Mine prospecting prediction method based on multi-agent technology

Abstract

The invention provides a prospecting prediction method based on a multi-agent technology. The method specifically comprises the steps that S11, keyword extraction is conducted on mineral problems input by a user; s12, establishing a vector database of the mineral knowledge text and a graph database of the mineral knowledge graph, and optimizing the extracted keywords to generate cue words; s13, constructing a large prospecting model to think the cue words, expanding the cue words, and obtaining a complete statement for professional expression of the user question again; s14, performing multi-task decomposition on the complete statement to obtain a plurality of sub-tasks, and scheduling the plurality of sub-tasks to corresponding agents based on a graph database of the mineral knowledge graph to perform step-by-step prediction; and S15, integrating prediction results of the intelligent agents and feeding back the integrated prediction results to a user. According to the method, prospecting prediction is promoted to be converted from manual driving to intelligent driving, and emerging technical support is provided for prospecting breakthrough.

Description

Technical Field

[0001] The present invention relates to the technical field of mineral prediction, and particularly to a prospecting prediction method based on multi-agent technology. Background Art

[0002] In the geological field, traditional geological analysis often relies on manual observation of limited samples and empirical judgment, which is inefficient and may be one-sided. Artificial intelligence, with its powerful algorithms, can quickly analyze massive geological data, discover hidden laws and features, promoting the vigorous development of artificial intelligence in the fields of mathematical geosciences and mineral exploration, providing more scientific and efficient technical means for prospecting work. The application of new generation information technologies such as big data, machine learning, and cloud computing in prospecting exploration work for prospecting prediction can improve the efficiency of prospecting. Looking at home and abroad, the emerging development trend of detecting mineral resources has been clearly presented. In future mineral resource detection and prediction work, more attention will be paid to integrating data resources in different spatio-temporal dimensions, using powerful computing capabilities and advanced model algorithms to discover more valuable prospecting information.

[0003] With years of mining and the continuous increase in exploration difficulty, many challenges are faced, such as the detection of deep mineral resources, the discovery of hidden ore deposits, and metallogenic prediction under complex geological conditions, etc. Developing prospecting prediction methods based on technologies such as RAG (Retrieval-Augmented Generation), knowledge graphs, large models, agents, big data, spatial computing, and scientific analysis has become a key direction. Developing new AI prospecting productive forces is of crucial significance for breaking through and solving major scientific problems and technical difficulties in resource evaluation and prediction. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide a prospecting prediction method based on multi-agent technology to improve the efficiency and accuracy of prospecting prediction through intelligent means.

[0005] To achieve the above-mentioned invention purpose, the present invention provides a prospecting prediction method based on multi-agent technology, and the method includes:

[0006] S11. Extract keywords from the mineral problems input by the user;

[0007] S12. Establish a vector database of mineral knowledge texts and a graph database of mineral knowledge graphs, and optimize the extracted keywords to generate prompt words;

[0008] S13. Construct a prospecting large model to think about the prompt words, expand the prompt words, and obtain a complete sentence for re-professional expression of the user's problem;

[0009] S14. Decompose the complete statement into multiple tasks to obtain several subtasks, and schedule the several subtasks to the corresponding agents for step-by-step prediction based on the graph database of the mineral knowledge graph;

[0010] S15. Integrate the results predicted by each agent and feedback them to the user.

[0011] Further, in step S11, extract keywords from the user's question, specifically including:

[0012] S21. Parse the mineral problem through natural language processing technology and match and classify it with the nodes and relationships in the mineral knowledge graph;

[0013] S22. Combine the mineral knowledge graph and the historical mineral question data to automatically generate keywords corresponding to the mineral problem.

[0014] Further, in step S12, establish a vector database of mineral knowledge texts and a graph database of the mineral knowledge graph, specifically including:

[0015] S31. Collect prospecting prediction papers, monographs, and geological report document materials in real time and store them;

[0016] S32. Convert the stored document data into a unified format and perform preprocessing;

[0017] S33. After dividing the processed document data into segments according to a preset length, input them into an embedding model to convert the document data into vectors, and construct a vector database of mineral knowledge texts;

[0018] S34. Perform knowledge extraction, relationship extraction, and attribute extraction on the document data divided into segments respectively, and fuse the extracted data to construct a graph database of the mineral knowledge graph.

[0019] Further, in step S12, optimize the extracted keywords to generate prompt words, specifically including:

[0020] S41. Construct a graph database query language and query statements, check the grammar of the keywords, and determine the specific graph structure and information that the user wants to query;

[0021] S42. Based on the analysis of the graph structure and the query requirements, formulate a specific query execution plan to determine the order of accessing nodes and relationships;

[0022] S43. Traverse the nodes and relationships, organize the qualified nodes, relationships, and related attribute information, and generate prompt words.

[0023] Further, in step S13, construct a prospecting large model, specifically including:

[0024] S51. Clean the stored data, select a word segmentation tool to perform word segmentation on the cleaned data, expand the vocabulary, and increase the diversity of samples through text data augmentation techniques;

[0025] S52. Divide the processed data into a training set, a validation set, and a test set according to a preset ratio, train, optimize, and evaluate the pre-trained large ore prospecting model to obtain the large ore prospecting model;

[0026] S53. Adjust the dimensions of the word embedding layer and the output layer of the large ore prospecting model according to the expanded vocabulary to be consistent with the size of the expanded vocabulary;

[0027] S54. Conduct supervised adjustment, incremental training, and ore prospecting instruction learning on the large ore prospecting model.

[0028] Furthermore, step S14 also includes:

[0029] Use natural language processing techniques to perform semantic understanding on the complete sentence to extract key information, and perform task decomposition. At the same time, establish a semantic label library for the algorithms corresponding to each agent. According to the matching degree between the task semantics and the algorithm semantic labels, intelligently select and arrange the algorithms to match the tasks and algorithms;

[0030] Construct a dynamic Bayesian network, with the geological characteristics of the ore prospecting area, data quality, and computing resources as nodes, and the dependence relationships between various factors as edges. Dynamically adjust the selection and execution order of the algorithms by updating the node information and the weights of the edges in real time;

[0031] Explore the optimal cooperation strategy of the agents through reinforcement learning, maintain real-time communication between the agents, and adjust the cooperation strategy between the agents according to the global reward signal;

[0032] Record the execution process and results of the algorithms, and give rewards to the agents corresponding to the algorithms whose cooperation degree reaches the preset threshold in the cooperation;

[0033] Real-time monitor the computing resource usage of each agent, and dynamically adjust the execution order of the algorithms and allocate resources according to the remaining resources and the resource requirements of the algorithms.

