A method for extracting mineralization anomalies based on a geological "big model"

Abstract

This invention belongs to the field of solid mineral exploration technology, specifically a method for extracting mineralization anomalies based on a geological "big model." Compared to traditional deep learning-based mineralization prediction methods, this method can address practical issues encountered in prospecting in unfamiliar work areas, such as low levels of expertise, limited geological data, and a small number of discovered mineral deposits. This method connects semantic space and feature space through spatial mapping, building a bridge between "big models" and "big data," thereby enabling knowledge- and data-driven delineation of prospecting areas with minimal sample size.

Description

Technical Field

[0001] The present invention belongs to the technical field of solid mineral exploration, and in particular relates to a method for extracting mineralization anomalies based on a geological "big model". Background Art

[0002] Geological data records the specific and final results of geological work, while geological big data is a more comprehensive and in-depth collection of geological information generated from this data through technical means such as data integration, mining, and analysis. Mineralization prediction, as a core component of mineral exploration, is an important means and method for achieving scientific prospecting and increasing resource reserves. This research involves multidisciplinary knowledge such as geology, geophysical exploration, geochemical exploration, and remote sensing, and utilizes ore deposit models and metallogenic system prediction models to assess mineral resources for the purpose of prospecting and prediction. Big data and artificial intelligence are significantly changing human cognition and research paradigms, particularly in the data-intensive field of mineral exploration. This not only poses enormous challenges for change, but also provides unprecedented opportunities for innovation and development. Academician Zhao Pengda, the founder of Chinese mathematical geology, has pointed out that geological big data analysis is an emerging interdisciplinary discipline. Based on geological science and information technology, it builds and applies various mathematical models to describe geological bodies, geological processes, and geological working methods. Using data science methods, it intelligently processes geological big data, analyzes and mines valuable core information and key data, forming concentrated digital knowledge and revealing information that traditional logic or methods may overlook. This significantly improves exploration efficiency, reduces uncertainty, lowers exploration costs and risks, and promotes the development of mineral exploration towards a more quantitative, refined, and visualized approach. For example, some scholars have applied machine learning methods such as artificial neural networks, support vector machines, and random forests to the field of geology, achieving promising results. Deep learning methods process and interpret information by constructing neural networks that mimic the analytical and learning processes of the human brain. Typical models include convolutional neural networks, recurrent neural networks, and stacked autoencoders. The application of machine learning methods, particularly deep learning methods, to mineralization prediction is an inevitable trend in the development of quantitative geology research. With the assistance of artificial intelligence technologies such as machine learning and deep learning, valuable information from geological big data can be further mined, thereby promoting the development of geological science and contributing to social and economic development.

[0003] In the field of mineral resource exploration, the application of geological big data is particularly prominent. By integrating geological, geophysical, geochemical, and remote sensing data, it significantly improves the accuracy of predictions of mineral resource distribution and reserves, providing a solid scientific foundation for ensuring the rational development and utilization of mineral resources. Integrating intelligent analysis techniques for geological big data into mineral resource exploration and assessment is an indispensable step in the development of mathematical geology and the initial stage in the transition to intelligent mineral prospecting. Mineralization prediction relies on the extraction and interpretation of multidisciplinary and multi-dimensional geological information. While big data-based approaches can subject all data to machine learning for processing at once, relying solely on data to analyze its own characteristics or its relationship to known mineral deposits is not sufficient. Mineralization prediction based on deep learning is essentially a mathematical and statistical analysis of multi-source geological information. Its advantage lies in its ability to extract feature representations at different levels of abstraction. While different types of mineral deposits can share the same geological genesis, the same type of deposit in the same study area can exhibit different deposit types due to differences in geological processes. Deep learning methods such as convolutional neural networks can be used to learn optimal representations, or supervised learning can be used to obtain supervised representations relevant to the mineral deposits in the study area. In theory, the same type of mineral deposits in the same study area should be learned to have the same representation. However, due to the multi-solution nature of geological information and the data dimensionality (the number of prediction elements), consistency in learned features is a core issue in deep learning metallogenic prediction. Deep learning tasks are plagued by a shortage of positive and negative samples, necessitating the study of deep learning metallogenic prediction with few or even zero samples. Summary of the Invention

[0004] Based on the above problems, the purpose of this application is to provide an innovative deep learning mineral resource prediction method that can solve practical problems in prospecting and prediction in unfamiliar working areas under conditions such as low work level, little geological data, or few discovered mineral points.

[0005] In order to achieve the above objectives, the technical solution of the present application provides a method for extracting mineralization anomalies based on a geological "big model", comprising the following steps:

[0006] S1. Constructing text and spatial databases: Determine the search topic for the target mining area for mineralization anomaly extraction, perform text and spatial big data discovery using data acquisition tools, and obtain geological corpus and geological maps, respectively. After preprocessing, a text database and a vector layer library are generated. The text database serves as the corpus, and the vector layer library is used for spatial connection.

