Human intelligence collaborative generation type interaction method for coal mining scene

Abstract

The present application relates to a kind of human intelligence collaborative generation type interaction methods of coal mining scene, belong to coal mine intelligent technology field.The present application includes: through coal mine multi-modal intention decoupling-dynamic Agent evaluation and fusion reasoning engine algorithm, the intention of coal mining process scene instruction is identified, subtask division and dependency analysis, obtain structured semantic intention, executable subtask single and task scheduling result and submit to orchestrator, select coal mine professional big model, external tool integrated Agent and decision model integrated Agent are routed by language model empowerment intention analysis database routing technology and are retrieved, obtain the preliminary search result of each Agent;Determine the reliability score of preliminary search result, generate and visual display final search result.The present application makes search result fully fuse multi-modal data, cooperatively consider the search result of different Agent, search result is richer, more accurate, improves the intelligence of human intelligence interaction.

Description

Technical Field

[0001] This invention relates to the field of intelligent coal mining technology, and in particular to a human-intelligent collaborative generative interaction method for coal mining scenarios. Background Technology

[0002] As a vital foundational industry of the national economy, the coal mining sector is also a traditionally high-risk industry. Intelligent transformation is the core path to improving its safety levels, production efficiency, and management effectiveness. With the rapid development of artificial intelligence technology, especially large-scale modeling, deeply integrating the natural language understanding, multimodal processing, and reasoning capabilities of large-scale models with the professional knowledge and multimodal heterogeneous data in the coal mining field has become a key breakthrough in propelling coal mine intelligence from single-point technology applications to system-level intelligent decision-making.

[0003] However, the current challenges include the difficulty in fusing multimodal heterogeneous data such as semantic text, underground video / images, and time-series sensor data; the heterogeneity of intelligent decision-making models such as hydraulic support control and pressure event prediction, and the lack of a unified natural language interaction interface; insufficient collaboration between external tools and decision-making model agents; and weak human-machine natural language interaction and intelligent feedback. Therefore, there is an urgent need to build an integrated system that combines a large professional coal mine model with a multi-agent collaborative architecture. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a human-intelligent collaborative generative interaction method for coal mining scenarios. The technical solution of this invention is as follows:

[0005] A human-intelligence collaborative generative interaction method for coal mining scenarios includes:

[0006] S1, obtain the coal mining process scenario instructions input by the user in natural language, and determine the instruction intent of the coal mining process scenario instructions based on the rule matching method of keywords;

[0007] S2, when the instruction intent of the coal mining process scenario command is a task-oriented instruction, the coal mine multimodal intent decoupling-dynamic agent evaluation and fusion reasoning engine algorithm is used to perform intent recognition, subtask segmentation and dependency analysis on the coal mining process scenario command to obtain structured semantic intent. Executable subtask list and task scheduling results ;

[0008] S3 will structure semantic intent Executable subtask list and task scheduling results Submit to the orchestrator Arranger Based on capability matching and rule gating mechanisms, three types of agents are selected: coal mine professional big model, external tool integration agent, and decision model integration agent. The selected agents are controlled to perform retrieval through language model-enabled intent parsing database routing technology to obtain preliminary retrieval results for each type of agent.

[0009] S4, determine the credibility score of the preliminary search results of the selected Agent, and determine the final search results based on the credibility score;

[0010] S5 provides a visual representation of the final search results.

[0011] Preferably, S2 includes:

[0012] S21, when the instruction intent of the coal mining process scenario instruction is a task-type instruction, the coal mining process scenario instruction is parsed into a form including intent nodes. Legal constraints Task dependency graph Structured semantic intent ;

[0013] S22, based on the intent node Generate executable subtask sheets ;

[0014] S23, based on the task dependency graph and executable subtasks Generate task scheduling results .

[0015] Preferably, S21 includes:

[0016] S211, when the instruction intent of the coal mining process scenario command is a task-oriented command, the coal mining process scenario command is input into the structured pre-trained model, and the structured pre-trained model outputs the initial structured triplet through generative modeling. and from legal structured triples Extracting intent nodes The structured triple includes a subject, an object, and the relationship between them; the intent node... The core task representing user needs;

[0017] S212, Extracting initial constraints from coal mining process scenario instructions through named entity recognition and slot filling model. And by pre-setting the physical entity of the coal mining field For initial constraints Perform validation to obtain valid constraints. ;

[0018] S213, using dependency parsing algorithm to analyze valid constraints. Processing is performed, and the intent nodes are processed through a graph neural network. The processing results of the dependency parsing algorithm generate a task dependency graph. The task dependency graph It includes the timing, condition, and causal dependencies between task nodes.

[0019] Preferably, step S211 involves passing through a pre-defined coal mine domain body. For the initial structured triples The validation process yields valid structured triples. At that time, including:

[0020] First, extract the initial structured triples. The subject and object in the context, and whether the subject and object belong to the pre-defined coal mine domain ontology. If a defined core entity category is not included, it is marked as illegal.

[0021] Secondly, verify the compliance of entity attributes;

[0022] Next, perform relational validity verification: extract the initial structured triples. The relationship in the middle, determine whether the relationship exists. Defined allowed sets of relations;

[0023] Next, verify the compatibility between the relationship and the entity;

[0024] Then, a logical consistency check is performed to examine whether the overall logic of the subject, object, and their relationship conforms to the rules of coal mine production.

[0025] Finally, the initial structured triples The illegal parts in the structured triples are corrected or removed to obtain valid structured triples that pass the verification. .

[0026] Preferably, S22 includes:

[0027] S221, Determine the intent node based on the coal mine task template library. Is it a high-frequency task node, a low-frequency task node, or a new task node?

[0028] S222, if the intent node For high-frequency task nodes, the intent node is determined based on the mapping relationship between task nodes and subtasks in the coal mine task template library. Executable subtask list ;

[0029] S223, if the intention node If it is a low-frequency task node or a new task node, then according to the subtask generative model and Generate intent nodes Executable subtask list ;

[0030] S224, Define executable subtasks using a structured pattern. Input sources for each subtask Output format This forms a standardized, executable subtask list for coal mine scenarios. .

[0031] Preferably, S23 includes:

[0032] S231, based on the task dependency graph Executable subtasks are sorted using a topological sorting algorithm. Each subtask in the graph is converted into a directed acyclic graph. The directed acyclic graph A sequential framework for the execution of subtasks;

[0033] S232, based on the coal mine's predefined data routing rule engine and directed acyclic graph. Generate data flow rules The data flow rules Define the path, format, and triggering conditions for data transfer between subtasks;

[0034] S233, based on the directed acyclic graph Differentiated retrieval strategies are generated based on the types of each subtask. ;

[0035] S234, if there is a directed acyclic graph If there are conflicts among the subtasks, then the directed acyclic graph (DAG) should be used. The features of each subtask are used to generate the task fusion result using a weighted fusion algorithm. ;

[0036] S235, Combinatorial Directed Acyclic Graph Data flow rules Differentiated search strategies Results of task integration Obtain the task scheduling result .