[0034] Furthermore, using natural language processing techniques to perform semantic understanding on the complete sentence to extract key information specifically includes:

[0035] S61. Identify the vocabulary in the complete sentence and the semantic roles between the vocabulary, and analyze the dependency relationships between the vocabulary in the complete sentence;

[0036] S62. Capture the sentence structure and sequential information in the complete sentence through BiLSTM, and extract the context semantic features in the complete sentence;

[0037] S63. Extract the key information in the complete sentence by combining the dependency relationships between words obtained in step S61 and the context semantic features extracted in step S62 with the graph database of the mineral knowledge graph.

[0038] Furthermore, explore the optimal cooperation strategy of the agents through reinforcement learning, specifically including:

[0039] S71. Define the local state, global environment, and global reward signal of each agent;

[0040] S72. Interact the observation information of all agents about the local state and global environment to obtain the current agent that incorporates the observation information of other agents;

[0041] S73. Process the observation information of different dimensions through semantic embedding and transform it into Q (Query), K (Key), and V (Value) sequences. Learn the features and their mutual relationships of the Q, K, and V sequences through the self-attention mechanism, and fuse them through the Transformer structure to generate a globally optimized hidden state;

[0042] S74. Use the fused hidden state to generate the local Q value of each agent through a multi-layer perceptron (MLP), and combine the local Q values of each agent into a global Q value through a value decomposition method;

[0043] S75. Update the policy network of each agent through a reinforcement learning algorithm using the global Q value and the global reward signal.

[0044] Furthermore, in step S15, integrate the results predicted by each agent and feedback them to the user, specifically including:

[0045] Integrate the results predicted by the agents into forms such as text, pictures, GIS maps, 3D models, and file reports and feedback them to the user.

[0046] Compared with the prior art, the beneficial effects of the present invention are:

[0047] A prospecting prediction method based on multi-agent technology provided by the present invention realizes keyword extraction of mineral problems through natural language processing technology, innovatively combines the vector database of mineral knowledge texts with the graph database of mineral knowledge graphs, optimizes the keywords to generate prompt words, thereby improving knowledge retrieval and enhancing performance, and uses a pre-trained prospecting large model to expand the prompt words to generate complete statements for prospecting problems. The intelligent arrangement and scheduling of the prospecting prediction algorithm are realized through multi-agent technology, the advantages of the prospecting large model and professional algorithm tools are integrated, and the intelligent level of tasks such as data query, mining, information extraction, prospecting prediction, target area optimization and decision support is improved; the prediction results are fed back through a multi-modal data structure to meet the diverse needs of users for mineral prediction. Compared with the traditional prospecting prediction method that is mainly manual and relies on decentralized tools, the present invention promotes the transformation of prospecting prediction from manual drive to intelligent drive and provides emerging technical support for prospecting breakthroughs. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present application, and for those of ordinary skill in the art, without creative efforts, other drawings can also be obtained according to the provided drawings.

[0049] Figure 1 It is a schematic flow chart of a prospecting prediction method based on multi-agent technology provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0050] The present invention will be further described in detail below with reference to the drawings and embodiments. It can be understood that the specific embodiments described herein are only used to explain the present invention, rather than limiting the present invention. In addition, it should be noted that for the convenience of description, only parts related to the present invention are shown in the drawings, rather than all structures.

[0051] Refer to Figure 1 , this embodiment provides a prospecting prediction method based on multi-agent technology, and the method includes:

[0052] S11. Perform keywords on the mineral problems input by the user. To ensure the security and privacy of data, a security mechanism and a permission management system are introduced. Through technologies such as identity authentication, access control, and data encryption, it is ensured that only authorized users can proceed with the subsequent prospecting prediction process. At the same time, the access records are audited in detail and the user behavior is tracked to promptly discover and handle potential security problems, providing reliable security guarantees for the prospecting prediction work.

[0053] In the operation process of the entire prospecting prediction method, keywords are extracted for various professional mineral problems related to prospecting prediction proposed by users. For example, mineral problems include but are not limited to: viewing the database of the prediction work area, analyzing and classifying mineral prediction types, determining the type of mineral prediction method, summarizing the metallogenic geological background and mapping; the extracted keywords include but are not limited to: mineral prediction, database, prediction type, prediction method.

[0054] In step S11, keywords are extracted from the user's questions, specifically including:

[0055] S21. After parsing the mineral problems through natural language processing technology, they are matched and classified with the nodes and relationships in the mineral knowledge graph.

[0056] S22. Combining the mineral knowledge graph and historical mineral question data, automatically generate keywords corresponding to the mineral problems.

[0057] In this embodiment, in order to extract keywords for mineral problems, a professional question classification and intelligent prompt technology is proposed. First, the mineral problems are classified based on the knowledge graph. The mineral knowledge graph in the field of prospecting prediction is innovatively utilized. When the user inputs a question, the natural language processing technology is used to parse the mineral problem and match it with the nodes and relationships in the mineral knowledge graph, so as to realize the intelligent classification of the mineral problem. For example, for the question of "analyzing and classifying mineral prediction types", through the mineral knowledge graph, it can be quickly determined that it belongs to the knowledge category of mineral prediction type classification, providing an accurate classification basis for subsequent question processing.

[0058] Secondly, an intelligent prompt word generation and optimization technology and question intention recognition and reasoning are proposed. According to the partial content input by the user or the general direction of the question, combined with the mineral knowledge graph and historical question data, relevant keywords are automatically generated and dynamically recommended to the front-end user for alternative use. For example, when the user inputs "view prediction", the present invention can prompt "view the database of the prediction work area", "view the prediction model parameters", etc., improving the efficiency and accuracy of question input. At the same time, the algorithm can continuously optimize the keyword generation strategy according to the user's feedback and usage habits. Finally, the multi-modal input support is optimized, in addition to text input, voice input, image input, etc., providing data support for subsequent prospecting prediction.

[0059] S12. Establish a vector database of mineral knowledge texts and a graph database of the mineral knowledge graph, and optimize the extracted keywords to generate prompt words. Use the papers and monographs on prospecting prediction as the knowledge base, and at the same time establish a vector database of mineral knowledge texts and a graph database of the mineral knowledge graph, perform knowledge enhancement on the keywords through retrieval, and after traversing the vector database and the graph database, perform context fusion to form high-quality prompt words that are more in line with the prospecting prediction business.

