A method for ore prospecting prediction based on multi-agent technology

Abstract

The present application provides a kind of prospecting prediction method based on multi-agent technology, the method specifically includes: S11, the key word extraction of mineral problem input by user is carried out;S12, the vector database of mineral knowledge text and the graph database of mineral knowledge graph are established, and the extracted key words are optimized to generate prompt words;S13, build prospecting big model to think about prompt words, expand prompt words, obtain complete sentence for re-professional expression of user question;S14, complete sentence is decomposed into several sub-tasks, and based on the graph database of mineral knowledge graph, several sub-tasks are scheduled to corresponding agent for step-by-step prediction;S15, the result of each agent prediction is integrated and fed back to user.The present application promotes the change from artificial driving to intelligent driving for prospecting prediction, and provides emerging technical support for prospecting breakthrough.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of mineral prediction, in particular to a mineral prediction method based on multi-agent technology. BACKGROUND

[0002] In the field of geology, traditional geological analysis often relies on manual observation and experience judgment of limited samples, which is low in efficiency and may be one-sided. Artificial intelligence, with its powerful algorithms, can quickly analyze massive geological data and mine hidden rules and characteristics, promoting the vigorous development of artificial intelligence in mathematical earth science and mineral exploration, and providing more scientific and efficient technical means for mineral exploration. The application of new generation information technologies such as big data, machine learning and cloud computing in mineral exploration and prediction can improve the efficiency of mineral exploration. Looking at the domestic and foreign situation, the emerging development trend of mineral resource exploration is clear. In the future, more attention will be paid to integrating data resources in different time and space dimensions, using powerful computing power and advanced model algorithms to mine more valuable mineral exploration information.

[0003] With years of mining and increasing exploration difficulty, many challenges are faced, such as deep mineral resource exploration, discovery of concealed ore deposits, and mineral prediction under complex geological conditions. The development of mineral prediction methods based on RAG (Retrieval Augmentation Generation), knowledge graph, large model, agent, big data, spatial computing, scientific analysis and other technologies has become a key direction. The development of AI mineral exploration new production force is of great significance to break through the major scientific problems and technical difficulties of resource evaluation and prediction. SUMMARY

[0004] Therefore, the purpose of the present application is to provide a mineral prediction method based on multi-agent technology, which improves the efficiency and accuracy of mineral prediction through intelligent means.

[0005] To achieve the above-mentioned purpose of the application, the present application provides a mineral prediction method based on multi-agent technology, which comprises:

[0006] S11, extracting keywords from the mineral problem input by the user;

[0007] S12, establishing a vector database of mineral knowledge text and a graph database of mineral knowledge graph, and optimizing the extracted keywords to generate prompt words;

[0008] S13, constructing a mineral exploration large model to think about the prompt words, expanding the prompt words, and obtaining complete sentences for professional expression of the user's problem;

[0009] S14, the complete sentence is multi-tasked and decomposed into several sub-tasks, and the graph database based on the mineral knowledge graph is used to schedule the several sub-tasks to the corresponding intelligent agent for step-by-step prediction;

[0010] S15, the results of each intelligent agent prediction are integrated and fed back to the user.

[0011] Further, in step S11, the keywords of the user's question are extracted, which specifically includes:

[0012] S21, the mineral question is analyzed by natural language processing technology and matched and classified with the nodes and relationships in the mineral knowledge graph;

[0013] S22, combined with the mineral knowledge graph and the mineral history question data, the keywords corresponding to the mineral question are automatically generated.

[0014] Further, in step S12, the vector database of the mineral knowledge text and the graph database of the mineral knowledge graph are established, which specifically includes:

[0015] S31, real-time collection of prospecting prediction papers, monographs and geological report document materials, and storage;

[0016] S32, the stored document data is converted into a unified format and preprocessed;

[0017] S33, after the processed document data is divided into segments according to the preset length, the document data is input into the embedding model to convert the document data into vectors, and the vector database of the mineral knowledge text is constructed;

[0018] S34, the document data divided into segments is respectively subjected to knowledge extraction, relationship extraction and attribute extraction, the extracted data is fused, and the graph database of the mineral knowledge graph is constructed.

[0019] Further, in step S12, the extracted keywords are optimized to generate prompt words, which specifically includes:

[0020] S41, the graph database query language and query statement are constructed, the keywords are syntax checked, and the specific graph structure and information that the user wants to query are determined;

[0021] S42, based on the analysis of the graph structure and the query requirements, a specific query execution plan is formulated, and the order of accessing nodes and relationships is determined;

[0022] S43, the nodes and relationships are traversed, the nodes, relationships and related attribute information that meet the conditions are sorted, and the prompt words are generated.

[0023] Further, in step S13, the prospecting large model is constructed, which specifically includes:

[0024] S51, clean the stored data, select a word segmentation tool to process the cleaned data, expand the word list, and increase the diversity of the sample through text data enhancement technology;

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

[0026] S53, adjust the dimensions of the word embedding layer and the output layer of the ore prospecting large model according to the expanded word list, and keep consistent with the size of the expanded word list;

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

[0028] Further, step S14 further includes:

[0029] Using natural language processing technology to understand the semantics of complete sentences and extract key information, and task decomposition, while establishing a semantic tag library for each agent's algorithm, according to the matching degree of task semantics and algorithm semantic tags, intelligently selecting and arranging algorithms to match tasks and algorithms;

[0030] Constructing a dynamic Bayesian network, taking the geological characteristics of the ore prospecting area, data quality and computing resources as nodes, and the dependency relationship between factors as edges, dynamically adjusting the selection and execution order of the algorithm by updating the node information and edge weight in real time;

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

[0032] Record the execution process and results of the algorithm, and reward the agent corresponding to the algorithm that meets the preset threshold in cooperation;

[0033] Real-time monitoring of the computing resource usage of each agent, dynamically adjusting the execution order and resource allocation of the algorithm according to the resource remaining amount and resource demand of the algorithm.