[0007] S2. Constructing a knowledge graph: Determine the ontology and entity types within the knowledge graph to be constructed, extract information from the text database using a pre-trained information extraction model, obtain "entity-relationship-entity" triples, construct nodes and edges of the knowledge graph, and form a knowledge graph; where the nodes are entities and the edges are relationships;

[0008] S3. Prediction model training: Use knowledge graph embedding technology to model the feature space of the obtained knowledge graph. Use convolutional neural networks to train the model, establish a mapping from data to knowledge space, and obtain a trained prediction model.

[0009] S4. Use the trained prediction model to perform mineralization prediction and mineralization anomaly extraction.

[0010] Furthermore, the specific construction method of the information extraction pre-training model in S2 is:

[0011] (1) Fine-tuning of geological corpus: Using Label Studio, randomly extract some text from the text database for data annotation to obtain a fine-tuning corpus;

[0012] (2) The fine-tuned corpus is divided into training set, validation set, and test set, and the fine-tuned geological corpus is obtained after training;

[0013] (3) The fine-tuned geological corpus is used to fine-tune the pre-trained model to obtain a fine-tuned information extraction pre-trained model.

[0014] Furthermore, the pre-trained model is an ERNIE-UIE open source model.

[0015] Furthermore, the specific steps of using knowledge graph embedding technology to perform feature space modeling on the obtained knowledge graph as described in S3 are: extracting all entities and relationships in the knowledge graph, using the knowledge graph embedding tool to train the vector representation of entities and relationships, mapping the embedding vectors of entities and relationships into the semantic space, and realizing feature space modeling.

[0016] Furthermore, the knowledge graph embedding tool is a KG2E model.

[0017] Furthermore, the specific method of using the convolutional neural network to train the model in S3 is:

[0018] (1) Pairing the data features of entities in the model after feature space modeling with corresponding semantic labels, constructing a sample pair set consisting of paired semantics and data as a training set, and using a convolutional neural network encoder (CNN) for training;

[0019] (2) Each time during training, one pair of samples with different semantics is selected from the training set, the data part is passed through the convolutional neural network to obtain the feature vector, and the cosine similarity of the semantic and data features is calculated to form a similarity matrix;

[0020] (3) The CNN encoder is updated by the gradient descent method, and the values ​​on the diagonal of the similarity matrix are optimized to be the largest in both the horizontal and vertical directions, that is, the similarity between samples belonging to the same sample pair is the largest, and the similarity between samples belonging to different sample pairs is the smallest.

[0021] (4) Repeat step (3) until the model converges, that is, the value of the loss function no longer decreases significantly, and the trained prediction model is obtained; wherein, during the training process, the model parameters with the best performance are saved.

[0022] Beneficial Effects: Compared to traditional deep learning-based mineralization prediction methods, this method can address practical challenges in prospecting in unfamiliar areas, where there is a lack of knowledge, limited geological data, or a small number of discovered mineral deposits. This method connects semantic space and feature space through spatial mapping, building a bridge between "big models" and "big data," thereby enabling knowledge- and data-driven delineation of prospecting areas with minimal sample size. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is the flow chart for constructing the knowledge graph of this application;

[0024] Figure 2 Embed a flow chart into the knowledge graph based on the KG2E model;

[0025] Figure 3 A technical flowchart for the model training process;

[0026] Figure 4 To call the prediction algorithm of CNNEncoder;

[0027] Figure 5 Visualize the results for similarity;

[0028] Figure 6 To overlay the prediction results with geochemical interpolation maps, mineralization strata distribution maps, structural and buffer zone maps;

[0029] Figure 7 It is the overlay map of similarity prediction results and geochemical comprehensive anomaly;

[0030] Figure 8 This is an overlay of the similarity prediction results, geochemical single elements, and the first principal component anomalies. DETAILED DESCRIPTION

[0031] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative efforts shall fall within the scope of protection of the present invention.

[0032] Unless otherwise specifically stated, the relative arrangement of the parts and steps, numerical expressions and numerical values ​​set forth in these embodiments do not limit the scope of the application. The technology, methods and equipment known to those of ordinary skill in the relevant art may not be discussed in detail, but in appropriate cases, the technology, methods and equipment should be considered as a part of the authorization specification. In all examples shown and discussed here, any specific value should be interpreted as being merely exemplary, rather than as a limitation. Therefore, other examples of the exemplary embodiments can have different values. It should be noted that similar numbers and letters represent similar items in the following drawings, and therefore, once an item is defined in one drawing, it does not need to be further discussed in subsequent drawings.