[0037] Preferably, when controlling the selected Agent to perform retrieval using language model-enabled intent parsing database routing technology to obtain preliminary retrieval results for various types of Agents, S3 includes:

[0038] S31, Arranger The selected agent will structure semantic intent. Executable subtask list and task scheduling results By using a coal mine intent-weighted multimodal feature enhancement query mechanism, searches are performed in the dynamic memory, topological cognition, and trend perception databases of the coal external knowledge base to obtain preliminary search results for various types of agents in the dynamic memory, topological cognition, and trend perception databases.

[0039] S32, programmable By using a cross-database query result feedback and iteration mechanism, the preliminary search results of various types of agents in the dynamic memory database, topology cognition database, and trend awareness database are linked together to form preliminary search results for various types of agents.

[0040] Preferably, S31 includes:

[0041] S311, programmable The structured semantic intent is extracted using a coal mine task sheet structured parsing algorithm. Executable subtask list and task scheduling results The query is parsed into a routable query plan, which uses structured fields to describe the set of databases to be accessed, the query type, the core entities, the time range constraints, the retrieval strategy, and the rollback strategy.

[0042] S312 calls the coal mine scenario SQL generation algorithm, the coal mine data retrieval vector generation and mapping algorithm, and the coal mine time series query statement generation algorithm respectively to translate the routable query plan into specific structured data query requests, text retrieval requests, and time series query requests;

[0043] S313, based on the structured data query request, text retrieval request, or time-series query request, perform a search in the dynamic memory, topological cognition, and trend perception database of the coal external knowledge base to obtain preliminary search results of the selected Agent types in the dynamic memory, topological cognition, and trend perception databases.

[0044] Preferably, S4 includes:

[0045] S41, via evaluator Determine the credibility score of the initial search results for the selected agent. If the initial search results returned by various types of agents conflict with each other and the credibility score is inconsistent, Below the dynamic threshold Then, a re-search-re-reasoning loop will be automatically initiated until the initial search results returned by various types of agents are consistent and the credibility score is satisfactory. Not lower than the dynamic threshold Or, when the iteration reaches the preset maximum number of iterations, the evaluator... The system will integrate all scores, consistency check results, and iteration process records to generate an assessment report in the standard coal mine assessment report format, and then input the assessment report into the results fusion processor. ;

[0046] S42, via the result fusion unit The fragmented search results of different types of agents are linked together in the form of a strong relation graph. After prioritizing the different search results, the final search results are output according to the standard format of the coal mining industry. .

[0047] Preferably, before step S41, the method further includes:

[0048] S411, calculate the fuzzy features of the coal mining process scenario instructions through the fuzziness function;

[0049] S412, Statistical Preset Coal Mine Domain Ontology The number of core entities;

[0050] S413, obtain the knowledge base data matching quantity that matches the coal mining process scenario instructions in the coal mine plug-in knowledge base, and obtain the relevant data quantity that matches the coal mining process scenario instructions in the trend perception library of the coal mine plug-in knowledge base, and calculate the data sparsity based on the knowledge base data matching quantity and the relevant data quantity;

[0051] S414, Calculate the dynamic threshold based on fuzzy features, the number of core entities, and data sparsity. .

[0052] All of the above-mentioned optional technical solutions can be combined arbitrarily, and the present invention will not provide a detailed description of the structure after each combination.

[0053] By means of the above solution, the beneficial effects of the present invention are as follows:

[0054] The coal mine multimodal intent decoupling-dynamic agent evaluation and fusion reasoning engine algorithm performs intent recognition, subtask segmentation and dependency analysis on coal mining process scenario instructions. It selects three types of agents—coal mine professional large model, external tool integrated agent, and decision model integrated agent—and uses language model-enabled intent parsing database routing technology for retrieval. This allows for retrieval through various types of agents in conjunction with the coal mine external knowledge base after obtaining coal mining process scenario instructions. The retrieval results fully integrate multimodal data and collaboratively consider the retrieval results of different types of agents, resulting in richer and more accurate retrieval results and improved intelligence of human-computer interaction.

[0055] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0056] Figure 1 This is a schematic diagram of the system composition of the human-intelligent collaborative generative interaction method for coal mining scenarios provided in this embodiment of the invention.

[0057] Figure 2 This is a flowchart of the human-intelligent collaborative generative interaction method for coal mining scenarios provided in this embodiment of the invention. Detailed Implementation

[0058] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0059] The human-intelligent collaborative generative interaction method for coal mining scenarios provided in this invention is based on Figure 1 The illustrated multi-agent human-intelligent interaction system for the coal mining industry is implemented. The core module of this embodiment is a coal mine multimodal intent decoupling-dynamic agent evaluation and fusion reasoning engine algorithm, and an intent parsing database routing technology empowered by a language model. Based on this, it uses three types of agents—a coal mine professional large model, an external tool integrated agent, and a decision model integrated agent—along with a coal mine external knowledge base to parse the natural language input by the user, and finally returns the final search results to the user.

[0060] Among them, the coal mine professional large model is a large language model, a multimodal large model, etc., fine-tuned by PEFT (Parameter-Efficient Fine-Tuning) technology, such as the Deepseek large model and the Qwen multimodal large model. This enables the large model to understand the specific legal documents, operating procedures, mine design parameters and other information of a certain coal mine, that is, to achieve cross-modal cognition and collaboration.

[0061] External Tool Integration Agent: Integrates with various external tools. For example, a data cleaning agent unifies the format of multi-source data (e.g., converting binary data from sensors into structured JSON, or converting units and coordinate systems), resolving compatibility issues with heterogeneous data; a GitHub agent handles user issues programmatically by calling GitHub code.

[0062] Decision Model Integration Agent: This agent integrates machine learning models from various coal mine industrial scenarios (such as the manual control decision model after automation of the central hydraulic support cluster, the dynamic forecasting model of spatiotemporal pressure events in the intelligent fully mechanized mining face, the intelligent prediction and control model of working face straightness, and the real-time prediction model of pressure bearing effect after initial support of the hydraulic support, etc.). This allows the model to return the predicted results and decision opinions obtained after analysis to the coal mine multimodal intent decoupling-dynamic agent evaluation and fusion reasoning engine algorithm when the database cannot be queried or when the user inputs a professional question.