[0034] Further, using natural language processing technology to understand the semantics of complete sentences and extract key information, specifically including:

[0035] 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;

[0036] S62, capture the sentence structure and sequence information in the complete sentence through BiLSTM, and extract the context semantic features in the complete sentence;

[0037] S63, the dependency relationship between the words obtained by step S61 and the context semantic features extracted by step S62 are combined with the graph database of the mineral knowledge graph to extract key information in the complete sentence.

[0038] Further, the optimal collaboration strategy of the agent is explored through reinforcement learning, specifically including:

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

[0040] S72, the observation information of the local state and the global environment of all agents is interacted to obtain the current agent fused with the observation information of other agents;

[0041] S73, the observation information of different dimensions is processed through semantic embedding and converted into Q (Query), K (Key) and V (Value) sequences, the features of the Q, K and V sequences and their mutual relationship are learned through a self-attention mechanism, and the global optimized hidden state is generated through a Transformer structure;

[0042] S74, using the fused hidden state, the local Q value of each agent is generated through a multi-layer perception (MLP), and the local Q value of each agent is combined into a global Q value through a value decomposition method;

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

[0044] Further, in step S15, the results predicted by each agent are integrated and fed back to the user, specifically including:

[0045] The results predicted by the agent are integrated into text, pictures, GIS maps, three-dimensional models and file reports to feed back to the user.

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

[0047] The application provides a mineral prospecting prediction method based on multi-agent technology, key words of a mineral problem are extracted through a natural language processing technology, a vector database of a mineral knowledge text and a graph database of a mineral knowledge graph are innovatively combined, prompt words are generated by optimizing the key words, so that the knowledge retrieval is improved and the performance is enhanced, a pre-trained mineral prospecting large model is used to expand the prompt words to generate complete sentences for the mineral prospecting problem. The intelligent arrangement and scheduling of the mineral prospecting prediction algorithm are realized through the multi-agent technology, the advantages of the mineral prospecting large model and professional algorithm tools are combined, and the intelligent level of data query, mining, information extraction, mineral prospecting prediction, target area optimization and decision support and other tasks is improved; the prediction result is fed back through a multi-modal data structure to meet the diversified needs of users for mineral prediction. Compared with the traditional mineral prospecting prediction method mainly relying on artificial and relying on decentralized tools, the application promotes the change of mineral prospecting prediction from artificial driving to intelligent driving, and provides emerging technical support for mineral prospecting breakthrough. BRIEF DESCRIPTION OF DRAWINGS

[0048] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of the provided drawings.

[0049] Figure 1 A mineral prospecting prediction method based on multi-agent technology provided by the embodiment of the present application is shown in the flowchart. DETAILED DESCRIPTION

[0050] The present application 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 application, but not to limit the present application. In addition, it should be noted that, for the convenience of description, only the parts related to the present application are shown in the drawings, not all the structures.

[0051] REFERENCE Figure 1 The embodiment provides a mineral prospecting prediction method based on multi-agent technology, and the method comprises the following steps:

[0052] S11, key words of a mineral problem input by a user are extracted. In order to protect the security and privacy of data, a security mechanism and a permission management system are introduced, through identity authentication, access control, data encryption and other technologies, it is ensured that only authorized users can perform the subsequent mineral prospecting prediction process. At the same time, the access record is audited in detail and the user behavior is tracked, so as to discover and handle potential security problems in time, and provide reliable security protection for mineral prospecting prediction work.

[0053] In the running process of the whole ore-prospecting prediction method, keywords are extracted for various professional mineral problems related to ore-prospecting prediction proposed by users, for example, the mineral problems include but are not limited to: viewing the prediction work area database, analyzing and dividing the mineral prediction type, determining the mineral prediction method type, summarizing the ore-forming 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 for the user's question, specifically including:

[0055] S21, after analyzing the mineral problem through natural language processing technology, the nodes and relationships in the mineral knowledge graph are matched and classified.

[0056] S22, combined with the mineral knowledge graph and the mineral historical question data, the keywords corresponding to the mineral problem are automatically generated.

[0057] In this embodiment, in order to extract the keywords of the mineral problem, the professional problem classification and intelligent prompt technology is proposed, first, based on the knowledge graph, the mineral problem is classified, and the mineral knowledge graph in the field of ore-prospecting prediction is innovatively used, when the user inputs the question, the mineral problem is analyzed through natural language processing technology, and the nodes and relationships in the mineral knowledge graph are matched, so as to realize the intelligent classification of the mineral problem. For example, for the problem of "analyzing and dividing the mineral prediction type", through the mineral knowledge graph, it can be quickly determined that it belongs to the knowledge category of mineral prediction type division, which provides accurate classification basis for subsequent problem processing.

[0058] Secondly, the intelligent prompt word generation and optimization technology and the problem intention recognition and reasoning are proposed, according to the user's input of part of the content or the general direction of the question, combined with the mineral knowledge graph and the historical question data, the related keywords are automatically generated and dynamically recommended to the front-end user for selection. For example, when the user inputs "viewing the prediction", the present application can prompt "viewing the prediction work area database" "viewing the prediction model parameters" and the like, so as to improve the efficiency and accuracy of the problem input. At the same time, the algorithm can also optimize the generation strategy of the keywords according to the user's feedback and use habit. Finally, the multi-modal input support is optimized, in addition to text input, voice input, image input and the like, which provides data support for subsequent ore-prospecting prediction.