[0033] Unless otherwise specified, the meanings of technical terms in this specification are the same as those generally understood by those skilled in the art. However, if there is any conflict, the definitions in this specification shall prevail.

[0034] The term "low level of work" in this application refers to a low level of systematic geological research in the target area, with no basic or key exploration work completed. The study area is in the preliminary / general survey stage (not the detailed survey or exploration stage), with only a 1:200,000 or 1:50,000 regional geological survey completed. No geophysical surveys (such as gravity and magnetic methods), geochemical surveys (such as stream sediment measurements), or drilling verification have been conducted.

[0035] The term "scarce geological data" refers to the incomplete coverage or insufficient accuracy of multi-source geological data (structure, rocks, geochemistry, etc.), which makes it difficult to support refined modeling.

[0036] The term “few discovered mineral deposits” refers to the low number of known mineral deposits (points) in the study area, or the fact that no industrial-grade mineral deposits have been discovered in the exploration history.

[0037] like Figure 1 As shown, the embodiment of the present application discloses a method for extracting mineralization anomalies based on a geological "big model", comprising the following steps:

[0038] S1. Constructing text and spatial databases: Determine the search topic for the target mining area for mineralization anomaly extraction, perform text and spatial big data discovery using data acquisition tools, and obtain geological corpus and geological maps, respectively. After preprocessing, a text database and a vector layer library are generated. The text database serves as the corpus, and the vector layer library is used for spatial connection.

[0039] S2. Constructing a knowledge graph: Determine the ontology and entity types within the knowledge graph to be constructed, extract information from the text database using a pre-trained information extraction model, obtain "entity-relationship-entity" triples, construct nodes and edges of the knowledge graph, and form a knowledge graph; where the nodes are entities and the edges are relationships;

[0040] S3. Prediction model training: Use knowledge graph embedding technology to model the feature space of the obtained knowledge graph. Use convolutional neural networks to train the model, establish a mapping from data to knowledge space, and obtain a trained prediction model.

[0041] S4. Use the trained prediction model to perform mineralization prediction and mineralization anomaly extraction.

[0042] In this embodiment, the data acquisition tool in S1 is a data retrieval tool commonly used in the prior art. The preferred method for discovering text big data and spatial big data by using the data acquisition tool is:

[0043] (1) After determining the search topic, use the geological dictionary to perform word segmentation to determine the search subject terms;

[0044] (2) Reading the geological thesaurus, using the knowledge tree based on the geological thesaurus to calculate and analyze the input keywords, and obtaining more related words;

[0045] (3) Continue searching with the obtained related words as input until the search is complete or other termination conditions are met. In the link retrieval iteration phase, the keywords given in the previous phase are used as topic inputs, and the initial URL seeds generated for each keyword are used by mainstream search engine APIs to conduct link analysis and data crawling;

[0046] (4) Using search engine network resources and algorithms to obtain massive URLs;

[0047] (5) By performing a topic relevance analysis on the URL page data returned by the search engine (including title, abstract, source, and related URLs), URLs related to geological topics are stored in the crawl URL queue library, and research data are crawled according to the size of the relevance;

[0048] (6) Remove invalid, erroneous, and duplicate data; merge data from different data sources and resolve inconsistencies and conflicts between data.

[0049] In this embodiment, in the knowledge graph construction described in S2, the information extraction pre-training model is a conventional pre-training model from the prior art, such as open-source models such as DeepDive and ERNIE-UIE. The geological text primarily consists of common modern Chinese vocabulary and specialized geological and mineral terminology. The characteristics of geological text are primarily reflected in its standardized compilation and accurate wording. Terms from geology and other disciplines, such as statistics and computer science, are widely used, and even exhibit a mix of Chinese and English. Because geological research typically focuses on specific regions, the text contains a large number of place names. Furthermore, there is a nesting phenomenon between specialized terms, which may be a requirement for refined professional descriptions. These characteristics reflect the characteristics of geological literature as an exact science while also reflecting the complexity of professional descriptions. These characteristics of geological text make it suitable for general information extraction frameworks. There are many common open-source tools for information extraction, such as DeepDive. In this embodiment, the open-source large-scale model ERNIE-UIE is preferably used. The reasons are: first, it is an open-source pre-training model that is easily accessible; second, due to its use of prompt learning, it has excellent small-sample fine-tuning capabilities, greatly reducing the workload of corpus annotation. Due to the specialized nature of geological data and the fact that pre-trained language models use a small amount of specialized data, fine-tuning the open-source pre-trained model is necessary to ensure the accuracy of knowledge graph extraction. In this implementation, Label Studio is preferably used to annotate randomly extracted geological texts. The data exported by Label Studio is converted into the form required for input to the model through a script. Then, appropriate structural patterns are selected for annotation based on the requirements to achieve fine-tuning of the geological data.