[0063] The coal mine external knowledge base consists of a dynamic memory base, a topological cognition base, and a trend perception base.

[0064] Dynamic memory: Stores continuously updated coal mine regulations, underground videos, and underground images to form a semantic cognitive system. Specifically, it utilizes the Milvus vector database to construct a real-time coal mine data knowledge base, primarily comprising four modules: document loading, segmentation, vectorization, and indexing, to achieve efficient storage of real-time coal mine data. First, real-time collected coal mine data (such as hydraulic support roof pressure data) and the latest files are segmented and processed using an embedded model to generate semantic vector representations. Then, the vectorized data is combined with structured metadata (such as support IDs and timestamps) to construct a hybrid index, which is finally written to the coal mine vector database, supporting efficient multimodal queries.

[0065] Topology Cognition Library: This library stores and inputs information on relationships between equipment, operational processes, and accident chains to create a relational reasoning system. First, entities are abstracted into nodes: equipment, process steps, and accidents / causes / consequences. Then, associations are abstracted into relationships: connection / control relationships between equipment, the sequence of process steps, and causal relationships in accident chains. Finally, the data is converted according to Cypher statement format and stored by executing statements through the Neo4J client or API.

[0066] Trend Awareness Library: Responsible for storing sensor data from fully mechanized mining faces and tunneling faces, thereby constructing a time-series evolution system. The specific process is as follows: First, the raw data from the fully mechanized mining and tunneling faces is cleaned; then, the data acquisition time is extracted as a timestamp, and fixed identifiers such as the face number and equipment ID are set as tags, while dynamic measurement values ​​such as current and gas concentration are set as fields; subsequently, the data is converted according to the InfluxDB row protocol format, and the converted data is written to the corresponding measurement table through client tools; finally, a data retention policy is configured, specifying the retention duration, number of copies, and expiration handling methods for different types of data, completing the entire data storage process.

[0067] Based on the above, such as Figure 2 As shown, the human-intelligent collaborative generative interaction method for coal mining scenarios provided in this embodiment of the invention can be implemented through the following steps S1 to S5:

[0068] S1: Obtain the coal mining process scenario instructions from the user's natural language input, and determine the instruction intent of the coal mining process scenario instructions based on the keyword rule matching method.

[0069] Specifically, the intent of instructions in the coal mining process scenario is divided into two categories: task-oriented instructions and investigation-oriented instructions. This embodiment of the invention pre-establishes a core keyword library for both task-oriented and investigation-oriented instructions. The keywords in the core keyword library for task-oriented instructions focus on "execution actions, task scheduling, and parameter adjustment," such as task, execution, operation, adjustment, start, stop, scheduling, operation, and allocation. The keywords in the core keyword library for investigation-oriented instructions focus on "status monitoring, situation investigation, and data feedback," such as patrol, investigation, monitoring, viewing, feedback, status, data, working conditions, and inspection. After establishing the core keyword library, this embodiment of the invention also assigns a weight to each keyword based on the degree of influence of the instruction, with the help of coal mining experts. For example, among the keywords for task-oriented instructions, "execution," "start," and "stop" have the highest weight (1.0); "scheduling," "operation," and "allocation" have the next highest weight (W=0.8); "adjustment" has a weight of 0.7, and so on. Among the keywords in investigative instructions, "monitoring," "operating conditions," and "data" have the highest weight, at 1.0; "patrol" and "inspection" have a weight of 0.8; "feedback" and "view" have a weight of 0.6, and so on. Based on this, the weights of each keyword are dynamically adjusted using historical statistical data from the coal mine. For example, if the matching accuracy of a keyword with the actual instruction type is >90%, the weight increases by 0.1; if the accuracy is <60%, the weight decreases by 0.1, and the adjusted weight is never lower than 0.3.

[0070] Based on the established core keyword library and assigned weights, when a user's natural language input command for a coal mining process scenario is obtained, the command is first segmented into words. Then, the frequency of occurrence of the two types of keywords is counted and a weighted score is calculated (the weights of the segmented words belonging to task-type commands and the weights of the segmented words belonging to survey-type commands are added together). If the score of the task-type segmented words is greater than the score of the survey-type segmented words, it is determined to be a task-type command; otherwise, it is determined to be a survey-type command.

[0071] S2, when the instruction intent of the coal mining process scenario command is a task-oriented instruction, the coal mine multimodal intent decoupling-dynamic agent evaluation and fusion reasoning engine algorithm is used to perform intent recognition, subtask segmentation and dependency analysis on the coal mining process scenario command to obtain structured semantic intent. Executable subtask list and task scheduling results .

[0072] In one specific embodiment, S2 includes S21 to S23.

[0073] S21, when the instruction intent of the coal mining process scenario instruction is a task-type instruction, the coal mining process scenario instruction is parsed into a form including intent nodes. Legal constraints Task dependency graph Structured semantic intent .

[0074] Specifically, S21 includes:

[0075] S211, when the instruction intent of the coal mining process scenario command is a task-oriented command, the coal mining process scenario command is input into the structured pre-trained model, and the structured pre-trained model outputs the initial structured triplet through generative modeling. And by pre-setting the physical entity of the coal mining field For the initial structured triples The validation process yields valid structured triples. and from legal structured triples Extracting intent nodes Structured triples include subject, object, and the relationship between them, and intent nodes. The core task that represents user needs.

[0076] Among them, the structured pre-trained model is the large-scale coal mine professional model mentioned above, which can be specifically adopted... For example, the Qwen multimodal large model is used for generative modeling to output structured triples. A pre-defined ontology for the coal mining domain is also provided. It includes core entities in the coal mining field (equipment, time, fault type, etc.) and legality rules, used to verify the legality of the results. Initial structured triples. Depend on ( Generates a preliminary semantic ternary relation containing user requirements (e.g., "<coal mining machine, query, fault data>").

[0077] In specific implementation, S211 involves passing through a pre-set coal mine domain body. For the initial structured triples The validation process yields valid structured triples. At that time, including:

[0078] First, extract the initial structured triples. The text defines the subject (S) and object (O) in the context of a coal mining machine query and fault data query, where "coal mining machine" is the subject and "fault data" is the object. It then determines whether the subject and object belong to the predefined coal mining domain ontology. First, define the core entity categories (e.g., "coal mining machine" belongs to the "mining equipment" category, "fault data" belongs to the "monitoring data" category). If an entity does not belong to these categories, it is marked as invalid. Second, verify the compliance of entity attributes (e.g., equipment numbers must conform to the "letter + number" format, and time must be within the effective monitoring range). Next, perform relationship validity verification: extract the initial structured triples. The relation (such as "query") in the query determines whether the relation exists. Define the allowed set of relationships (e.g., "query," "adjust," and "start" are legal, while "destroy" and "tamper" are illegal); next, verify the compatibility of relationships with entities (e.g., "coal mining machine" can be associated with "query" and "start," but cannot be associated with non-coal mining scenarios such as "irrigation"); then, perform a logical consistency check to examine whether the overall logic of the subject, object, and their relationship conforms to coal mine production rules (e.g., <conveyor, start, gas concentration > 1.5%> indicates a logical conflict due to excessive gas levels prohibiting equipment startup); finally, convert the initial structured triplet... The illegal parts in the structured triples are corrected (e.g., missing entity attributes are added) or removed to obtain valid structured triples that pass the validation. .

