on point data input to prepare the dataset; Inverse distance weighted interpolation is performed on the original point data to obtain raster data. All raster data are converted into grayscale images as a mineralization evidence layer based on image data input. The mineralization evidence layer based on image data input is cropped according to the locations of representative mineralization points and non-mineralization points to determine the labeled training data. Candidate unlabeled data are obtained by dividing the mineralization evidence layer into blocks.
3. The semi-supervised mineral resource prediction method according to claim 1, characterized in that, The steps of evaluating the similarity between candidate unlabeled sample data and known deposit or mineralized area labeled data using a similarity strategy based on principal component analysis and cosine distance, ranking all candidate unlabeled samples within the study area according to the similarity score, and selecting representative unlabeled samples include: Principal component analysis was used to reduce the dimensionality of the mineralization evidence layer data, resulting in redundancy-free evidence layer data. Cosine distance is used to evaluate the similarity between unlabeled samples and samples from known deposits or mineralized regions, where two sample vectors... and The cosine distance is calculated as follows: in, Representing vectors A and B The spatial angle, A and B belong n dimensional vector; use This represents a set of known mineral deposits or mineralized areas. This represents an unlabeled sample. and The average similarity of all known deposits or mineralized areas in the region is: All candidate unlabeled samples within the study area are ranked according to similarity scores, and representative unlabeled samples are selected.
4. The semi-supervised mineral resource prediction method according to claim 1, characterized in that, The base classifiers include: random forest model and support vector machine model.
5. A semi-supervised mineral resource prediction device, characterized in that, A method for implementing any one of claims 1 to 4 includes the following modules: The dataset construction module is used to trim and segment the mineralization evidence layer, and to establish standardized labeled data and candidate unlabeled data. The similarity screening module is used to evaluate the similarity between candidate unlabeled sample data and known mineral deposit or mineralized area labeled data based on a similarity strategy based on principal component analysis and cosine distance. Based on the similarity score, all candidate unlabeled samples in the study area are sorted and representative unlabeled samples are selected. The modeling and training module is used to build supervised machine learning and semi-supervised machine learning mineral prediction models. Labeled data is input into the supervised machine learning model for training, and labeled data and representative unlabeled data are input into the semi-supervised machine learning model for training. After training, the optimal mineral prediction model is obtained. The prediction and evaluation module is used to predict and evaluate the mineral prospects of the study area using the optimal mineral prediction model.
6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the semi-supervised mineral resource prediction method as described in any one of claims 1-4.
7. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the semi-supervised mineral resource prediction method as described in any one of claims 1-4.