[0050] In a further embodiment, a preferred method for fine-tuning the ERNIE-UIE pre-trained model using Label Studio is disclosed:

[0051] (1) Based on the constructed text database, a certain number of geological texts are randomly selected and annotated to form a geological corpus;

[0052] (1.1) Random sampling: Randomly select a certain number of geological texts from the text database and preprocess the selected texts, including removing format information, unifying fonts, and sentence separation.

[0053] (1.2) Annotation standards: clarify the entity types in the geological field, define the relationships between entities, and formulate annotation standards, including the use of annotation tools, annotation principles, and ambiguity handling.

[0054] (1.3) Corpus annotation: Use annotation tools to annotate entities and relationships.

[0055] (2) For the task of knowledge extraction, a language model is constructed and used to extract prospecting-related knowledge from text data; the extracted knowledge is stored in the form of a graph database.

[0056] (2.1) Data preparation: Divide the corpus into training set, validation set and test set.

[0057] (2.2) Model Selection and Training: Select a pre-trained language model suitable for geological text, such as BERT or GPT. Based on the pre-trained model, fine-tune the model using a geological corpus to adapt it to the characteristics of the geological field. Optimize model performance by adjusting parameters such as the learning rate and batch size.

[0058] (2.3) Knowledge extraction: Use the trained model to identify geological entities from the text; identify the relationships between entities and classify them into predefined relationship types.

[0059] (2.4) Knowledge storage: Convert the extracted knowledge into a format supported by the graph database; use the import tools or API provided by the graph database to import the knowledge data into the graph database.

[0060] In this embodiment, the specific steps of using knowledge graph embedding technology to perform feature space modeling on the obtained knowledge graph described in S3 are: extracting all entities and relationships in the knowledge graph, using the knowledge graph embedding tool to train the vector representation of entities and relationships, mapping the embedded vectors of entities and relationships into the semantic space, and realizing feature space modeling.

[0061] In a further embodiment, the specific method of using a convolutional neural network to train the model in S3 is:

[0062] (1) Pairing the data features of entities in the model after feature space modeling with corresponding semantic labels, constructing a sample pair set consisting of paired semantics and data as a training set, and using a convolutional neural network encoder (CNN) for training;

[0063] (2) Each time during training, one pair of samples with different semantics is selected from the training set, the data part is passed through the convolutional neural network to obtain the feature vector, and the cosine similarity of the semantic and data features is calculated to form a similarity matrix;

[0064] (3) The CNN encoder is updated by the gradient descent method, and the values ​​on the diagonal of the similarity matrix are optimized to be the largest in both the horizontal and vertical directions, that is, the similarity between samples belonging to the same sample pair is the largest, and the similarity between samples belonging to different sample pairs is the smallest.

[0065] (4) Repeat step (3) until the model converges, that is, the value of the loss function no longer decreases significantly, and the trained prediction model is obtained; wherein, during the training process, the model parameters with the best performance are saved.

[0066] The Xiarihamu area will be selected as the research area below, and the technical means and technical effects of the mineralization anomaly extraction method of the geological "big model" of this application will be explained in detail.

[0067] Although the study area in the Xiarihamu region has a high degree of geological work and is partially covered by 1:200,000 stream sediment geochemical data, only one known mineral occurrence has been obtained, thus meeting the criterion of “few discovered mineral occurrences” in this application.

[0068] S1. Constructing a text and spatial database: Aiming at the specific problems of the mining area, we used "Xiarihamu nickel-cobalt mine" as the search theme, collected a variety of data on Xiarihamu nickel-cobalt mine, including geological reports, academic papers, and geological maps, through LAN distributed search and WAN special crawler and double-iteration website discovery, and made open access to the national mineral deposit database at the National Geological Information Center. Through LAN data discovery, we found and obtained 14 geological reports, documents, monographs, etc. in the local LAN, and the 1:250,000 structural structure map drawn by the Qinghai Provincial Geological Survey Institute. Figure 2 The research team collected 100 pieces of geochemical data (Bruntai piece and Dazaohuo piece), 1 statistical table of geochemical data of stream sediments in the study area (containing 30 elements and 1779 sampling points in total), 30 single-element geochemical maps drawn based on the national regional geochemical database; and crawled 197 related papers in batches from China National Knowledge Infrastructure through wide area network big data discovery; and obtained open access to a national mineral deposit database at the National Geological Information Center.