[0079] For example, when the instruction in the coal mining process scenario is "Query yesterday's fault data for the coal mining machine with equipment ID K12", the extracted intent node The function is to "query coal mining machine fault data".

[0080] S212, Extracting initial constraints from coal mining process scenario instructions through named entity recognition and slot filling model. And by pre-setting the physical entity of the coal mining field For initial constraints Perform validation to obtain valid constraints. .

[0081] Among them, the named entity recognition and slot filling model is the large-scale coal mine model mentioned above, which can be specifically adopted as follows: For example, the BERT+CRF model is used to extract constraints from instructions in a coal mining scenario, including original constraint information such as equipment number and time. (Legal constraints) That is, the pre-set coal mine domain body Verified constraint information. Taking the above coal mining process scenario instruction as an example, the legal constraint conditions... The values ​​are: "Device ID=K123" and "Time=Yesterday".

[0082] S213, using dependency parsing algorithm to analyze valid constraints. Processing is performed, and the intent nodes are processed through a graph neural network. The processing results of the dependency parsing algorithm generate a task dependency graph. The task dependency graph It includes the timing, condition, and causal dependencies between task nodes.

[0083] Specifically, based on the instructions in the coal mining process scenario, the legal constraints are first parsed using a dependency parsing algorithm. Extract the implicit task-related logical connections, and then combine them with graph neural networks to analyze the intent nodes. Deep association mining was conducted to ultimately construct a directed dependency graph containing temporal, conditional, and causal dependencies between task nodes.

[0084] S22, based on the intent node Generate executable subtask sheets .

[0085] In one specific embodiment, S22 includes:

[0086] S221, Determine the intent node based on the coal mine task template library. Is it a high-frequency task node, a low-frequency task node, or a new task node?

[0087] Specifically, the coal mining sector task template library ( The system stores the mapping rules between high-frequency task nodes and subtasks (e.g., "fault data query" corresponds to the subtask chain of "data retrieval - anomaly filtering - result processing"). This embodiment of the invention distinguishes task node types through a "dual determination of historical frequency statistics + template matching degree" mechanism. The specific process is as follows: First, a task frequency database is constructed to record the number of times historical task nodes appear in the system (…). ), and associated with the coal mining task template library ( The system pre-defines high-frequency task tags such as "fault query" and "parameter adjustment"; then, it sets high-frequency task judgment thresholds based on historical data statistics. If it appears ≥50 times in the past 3 months, if the intent node ( )of And in If a complete mapping rule exists, that is, if a similarity calculation is performed with high-frequency tags (matching degree ≥ 90%), then it is determined to be a high-frequency task node; if of But If a node contains some mapping rules (with a 60%-90% match rate with high-frequency tags) and its core entities such as device type and action are consistent with the template, it is determined to be a low-frequency task node; if exist There are no matching rules (match rate < 60%), or it contains an ontology related to the coal mining industry. New entities such as new equipment or new operational terms that are not defined in the task node are identified as new task nodes.

[0088] S222, if the intent node For high-frequency task nodes, the intent node is determined based on the mapping relationship between task nodes and subtasks in the coal mine task template library. Executable subtask list .

[0089] Specifically, if the intent node For high-frequency task nodes, the task template library in the coal mining field will be matched with the intent node. The subtask corresponding to the matched task node is determined as the intent node. Executable subtask list .

[0090] S223, if the intention node If it is a low-frequency task node or a new task node, then according to the subtask generative model and Generate intent nodes Executable subtask list .

[0091] Specifically, if the intent node For low-frequency task nodes or new task nodes, the subtask generative model is used. )and Generate intent nodes Executable subtask list Here, For Generating subtasks provides semantic support in the coal mining field (such as clearly defining the monitoring data types associated with "coal mining machine"). Subtask-based generative models ( ), used to dynamically generate subtasks for low-frequency or new task nodes.

[0092] S224, Define executable subtasks using a structured pattern. Input sources for each subtask Output format This forms a standardized, executable subtask list for coal mine scenarios. .

[0093] Specifically, the structured pattern defines the input sources for subtasks ( ) and output format ( )rule. For a single subtask ( The input source is something like "Device ID=K123" or "Time Series Data (Yesterday 00:00-24:00)". Taking the above coal mining process scenario command as an example, The output format for a single subtask is as follows: "Time Sequence Anomaly Subtask" outputs "List of Anomaly Periods (Format: YYYY-MM-DD HH:MM)", and "Procedure Retrieval Subtask" outputs "Relevant Procedure Clauses (Number + Content)".

[0094] S23, based on the task dependency graph and executable subtasks Generate task scheduling results .

[0095] In one specific embodiment, S23 includes:

[0096] S231, based on the task dependency graph Executable subtasks are sorted using a topological sorting algorithm. Each subtask in the graph is converted into a directed acyclic graph. The directed acyclic graph A framework for the sequential execution of subtasks.

[0097] Among them, topological sorting algorithm ( ) is used to execute subtasks Each subtask in the graph is converted into a directed acyclic graph.

[0098] S232, based on the coal mine's predefined data routing rule engine and directed acyclic graph. Generate data flow rules The data flow rules Define the path, format, and triggering conditions for data transfer between subtasks.

[0099] Among them, the coal mine predefined data routing rule engine is used to clarify the data flow rules between subtasks.

[0100] S233, based on the directed acyclic graph Differentiated retrieval strategies are generated based on the type of each subtask (e.g., retrieval-based, analysis-based). .

[0101] Among them, differentiated search strategies It includes retrieval logic adapted to different subtasks (such as retrieval and calculation). For example, the "procedure retrieval subtask" adopts a "keyword + ontology semantic matching" strategy, while the "real-time data calculation subtask" adopts an "incremental retrieval + cache update" strategy.

[0102] S234, if there is a directed acyclic graph If there are conflicts among the subtasks, then the directed acyclic graph (DAG) should be used. The features of each subtask are used to generate the task fusion result using a weighted fusion algorithm. .