[0059] S12, the vector database of the mineral knowledge text and the graph database of the mineral knowledge graph are established, and the extracted keywords are optimized to generate prompt words. The papers and monographs of ore-prospecting prediction are used as knowledge base, and the vector database of the mineral knowledge text and the graph database of the mineral knowledge graph are established, the keywords are retrieved and knowledge is enhanced, after traversing the vector database and the graph database, the context is fused, and high-quality prompt words more suitable for the ore-prospecting prediction business are formed.

[0060] In this embodiment, firstly, the knowledge retrieval and verification technology based on graph structure enhancement is proposed. In the RAG architecture, the knowledge is organized and stored based on graph structure, where the nodes represent knowledge entities and the edges represent the relationships between entities. When the large model generates content, knowledge retrieval is performed based on the graph structure, not only retrieving entities related to the question, but also retrieving the associated relationships between these entities. After the retrieval results are returned, the retrieved knowledge is verified using the relationship information in the graph to ensure that the knowledge relied on for generating the content is accurate and interrelated. Secondly, the 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 embedded representation of the graph. When the large model generates content, the generation process is associated with the graph embedding to ensure that the generated content conforms to the knowledge structure and logical relationships implied in the graph by constraining the consistency of the vector representation of the generated content and the graph embedding. Through graph embedding constraint, the relationships between concepts in the generated content are ensured to be consistent with the scientific knowledge system represented in the graph.

[0061] Thirdly, the knowledge enhancement technology based on multi-modal graph fusion is proposed. A multi-modal graph is constructed by combining information from multiple modalities such as text, images, audio, and spatial data. 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 the knowledge. In addition to textual descriptions, spatial data can be combined to more accurately predict the geological features of the prediction area, the geological features of the mineral resources, and the metallogenic regularity, thereby reducing the occurrence of hallucinations.

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

[0063] In step S12, the vector database of the mineral knowledge text and the graph database of the mineral knowledge graph are established, which specifically includes:

[0064] S31, real-time collection of prospecting prediction papers, monographs, and geological report document materials, and storage;

[0065] S32, unified format conversion and preprocessing of the stored document data;

[0066] S33, after the processed document data is divided into segments according to the preset length, the document data is input into an embedding model to convert the document data into vectors, and a vector database of mineral knowledge text is constructed;

[0067] S34, the document data divided into segments is respectively subjected to knowledge extraction, relation extraction and attribute extraction, the extracted data is fused, and a graph database of the mineral knowledge graph is constructed.

[0068] In this embodiment, the vector database of the mineral knowledge text and the graph database of the mineral knowledge graph are established. First, the documents such as papers, monographs and geological reports related to prospecting prediction are stored in the personal user workspace. Secondly, all the document data such as papers, monographs and geological reports in the personal user workspace are extracted and analyzed. Thirdly, the document data is preprocessed, including but not limited to: file analysis, noise elimination, redundancy, irrelevance and potential harmful data, converting the document data in different formats into the required format txt or json for training data, and the PDF file in picture format needs to be analyzed by relying on the optical character recognition (OCR) technology in the visual algorithm. The data is clarified to improve the quality of the document data by using document cutting, special format processing, short sentence removal, paragraph splicing, deduplication, semantic paragraph deletion and text reference replacement.

[0069] Then the document data is divided into smaller segments to obtain processed document data. Finally, the processed document data is converted into a vector representation form, such as mapping words, sentences or documents into high-dimensional vector space through word embedding technology, so that the machine can understand and process these data, capture the semantic information and relationship in the text, for example, using the Transformer architecture to deeply encode the text in both directions, learning more rich semantic information and context relationship. The vector database of the mineral knowledge text first performs retrieval, queries and retrieves the similarity between the document data according to the keywords, selects the most relevant document data from the pre-constructed personal user workspace, for example, using the hybrid retrieval method (Hybrid Retrieval), first using the sparse method to quickly screen out the candidate document data, and then using the dense method to accurately locate the most relevant document data.

[0070] Secondly, re-ranking is performed. After the initial retrieval, the re-ranking stage is used to improve the relevance of the retrieved document data, to ensure that the most relevant information is displayed at the top of the list, and more accurate and time-consuming methods are used to effectively reorder the document data, thereby improving the similarity between the query and the document data at the top of the ranking. Re-ranking prioritizes performance and efficiency, mainly including two methods of DLMReranking and TILDEReranking.

[0071] Then the document data repacking is performed, and the performance of the LLM response generation can be affected by the provided document data. There are currently three repacking methods for the document data repacking process after reordering: "forward", "reverse", and "sides". The "forward" method repacks the document data by descending order of relevance scores from the re-ranking stage. The "reverse" method arranges the document data in ascending order. The "sides" method can achieve the best performance when the relevant information is placed at the head or tail of the input.

[0072] A vector database is a database for storing, managing, and retrieving vector data. In the RAG scenario, document data is first converted into a vector representation, mapping words, sentences, or documents into a high-dimensional vector space. These vector data are stored in the vector database. In the face of massive text data, the vector database can store the corresponding vectors in order and establish an effective indexing mechanism, facilitating subsequent rapid searching and matching, so that the information related to the user's question can be quickly located in the vector.