[0060] In this embodiment, a knowledge retrieval and verification technology with enhanced graph structure is first proposed. In the RAG architecture, a graph structure is used to organize and store knowledge. Nodes represent knowledge entities, and edges represent the relationships between entities. When the large model generates content, knowledge retrieval is performed based on the graph structure, retrieving not only the entities related to the question but also the association relationships between these entities. After the retrieval results are returned, the relationship information in the graph is used to verify the retrieved knowledge to ensure that the knowledge relied on by the generated content is accurate and interrelated. Secondly, a generation constraint technology based on graph embedding is proposed. The nodes and edges in the graph structure are embedded into a low-dimensional vector space to obtain the embedding representation of the graph. When the large model generates content, the generation process is associated with the graph embedding. By constraining the consistency between the vector representation of the generated content and the graph embedding, it is ensured that the generated content conforms to the knowledge structure and logical relationships contained in the graph. Through graph embedding constraints, it is ensured that the relationships between various concepts in the generated content are consistent with the scientific knowledge system represented in the graph.

[0061] Thirdly, a knowledge enhancement technology for multi-modal graph fusion is proposed. Information from multiple modalities such as text, images, audio, and spatial data is combined to construct a multi-modal graph. In the RAG process, not only text knowledge is used for retrieval and generation, but also information from other modalities is fused to enhance the accuracy and completeness of knowledge. In addition to text descriptions, spatial data can also be combined to more accurately propose the regional geological characteristics, mineral geological characteristics, and metallogenic laws of the prediction area, thereby reducing the occurrence of hallucinations.

[0062] Then, a dynamic graph update and feedback mechanism is proposed. A dynamic graph update mechanism is established. According to the generation results of the large model, user feedback, and the maintenance and update of the knowledge base and spatial database, the vector database and graph database are updated in real time. When hallucinations or errors are found in the generated content, by analyzing the reasons for the errors, the relevant knowledge in the graph is corrected or supplemented to improve the generation quality of the large model and reduce the occurrence of hallucinations. Finally, a knowledge focusing technology based on graph attention mechanism is proposed. In the RAG process, the graph attention mechanism is introduced, and according to the characteristics of the question and the requirements of the generation process, key knowledge nodes and edges in the graph are dynamically focused on.

[0063] In step S12, a vector database for mineral knowledge texts and a graph database for mineral knowledge graphs are established, specifically including:

[0064] S31. Real-time collect prospecting prediction papers, monographs, and geological report document materials and store them;

[0065] S32. Convert the stored document data into a unified format and perform preprocessing;

[0066] S33. After chunking the processed document data into segments according to a preset length, input the document data into an embedding model to convert it into vectors, and construct a vector database for mineral knowledge texts.

[0067] S34. Perform knowledge extraction, relation extraction, and attribute extraction on the document data chunked into segments respectively, and fuse the extracted data to construct a graph database for a mineral knowledge graph.

[0068] In this embodiment, to establish a vector database for mineral knowledge texts and a graph database for a mineral knowledge graph, first, store document materials such as papers, monographs, and geological reports related to prospecting prediction in the personal user's workspace. Second, extract and parse all the document data in the personal user's workspace, including papers, monographs, geological reports, etc. Third, preprocess the document data, including but not limited to: file parsing, eliminating noise, redundancy, irrelevant, and potentially harmful data, and converting document data in different formats into the required formats txt or json for training data. PDF files in picture format need to be parsed relying on the optical character recognition (OCR) technology in visual algorithms. Use methods such as document cutting, special format processing, short sentence removal, paragraph splicing, deduplication, meaningless paragraph deletion, and body citation replacement to improve the quality of the document data.

[0069] Then, chunk the document data into smaller segments to obtain the processed document data. Finally, convert the processed document data into a vector representation form. For example, map words, sentences, or documents into a high-dimensional vector space through techniques such as word embedding, so that the machine can understand and process this data, capture semantic information and relationships in the text. For example, adopt the Transformer architecture to perform deep bidirectional encoding on the text and learn richer semantic information and context relationships. For the vector database of mineral knowledge texts, first perform retrieval, query and retrieve the similarity between document data according to keywords, and select the most relevant document data from the pre-constructed personal user's workspace. For example, adopt the Hybrid Retrieval method. First, use a sparse method to quickly screen out candidate document data, and then accurately locate the most relevant document data through a dense method.

[0070] Second, perform re-ranking. After the initial retrieval, use a re-ranking stage to improve the relevance of the retrieved document data, ensure that the most relevant information is displayed at the top of the list, and effectively reorder the document data using a more accurate and time-consuming method, thereby improving the similarity between the query and the top-ranked document data. Re-ranking gives priority to performance and efficiency, mainly including two methods: DLMReranking and TILDEReranking.

[0071] Then, the document data is repackaged. The performance of LLM response generation may be affected by the provided document data. In the process of repackaging the document data after reordering, there are currently three main repackaging methods: "forward", "reverse", and "sides". The "forward" method repackages the document data by descending the relevance scores starting from the reranking stage. The "reverse" method arranges the document data in ascending order; when the relevant information is placed at the head or tail of the input, the "sides" method can achieve the best performance.

[0072] A vector database is a database used to store, manage, and retrieve vector data. In the RAG scenario, document data will first be converted into vector representation forms, mapping words, sentences, or documents into a high-dimensional vector space, and these vector data are stored in the vector database. Facing a large amount of text data, the vector database can store the corresponding vectors in an orderly manner and establish an effective indexing mechanism to facilitate subsequent rapid searching and matching, enabling it to quickly locate the vectors corresponding to the information related to the user's question in RAG applications.

[0073] The vector database can find the text content corresponding to the vectors with relatively high similarity in an extremely short time, providing a key guarantee for accurately retrieving relevant knowledge from the knowledge base in RAG, thereby improving the quality and accuracy of the generated answers. The vector database has four capabilities, including multiple index types, support for billions of vectors, hybrid search, and cloud-native capabilities. Multiple index types provide the flexibility to optimize searches according to different data characteristics and use cases, billions of vectors handle large datasets, hybrid search combines vector search with traditional keyword search to improve retrieval accuracy, and cloud-native capabilities ensure seamless integration, scalability, and management in cloud environments.

[0074] The graph database of the mineral knowledge graph is used for the overall planning of the prospecting prediction knowledge graph. First, determine the entity types and sort out the core entities in the field of prospecting prediction. For example, the entity categories in the prospecting prediction knowledge graph include but are not limited to: "deposit model", "prospecting prediction model", "genetic model", "deposit geological characteristics". Clearly defining the core entities is the basis for constructing the graph relationships subsequently.