[0069] S2. Constructing a knowledge graph: As a technical method for describing knowledge using a graph model, the concept of knowledge graph comes from semantic network. This paper mainly uses the method of constructing a domain knowledge graph to condense the knowledge contained in the geological literature of the study area. The technical process is as follows: Figure 1As shown in the figure, after determining the research objectives, text and spatial big data discovery and preprocessing are used to generate a text database and vector layer library. The text database serves as the corpus for constructing the knowledge graph, while the vector layer library, as attribute information, is spatially connected to the knowledge graph, laying the foundation for the next step of prediction. By organizing the core research requirements, the ontology of the knowledge graph is determined. The entity types within the ontology are then determined, which are collections of entities with the same characteristics or attributes. Next, we move into the entity extraction (named entity recognition) and relationship extraction stages, constructing nodes and edges in the knowledge graph, respectively. This stage can also be collectively referred to as information extraction. Using knowledge extraction techniques, we retrieve triples, the smallest units contained in the text. Then, through knowledge fusion and knowledge representation, we construct a complete knowledge graph. Geological texts primarily consist of common modern Chinese vocabulary and specialized geological and mineral terminology. The main characteristics of geological texts are their standardized and accurate use of terms. Terms from geology, statistics, computer science, and other disciplines are widely used, and some even exhibit a mix of Chinese and English. Because geological research typically focuses on specific regions, texts contain numerous place names. Furthermore, there is nesting between specialized terms, likely necessitating a refined description. These characteristics reflect the nature of geological literature as an exact science, while also reflecting the complexity of specialized descriptions. These characteristics of geological text make it suitable for general information extraction frameworks. Many common open-source tools for information extraction exist, such as DeepDive. For this study, we chose the open-source large-scale model ERNIE-UIE because, first, it is an open-source pre-trained model that is readily available; and second, due to its use of prompt learning, it possesses excellent small-sample fine-tuning capabilities, significantly reducing the workload of corpus annotation.

[0070] Due to the specialized nature of geological data and the fact that pre-trained language models use a limited amount of specialized data, we needed to fine-tune the open-source pre-trained model to ensure accurate knowledge graph extraction. We used Label Studio to annotate randomly extracted geological text. Using a script, we converted the exported data from Label Studio into the format required for model input. We selected appropriate structural patterns for annotation based on our needs. For example, in the Xia Rihamu research example, we selected the structural pattern "[spot] entity category [asso] relationship category [text]."

[0071] 1. Spotting: Locating target information fragments. In the case of research, the information that needs to be located is mainly entities such as the name of the mineral deposit, geographical location, geological background of the metallogenic environment, mineralization age, genetic type, mineralization type, geology, geophysics, geochemistry, remote sensing, etc.

[0072] 2. Associating: Identifying relationships between target information fragments. In this case study, to characterize prospecting knowledge for the Xiarihamu deposit, we only need to understand its geographic location, its geological setting, geologic age, genetic type, and mineralization type, as well as which geophysical and remote sensing information is beneficial for its exploration. Therefore, in this case study, we chose to identify three types of associations: "located in," "belongs to," and "favorable."

[0073] In this extraction task, on the label-studio platform, we constructed entity type labels including: ore deposit name, genetic type, mineralization type, mineralization age, mineralization geological background, geology, geophysics, geochemistry, remote sensing and geographical location; and constructed relationship type labels including: belongs to, located in and favorable. The labeling operation can be mainly described as: labeling the subject and object in the triple with entity type labels; adding relationship lines with arrows pointing from the subject to the object; selecting relationship type labels for the relationship. Based on the defined entity and relationship type labels, we annotated the text data, annotating a total of 1,883 sentences and generating 16,424 corpus. The sentences were randomly distributed in a ratio of 8:1:1 into a training set (1,507 sentences, 12,913 items), a validation set (188 sentences, 1,877 items), and a test set (188 sentences, 1,634 items).

[0074] After obtaining the correctly formatted geological corpus, we fine-tuned the ERNIE-UIE pre-trained model on the downstream task of geological information extraction and used the fine-tuned model to complete information extraction from the collected text database. The extracted information is output in the form of a structured extraction language, which can be organized into a large number of triples. In this study, we used UIE-base (containing 12 layers, 768 hidden units, and 12 attention heads) as the initial pre-trained model and fine-tuned it. During the evaluation process, we used a single-stage evaluation method and evaluated each positive example category separately. The validation / test set uses all labels at the same level to generate corresponding negative examples. Table 1 shows the performance of the model on geological text before and after fine-tuning on the geological corpus. It can be found that the performance of the fine-tuned model in geological text information extraction has been significantly improved.