[0073] The vector database can quickly find the text content corresponding to the vector with high similarity in a very short time, providing a key guarantee for RAG to accurately extract relevant knowledge from the knowledge base, and thus improving the quality and accuracy of the generated answers. The vector database has four capabilities, including multiple index types, support for billion-level vectors, hybrid search, and cloud-native capabilities. Multiple index types provide flexibility for optimizing searches based on different data characteristics and use cases. Billion-level vector processing handles large data sets. Hybrid search combines vector search with traditional keyword search to improve retrieval accuracy. Cloud-native capabilities ensure seamless integration, scalability, and management in a cloud environment.

[0074] The graph database of the mineral knowledge graph is used for overall planning of the ore-prospecting prediction knowledge graph. First, the entity types are determined, and the core entities in the field of ore-prospecting prediction are combed, such as the entity categories in the ore-prospecting prediction knowledge graph including but not limited to "ore deposit model", "ore-prospecting prediction model", "genesis model", and "ore deposit geological characteristics". The core entities are the basis for subsequent construction of graph relationships.

[0075] Secondly, the relationship types are defined, and the various connections between the entities in the mineral knowledge graph are analyzed, such as the "containment" relationship between "predictive factors" and "ore-controlling structures", "rock bodies", and "gravity anomalies", and the "anomaly combination" relationship between "Au element anomalies" and "As element anomalies", and so on. Accurate and clear definition of these relationships can build a complete knowledge network. Then, the pattern layer is constructed to describe the entities and relationships related to ore prediction in a structured manner, and the graph model (such as RDF) or ontology language is used for expression, which is equivalent to drawing the "blueprint" of the knowledge graph.

[0076] According to the entity types, relationship types, and patterns determined by the mineral knowledge graph planning, the knowledge extraction is performed, including entity extraction, relationship extraction, and attribute extraction. Then, the relationship extraction is performed to deeply analyze the semantics of the document data. Then, the attribute extraction is performed to extract the related attributes of the entities, further enrich the connotation of the entities, and make the knowledge in the mineral knowledge graph more complete and detailed. Finally, the knowledge fusion is performed, including entity alignment and knowledge merging. The entity alignment ensures the uniqueness and accuracy of the entities in the mineral knowledge graph, and the knowledge merging is used to integrate the knowledge extracted from different sources, and remove the redundant and contradictory parts.

[0077] The constructed mineral knowledge graph is stored in a graph database. The graph database is based on graph theory, and the data is represented as nodes (Node), edges (Edge), and their carried attributes (Property). The nodes are usually used to represent the entities related to mineral prediction, such as "ore deposit geological characteristics", "ore-forming pattern", "predictive factor", "ore-controlling structure", "rock body", "gravity anomaly", "magnetic anomaly", and so on. The edges are used to describe the relationships between the mineral prediction entities, such as the "parallel unconformity" relationship between "strata", and the "main and auxiliary" relationship between "faults".

[0078] In step S12, the extracted keywords are optimized to generate prompt words, which specifically include:

[0079] S41, a graph database query language and a query statement are constructed to perform syntax checking on the keywords, and to determine the specific graph structure and information that the user wants to query;

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

[0081] S43, the nodes and relationships are traversed, and the nodes, relationships, and related attribute information that meet the conditions are sorted to generate prompt words.

[0082] In this embodiment, first, a large model is used to construct a graph database query language and a query statement; second, the query statement is parsed for syntax checking to check whether the statement conforms to the query language syntax rules specified thereby, and on the basis of correct syntax, the system further analyzes the semantics of the statement to determine the specific graph structure and information that the user wants to query. Third, a 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 which nodes to access first, along which relationships to traverse, and other operation sequences. Then, the query operation is executed, and according to the planned execution path, actual node and relationship traversal operations are performed, and the traversal process will search for nodes that meet the conditions along the edges in the graph according to the set search strategy; finally, the query result is returned, and after completing the query operation, the found nodes, relationships, and related attributes and other information that meet the conditions are sorted to generate prompt words, and the forms of prompt words include but are not limited to table form or JSON format.

[0083] S13, the prospecting large model is constructed to think about the prompt words, and the prompt words are expanded to obtain complete sentences that re-professionally express the user's question. Incremental training is performed on the basis of the general large model to obtain the prospecting large model, and the language learning ability, summarization ability, text generation ability, and text modification ability of the prospecting large model are improved as a whole, so that the prospecting large model has comprehensive research ability in the prospecting prediction business process.

[0084] In step S13, the prospecting large model is constructed, specifically including:

[0085] S51, the stored data is cleaned, a word segmentation tool is selected to perform word segmentation processing on the cleaned data, the word table is expanded, and the diversity of samples is increased through a text data enhancement technique;

[0086] S52, the processed data is divided into a training set, a validation set, and a test set according to a preset ratio, the pre-trained prospecting large model is trained, optimized, and evaluated, and the prospecting large model is obtained;

[0087] S53, the dimensions of the word embedding layer and the output layer of the prospecting large model are adjusted according to the expanded word table to be consistent with the size of the expanded word table;

[0088] S54, the prospecting large model is supervised, adjusted, incrementally trained, and learned for prospecting instructions.

[0089] In this embodiment, the ore-prospecting large model module uses an open-source general large model as a base LLM-3, such as Qwen72B, Llama3.2, which has learned general knowledge generation capabilities through training on open datasets. On the basis of the general large model LLM-3, an ore-prospecting dataset is added to the personal user workspace for incremental pre-training of a large model in the mineral vertical field, i.e., the ore-prospecting large model. Specifically, first, the added ore-prospecting dataset is cleaned, including but not limited to document parsing, noise elimination, redundancy removal, and irrelevant content; second, the data is processed, a word segmentation tool is selected to process the ore-prospecting knowledge text, and the word table is expanded to adapt to the vocabulary in the ore-prospecting prediction field; through synonym conversion, data back-translation, random insertion and deletion of words, and other text data enhancement techniques, the diversity of the sample is increased, and the generalization ability of the model is improved; the processed dataset is divided into a training set, a validation set, and a test set, usually segmented according to a certain ratio, such as 8:1:1, for model training, tuning, and evaluation.