[0075] Secondly, define the relationship types and analyze various connections existing between entities in the mineral resources knowledge graph. For example, there is an "inclusion" relationship between "prediction elements" and "ore-controlling structures", "rock masses", and "gravity anomalies", and there is an "anomaly combination" relationship between "Au element anomaly" and "As element anomaly", etc. Defining these relationships accurately and clearly can build a complete knowledge network. Then construct the schema layer to describe the entities and relationships related to ore prospecting prediction in a structured way, and use graph models (such as RDF, etc.) or ontology languages for expression, which is equivalent to drawing the "blueprint" of the knowledge graph.

[0076] According to the entity types, relationship types, and schemas determined by the mineral resources knowledge graph plan. First, conduct knowledge extraction, including entity extraction, relationship extraction, and attribute extraction; secondly, conduct relationship extraction to deeply analyze the semantics of document data; then conduct attribute extraction to extract relevant attributes of entities, further enriching the connotation of entities and making the knowledge in the mineral resources knowledge graph more complete and detailed. Finally, conduct knowledge fusion, including entity alignment and knowledge merging. Entity alignment ensures the uniqueness and accuracy of entities in the mineral resources knowledge graph, and knowledge merging is used to integrate knowledge extracted from different channels and remove duplicate and contradictory parts.

[0077] Store the constructed mineral resources knowledge graph in a graph database. The graph database is based on graph theory and represents data in the form of nodes (Node), edges (Edge), and the attributes (Property) they carry. Nodes are usually used to represent entities related to mineral prediction, such as "ore deposit geological characteristics", "ore-forming patterns", "prediction elements", "ore-controlling structures", "rock masses", "gravity anomalies", "magnetic anomalies", etc.; edges are used to describe the relationships between mineral prediction entities, such as the "parallel unconformity" relationship between "strata" and the "main-secondary" relationship between "faults", etc.

[0078] In step S12, optimize the extracted keywords to generate prompt words, specifically including:

[0079] S41. Construct a graph database query language and query statements, perform a syntax check on the keywords, and determine the specific graph structure and information that the user wants to query;

[0080] S42. Based on the analysis of the graph structure and query requirements, formulate a specific query execution plan to determine the order of accessing nodes and relationships;

[0081] S43. Traverse the nodes and relationships, organize the nodes, relationships, and relevant attribute information that meet the conditions, and generate prompt words.

[0082] In this embodiment, first, a large model is used to construct a graph database query language and query statements; second, the query statements are parsed for syntax checking to check whether the statements conform to the specified query language syntax rules. On the basis of correct syntax, the system further analyzes the semantics of the statements to determine the specific graph structure and information that the user wants to query. Third, the query execution path is planned. Based on the analysis of the graph structure and the query requirements, a specific query execution plan is formulated to determine the operation sequence such as which nodes to visit first and which relationships to traverse along. Then, the query operation is executed. According to the planned execution path, actual node and relationship traversal operations are performed. During the traversal process, according to the set search strategy, edges in the graph are used to find nodes that meet the conditions; finally, the query results are returned. After the query operation is completed, the nodes, relationships, and related attribute information that meet the conditions are sorted out to generate prompt words. The forms of the prompt words include but are not limited to being expressed in the form of a table or in JSON format.

[0083] S13. Construct a prospecting large model to think about the prompt words, expand the prompt words, and obtain a complete statement that re-expresses the user's question professionally. An incremental training is carried out on the basis of the general large model to obtain the prospecting large model, which overall improves the language learning ability, summarization ability, text generation ability, and text modification ability performance, enabling the prospecting large model to have the comprehensive research ability in the prospecting prediction business process.

[0084] In step S13, constructing the prospecting large model specifically includes:

[0085] S51. Clean the stored data, select a word segmentation tool to perform word segmentation on the cleaned data, expand the vocabulary, and increase the diversity of samples through text data augmentation technology;

[0086] S52. Divide the processed data into a training set, a validation set, and a test set according to a preset ratio, train, optimize, and evaluate the pre-trained prospecting large model to obtain the prospecting large model;

[0087] S53. Adjust the dimensions of the word embedding layer and the output layer of the prospecting large model according to the expanded vocabulary to be consistent with the size of the expanded vocabulary;

[0088] S54. Perform supervised adjustment, incremental training, and prospecting instruction learning on the prospecting large model.

[0089] In this embodiment, the prospecting large model module uses open-source general large models as the base LLM-3, such as Qwen72B and Llama3.2, and has acquired general knowledge generation capabilities through training on an open dataset. Based on the general large model LLM-3, prospecting datasets are appended on the basis of the personal user workspace, and an incremental pre-training is carried out to obtain a large model in the vertical field of minerals, that is, the prospecting large model. Specifically, it includes: First, clean the appended prospecting datasets, including but not limited to document parsing, noise elimination, removal of redundancy and irrelevant content; Second, perform data processing, select a tokenization tool to tokenize the prospecting knowledge text, and expand the vocabulary to adapt to the vocabulary in the prospecting prediction field; Through text data augmentation techniques such as synonym conversion, data back-translation, random insertion and deletion of words, increase the diversity of samples and improve the generalization ability of the model; Divide the processed dataset into a training set, a validation set and a test set, usually split according to a certain ratio, such as 8:1:1, for model training, tuning and evaluation.

[0090] Then perform model adjustment. The model dimension adjustment is based on the size of the expanded vocabulary, and the dimensions of the word embedding layer and the output layer of the large language model are re-adjusted to be consistent with the number of new vocabulary. Finally, perform incremental training. Select a training framework and determine three training strategies according to computing resources and training requirements. One is to first pre-train with general corpus such as encyclopedias, books, reports, papers related to large-scale prospecting predictions, and then perform secondary training with domain corpus such as prospecting prediction monographs and papers published by small-scale academicians and authoritative experts; The second is to directly pre-train with large-scale domain corpus; The third is to mix the general corpus and the domain corpus in a certain proportion and train them simultaneously.

[0091] Add prospecting-related question-and-answer pairs for prospecting instruction learning, including but not limited to: metallogenic geological background knowledge, metallogenic regularity knowledge, geophysical feature knowledge, geochemical feature knowledge. Supervise and adjust the prospecting large model. First, pre-train the large language model to obtain the parameter results on the general dataset, and then use the cross-entropy loss to adjust the prospecting large model on the labeled dataset of specific tasks, so that the parameters of the prospecting large model are adjusted on the basis of the initial parameters to minimize the loss function of the supervised task.

[0092] Adjust the instructions for the prospecting large model. First, collect or construct instances in the instruction format. Each instance in the instruction dataset consists of three elements: instruction, context supplement information (optional), and input-output based on the instruction. The instruction includes detailed information such as the name, description, precautions, positive and negative examples of the task. By means of manual integration and transformation or generation by large language models, a dataset in the format of "instruction, output" is obtained. Then, adjust the LLM in a supervised manner. After obtaining the general instruction dataset, it can be adjusted on the pre-trained prospecting large model. After the instruction adjustment, the prospecting large model can demonstrate excellent capabilities to generalize to unseen tasks. Instruction adjustment stimulates the language model's understanding ability, makes better use of prior knowledge, and enables the model to understand and respond correctly after giving more obvious instructions, generalizing to various tasks.