[0075] Table 1 Performance of the model on geological texts before and after fine-tuning with geological corpus

[0076]

[0077] Table 1 shows that before fine-tuning, UIE-base had a certain degree of recognition for mineral deposit names, metallogenic eras, and geographic locations, but its overall recognition performance was poor. After fine-tuning with the geological corpus, its overall extraction performance significantly improved. Finally, we stored the organized triples in a graph database. This study used Neo4j as a knowledge storage tool. Neo4j is a popular graph database due to its high scalability, efficient representation and storage methods, and low hardware requirements.

[0078] S3. Prediction model training;

[0079] First, the neural system is used to solve the reasoning problem of the symbolic system, that is, the knowledge graph embedding technology is used to realize the semantic space modeling of the prospecting knowledge in the study area; then the neural system is improved with the help of the symbolic system, that is, the principle of using semantic space and feature space to jointly express the same entity is used to realize the feature space modeling of the comprehensive prospecting information in the study area by establishing a mapping.

[0080] In the knowledge graph representation stage, the example uses the KG2E model. According to the principle of the KG2E model, the pseudo code for implementing the knowledge graph embedding based on the KG2E model is shown in Table 2, and the technical flow chart is shown in Figure 2 shown.

[0081] Table 2 Learning algorithm of KG2E model

[0082]

[0083] Existing technologies generally construct prospecting concept models for the study area based on the knowledge extracted from the knowledge graph, further guiding prospecting prediction work in the study area. However, in this application, we aim to obtain the semantic embeddings of the knowledge in the National Mineral Deposit Information Database as a training set. The semantic embeddings of the National Mineral Deposit Information Database serve as the semantic training set for the model used for two-dimensional prospecting area delineation, driven by both knowledge and data, while the semantic embeddings of the prospecting knowledge graph for the study area serve as the data used for prediction. The knowledge in the National Mineral Deposit Information Database is structured and therefore easily converted into triples. Connecting the two knowledge graphs yields a total of 8,767 entities and three relationships, with a training set size of 36,651 and a test and validation set size of 4,072 each. Embedding is then performed using the learning algorithm of our KG2E model, resulting in a Gaussian distribution of all entities and relationships as the final semantic embedding result. We use the MR value (Mean Rank) as a parameter to evaluate model performance. This can be understood as the average ranking of the true values ​​of all triples in the list of predicted values, so the smaller the value, the better. The evaluation process can be understood as constructing incorrect triple entities and calculating the similarity between the head and tail entities. By comparing the similarity and ranking of correct and incorrect triples, the quality of the knowledge graph representation vector can be evaluated. Ideally, the score of the correct triple should be lower than that of the incorrect triple and the ranking should be higher. Our evaluation results show that after 90 rounds of training, the average ranking of the model has dropped from 860 to around 650.

[0084] On this basis, knowledge and data need to be integrated and a data-to-knowledge mapping established through training. Through the above experiments, we successfully obtained Gaussian embeddings for all entities and relations in the triples and derived a set of semantic vectors for all entities. Considering the subsequent model training, we will generate sample pairs required for training the four mappings. These pairs will serve as positive sample pairs in the training process.

[0085] Taking the genetic type as an example, the process of making these sample pairs is briefly described: obtaining the number of unique values ​​of the genetic type, in other words, counting how many different genetic types there are in the statistical semantic vector; counting the number of "ore deposit name" type entities belonging to each genetic type; removing the genetic types with too few entities, and retaining the genetic types with sufficient number of entities for training; querying all the ore deposit name entities connected to the retained genetic type entities; querying the comprehensive information attributes of these ore deposit name entities obtained through spatial connection; using the semantic vectors and the comprehensive information attributes of the genetic type entities connected to each ore deposit name entity to form a set of semantic and data sample pairs for training the genetic type CNN encoder.

[0086] The convolutional neural network encoder (CNNEncoder) is a key component in the deep learning architecture. Its main responsibility is to extract and encode the features of the input data in the convolutional neural network. In the convolutional neural network, the encoder usually refers to the first few layers of the network, which reduce the spatial dimension of the data through convolution and pooling operations while retaining key feature information. To simplify the training process, we designed a CNN encoder structure consisting of 5 convolutional layers and activation functions, 1 maximum pooling layer, and 1 fully connected layer, as shown in Table 3. The technical flow chart is as follows: Figure 3 As shown. The convolutional layer, activation function and maximum pooling layer of this structure are responsible for the conversion from comprehensive information to feature space, while the fully connected layer is responsible for the mapping from data to knowledge space. When training the CNN encoder, we use the following steps: (1) Construct a set of sample pairs consisting of paired semantics and data as the training set; (2) In each training, select one sample pair with different semantics from the training set, pass the data part through the CNN encoder to obtain the feature vector, and calculate the cosine similarity of the semantic and data features to form a similarity matrix; (3) Update the CNN encoder through the gradient descent method so that the values ​​on the diagonal of the similarity matrix are maximized in both the horizontal and vertical directions; (4) Repeat step 3 until convergence and save the best model. After training, we will use the best model to predict the comprehensive information of the study area.