[0090] Then the model is adjusted. The dimensions of the model are adjusted according to the size of the expanded word table, 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 word tables; finally, incremental training is performed, and the training framework is selected according to the computing resources and training requirements to determine the training strategy, which adopts three training strategies: first, pre-training with large-scale ore-prospecting prediction-related encyclopedias, books, reports, papers, and other general corpora, and then secondary training with small-scale field corpora such as papers and monographs published by academicians and authoritative experts; second, directly pre-training with large-scale field corpora; third, mixing general corpora and field corpora in a certain proportion and training them simultaneously.

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

[0092] The instruction adjustment of the ore-finding large model first collects or constructs instances of instruction format. Each instance of the instruction data set consists of three elements: instruction, context (optional) and input and output based on instruction. The instruction includes detailed information such as the name, description, notes, positive and negative examples of the task. The data set with the format of "instruction, output" is obtained by artificial integrated conversion or large language model generation. Then, the LLM is adjusted in a supervised manner. After obtaining the general instruction data set, the ore-finding large model can be adjusted on the pre-trained ore-finding large model. After instruction adjustment, the ore-finding large model can exhibit excellent ability to generalize to unseen tasks. Instruction adjustment stimulates the understanding ability of the language model and makes full use of prior knowledge. After giving more obvious instructions, the model understands and responds correctly, and generalizes to multiple tasks.

[0093] In the ore-finding large model, the ore-finding preference pair is added to learn according to the preferences of ore-finding experts, to solve the risk of inaccurate, misleading or even harmful information in the content generated by the ore-finding large model, and to ensure 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 free of bias, and reflects the values and ethical standards generally recognized by society. The purpose of alignment is to create a model that can understand and generate human language, and also reflect the importance of fairness, transparency and responsibility in its decision-making, reducing the potential negative impact. The ore-finding large model uses the reinforcement learning based on human feedback RLHF method, which is divided into pre-training and supervised adjustment model, training reward model and using reinforcement learning to adjust the model.

[0094] First, starting from supervised adjustment, the ore-finding large model is adjusted in the annotator using a supervised learning method, a supervised data set containing input prompts (instructions) and required outputs, i.e. ore-finding prediction knowledge preference pairs, is collected to adjust the ore-finding large model. These ore-finding prediction knowledge preference pairs are written by human annotators for certain specific tasks while ensuring task diversity. Second, the human feedback training reward model is obtained using the human annotator. The human annotator sorts the output of the reward model according to its consistency with the expected behavior. Specifically, the reward model is input with sampled prompts (from the supervised data set or human-generated prompts) to generate a certain number of output texts, and then human annotators label the preferences for these input-output pairs. Then, the reward model is trained to predict the output of human preferences.

[0095] Finally, the reinforcement learning algorithm is used in combination with the trained reward model. The strategy 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, and the state is the currently generated token sequence. The reward is provided by the reward model. In order to avoid the ore-prospecting large model deviating significantly from the initial (before adjustment) model, a penalty term is usually included in the reward function, and multiple iterations are used to better align the ore-prospecting large model. 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 is more in line with human preferences, and the instruction following ability of the ore-prospecting large model is improved.

[0096] In order to guide the ability of the ore-prospecting large model to perform complex reasoning, the ore-prospecting large model is expanded by using prompt engineering to guide and optimize the ability of the ore-prospecting large model and expand the range of tasks that the ore-prospecting large model can effectively perform. For the input prompt word, the ore-prospecting large model will give the corresponding output answer. The prompt word has three main content types: input, context and example. The input specifies the information that the model needs to generate a response. The context and example are optional parts of the prompt. The context provides instructions on the behavior of the model, and the example is an input and output pair in the prompt that demonstrates the expected response.

[0097] S14, the complete sentence is decomposed into several sub-tasks, and the graph database based on the mineral knowledge graph is used to schedule the several sub-tasks to the corresponding intelligent agents for step-by-step prediction.

[0098] In this embodiment, the complete sentence (complex ore-prospecting prediction task) is decomposed into several sub-tasks, and the ore-prospecting prediction workflow sub-tasks are scheduled to the corresponding intelligent agents to achieve the final ore-prospecting prediction goal. The intelligent agent can generate a plan without adjusting through different reasoning strategies, or adjust the generated plan according to external feedback. The present application decomposes the complete sentence into several sub-tasks and strategically arranges the corresponding intelligent agents to ensure the seamless performance of the ore-prospecting prediction task workflow generated by the planning.

[0099] The ore-prospecting multi-agent (hereinafter referred to as intelligent agent) has a high degree of intelligent understanding and generation ability, and is an intelligent computing entity that can accurately perceive the environment, make decisions and execute actions. The ore-prospecting micro-service component completes specific tasks such as data query, data mining, information extraction, etc. in the ore-prospecting business process.

[0100] First, record the history of thinking, action and environmental observation generated during the execution of the ore-prospecting prediction agent, such as reading topographic and remote sensing data from the database, using geophysical processing analysis tools, geochemical processing analysis tools, etc., to the use of ore-prospecting big model to select prediction factors, etc. Based on the accumulation of various ore-prospecting prediction memories, the agent can re-examine and use previous records and experiences to more effectively handle more complex ore-prospecting prediction tasks.