[0103] Specifically, the determination of conflicts among sub-tasks is as follows: when there is a directed acyclic graph... When multiple sources of data converge, such as the output results of different subtasks, knowledge base data matching the coal mining process scenario instructions in the coal mine plug-in knowledge base, and relevant data matching the coal mining process scenario instructions in the trend perception library, consistency is determined through a three-dimensional consistency check: "quantified values, core semantics, and compliance rules." For example, if the difference between numerical data such as temperature, gas concentration, and risk probability exceeds a preset threshold (e.g., relative error > 10% or exceeds the equipment accuracy range, for example, subtask 1 outputs "equipment K123 bearing temperature 78℃," while real-time monitoring data shows "70℃," with a relative error of approximately 11.4% > 10%), or if the core semantics of semantic / conclusion data such as fault causes and compliance determinations are contradictory and cannot pass the preset coal mine domain ontology, then consistency is determined. A conflict is defined as a conflict between two subtasks: one is defined as "equipment operation meets standards" (e.g., subtask B determines "equipment operation meets standards," while subtask C determines "equipment operation does not meet standards," and both provide some supporting data, but cannot be reconciled through domain rules). This conflict arises when the execution logic of multi-source data violates predefined rules / safety regulations in coal mines, and the conclusions conflict, exceeding the allowable error range or interpretation boundary of the coal mine scenario. For example, the coal mine external knowledge base data shows "equipment startup is prohibited when gas concentration > 1.0%", while real-time monitoring data shows "gas concentration 0.9%". However, one subtask outputs "it is recommended to start the equipment," while another subtask outputs "wait for the gas concentration to drop below 0.8%." The conclusions based on compliance rules conflict, thus constituting a data conflict. Subsequently, a system based on subtask confidence (…) is initiated. The weighted fusion algorithm handles conflicts, that is, it uses a weighted fusion algorithm to generate the task fusion result. .

[0104] in, , The contextual features of the i-th subtask are extracted using a feature extraction algorithm. , The confidence level of the i-th subtask is assigned by the expert, and n is the number of conflicting subtasks.

[0105] S235, Combinatorial Directed Acyclic Graph Data flow rules Differentiated search strategies Results of task integration Obtain the task scheduling result In other words, .

[0106] S3 will structure semantic intent Executable subtask list and task scheduling results Submit to the orchestrator Arranger Based on capability matching and rule gating mechanisms, three types of agents are selected: a coal mine professional big model, an external tool integration agent, and a decision model integration agent. The selected agents are then controlled to perform retrieval through language model-enabled intent parsing database routing technology to obtain preliminary retrieval results for each type of agent.

[0107] Among them, the arranger Based on ability matching ( ) and rule gating ( When selecting one of the three types of agents—a large-scale coal mine professional model, an external tool integration agent, and a decision model integration agent—the first step is to match capabilities. In this phase, an initial matching process is conducted based on the core requirements of the sub-tasks and the core capabilities of the three types of agents. For example: for the cleaning / alignment / anomaly detection (time series / video) sub-tasks, the high-throughput, low-cost structured computing capabilities of external tool-integrated agents are matched; for the semantic understanding / retrieval-based generation / summarization / suggestion sub-tasks, the natural language reasoning, RAG, and procedure interpretation capabilities of large-scale coal mine professional models are matched; for the risk probability / compliance determination / control strategy sub-tasks, the numerical prediction, control, and analysis capabilities of decision-making model-integrated agents are matched. Next, rule gating is implemented. In this stage, secondary screening and confirmation are performed using typical triggering conditions as gating rules. For example: if a subtask has issues with raw data or data quality, or requires high-throughput / low-cost structured computation, the external tool integration agent invocation rule is triggered; if a subtask requires natural language inference, RAG, or procedure interpretation, the coal mine professional large-scale model invocation rule is triggered; if a subtask requires numerical prediction or control and meets safety compliance requirements, the decision model integration agent invocation rule is triggered, ensuring that only agents that simultaneously meet capability matching and rule gating are selected. Finally, the orchestrator... Based on the task dependency graph between subtasks (e.g., data cleaning should be performed sequentially over semantic understanding, while independent anomaly detection and initial risk assessment can be performed in parallel), the selected Agent is scheduled to perform retrieval in parallel / serial order.

[0108] Table 1 shows the correspondence between subtask types, capability matching, and rule gating.

[0109]

[0110] In one specific embodiment, when controlling the selected Agent to perform retrieval using language model-enabled intent parsing database routing technology and obtain preliminary retrieval results for various types of Agents, S3 includes:

[0111] S31, Arranger The selected agent will format the semantic intent. Executable subtask list and task scheduling results By using a coal mine intent-weighted multimodal feature enhancement query mechanism, searches are performed in the dynamic memory, topological cognition, and trend perception databases of the coal external knowledge base to obtain preliminary search results for various types of agents in the dynamic memory, topological cognition, and trend perception databases.

[0112] In specific implementation, S31 includes:

[0113] S311, programmable The coal mine task order structure parsing algorithm is used to analyze the structured semantic intent and executable task order. and task scheduling results The query is parsed into a routable query plan, which uses structured fields to describe the set of databases to be accessed, the query type, the core entities, the time range constraints, the retrieval strategy, and the rollback strategy.

[0114] Among them, the coal mine task sheet structured parsing algorithm is used to extract structured semantic intent. Executable subtask list and task scheduling results Parse into a structured, routable query plan (structured JSON).

[0115] S312 calls the coal mine scenario SQL generation algorithm, the coal mine data retrieval vector generation and mapping algorithm, and the coal mine time series query statement generation algorithm respectively to translate the routable query plan into specific structured data query requests, text retrieval requests, and time series query requests.

[0116] The SQL generation algorithm for coal mine scenarios includes a rule template library (such as equipment parameter query templates) and a pre-trained language model (such as the BART model for coal mine SQL generation). The coal mine data retrieval vector generation and mapping algorithm is based on a coal mine domain feature encoding model (such as coal mine BERT) to convert textual intent into retrieval vectors. The coal mine time-series query statement generation algorithm supports query syntax generation for time-series databases (such as InfluxDB and TimescaleDB), including parameter rules such as sliding window and sampling frequency.

[0117] Among them, the structured data query request is an SQL statement output by the coal mine scenario SQL generation algorithm, such as "SELECT vibration value, temperature FROM coal mining machine monitoring table WHERE device ID='K123' AND time BETWEEN '2025-10-20 08:00' AND '2025-10-20 18:00'".

[0118] Text retrieval request: The retrieval vector output by the coal mine data retrieval vector generation and mapping algorithm, such as the 768-dimensional vector encoded based on "coal mining machine K123 October 20th fault" and the vector library retrieval instruction.