[0087] Table 3 Learning algorithm of CNN encoder

[0088]

[0089] Based on in-depth research on metallogenic age, metallogenic geological setting, genetic type, and mineralization type, we developed a Python script using training and validation sets to train four different convolutional neural network (CNN) encoders. These four encoders focus on the aforementioned geological feature classification tasks. Classification tests were conducted using the test set data. The average values ​​of the three evaluation metrics, precision, recall, and F1 score, were obtained for metallogenic age, metallogenic geological setting, genetic type, and mineralization type, as shown in Table 4. Analysis of the model's performance in classifying different geological features revealed that metallogenic geological setting achieved the best classification performance across all three metrics. Overall, metallogenic geological setting achieved the most reliable classification results of the four categories, performing the best across all training metrics. This is likely due to the fact that it utilized the most training samples and had the fewest classifications. The classification performance of metallogenic age was second only to that of metallogenic geological setting in terms of precision and recall, and the F1 score also showed a high performance. This indicates that the model performs well in metallogenic age classification, accurately predicting and capturing positive samples. This is consistent with its training performance. We found that it used the second-most training samples. The significant drop in performance, even with the same number of classifications as the mineralization geological background, may be related to the sharp reduction in training samples. The classification performance for genetic type was similar in accuracy to that for mineralization age, but slightly lower than that for mineralization geological background. The F1 score showed a similar trend, indicating that the model performed well for genetic type classification but may have missed some positive examples. The genetic type encoder used fewer training samples and had more categories to distinguish, but its accuracy was comparable to that for mineralization age, rarely predicting negative examples as positive. This may indicate that geochemical data better characterizes genetic type than mineralization age. Mineralization type classification performance was the worst across all three metrics, indicating that the classification results are the least reliable of the four categories. The number of categories gradually increases from 14 for mineralization geological background to 19 for mineralization type, potentially indicating that the model's classification difficulty increases with the number of categories, which may contribute to the poor performance of mineralization type classification.

[0090] From the above analysis, we can see that the model's performance varies across different geological features. Metallogenic geological background performs best, followed by metallogenic age and genetic type, and mineralization type performs worst. Furthermore, as the number of categories increases and the number of samples decreases, the model's classification performance also decreases.

[0091] Table 4 Comparison of model evaluation indicators

[0092]

[0093] S4. Use the trained prediction model to predict mineralization and extract mineralization anomalies

[0094] We will train and predict the four attributes of metallogenic geological background, metallogenic age, genetic type and mineralization type. Through training and prediction, we will obtain the prediction algorithm pseudo code of calling CNN encoder for each grid cell in the study area as shown in Table 5, and the prediction algorithm flowchart of calling CNNEncoder as shown in Figure 4 The predicted mineralization probability is calculated using the average of the four mineralization probabilities (shown in Figure 2). These models do not require retraining and can be used to predict other types of minerals in the study area. To predict other types of minerals, we simply need to modify the semantic vectors used during prediction.

[0095] Table 5 Pseudocode of the prediction algorithm calling the CNN encoder

[0096]

[0097] The data of the study area is passed through four trained CNN encoders in grid units, and the average of the four similarities is taken as the similarity of the grid unit to the "Xiarihamu" standard point. The visualization results are as follows: Figure 5 As shown in the figure, the visualization results can be regarded as the extraction results of geochemical anomalies of stream sediments in the study area (extremely strong similarity: top 5%, strong similarity: top 15%, moderate similarity: top 25%, weak similarity: similarity greater than 0 but not in the top 25%, dissimilar area: similarity less than 0).

[0098] Our analysis reveals that within the study area, the known Xiarihamu copper-nickel-cobalt deposit is located within a moderately similar region, falling within the top 15%-25% of anomalies. Four significant anomaly regions were observed, three in the northern part of the study area and one in the southwest. In the southern part of the study area, strong anomalies exhibit a more continuous distribution pattern, while in the central region, strong anomalies are primarily scattered and point-like. The central region of the study area is primarily composed of large areas of moderately similarity. Furthermore, similarity is generally lower in the eastern part of the study area.