[0093] Add prospecting preference pairs to the prospecting large model, learn according to the preferences of prospecting experts, and address the risks of inaccuracy, misleadingness, or even harmful information that may exist when the prospecting large model generates content, ensuring that the behavior of the large language model is consistent with human values. Human preference alignment adjusts and optimizes the decision-making process of the large language model to ensure that its output is not only accurate but also follows ethical norms, is unbiased, and reflects the generally recognized values and ethical standards of society. The purpose of alignment is to create a model that can both understand and generate human language and also demonstrate an emphasis on fairness, transparency, and responsibility in its decision-making, reducing potential negative impacts. The prospecting large model adopts the Reinforcement Learning from Human Feedback (RLHF) method, which is divided into pre-training and supervised adjustment of the model, training the reward model, and using reinforcement learning to adjust the model.

[0094] First, start with supervised adjustment. Use the supervised learning method to adjust the prospecting large model in the annotator, and collect a supervised dataset containing input prompts (instructions) and the required outputs, that is, prospecting prediction knowledge preference pairs, to adjust the prospecting large model. These prospecting prediction knowledge preference pairs are written by human annotators for certain specific tasks while ensuring task diversity. Secondly, use the human feedback obtained from the human annotator to train the reward model. The human annotator ranks the output of the reward model according to the consistency between the reward model output and the expected behavior. Specifically, input sampled prompts (from the supervised dataset or human-generated prompts) into the reward model to generate a certain number of output texts, and then the human annotator annotates the preferences for these input-output pairs. Then, train the reward model to predict the output preferred by humans.

[0095] Finally, the reinforcement learning algorithm is combined with the trained reward model. The policy of the reinforcement learning problem is given by the pre-trained ore prospecting large model. The action space is the vocabulary of the language model, the state is the currently generated token sequence, and the reward is provided by the reward model. To prevent the ore prospecting large model from deviating significantly from the initial (before adjustment) model, a penalty term is usually incorporated into the reward function, and the ore prospecting large model is better aligned through multiple iterations. The PPO algorithm is used to further adjust the ore prospecting large model according to the received human feedback, so that the ore prospecting large model better conforms to human preferences and improves the instruction-following ability of the ore prospecting large model.

[0096] To guide the ore prospecting large model's ability to perform complex reasoning, the ore prospecting large model expands its capabilities and the scope of tasks it can effectively execute by using prompt engineering to achieve guidance and optimization. For the input prompt words, the ore prospecting large model will give corresponding output answers. Prompt words have three main content types: input, context, and examples. The input specifies the information for which the model needs to generate a response; context and examples are optional parts of the prompt. The context provides instructions on the model's behavior, and examples are input-output pairs in the prompt used to demonstrate the expected response.

[0097] S14. Decompose the complete statement into multiple tasks to obtain several subtasks, and schedule the several subtasks to the corresponding agents for step-by-step prediction based on the graph database of the mineral knowledge graph.

[0098] In this embodiment, the complete statement (complex ore prospecting prediction work task) is decomposed into several subtasks, and the subtasks of the ore prospecting prediction workflow are scheduled to the corresponding agents to achieve the final ore prospecting prediction goal. The agent can generate a plan without adjusting through different reasoning strategies, or adjust the generated plan according to external feedback. The present invention decomposes the complete statement into several subtasks through multitasking and strategically arranges the corresponding agents to ensure the seamless progress of the planned ore prospecting prediction task workflow.

[0099] The ore prospecting prediction multi-agent (hereinafter referred to as the agent) is an intelligent computing entity with highly intelligent understanding and generation capabilities, and with ore prospecting prediction microservice components capable of accurately perceiving the environment, making decisions, and executing actions, to complete specific tasks such as data query, data mining, and information extraction in the ore prospecting prediction business process.

[0100] First, record the historical thoughts, actions, and environmental observations generated during the execution of the prospecting prediction agent. For example, remember that the user reads geo-physical, geo-chemical, remote sensing data from the database, then uses tools such as geophysical processing and analysis tools, geochemical processing and analysis tools, and then uses the prospecting large model to select prediction elements. Based on the accumulated memories of various prospecting predictions, the agent can re-examine and utilize previous records and experiences, thus more effectively handling more complex prospecting prediction tasks.

[0101] The agent's memory can be divided into short-term memory and long-term memory according to the memory duration. The short-term memory is integrated into the agent to enhance its ability to maintain the trajectory of the current ongoing prospecting prediction task and is often used in multi-round interactions. The long-term memory is used to remember the valuable experiences of historical prospecting prediction tasks and is recalled by the prospecting prediction agent when solving unseen tasks. Due to the extensive trajectory, the long-term memory usually uses distillation techniques or only stores key information. Secondly, receive information from the prospecting prediction workspace environment. The prospecting prediction agent can perceive multi-modal inputs, such as text input, visual input, and auditory input. All historical prospecting prediction data is stored in text form, and prospecting prediction text can flexibly express intentions, information, and knowledge.

[0102] Thirdly, interact with and influence the external environment. The important mechanism is to control and utilize external prospecting prediction microservice algorithm tools to expand the inherent functions of the LLM by accessing more external resources and extending the operation space beyond individual text interactions. The prospecting prediction multi-agent includes decision support agents, target area optimization agents, prospecting prediction agents, information extraction agents, big data agents, GIS tool agents, mathematical geology agents, geostatistics agents, classical prediction agents, machine learning agents, deep learning agents, etc. Each agent automatically schedules and executes the microservices at the backend when executing to complete the corresponding subtasks. All microservices are connected to the geo-physical, geo-chemical, remote sensing database to achieve two-way interconnection, dynamic access, and storage of source data, information extraction data, prospecting prediction data, target area optimization data, and decision support data for prospecting prediction tasks.

[0103] The decision support agent is responsible for integrating, analyzing, and processing information from various data sources to provide intelligent support and suggestions for users. By considering multiple factors, it estimates the potential value of the target area and provides a scientific basis for users, including historical exploration data, market demand, environmental impact, and economic feasibility, etc. Continuously feedback and improve the decision-making model through machine learning technology. Through its adaptive characteristics, the decision support agent can maintain efficient and effective decision-making capabilities in a changing environment and geological conditions with high uncertainty.

