Knowledge base construction and adaptive question answering method for industrial manufacturing

Abstract

The present application relates to the technical field of intelligent industrial manufacturing, more specifically, to a knowledge base construction and adaptive question-answering method for industrial manufacturing, which is used to solve the problem that the prior art cannot integrate user questioning and multi-round context to realize continuous interaction, cannot combine real-time working conditions and abnormal events to enhance field perception, cannot accurately match entities and identify multi-hop paths to mine deep intentions, and is difficult to realize reliable conversion from natural language to executable queries; the present application keeps interaction continuous by integrating user questioning and multi-round context, enhances field perception by combining real-time working conditions and abnormal events, matches entities and identifies multi-hop paths by using a knowledge graph to mine deep intentions, associates time and space with equipment information to ensure accurate understanding, and guarantees structured output to be legal and complete through a rule constraint and repair mechanism, finally realizing reliable conversion from natural language to executable queries.

Description

Technical Field

[0001] This invention relates to the field of intelligent industrial manufacturing technology, and more specifically, to an adaptive question-answering method for knowledge base construction and intelligent agents in industrial manufacturing. Background Technology

[0002] In industrial manufacturing, equipment operation logs, process documents, and maintenance records are scattered across heterogeneous systems such as MES, PLM, and SCADA, forming information silos and making knowledge integration difficult. Frequent iterations of production line equipment and continuous changes in processes make it difficult to update traditional static knowledge bases in a timely manner, compromising the timeliness of knowledge. On-site problems often involve multimodal data such as vibration spectra, AR images, and natural language processing, and existing systems lack cross-modal understanding capabilities, relying on manual handling for complex operating conditions. General large models suffer from problems such as generating incorrect information and lack of interpretability, and do not embed compliance verification mechanisms for standards such as AS9100 and IATF16949, which can easily lead to the output of operational suggestions that violate safety regulations. Existing methods either rely on rigid rule bases, resulting in poor adaptability, or use black-box models, resulting in low credibility, making it difficult to meet the intelligent question-and-answer requirements of industrial scenarios for dynamic updates, multimodal understanding, and standard compliance.

[0003] Reference patent application CN119961418A discloses a question-answering method and system integrating multi-granularity knowledge graphs and adaptive intelligent agents. It includes: knowledge governance and extraction, knowledge generation and expansion, knowledge retrieval and management, and knowledge pattern discovery and adaptive retrieval. Through the collaboration of multi-granularity knowledge graphs and adaptive intelligent agents, it can more comprehensively understand user questions and retrieve more relevant knowledge from the knowledge base, thereby significantly improving the accuracy of question answering and reducing the generation of incorrect answers. Simultaneously, the multi-granularity knowledge graph can support more complex knowledge application needs, thereby enhancing the application capabilities of the question-answering system and enabling it to handle more complex question-answering tasks. However, while the aforementioned reference patents improve the accuracy, complex task handling, and multi-turn interaction capabilities of question-answering systems by integrating multi-granularity knowledge graphs and adaptive intelligent agents, thus providing users with a better user experience, they cannot integrate user questions with multi-turn contexts to achieve coherent interaction, cannot combine real-time operating conditions and abnormal events to enhance on-site perception, cannot accurately match entities and identify multi-hop paths to uncover deeper intentions, and are difficult to reliably convert natural language into executable queries. Furthermore, they cannot achieve comprehensive perception of industrial scenarios, cannot ensure that reasoning conforms to industrial logic, cannot generate optimal solutions under resource constraints, and are difficult to construct an adaptive reasoning mechanism from strategy generation and dynamic execution to closed-loop optimization.

[0004] To address these issues, we propose an adaptive question-answering method for knowledge base construction and intelligent agents in industrial manufacturing. Summary of the Invention

[0005] The purpose of this invention is to provide an adaptive question-answering method for knowledge base construction and intelligent agents in industrial manufacturing. It solves the problems of existing technologies, such as the inability to integrate user questions with multi-turn context to achieve coherent interaction, the inability to combine real-time operating conditions and abnormal events to enhance on-site perception, the inability to accurately match entities and identify multi-hop paths to uncover deep intentions, the difficulty in achieving reliable conversion from natural language to executable queries, the inability to achieve comprehensive perception of industrial scenarios, the inability to ensure that reasoning conforms to industrial logic, the inability to generate optimal solutions under resource constraints, and the difficulty in constructing an adaptive reasoning mechanism from policy generation and dynamic execution to closed-loop optimization.