[0101] Agent memory can be divided into short-term memory and long-term memory according to memory duration. Short-term memory is integrated into the agent to enhance its ability to maintain the current ongoing ore-prospecting prediction task trajectory, which is often used in multi-round interactions. Long-term memory is used to remember valuable experiences from previous ore-prospecting prediction tasks, which are recalled by the ore-prospecting prediction agent when solving unseen tasks. Due to the wide range of trajectories, long-term memory usually uses distillation technology or only stores key information. Second, receive information from the ore-prospecting workspace environment. The ore-prospecting prediction agent can perceive multi-modal input such as text input, visual input and auditory input. Ore-prospecting history data is stored in text form, and ore-prospecting text can flexibly express intent, information and knowledge.

[0102] Third, interact with and affect the external environment. The important mechanism is to control and utilize external ore-prospecting micro-service algorithm tools to expand the inherent functions of LLM by accessing more external resources and expanding the operation space beyond individual text interactions. The ore-prospecting multi-agent includes decision support agent, target area optimization agent, ore-prospecting agent, information extraction agent, big data agent, GIS tool agent, mathematical geology agent, geostatistics agent, classical prediction agent, machine learning agent, deep learning agent, etc. Each agent automatically schedules the execution of the backend micro-service when executing to complete the corresponding subtask. All micro-services interface with topographic and remote sensing databases to realize the bidirectional interconnection, dynamic access and storage of source data, information extraction data, ore-prospecting prediction data, target area optimization data and decision support data.

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

[0104] The target area optimization agent identifies blocks with exploration value from massive geological data, including comprehensive analysis of geological, geophysical, geochemical, and remote sensing information. Meanwhile, the target area optimization agent performs risk assessment and economic feasibility analysis, considering factors such as potential value of mineral deposits, mining costs, environmental impact, and legal restrictions, to ensure maximum reduction of environmental impact and compliance with sustainable development principles. The target area optimization agent not only relies on historical data and expert experience, but also incorporates dynamic data, with real-time updating and feedback functions. When new geological survey data or market dynamics emerge, the target area optimization agent can quickly integrate these new information into its optimization model and adjust the target area recommendation results in a timely manner.

[0105] The ore prediction agent uses different algorithms and techniques to help accurately locate potential mineral resource areas and delineate ore 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 various geological factors and known mineral distribution based on information content theory. By calculating the information content of each factor, the contribution of each factor to mineralization is measured, and the information is comprehensively considered to judge the mineralization potential of different areas in the study area, thereby providing a focus for subsequent exploration work.

[0106] The evidence weight method mineral prediction agent, based on Bayesian theory, treats various geological evidence related to mineralization as independent "evidence layers". The weights of each evidence layer on mineralization are determined, including stratigraphic lithology, structural features, and geochemical anomalies. After complex statistical analysis, the weight values of each evidence layer are obtained, and the mineralization probabilities of different areas are evaluated based on these weights to predict favorable areas for mineralization, making the mineral prediction work more scientific and improving the accuracy of ore prospecting.

[0107] The characteristic analysis method mineral prediction agent focuses on various characteristics of geological bodies, collects data on mineral composition, structural features, and geophysical field characteristics, and conducts in-depth analysis and mining of these characteristics to find out the characteristic combinations closely related to mineralization. By establishing a characteristic model, the agent can identify areas with potential mineral resources based on the characteristics of geological bodies in the study area, helping exploration personnel to narrow down the ore prospecting range and focus on areas with the highest potential for mineralization.

[0108] Random Forest method mineral prediction agent, an ensemble learning algorithm composed of multiple decision trees, the corresponding mineral prediction agent uses its advantages to carry out work, taking a large amount of geological data as input, including geological structure, geochemical exploration data, geophysical data, etc. Each decision tree grows and makes metallogenic prediction based on part of the data and random features. After the results of numerous decision trees are aggregated, the final metallogenic prediction result is determined through voting mechanism, which effectively reduces the error of a single model, improves the reliability and stability of the prediction, and can also work well under complex geological conditions.

[0109] Feature analysis method mineral prediction agent, which focuses on various features of geological bodies, collects data such as mineral composition, structural characteristics, and geophysical field characteristics, and conducts in-depth mining and analysis of these features to find out the feature combinations closely related to mineralization. By establishing a feature model, the agent can identify areas where minerals may exist based on the characteristics of the geological bodies in the study area, helping exploration personnel narrow down the prospecting range and focus on the most likely mineralized areas.

[0110] Support Vector Machine method mineral prediction agent, which is good at handling classification and regression problems. In mineral prediction, the agent maps geological features and other data to high-dimensional space, finds an optimal classification hyperplane to distinguish between mineralized and non-mineralized areas, and can handle nonlinear relationships in data. Even in the face of complex variable relationships in geological data, it can accurately extract potential mineralization rules and accurately predict whether minerals exist in different areas, providing strong decision support for mineral exploration.

[0111] Convolutional Neural Network method mineral prediction agent, which has strong automatic feature extraction capability. This agent uses CNN to process geological data, geophysical data, and geochemical data, can automatically capture features such as strata texture and abnormal morphology, and continuously refines key information through layer-by-layer convolution and pooling operations. Based on these features, it can judge the likelihood of mineralization and open up new perspectives for prospecting work.

[0112] Graph Neural Network method mineral prediction agent, which is born in response to the complex network relationships in geological phenomena, such as graphical structural geology and spatial correlation of mineralization alteration. This agent constructs a graph structure data from geological entities and their relationships, updates information through node and edge transmission, and extracts hidden mineralization information. It can accurately grasp the mineralization rules under complex geological relationships and effectively predict areas with mineralization potential, adapting to the needs of modern geological research for complex relationship analysis.