[0119] Time series query request: A time series database query statement output by the coal mine time series query statement generation algorithm, such as "InfluxQL: SELECT mean (vibration value) FROM coal mining machine - K123 WHERE time >= '2025-10-20T08:00:00Z' AND time <= '2025-10-20T18:00:00Z' GROUP BY time (10m)".

[0120] S313, based on the structured data query request, text retrieval request, or time-series query request, perform a search in the dynamic memory, topological cognition, and trend perception database of the coal external knowledge base to obtain preliminary search results of the selected Agent types in the dynamic memory, topological cognition, and trend perception databases.

[0121] S32, programmable By using a cross-database query result feedback and iteration mechanism, the preliminary search results of various types of agents in the dynamic memory database, topology cognition database, and trend awareness database are linked together to form preliminary search results for various types of agents.

[0122] Specifically, the input to step S32 is the preliminary retrieval result obtained in S31. In practice, the coal mine entity association topology reasoning algorithm infers the associated components and dependencies between devices from the topology cognitive base; the coal mine time series data range limitation and anomaly detection algorithm narrows the time series data query range based on the preliminary retrieval results and detects abnormal time periods; the coal mine semantic range focusing and case matching retrieval algorithm narrows the semantic retrieval range of historical cases based on abnormal time periods to improve matching accuracy; and the mutual feedback iterative controller coordinates the transmission of preliminary retrieval results between various knowledge bases (dynamic memory base, topology cognitive base, and trend perception base) and sets iteration termination conditions (such as query boundary convergence or reaching the maximum number of iterations of 3).

[0123] The results obtained in step S32 include:

[0124] 1. Intermediate result chain for cross-database feedback:

[0125] Topology Cognition Library Output (R1): For example, "The directly related components of device A are B (drive motor) and C (conveyor belt), and the dependency relationship is B→A (B failure causes A to stop)".

[0126] Trend Awareness Library Output (R2): For example, "Component B's current value exceeded the threshold (normal range 50-80A, actual 95-110A) between 10:00 and 12:00 on October 20, 2025, and was determined to be an abnormal period."

[0127] Dynamic memory output (R3): such as "2024-05-12 Historical case of component B (same model) during the current over-threshold period: due to bearing wear, the corresponding measure is to stop the machine and replace the bearing, case matching degree 92%".

[0128] 2. Core Output:

[0129] (1) Comprehensive search results: Integrate the intermediate result chain to form a structured analysis report, including:

[0130] Anomaly cause identification: "The anomaly in equipment A is caused by bearing wear in associated component B (based on R1+R2+R3)";

[0131] Historical similar cases: "Matching case number 2024-05-12-B, key features: same model of component, same current over-threshold mode";

[0132] Recommended response: "1. Immediately stop the machine and inspect bearing B of component; 2. Replace the bearing and then run it under no-load for 30 minutes according to the procedure."

[0133] In this embodiment of the invention, the intermediate results of queries from the coal mine external knowledge base are used to form a "mutual feedback closed loop" between the query content and the intermediate data results through a cross-database query result mutual feedback iteration mechanism. For example, the topology cognition database first queries "the associated component B of equipment A" using the coal mine entity association topology reasoning algorithm, and feeds the result back to the trend perception database. The coal mine time series data range limitation and anomaly detection algorithm limit the query to "the operating data of component B in the last 7 days" to narrow the time range. The trend perception database returns "the abnormal fluctuation period of component B", which is then fed back to the dynamic memory database. The coal mine semantic range focusing and case matching retrieval algorithm limits the query to "historical cases related to the anomaly of component B within this period" to narrow the semantic range. Thus, through multiple rounds of mutual feedback iteration, the query boundary is gradually focused, invalid calculations are reduced, and finally, a comprehensive retrieval result of "analysis of the abnormal causes of the associated component B of equipment A, historical similar cases and response suggestions" is output.

[0134] S4, determine the credibility score of the preliminary search results of the selected Agent, and determine the final search results based on the credibility score.

[0135] In one specific embodiment, S4 includes:

[0136] S41, via evaluator Determine the credibility score of the initial search results of the selected agent. If the initial search results returned by various types of agents conflict with each other and the credibility score is inconsistent, Below the dynamic threshold Then, a re-search-re-reasoning loop will be automatically initiated until the initial search results returned by various types of agents are consistent and the credibility score is satisfactory. Not lower than the dynamic threshold Or, when the iteration reaches the preset maximum number of iterations, the evaluator... The system will integrate all scores, consistency check results, and iteration process records to generate an assessment report in the standard coal mine assessment report format, and then input the assessment report into the results fusion processor. .

[0137] Among them, the credibility score ( The data reliability score is determined by constructing a quantitative system (total score 0-100), with each dimension containing specific scoring rules and weights. (Weight 0.3) The calculation of the preliminary retrieval results for each type of Agent requires determining the data integrity (15 points for a missing rate < 5%, deducting 5 points for every 5% increase) and the reliability of data sources. The sum of these two factors constitutes the data reliability score. The grading rules are set based on the characteristics of Agent data sources in the coal mine scenario: 13-15 points for Agent data from a real-time coal mine monitoring system (e.g., PLC control system); 9-12 points for data from a historical coal mine database (e.g., equipment history database, fault case database); 5-8 points for data from external tool integration or cross-scenario migration (e.g., data from other mines of the same model without clear verification records); and 0-4 points for data from manual entry (e.g., digital records of underground inspections) or temporary data without fixed collection standards. Model confidence score ( (Weight 0.4) The model confidence score is calculated based on the confidence level of each agent type's output (e.g., 40 points for a large model generation probability ≥ 0.9, 25 points for 0.7-0.9) and the model's historical accuracy (40 points for accuracy ≥ 90% in the last 100 tasks, deducting 10 points for every 10% decrease). The result is the matching score. (Weight 0.3) The score is calculated by considering the matching degree with the coal mine knowledge base (≥90% for 30 points, 70%-90% for 20 points) and the deviation rate from the results of similar historical tasks (<5% for 30 points, deducting 10 points for every 5% increase). The sum of these two factors is the result matching degree score. Finally, through... Calculate credibility score .

[0138] Specifically, when determining whether the preliminary search results returned by various types of agents conflict, a consistency check is performed using three dimensions: "quantified numerical values, core semantics, and compliance rules." For example, if the preliminary search result from the coal mine professional big data model is a temperature greater than 60℃, while the preliminary search result from the external tool integrated agent is a temperature of 50℃, then there is a conflict between the two.

[0139] Through task scheduling results Fusion results of the fusion task It can provide the evaluator with a unified and focused evaluation object, avoiding the evaluation engine from directly facing scattered and conflicting raw data, such as conflicting conclusions from multiple subtasks, which can lead to chaotic scoring logic.