[0104] The target area optimization agent identifies plots with exploration value from a vast amount of geological data, including a comprehensive analysis of geological, geophysical, geochemical, and remote sensing information. At the same time, the target area optimization agent performs risk assessment and economic feasibility analysis, considering factors such as the potential value of the mineral deposit, mining costs, environmental impacts, and legal restrictions, to ensure minimizing environmental impacts and complying with the principles of sustainable development. The target area optimization agent not only relies on historical data and expert experience but also incorporates dynamic data and has real-time update and feedback functions. When new geological survey data or market dynamics emerge, the target area optimization can quickly integrate this new information into its optimization model and timely adjust the target area recommendation results.

[0105] The prospecting prediction agent uses different algorithms and technical means to assist in more accurately exploring the possible locations of mineral resources and delineating prospecting prediction prospective areas and target areas. The positioning prediction agent in mineral resource prediction includes: the information content method mineral prediction agent, which analyzes the correlation between many geological factors and the known mineral distribution based on the information content theory. By calculating the amount of information carried by each factor, it measures its contribution to mineralization. Considering these information comprehensively, it judges the mineralization possibility of different sections in the study area and then delineates areas with greater prospecting potential, providing key directions for subsequent exploration work.

[0106] The weights-of-evidence method mineral prediction agent is based on Bayesian theory. It regards various geological evidences related to mineralization as independent "evidence layers". First, it determines the weights of each evidence layer for mineralization, including evidence factors such as stratigraphic lithology, structural characteristics, and geochemical anomalies. After complex statistical analysis, it obtains their respective weights of influence on mineralization. Then, based on these weights, it comprehensively evaluates the mineralization probability of different regions and makes predictions on favorable mineralization areas, making the mineral prediction work more scientific and improving the accuracy of prospecting.

[0107] The characteristic analysis method mineral prediction agent focuses on various characteristics of geological bodies, collects data on the mineral composition, texture and structure characteristics of rocks, geophysical field characteristics, etc. It deeply excavates and analyzes these characteristics, finds out the characteristic combinations closely related to mineralization. By establishing a characteristic model, the agent can identify areas where minerals may exist based on the characteristics of geological bodies presented in the study area, helping exploration personnel narrow the prospecting scope and focus on the most likely mineralization zones.

[0108] The random forest method mineral prediction agent is an ensemble learning algorithm composed of multiple decision trees. The corresponding mineral prediction agent utilizes its advantages to carry out work, taking a large amount of geological data as input, including geological structures, geochemical data, geophysical data, etc. Each decision tree grows based on partial data and random features and makes a mineralization prediction judgment. After aggregating the results of numerous decision trees, the final mineralization prediction result is determined through mechanisms such as voting, effectively reducing the error of a single model and enhancing the reliability and stability of the prediction, and can also function well under complex geological conditions.

[0109] The feature analysis method mineral prediction agent focuses on various features of geological bodies, collects data from multiple aspects such as the mineral composition, texture and structure characteristics of rocks, and geophysical field characteristics, deeply excavates and analyzes these features, finds out the feature combinations closely related to mineralization, and by establishing a feature model, based on the characteristics presented by geological bodies in the study area, identifies areas where minerals may exist, helping exploration personnel narrow down the prospecting scope and focus on the areas most likely to mineralize.

[0110] The support vector machine method mineral prediction agent is good at dealing with classification and regression problems. In mineral prediction, the agent maps data such as geological features into a high-dimensional space, searches for an optimal classification hyperplane to distinguish between mineralized areas and non-mineralized areas, can well handle the non-linear relationship of data, and can accurately mine potential mineralization laws even in the face of complex variable relationships in geological data, thus accurately predicting whether minerals exist in different areas and providing strong decision-making support for mineral exploration.

[0111] The convolutional neural network method mineral prediction agent has a powerful automatic feature extraction ability. This agent uses CNN to process geological data, geophysical data, geochemical data, etc., can automatically capture features such as formation textures and abnormal forms contained therein, continuously refines key information through operations such as convolutional layers and pooling layers, and judges the possibility of mineralization based on these features, demonstrating unique advantages and opening up new perspectives for prospecting work.

[0112] The graph neural network method mineral prediction agent emerged due to the existence of many complex network relationships in geological phenomena, such as graphic-structured geological structures and spatial associations of mineralization alteration. The graph neural network method mineral prediction agent constructs geological entities and the relationships between them into graph-structured data, mines the hidden mineralization information through information transmission and update of nodes and edges, can accurately grasp the mineralization laws under complex geological relationships, effectively predicts areas with mineralization potential, and meets the requirements of modern geological research for the analysis of complex relationships.

[0113] The adversarial neural network method for mineral prediction intelligent agent. The mineral prediction intelligent agent utilizes the characteristics of two mutually adversarial and collaborative architectures, namely the generator and the discriminator. The generator can generate the geological data distribution similar to the ore-forming pattern, and the discriminator is used to judge its authenticity. The two continuously play games, making the results generated by the generator increasingly approximate the real ore-forming situation. Through this method, more potential ore-bearing units can be simulated to assist users in analyzing the ore-forming possibility in the study area from different perspectives and more comprehensively. These different intelligent agents start from their respective principles and methods, complement each other, cooperate synergistically, and jointly provide diverse technical support for the complex and important work of mineral resource prediction, helping to improve the efficiency and success rate of prospecting prediction.

[0114] The information extraction intelligent agent, one of the key components in the prospecting prediction multi-intelligent agent system, is responsible for extracting valuable information from various data sources. The information extraction intelligent agent usually collaborates closely with other intelligent agents and transfers the extracted information to the prospecting prediction intelligent agent, the target area optimization intelligent agent, and the decision support intelligent agent through the way of API or microservices. This interconnected mechanism ensures the flow and sharing of information, making the entire prospecting prediction system form an efficient and coordinated working unit. The big data intelligent agent is responsible for managing and analyzing a large amount of exploration data, extracting trends and patterns, and providing support for other intelligent agents. The GIS tool intelligent agent utilizes geographic information system technology to analyze spatial data, helps the intelligent agents understand the relationship between the geographical environment and prospecting prediction, and provides support for other intelligent agents.

[0115] The mathematical geology intelligent agent applies mathematical models to explain the phenomenon of ore deposit distribution, helps predict the geological structure where the ore deposit is located, and provides support for other intelligent agents. The geostatistical intelligent agent is responsible for using statistical methods to analyze sample data, evaluate the confidence level and geological uncertainty, and provides support for other intelligent agents. The classical prediction intelligent agent conducts a preliminary assessment using traditional geological prediction methods, compares and corrects with the calculation results of other intelligent agents, and provides support for other intelligent agents. The machine learning intelligent agent utilizes machine learning technology to automatically learn patterns from data, provides additional basis for decision support, and provides support for other intelligent agents. The deep learning intelligent agent is a further development of machine learning, processes more complex sample data through neural network models, improves the accuracy of prediction, and provides support for other intelligent agents.