[0099] In order to gain a deeper understanding of the specific element anomalies revealed by these abnormal extraction results, we conducted an overlapping analysis of the extracted abnormal results and the geochemical data of each element in the study area after interpolation processing using the inverse distance weighted method. According to existing academic research, the magmatic melt-type copper-nickel sulfide deposits in Xiarihamu are closely related to the anomalies of the water system sediment measurements of nickel (Ni), cobalt (Co), copper (Cu) and chromium (Cr). Figure 6:The prediction results are superimposed with the geochemical interpolation map, the ore-bearing stratum distribution map, the structure and buffer zone map, Figure 7 : As shown in the overlay map of the similarity prediction results and the geochemical comprehensive anomaly, through overlay analysis, we found that the predicted high-value areas are extremely consistent with the high-value areas of Ni, Co, and Cu elements. They also show good consistency with the high-value areas of elements such as manganese (Mn), titanium (Ti), vanadium (V), and iron oxide (Fe2O3). In addition, the median area of ​​the predicted results (between the top 15%-25%) shows a relatively consistent correspondence with the spatial distribution and structural characteristics of the Jinshuikou Group, the mineral-bearing strata in the area, and its 1,500-meter buffer zone. This discovery further enhances our understanding of the distribution pattern of mineral points in the study area.

[0100] like Figure 8 The figure shows the superposition of the similarity prediction results with the geochemical single element and the first principal component anomalies: the similarity prediction results are superimposed with the three single element anomalies of Cu, Ni, and Co that are closely related to mineralization, and the Fac1 anomaly that may reflect the influence of magmatic activity in the study area. It was found that the similarity prediction results identified the high-value anomalies of the three single elements in the western part of the study area, but did not identify the high-value area in the southeastern part of the study area as an anomaly favorable for mineralization. The similarity prediction results also have a certain degree of superposition with the Fac1 anomaly. The anomaly distribution in the central part of the study area mentioned above is well superimposed with the Fac1 anomaly in the central part. The above phenomena may indicate that the similarity prediction results are not a simple stacking of high-value areas of geochemical anomalies, but rather a dimensionality reduction representation of high-dimensional geochemical anomalies of a certain type of mineral deposit.

[0101] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A method for extracting mineralization anomalies based on a geological "big model", characterized in that: The steps include: S1. Constructing text and spatial databases: Determine the search topic for the target mining area for mineralization anomaly extraction, perform text and spatial big data discovery using data acquisition tools, and obtain geological corpus and geological maps, respectively. After preprocessing, a text database and a vector layer library are generated. The text database serves as the corpus, and the vector layer library is used for spatial connection. S2. Constructing a Knowledge Graph: Determine the ontology and entity types within the knowledge graph to be constructed. Use a pre-trained information extraction model to extract information from a text database to obtain "entity-relationship-entity" triples. These triples are then constructed to form nodes and edges within the knowledge graph, where the nodes are entities and the edges are relationships. S3. Prediction model training: Use knowledge graph embedding technology to model the feature space of the obtained knowledge graph, train the convolutional neural network model, establish a mapping from data to knowledge space, and obtain the trained prediction model; S4. Use the trained prediction model to perform mineralization prediction and mineralization anomaly extraction; The specific steps of using the knowledge graph embedding technology to perform feature space modeling on the obtained knowledge graph in S3 are: extracting all entities and relationships in the knowledge graph, using the knowledge graph embedding tool to train the vector representation of entities and relationships, mapping the embedding vectors of entities and relationships into the semantic space, and realizing feature space modeling; The specific method for training the convolutional neural network model described in S3 is: (1) Pairing the data features of entities in the model after feature space modeling with corresponding semantic labels, constructing a sample pair set consisting of paired semantics and data as a training set, and training the convolutional neural network encoder CNN; (2) Each time during training, one pair of samples with different semantics is selected from the training set, the data part is passed through the convolutional neural network to obtain the feature vector, and the cosine similarity of the semantic and data features is calculated to form a similarity matrix; (3) Update the CNN encoder by gradient descent method, and optimize the diagonal values ​​in the similarity matrix to maximize both horizontal and vertical directions, that is, the similarity between samples belonging to the same sample pair is maximized, and the similarity between samples belonging to different sample pairs is minimized; (4) Repeat step (3) until the model converges, that is, the value of the loss function no longer decreases significantly, and the trained prediction model is obtained; wherein, during the training process, the model parameters with the best performance are saved.

2. The extraction method according to claim 1, wherein The specific construction method of the information extraction pre-training model in S2 is: (1) Fine-tuning of geological corpus: Using Label Studio, randomly extract some text from the text database for data annotation to obtain a fine-tuning corpus; (2) The fine-tuned corpus is divided into training set, validation set, and test set, and the fine-tuned geological corpus is obtained after training; (3) The fine-tuned geological corpus is used to fine-tune the pre-trained model to obtain a fine-tuned information extraction pre-trained model.

3. The extraction method according to claim 2, characterized in that The pre-trained model is the ERNIE-UIE open source model.

4. The extraction method according to claim 1, wherein The knowledge graph embedding tool is the KG2E model.

Images