[0140] It should be noted that, before step S41, the dynamic threshold needs to be determined first. Specifically, in determining the dynamic threshold At that time, including S411 to S414:

[0141] S411 calculates the fuzzy features of the coal mining process scenario instructions using a fuzziness function.

[0142] Among them, fuzzy features , Represents the ambiguity function. This refers to commands for the coal mining process.

[0143] S412, Statistical Preset Coal Mine Domain Ontology The number of core entities.

[0144] Among them, the number of core entities , This is a counting function.

[0145] S413, obtain the knowledge base data matching quantity that matches the coal mining process scenario instructions in the coal mine plug-in knowledge base, and obtain the relevant data quantity that matches the coal mining process scenario instructions in the trend perception library of the coal mine plug-in knowledge base, and calculate the data sparsity based on the knowledge base data matching quantity and the relevant data quantity.

[0146] Among them, data sparsity ;in, This indicates the number of data matches in the knowledge base. This indicates the amount of relevant data. This represents the maximum data volume threshold for the knowledge base; This indicates the maximum data volume threshold for real-time monitoring; both are preset values. It reflects the sufficiency of the current task data.

[0147] S414, Calculate the dynamic threshold based on fuzzy features, the number of core entities, and data sparsity. .

[0148] Specifically, in calculating dynamic thresholds First, combine the fuzzy features, the number of core entities, and the data sparsity with a coal mine task type correction coefficient (empirical value), and then substitute them into the random forest model mapping function. Generate the basic dynamic threshold for the current round, expressed by the formula: ,in, This is the mapping function for the random forest model. This is a correction coefficient for coal mine task types. Next, the initial threshold is calculated. ,in, Weighting based on expert experience; The baseline threshold was determined by combining the experience of experts in the coal mining field with statistical analysis of historical coal mine task data; The average effective threshold is obtained based on statistical analysis of historical task data. Next, using the initial threshold... Using the initial value as a reference, and combining the key feedback data from the previous iteration—inference accuracy (Acc) and iteration cost (Cost)—the reward value R is calculated through the reward function of the Deep Q-Network (DQN). The DQN adjusts the threshold parameter based on this reward value: if the previous Acc was high and the Cost was low, the threshold is slightly increased to enhance the reliability of the result; if the previous Acc was low or the Cost was high, the threshold is appropriately decreased to balance efficiency and the probability of achieving the target. ( (where Acc is the iteration cost), Cost is the inference accuracy at the current threshold, and Cost is the time and resource cost of iterative retrieval / inference. , Weighting coefficients ( It ensures a balance between security and efficiency, and adapts to dynamic changes in scenarios.

[0149] S42, via the result fusion unit The fragmented search results of different types of agents are linked together in the form of a strong relation graph. After prioritizing the different search results, the final search results are output according to the standard format of the coal mining industry. .

[0150] Specifically, the result fusion A multi-source heterogeneous data dynamic fusion and suggestion enhancement algorithm is used for data fusion. Nodes in a graph constructed with strong relationships are also considered. Representing result units and entities, edges Represents evidence correlation, temporal synchronization, or causal inference, including weighting. The result is that the fusion generator F constructs a graph based on the connectivity and total edge weights of strongly correlated graphs. The results are prioritized based on the credibility score given by evaluator E, and the final search results are output according to the standard format for the coal mining industry. .

[0151] It should be noted that the orchestrator evaluator and results fusion The relevant algorithms executed also belong to the coal mine multimodal intent decoupling-dynamic agent evaluation and fusion reasoning engine algorithm.

[0152] S5 provides a visual representation of the final search results.

[0153] This completes the feedback of instructions for the coal mining process.

[0154] The method provided in this embodiment of the invention has the following beneficial effects:

[0155] (1) It can solve the problems of unified representation, feature alignment and precise decoupling of intent for multimodal heterogeneous data in coal mines, laying the foundation for data and demand understanding for subsequent technology implementation; on this basis, a large professional coal mine model based on offline fine-tuning and online RAG is constructed, and its closed-loop evolutionary learning is realized to create the core intelligent engine of the system.

[0156] (2) External tools integrate Agent and decision model integration Agent task orchestration, efficient communication and dynamic scheduling technology, build a multi-subject collaborative architecture, and complete the structured storage of coal mine plug-in knowledge base including semantic, relational and temporal dimensions. Optimize intent weighted multimodal enhanced query and cross-database iterative update mechanism to provide solid knowledge support for large model and Agent.

[0157] (3) Coal Mine Multimodal Intent Decoupling - Dynamic Agent Evaluation and Fusion Reasoning Engine Algorithm, ensuring the real-time accuracy of multi-source information integration and reasoning.

[0158] (4) Integrate the natural language interaction between the large coal mine professional model and decision-making models in fields such as hydraulic support cluster automation and pressure event prediction, improve the intent mapping and feedback mechanism, and open up the entire technical link of "data-model-collaboration-knowledge-reasoning-application".

[0159] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A human-intelligent collaborative generative interaction method for coal mining scenarios, characterized in that, include: S1, obtain the coal mining process scenario instructions input by the user in natural language, and determine the instruction intent of the coal mining process scenario instructions based on the rule matching method of keywords; S2, when the instruction intent of the coal mining process scenario command is a task-oriented instruction, the coal mine multimodal intent decoupling-dynamic agent evaluation and fusion reasoning engine algorithm is used to perform intent recognition, subtask segmentation and dependency analysis on the coal mining process scenario command to obtain structured semantic intent. Executable subtask list and task scheduling results ; S3 will structure semantic intent Executable subtask list and task scheduling results Submit to the orchestrator Arranger Based on capability matching and rule gating mechanisms, three types of agents are selected: coal mine professional big model, external tool integration agent, and decision model integration agent. The selected agents are controlled to perform retrieval through language model-enabled intent parsing database routing technology to obtain preliminary retrieval results for each type of agent. S4, determine the credibility score of the preliminary search results of the selected Agent, and determine the final search results based on the credibility score; S5 provides a visual representation of the final search results.

2. The human-intelligent collaborative generative interaction method for coal mining scenarios according to claim 1, characterized in that, S2 includes: S21, when the instruction intent of the coal mining process scenario instruction is a task-type instruction, the coal mining process scenario instruction is parsed into a form including intent nodes. Legal constraints Task dependency graph Structured semantic intent ; S22, based on the intent node Generate executable subtask sheets ; S23, based on the task dependency graph and executable subtasks Generate task scheduling results .