[0116] Step S14 also includes:

[0117] Utilize natural language processing technology to conduct semantic understanding of the complete sentence, extract key information, and decompose tasks. At the same time, establish a semantic tag library for the algorithms corresponding to each intelligent agent, and intelligently select and arrange algorithms according to the matching degree between the task semantics and the algorithm semantic tags to match the tasks and algorithms.

[0118] Construct a dynamic Bayesian network, taking the geological characteristics of the prospecting area, data quality, and computing resources as nodes, and the dependencies between various factors as edges. By updating the node information and the weights of the edges in real time, dynamically adjust the selection and execution order of the algorithms.

[0119] Explore the optimal cooperation strategy of agents through reinforcement learning, maintain real-time communication between agents, and adjust the cooperation strategy between agents according to the global reward signal.

[0120] Record the execution process and results of the algorithms, and give rewards to the agents corresponding to the algorithms whose cooperation degree reaches the preset threshold in the cooperation.

[0121] Monitor the computing resource usage of each agent in real time, and dynamically adjust the execution order of the algorithms and allocate resources according to the remaining resources and the resource requirements of the algorithms.

[0122] Use natural language processing technology to perform semantic understanding on complete sentences and extract key information, specifically including:

[0123] S61. Identify the words in the complete sentence and the semantic roles between the words, and analyze the dependency relationship between the words in the complete sentence.

[0124] S62. Capture the sentence structure and sequential information in the complete sentence through BiLSTM, and extract the context semantic features in the complete sentence.

[0125] S63. Combine the dependency relationship between the words obtained in step S61 and the context semantic features extracted in step S62, and extract the key information in the complete sentence in combination with the graph database of the mineral knowledge graph.

[0126] In this embodiment, analyze the dependency relationship between the words in the complete sentence, that is, determine the syntactic structure relationship between the words in the sentence. The dependency relationship reflects the "dominance" and "subordination" relationships between the words, and is used to understand the syntactic structure of the sentence. The output of the dependency relationship analysis is a dependency tree, where each word has a pointer pointing to its "dominant word". For example, identify the entities (ore body, ore grade) in the report and their relationships (the ore body is located in...). Extract the context semantic features in the complete sentence, consider the specific meaning of the words in the context, pay attention to the semantic roles, semantic associations of the words, and the overall semantics of the sentence. The output of the context semantic feature extraction is the semantic representation of the words or the semantic structure of the sentence. Combine the dependency relationship and the context semantic features with the mineral knowledge graph, and use the background knowledge in the knowledge graph to further enhance the semantic understanding to extract the key information in the complete sentence, such as the attributes of the ore type, the type of geological structure, exploration methods, etc.

[0127] Explore the optimal cooperation strategy of agents through reinforcement learning, specifically including:

[0128] S71. Define the local state, global environment, and global reward signal for each agent;

[0129] S72. Interact the observation information of all agents about the local state and global environment to obtain the current agent that incorporates the observation information of other agents;

[0130] S73. Process the observation information of different dimensions through semantic embedding, and transform it into Q (Query), K (Key), and V (Value) sequences. Learn the features and their mutual relationships of the Q, K, and V sequences through the self-attention mechanism, and fuse them through the Transformer structure to generate a globally optimized hidden state;

[0131] S74. Use the fused hidden state to generate the local Q value of each agent through a multi-layer perceptron (MLP), and combine the local Q values of each agent into a global Q value through the value decomposition method;

[0132] S75. Update the policy network of each agent through the reinforcement learning algorithm using the global Q value and the global reward signal.

[0133] In this embodiment, in a multi-agent system, the observation information of each agent is usually local, while the global information is crucial for optimizing the cooperation strategy. Through the self-attention mechanism, the observation information of the current agent can be interacted with the observation information of other agents. The reinforcement learning optimizes the cooperation strategy of the agents through the global reward signal, enabling the agents to better understand the global environment and make better cooperation decisions. The observation information of each agent comes from different dimensions, such as position, speed, state, etc. In order to enable this information to be effectively processed in the self-attention mechanism, it is necessary to first map them into a unified vector space through semantic embedding. Then, it is transformed into Q, K, and V sequences, and their features and mutual relationships are learned through the self-attention mechanism, thereby realizing the information interaction between agents and the improvement of the global perception ability. The Q, K, and V sequences processed by the self-attention mechanism contain the local observation information of the current agent and the observation information of other agents. For example, through the self-attention mechanism, the similarity between the Query and the Key is calculated, and these similarities are used to perform weighted summation on the Value, thereby learning the features and their mutual relationships of these sequences and enhancing the agents' perception ability of the global environment.

[0134] S15. Integrate the results predicted by each agent and feedback them to the user. In step S15, integrating the results predicted by each agent and feedback them to the user specifically includes:

[0135] Integrate the results predicted by the agents into forms such as text, pictures, GIS maps, 3D models, and file reports and feedback them to the user.

[0136] In this embodiment, a multi-modal data fusion and intelligent generation technology is first proposed, including cross-modal information integration and conversion technology, which realizes the deep fusion of multi-modal data such as text, pictures, GIS maps, 3D geological models, and document reports. It can integrate information obtained from different data sources. For example, when dealing with the problem of "delimiting the prospective area for ore-forming prediction", the system can not only extract relevant theories and methods from the text, but also combine the geographical information in the GIS map and the spatial form of the ore body shown in the 3D model to generate a comprehensive and intuitive answer. At the same time, it has the ability of cross-modal conversion, which can convert text information into forms such as pictures, GIS maps or 3D models to meet different needs and understanding methods of users.

[0137] Secondly, a multi-modal generation model based on deep learning is proposed, which uses technologies such as Transformer and GAN to train according to the characteristics of different modal data. For text-based answers, the model can generate logically clear and content-accurate text expressions according to the questions and retrieved knowledge; for geological maps and 3D models, the model can automatically generate high-quality maps and models according to the input geological data and analysis results. For example, when generating a 3D model of an ore deposit, the model can accurately construct the 3D shape and internal structure of the ore body based on multi-source data such as geophysics and geochemistry, and can update and optimize the model in real time to reflect the latest ore prospecting information.