[0006] The objective of this invention is achieved through the following technical solution: An adaptive question-answering method for knowledge base construction and intelligent agents in industrial manufacturing includes the following steps: Step 1: Collect multi-source observation data during equipment operation, process execution, environmental monitoring, quality inspection, and production management. Clean the collected data, handle missing values, filter noise, standardize, and align the data over time. Generate structured abnormal events based on preset thresholds. Step 2: Based on maintenance records, process specifications, and FMEA documents, extract entities and relationships of equipment, processes, and failure modes through natural language processing, construct a knowledge graph, and perform incremental updates based on the frequency and pattern of abnormal events; Step 3: Based on user questions and multi-turn dialogue context, integrate real-time operating data and structured abnormal events, combine entities and relationships in the knowledge graph to perform semantic understanding and intent recognition, complete and reconstruct ambiguous or incomplete questions, and output structured question representations; Step 4: Based on the structured question representation generated by semantic understanding, combined with knowledge graph and real-time working condition data, generate an initial reasoning strategy, execute multi-task parallel reasoning, dynamically adjust the reasoning path and task priority, and update the reasoning model parameters and strategy rules using user feedback. Step 5: Respond to the user's natural language question, maintain the multi-turn dialogue context, initiate the semantic understanding and reasoning process, return the results, and record the user feedback information.

[0007] In a preferred embodiment of the present invention, step two, based on maintenance records, process specifications, and FMEA documents, involves extracting entities and relationships of equipment, processes, and failure modes through natural language processing to construct a knowledge graph, and then incrementally updating it according to the frequency and pattern of abnormal events. Entities and relationships of equipment, processes, and failure modes are extracted from maintenance records, process specifications, and FMEA documents to construct a knowledge graph. Incremental updates are performed based on the frequency and pattern changes of abnormal events. Named entity recognition is performed to extract entities of equipment, components, processes, failure modes, detection indicators, and operators, generating entity sets. Dependency parsing is performed to identify semantic relationships between entities, generating sets of triples. Entities are used as nodes and triples as edges to establish a graph structure. Different representations of the same entity are normalized. The graph structure is stored in a graph database, completing the initialization of the knowledge graph. The absolute value of the difference between the frequency of events occurring in the current time window and the historical average is taken. If the absolute value is greater than the product of the historical standard deviation and the fixed sensitivity coefficient, the update mechanism is triggered. The newly occurring abnormal event sequence is scanned, entity combinations not recorded in the graph are identified, the attribute information of the affected entities is updated, and the changes are written to the graph database to complete the incremental update of the knowledge graph.

[0008] In a preferred embodiment of the present invention, step three, which involves fusing user questions with multi-turn dialogue context and integrating real-time operating data with structured abnormal events, includes: The system obtains the semantic representation of the user's current question and the semantic representation of each statement in the multi-turn dialogue history. It calculates the median value of the association strength between the question and the historical statements using a formula. The calculation result reflects the comprehensive strength of semantic matching and positional relationship. The calculation result is multiplied by the time decay factor, and the product result is normalized to generate the weighting coefficient of the historical statements. The semantic representation of each historical statement is multiplied by the corresponding weighting coefficient and summed to obtain the historical context representation. The semantic representation of the current question is merged with the historical context representation to generate the dialogue context fusion result. Obtain semantic representations of real-time operating condition data and structured abnormal events. Map the dialogue context fusion results, real-time operating condition semantic representations, and structured abnormal event semantic representations to a unified semantic space to generate modal gating vectors. Using the dialogue context fusion result as the query, and the real-time working condition semantic representation and the structured abnormal event semantic representation as the key and value, perform cross-modal attention operation, add the cross-modal attention output to the dialogue context fusion result, generate an enhanced context representation, and output the enhanced context representation.