3. The human-intelligent collaborative generative interaction method for coal mining scenarios according to claim 2, characterized in that, S21 includes: S211, when the instruction intent of the coal mining process scenario command is a task-oriented command, the coal mining process scenario command is input into the structured pre-trained model, and the structured pre-trained model outputs the initial structured triplet through generative modeling. And by pre-setting the physical entity of the coal mining field For the initial structured triples The validation process yields valid structured triples. and from legal structured triples Extracting intent nodes The structured triple includes a subject, an object, and the relationship between them; the intent node... The core task representing user needs; S212, Extracting initial constraints from coal mining process scenario instructions through named entity recognition and slot filling model. And by pre-setting the physical entity of the coal mining field For initial constraints Perform validation to obtain valid constraints. ; S213, using dependency parsing algorithm to analyze valid constraints. Processing is performed, and the intent nodes are processed through a graph neural network. The processing results of the dependency parsing algorithm generate a task dependency graph. The task dependency graph It includes the timing, condition, and causal dependencies between task nodes.

4. The human-intelligent collaborative generative interaction method for coal mining scenarios according to claim 3, characterized in that, S211 is performed by passing through the pre-set coal mine area body. For the initial structured triples The validation process yields valid structured triples. At that time, including: First, extract the initial structured triples. The subject and object in the context, and whether the subject and object belong to the pre-defined coal mine domain ontology. If a defined core entity category is not included, it is marked as illegal. Secondly, verify the compliance of entity attributes; Next, perform relational validity verification: extract the initial structured triples. The relationship in the middle, determine whether the relationship exists. Defined allowed sets of relations; Next, verify the compatibility between the relationship and the entity; Then, a logical consistency check is performed to examine whether the overall logic of the subject, object, and their relationship conforms to the rules of coal mine production. Finally, the initial structured triples The illegal parts in the structured triples are corrected or removed to obtain valid structured triples that pass the verification. .

5. The human-intelligent collaborative generative interaction method for coal mining scenarios according to claim 2, characterized in that, S22 includes: S221, Determine the intent node based on the coal mine task template library. Is it a high-frequency task node, a low-frequency task node, or a new task node? S222, if the intent node For high-frequency task nodes, the intent node is determined based on the mapping relationship between task nodes and subtasks in the coal mine task template library. Executable subtask list ; S223, if the intent node If it is a low-frequency task node or a new task node, then according to the subtask generative model and Generate intent nodes Executable subtask list ; S224, Define executable subtasks using a structured pattern. Input sources for each subtask Output format This forms a standardized, executable subtask list for coal mine scenarios. .

6. The human-intelligent collaborative generative interaction method for coal mining scenarios according to claim 2, characterized in that, S23 includes: S231, based on the task dependency graph Executable subtasks are sorted using a topological sorting algorithm. Each subtask in the graph is converted into a directed acyclic graph. The directed acyclic graph A sequential framework for the execution of subtasks; S232, based on the coal mine's predefined data routing rule engine and directed acyclic graph. Generate data flow rules The data flow rules Define the path, format, and triggering conditions for data transfer between subtasks; S233, based on the directed acyclic graph Differentiated retrieval strategies are generated based on the types of each subtask. ; S234, if there is a directed acyclic graph If there are conflicts among the subtasks, then the directed acyclic graph (DAG) should be used. The features of each subtask are used to generate the task fusion result using a weighted fusion algorithm. ; S235, Combinatorial Directed Acyclic Graph Data flow rules Differentiated search strategies Results of task integration Obtain the task scheduling result .

7. The human-intelligent collaborative generative interaction method for coal mining scenarios according to claim 1, characterized in that, When S3 controls the selected Agent to perform retrieval using language model-enabled intent parsing database routing technology and obtains preliminary retrieval results for various types of Agents, it includes: S31, Arranger The selected agent will structure semantic intent. Executable subtask list and task scheduling results By using a coal mine intent-weighted multimodal feature enhancement query mechanism, searches are performed in the dynamic memory, topological cognition, and trend perception databases of the coal external knowledge base to obtain preliminary search results for various types of agents in the dynamic memory, topological cognition, and trend perception databases. S32, programmable By using a cross-database query result feedback and iteration mechanism, the preliminary search results of various types of agents in the dynamic memory database, topology cognition database, and trend awareness database are linked together to form preliminary search results for various types of agents.

8. The human-intelligent collaborative generative interaction method for coal mining scenarios according to claim 7, characterized in that, S31 includes: S311, programmable The structured semantic intent is extracted using a coal mine task sheet structured parsing algorithm. Executable subtask list and task scheduling results The query is parsed into a routable query plan, which uses structured fields to describe the set of databases to be accessed, the query type, the core entities, the time range constraints, the retrieval strategy, and the rollback strategy. S312 calls the coal mine scenario SQL generation algorithm, the coal mine data retrieval vector generation and mapping algorithm, and the coal mine time series query statement generation algorithm respectively to translate the routable query plan into specific structured data query requests, text retrieval requests, and time series query requests; S313, based on the structured data query request, text retrieval request, or time-series query request, perform a search in the dynamic memory, topological cognition, and trend perception database of the coal external knowledge base to obtain the preliminary search results of the selected Agent types in the dynamic memory, topological cognition, and trend perception databases.

9. The human-intelligent collaborative generative interaction method for coal mining scenarios according to claim 1, characterized in that, S4 includes: S41, via evaluator Determine the credibility score of the initial search results of the selected agent. If the initial search results returned by various types of agents conflict with each other and the credibility score is inconsistent, Below the dynamic threshold Then, a re-search-re-reasoning loop will be automatically initiated until the initial search results returned by various types of agents are consistent and the credibility score is satisfactory. Not lower than the dynamic threshold Or, when the iteration reaches the preset maximum number of iterations, the evaluator... The system will integrate all scores, consistency check results, and iteration process records to generate an assessment report in the standard coal mine assessment report format, and then input the assessment report into the results fusion processor. ; S42, via the result fusion unit The fragmented search results of different types of agents are linked together in the form of a strong relation graph. After prioritizing the different search results, the final search results are output according to the standard format of the coal mining industry. .

10. The human-intelligent collaborative generative interaction method for coal mining scenarios according to claim 9, characterized in that, Before S41, it also includes: S411, calculate the fuzzy features of the coal mining process scenario instructions through the fuzziness function; S412, Statistical Preset Coal Mine Domain Ontology The number of core entities; S413, obtain the knowledge base data matching quantity that matches the coal mining process scenario instructions in the coal mine plug-in knowledge base, and obtain the relevant data quantity that matches the coal mining process scenario instructions in the trend perception library of the coal mine plug-in knowledge base, and calculate the data sparsity based on the knowledge base data matching quantity and the relevant data quantity; S414, Calculate the dynamic threshold based on fuzzy features, the number of core entities, and data sparsity. .

Images