[0113] The mineral prediction agent uses the generator and discriminator architecture to generate geological data distribution similar to the ore-forming pattern. The discriminator judges the authenticity of the generated data, and the two parties constantly compete to make the generator's results more and more close to the real ore-forming situation. Through this way, more potential ore-bearing units can be simulated to help users analyze and study the mineralization possibility from different angles and more comprehensively. These different agents complement and cooperate with each other based on their respective principles and methods, providing diversified technical support for mineral resource prediction, a complex and important task, and improving the efficiency and success rate of prospecting and prediction.

[0114] The information extraction agent is one of the key components of the multi-agent system for prospecting and prediction, responsible for extracting valuable information from various data sources. The information extraction agent usually works closely with other agents, passing the extracted information to the prospecting and prediction agent, the target area optimization agent, and the decision support agent through API or microservice. This interconnected mechanism ensures the flow and sharing of information, making the entire prospecting and prediction system a highly coordinated work unit. The big data agent is responsible for managing and analyzing massive exploration data, extracting trends and patterns, and providing support for other agents. The GIS tool agent uses geographic information system technology to analyze spatial data, helping agents understand the relationship between geographic environment and prospecting and prediction, and providing support for other agents.

[0115] The mathematical geology agent applies mathematical models to explain ore distribution phenomena and help predict the geological structure of ore deposits, providing support for other agents. The geostatistics agent is responsible for analyzing sample data using statistical methods to assess confidence and geological uncertainty, providing support for other agents. The classical prediction agent uses traditional geological prediction methods for preliminary evaluation and comparison and correction with the calculation results of other agents, providing support for other agents. The machine learning agent uses machine learning techniques to automatically learn patterns from data, providing additional evidence for decision support, and providing support for other agents. The deep learning agent is a further development of machine learning, using neural network models to process more complex sample data and improve prediction accuracy, providing support for other agents.

[0116] Step S14 also includes:

[0117] The natural language processing technology is used to understand the semantics of complete sentences and extract key information, and task decomposition is performed. At the same time, a semantic tag library is established for each agent's algorithm, and the matching degree of task semantics and algorithm semantic tags is used to intelligently select and arrange algorithms to match tasks and algorithms.

[0118] A dynamic Bayesian network is constructed, taking the geological characteristics of the ore prospecting area, data quality, and computing resources as nodes, and the dependency relationships between the factors as edges. The selection and execution order of the algorithm are dynamically adjusted by updating the node information and edge weights in real time.

[0119] An optimal collaboration strategy for the agent is explored through reinforcement learning, real-time communication between agents is maintained, and the collaboration strategy between agents is adjusted according to the global reward signal.

[0120] The execution process and results of the algorithm are recorded, and the agent corresponding to the algorithm with a preset threshold of cooperation degree is rewarded.

[0121] The use of computing resources by each agent is monitored in real time, and the execution order and resource allocation of the algorithm are dynamically adjusted based on the remaining resources and resource requirements of the algorithm.

[0122] Natural language processing techniques are used to understand the semantics of complete sentences and extract key information, including:

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

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

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

[0126] In this embodiment, the dependency relationship between the vocabulary in the complete sentence is analyzed, i.e., the grammatical structure relationship between the vocabulary in the sentence is determined, the dependency relationship reflects the "domination" and "dependence" relationship between the vocabulary, and is used to understand the grammatical structure of the sentence. The output of the dependency relationship analysis is a dependency tree, in which each vocabulary has a pointer to its "dominating word". For example, identify the entities in the report (ore body, ore grade) and their relationships (ore body is located in …). Extract the context semantic features in the complete sentence, consider the specific meaning of the vocabulary in the context, focus on the semantic role, semantic association of the vocabulary, and the overall semantics of the sentence. The output of the context semantic feature extraction is the semantic representation of the vocabulary 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 semantic understanding to extract key information in the complete sentence, such as the properties of the mineral, the types of geological structure, and the exploration method.

[0127] An optimal collaboration strategy for the agent is explored through reinforcement learning, specifically including:

[0128] S71, define the local state of each agent, the global environment, and the global reward signal;

[0129] S72, interact the observation information of the local state and the global environment of all agents to obtain the current agent fused with the observation information of other agents;

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

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

[0132] S75, use the global Q value and the global reward signal to update the policy network of each agent through the reinforcement learning algorithm.

[0133] In the multi-agent system, the observation information of each agent is usually local, while the global information is crucial for optimizing the cooperative strategy. Through the self-attention mechanism, the observation information of the current agent can be interacted with the observation information of other agents. Reinforcement learning optimizes the cooperative strategy of the agent through the global reward signal, so that the agent can better understand the global environment and make better cooperative decisions. The observation information of each agent comes from different dimensions, such as position, velocity, state, etc. In order to make these information can be effectively processed in the self-attention mechanism, they need to be mapped to a unified vector space through semantic embedding first. Then, they are converted into Q, K, and V sequences, and their features and relationships are learned through the self-attention mechanism, so as to realize the information interaction between agents and the improvement of 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 Query and Key is calculated, and the Value is weighted and summed using these similarities, so as to learn the features and their relationships of these sequences, and enhance the perception ability of the agent to the global environment.

[0134] S15, integrate the results predicted by each agent and feed them back to the user. In step S15, the results predicted by each agent are integrated and fed back to the user, which specifically includes:

[0135] The results predicted by the agent are integrated into text, pictures, GIS maps, three-dimensional models, and file reports, and fed back to the user.