[0009] In a preferred embodiment of the present invention, step three, which involves semantic understanding and intent recognition based on entities and relationships in a knowledge graph, includes: Get the text content of the user's current question, match semantically related candidate entities in the knowledge graph based on the question content, obtain the contextual semantic representation of the multi-turn dialogue, align the contextual semantic representation with the candidate entities, confirm the entity association, obtain the entity relationship structure in the knowledge graph, and extract the directly associated adjacency subgraph with the candidate entity as the center. Identify multi-hop connection paths between entities in the subgraph to form a set of path patterns. Fuse the path patterns with the contextual semantics to generate a joint semantic representation. Determine the intent category corresponding to the question based on the matching strength between the path structure and entity relationships in the joint semantic representation. The system retrieves the time information mentioned in the question and associates it with the time entity in the knowledge graph; it retrieves the geographical location information mentioned in the question and associates it with the location entity in the knowledge graph; it retrieves the device identification information mentioned in the question and associates it with the device entity in the knowledge graph; and generates a structured semantic understanding result that includes intent category, time information, geographical location information, device identification information, and associated paths.

[0010] In a preferred embodiment of the present invention, step three, which involves completing and reconstructing the fuzzy or incomplete problem and outputting a structured problem representation, includes: The system obtains the text content of the user's current question, the contextual semantic representation of the multi-turn dialogue, the entity set and relation set in the knowledge graph, the initial structured representation of the user's question, and based on the contextual semantic representation and the entity relations in the knowledge graph, identifies the missing semantic components in the question and uses the association information in the knowledge graph to complete the missing fields in the initial structured representation. The completed structured representation is matched with the preset syntax and semantic rules to determine whether the completed structured representation meets all constraints. For structured representations that do not meet the constraints, a repair operation is performed. The repair process aims to minimize the scope of modification and constraint conflicts, and outputs the repaired structured problem representation.

[0011] In a preferred embodiment of the present invention, the process of generating an initial reasoning strategy in step four, which involves combining the structured question representation generated based on semantic understanding with knowledge graphs and real-time operational data, includes: Obtain the structured question representation output from the semantic understanding stage, obtain the device functions and operation specifications from the knowledge graph, determine the operation dependencies between entities based on the entities and operation specifications in the structured question representation, obtain sensor readings and system status flags from real-time operating data, and assign dynamic confidence weights to each entity in the structured question representation in combination with real-time operating data. Based on operation dependencies and dynamic confidence weights, a set of candidate inference strategies is constructed. Each strategy defines a mapping from the current system state to the action sequence. The expected benefits and execution costs of each candidate strategy under the current state distribution are evaluated, an initial set of inference strategies is generated, and the generated results are output as the initial inference strategy.

[0012] In a preferred embodiment of the present invention, step four, which involves performing multi-task parallel inference, dynamically adjusting the inference path and task priority, and updating the inference model parameters and policy rules using user feedback, includes: The system obtains the execution status of the current inference task, the latest confidence level of each entity and the urgency information of newly arrived tasks, and recalculates the priority of each inference task based on the changes in confidence level, task progress and urgency. Based on the updated priority, the system adjusts the task execution sequence, obtains the available parallel slots and hardware resource status of the task scheduling system, and generates an updated task scheduling plan based on the new execution sequence and resource status. The execution engine dynamically adjusts task execution according to the updated plan. Obtain user feedback data, update inference model parameters based on user feedback, update relevant rules in the policy rule base, perform rule addition, replacement or revocation operations, perform consistency verification on the updated rule set, write the verified model parameters and policy rules to persistent storage, and output the updated inference model parameters and policy rules.

[0013] In a preferred embodiment of the present invention, step five, which involves responding to a user's natural language question, maintaining the multi-turn dialogue context, initiating the semantic understanding and reasoning process, returning the result, and recording user feedback information, includes: Receive user natural language questions, obtain the existing dialogue state, merge the semantic parsing results of the current question into the current slot, update the dialogue state, transform the updated dialogue state into a formal query expression, obtain entity attributes, domain rules and system dynamic state from the knowledge base, perform knowledge retrieval based on formal query, generate a matching fact set, apply reasoning rules to deduce conclusions from premises, continuously expand until no new propositions are generated, and generate a reasoning result set. The reasoning results are organized into structured conclusions, the structured conclusions are converted into natural language response text, response information is returned to the user, the user's feedback behavior after this interaction is obtained, the feedback information is written into the user behavior log, and the interaction completion indicator is output.