[0138] Thirdly, an intelligent document report generation technology is proposed, including report content generation based on a mineral knowledge graph and a large model; the ore prospecting large model automatically extracts relevant information according to the questions and analysis results and organizes them into a logically clear and content-complete report text. For example, when compiling a mineral potential evaluation report, the system can obtain relevant knowledge such as the metallogenic geological background and mineral resource characteristics from the mineral knowledge graph, combine the analysis results of ore prospecting prediction, and automatically generate the content of each chapter of the report according to the specifications for writing geological investigation reports.

[0139] Finally, a personalized report customization and optimization technology is proposed. Users can select the content, format, key points, etc. of the report according to their own needs and preferences, automatically adjust the report generation strategy, and generate a report that meets the user's requirements. At the same time, it has a report optimization function, which can perform operations such as grammar checking, logic verification, and content improvement on the generated report to ensure the accuracy and professionalism of the report.

[0140] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A prospecting prediction method based on multi-agent technology, characterized in that: The method comprises: S11, extracting keywords from the mineral resources question input by the user; S12, establishing a vector database of mineral knowledge text and a graph database of mineral knowledge graph, optimizing the extracted keywords to generate prompt words; S13, constructing a large prospecting model to think about the prompt words, expand the prompt words, and obtain a complete sentence that professionally re-expresses the user's question; S14, decompose the complete sentence into multiple tasks to obtain several subtasks, and dispatch the several subtasks to corresponding intelligent agents for step-by-step prediction based on the graph database of the mineral knowledge graph; S15. Integrate the prediction results of each intelligent agent and feed them back to the user.

2. The prospecting prediction method based on multi-agent technology according to claim 1 is characterized in that: In step S11, keywords are extracted from the user's question, specifically including: S21. After parsing the mineral issues through natural language processing technology, they are matched and classified with the nodes and relationships in the mineral knowledge graph; S22. Combine the mineral knowledge graph and mineral historical question data to automatically generate keywords corresponding to mineral questions.

3. The prospecting prediction method based on multi-agent technology according to claim 1 is characterized in that: In step S12, a vector database of mineral knowledge text and a graph database of mineral knowledge graph are established, which specifically includes: S31. Real-time collection and storage of prospecting prediction papers, monographs and geological report documents; S32, converting the stored document data into a unified format and performing preprocessing; S33, dividing the processed document data into segments according to a preset length, inputting the embedding model to convert the document data into vectors, and constructing a vector database of mineral knowledge texts; S34. Perform knowledge extraction, relationship extraction and attribute extraction on the document data divided into segments, fuse the extracted data and build a graph database of the mineral knowledge graph.

4. The prospecting prediction method based on multi-agent technology according to claim 1 is characterized in that: In step S12, the extracted keywords are optimized to generate prompt words, which specifically includes: S41, constructing a graph database query language and query statements, performing a syntax check on keywords, and determining the specific graph structure and information that the user wants to query; S42. Based on the analysis of the graph structure and the query requirements, a specific query execution plan is formulated to determine the order of accessing nodes and relationships; S43, traversing the nodes and relationships, sorting out the nodes, relationships and related attribute information that meet the conditions, and generating prompt words.

5. The prospecting prediction method based on multi-agent technology according to claim 3 is characterized in that: In step S13, a large prospecting model is constructed, which specifically includes: S51. Clean the stored data, select a word segmentation tool to perform word segmentation on the cleaned data, expand the vocabulary, and increase the diversity of samples through text data enhancement technology; S52, dividing the processed data into a training set, a validation set and a test set according to a preset ratio, training, tuning and evaluating the pre-trained prospecting model to obtain a prospecting model; S53, adjusting the dimensions of the word embedding layer and the output layer of the prospecting model according to the expanded vocabulary to be consistent with the size of the expanded vocabulary; S54. Conduct supervision and adjustment, incremental training and prospecting instruction learning on the large prospecting model.

6. The prospecting prediction method based on multi-agent technology according to claim 1 is characterized in that: Step S14 also includes: Use natural language processing technology to understand the semantics of complete sentences, extract key information, and decompose tasks. At the same time, establish a semantic tag library for the algorithm corresponding to each intelligent agent. According to the matching degree between the task semantics and the algorithm semantic tags, intelligently select and arrange algorithms to match tasks and algorithms. Construct a dynamic Bayesian network, taking the geological characteristics, data quality, and computing resources of the prospecting area as nodes, and the dependencies between the factors as edges. By updating the node information and edge weights in real time, the selection and execution order of the algorithm can be dynamically adjusted. Explore the optimal collaboration strategy of intelligent agents through reinforcement learning, maintain real-time communication between intelligent agents, and adjust the collaboration strategy between intelligent agents based on global reward signals; The execution process and results of the algorithm are recorded, and the agents corresponding to the algorithm whose cooperation degree reaches the preset threshold in the collaboration are rewarded; Monitor the computing resource usage of each intelligent agent in real time, and dynamically adjust the execution order of the algorithm and allocate resources according to the remaining resources and the resource requirements of the algorithm.

7. The prospecting prediction method based on multi-agent technology according to claim 6 is characterized in that: Use natural language processing technology to understand the semantics of complete sentences and extract key information, including: S61, identifying the words in the complete sentence and the semantic roles between the words, and analyzing the dependency relationship between the words in the complete sentence; S62, capturing the sentence structure and order information in the complete sentence through BiLSTM, and extracting the contextual semantic features in the complete sentence; S63. The dependency relationship between words obtained in step S61 and the contextual semantic features extracted in step S62 are combined with the graph database of the mineral knowledge graph to extract the key information in the complete sentence.

8. The prospecting prediction method based on multi-agent technology according to claim 6 is characterized in that: Explore the optimal collaboration strategy of intelligent agents through reinforcement learning, including: S71. Define the local state, global environment, and global reward signal of each agent; S72, interacting the observation information of all agents on the local state and the global environment, and obtaining a current agent that integrates the observation information of other agents; S73, the observation information of different dimensions is processed through semantic embedding and converted into Q (Query), K (Key), V (Value) sequences. The features of Q, K, V sequences and their mutual relationships are learned through the self-attention mechanism, and fused through the Transformer structure to generate a globally optimized hidden state; S74, using the fused hidden state, generating a local Q value for each agent through a multi-layer perceptron (MLP), and combining the local Q value of each agent into a global Q value through a value decomposition method; S75. Use the global Q value and global reward signal to update the policy network of each agent through the reinforcement learning algorithm.

9. The prospecting prediction method based on multi-agent technology according to claim 1, characterized in that: In step S15, the prediction results of each agent are integrated and fed back to the user, which specifically includes: The results of the intelligent agent's prediction are integrated into the form of text, pictures, GIS maps, 3D models and file reports and fed back to users.

Images