[0136] In this embodiment, multimodal data fusion and intelligent generation technology is first proposed, including cross-modal information integration and conversion technology, which realizes the deep fusion of multimodal data such as text, pictures, GIS maps, three-dimensional geological models, and file reports. It can obtain information from different data sources. For example, when dealing with the problem of "delineating prospective mineralization prediction areas", the system can not only extract relevant theories and methods from the text, but also combine the geographic information in the GIS map and the spatial form of the ore body displayed by the three-dimensional model to generate a comprehensive and intuitive answer. At the same time, it has cross-modal conversion capabilities and can convert text information into pictures, GIS maps, or three-dimensional models to meet the different needs and understanding methods of users.

[0137] Secondly, a multimodal generative model based on deep learning was proposed. Using technologies such as Transformers and GANs, it is trained based on the characteristics of data from different modalities. For text-based answers, the model generates logically clear and accurate text representations based on the question and retrieved knowledge. For geological maps and 3D models, the model automatically generates high-quality maps and models based on input geological data and analysis results. For example, when generating a 3D model of a mineral deposit, the model accurately constructs the 3D morphology and internal structure of the ore body based on multiple sources of geophysical and geochemical data. It also updates and optimizes the model in real time to reflect the latest prospecting information.

[0138] Thirdly, intelligent document report generation technology was proposed, including report content generation based on the mineral knowledge graph and large model. The prospecting large model automatically extracts relevant information based on the questions and analysis results and organizes it into a logically clear and complete report text. For example, when compiling a mineral potential evaluation report, the system can obtain relevant knowledge such as the mineralization geological background and mineral resource characteristics from the mineral knowledge graph, combine it with the analysis results of the prospecting prediction, and automatically generate the content of each chapter of the report according to the geological survey report compilation standards.

[0139] Finally, we propose personalized report customization and optimization technology. Users can select report content, format, and focus based on their needs and preferences, automatically adjusting report generation strategies to produce reports that meet user requirements. Furthermore, we also provide report optimization capabilities, enabling syntax checking, logic verification, and content enhancement on generated reports to ensure accuracy and professionalism.

[0140] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection 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 resource questions input by the user; S12. Establish a vector database of mineral knowledge text and a graph database of mineral knowledge graph, and optimize the extracted keywords to generate prompt words; Establish a vector database of mineral knowledge text and a graph database of mineral knowledge graph, specifically including: 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 segment into an embedding model to convert the document data into vectors, and constructing a vector database of mineral knowledge text; S34, performing knowledge extraction, relationship extraction, and attribute extraction on the document data divided into segments, fusing the extracted data, and constructing a graph database of the mineral knowledge graph; Optimize the extracted keywords to generate prompt words, including: S41. Construct a graph database query language and query statement, perform a syntax check on keywords, and determine 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 developed 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; S13. Build a large prospecting model to consider the prompt words, expand the prompt words, and obtain a complete sentence that professionally expresses the user's question; S14. Decompose the complete statement into multiple tasks to obtain several subtasks. Based on the graph database of the mineral knowledge graph, dispatch the subtasks to the corresponding intelligent agents for step-by-step prediction. Step S14 also includes: Natural language processing technology is used to semantically understand complete sentences, extract key information, and perform task decomposition. At the same time, a semantic tag library is established for the algorithm corresponding to each intelligent agent. Based on the matching degree between the task semantics and the algorithm semantic tags, algorithms are intelligently selected and arranged to match tasks and algorithms. Construct a dynamic Bayesian network, using the geological characteristics of the prospecting area, data quality, and computing resources as nodes, and the dependencies between these factors as edges. By updating node information and edge weights in real time, the selection and execution order of the algorithm can be dynamically adjusted. Explore the optimal collaborative strategy of intelligent agents through reinforcement learning, maintain real-time communication between intelligent agents, and adjust the collaborative 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 reaches the preset threshold in the collaboration are rewarded; Monitor the computing resource usage of each agent in real time, and dynamically adjust the execution order of the algorithm and allocate resources based on the remaining resources and the resource requirements of the algorithm; Use natural language processing technology to understand the semantics of complete sentences and extract key information, including: S61, identifying the words in a complete sentence and the semantic roles between the words, and analyzing the dependency relationships between the words in the complete sentence; S62, using BiLSTM to capture the sentence structure and order information in the complete sentence and extract the contextual semantic features in the complete sentence; S63, extracting key information from the complete sentence by combining the dependency relationship between the words obtained in step S61 and the contextual semantic features extracted in step S62 with the graph database of the mineral knowledge graph; Explore the optimal collaboration strategy for 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 to obtain a current agent that integrates the observation information of other agents; S73. Process observation information of different dimensions through semantic embedding and convert it into Q, K, V sequences. Learn the features of Q, K, V sequences and their relationships through the self-attention mechanism, and fuse them through the Transformer structure to generate a globally optimized hidden state. S74. Using the fused hidden state, generate a local Q value for each agent through a multi-layer perceptron, and combine 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; 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, characterized in that: In step S11, keywords are extracted from the user's question, specifically including: S21. Analyze mineral issues using natural language processing technology and then match and classify them with 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 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 the 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 keep them consistent with the size of the expanded vocabulary; S54. Supervise and adjust the large prospecting model, conduct incremental training, and learn prospecting instructions.

4. The prospecting prediction method based on multi-agent technology according to claim 1 is characterized in that: In step S15, the prediction results of each agent are integrated and fed back to the user, specifically including: 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