[0014] Compared with the prior art, the advantages of this invention are: (1) In this invention, user questions and multi-turn contexts are integrated to maintain interactive coherence, real-time working conditions and abnormal events are combined to enhance on-site perception, cross-modal alignment is used to improve semantic consistency, knowledge graphs are used to match entities and identify multi-hop paths to mine deep intentions, spatiotemporal and equipment information is associated to ensure accurate understanding, knowledge is used to complete missing semantics to deal with fuzzy input, and rule constraints and repair mechanisms are used to ensure the legality and completeness of structured output. Finally, a reliable conversion from natural language to executable queries is achieved, which effectively improves the accuracy and robustness of industrial intelligent question answering. (2) In this invention, by integrating structured problems, knowledge graphs and real-time working conditions, a comprehensive perception of industrial scenarios is achieved, operational dependencies are constructed to ensure that reasoning conforms to industrial logic, confidence weights are dynamically allocated to improve the reliability of information fusion, candidate strategies are generated and the benefits and costs are weighed, the optimal solution under resource constraints is output, parallel execution improves reasoning efficiency, dynamic scheduling ensures priority response to key tasks, and the model and rules are continuously optimized in combination with user feedback to maintain knowledge consistency. Finally, an adaptive reasoning mechanism from strategy generation, dynamic execution to closed-loop optimization is constructed, which effectively improves the decision-making accuracy and response flexibility of the agent in complex industrial environments. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the steps of the knowledge base construction and intelligent agent adaptive question answering method in this invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0017] Example 1: As Figure 1 As shown, the adaptive question-answering method for knowledge base construction and intelligent agents in industrial manufacturing proposed in this invention includes the following steps: Step 1: Collect multi-source observation data during equipment operation, process execution, environmental monitoring, quality inspection, and production management, including equipment temperature, vibration frequency, processing time, process cutting force, ambient temperature, ambient relative humidity, product dimensional tolerance, and surface roughness. Clean the collected data, handle missing values, filter noise, standardize, and align the data with time. Generate structured abnormal events based on preset thresholds as context input for subsequent knowledge extraction and agent reasoning. Obtain a preset threshold table, in which each type of parameter corresponds to a normal range and a graded threshold. Traverse each aligned observation record, extract the timestamp, device identifier, parameter name, and measured value. Find the corresponding threshold interval based on the parameter name, determine the anomaly type and severity level corresponding to the interval to which the measured value belongs. If the measured value exceeds the normal range, generate an anomaly event. The anomaly event includes: timestamp, device identifier, anomaly type, and severity level. Mark the anomaly event as pending confirmation. Obtain the current equipment operating condition context, verify the validity of the anomaly event in combination with the operating condition context, update the status of the valid anomaly event to confirmed, and output all confirmed structured anomaly events. By covering multi-source data from equipment, processes, environment, quality, and management, and through standardized preprocessing, comprehensive governance of observational data is achieved. An interval matching mechanism based on a preset threshold table enables rule-based judgment and classification of anomalies. Operating condition context verification and status confirmation mechanisms improve the accuracy and reliability of abnormal events. By generating structured events containing time, equipment, anomaly type, and severity level, the output format is unified, semantically clear, and directly usable for downstream knowledge extraction and reasoning. Through a model-free, configurable threshold, and modular process design, the method achieves high maintainability, scalability, and industrial scenario adaptability.

[0018] Step 2: Based on maintenance records, process specifications, and FMEA documents, extract entities and relationships of equipment, processes, and failure modes through natural language processing, construct a knowledge graph, and perform incremental updates based on the frequency and pattern of abnormal events; Step two involves extracting entities and relationships related to equipment, processes, and failure modes using natural language processing based on maintenance records, process specifications, and FMEA documents. This process constructs a knowledge graph and performs incremental updates based on the frequency and patterns of abnormal events. Entities and relationships of equipment, processes, and failure modes are extracted from maintenance records, process specifications, and FMEA documents to construct a knowledge graph. Incremental updates are performed based on the frequency and pattern changes of abnormal events. Named entity recognition is performed to extract entities of equipment, components, processes, failure modes, detection indicators, and operators, generating entity sets. Dependency parsing is performed to identify semantic relationships between entities, with relationship types limited to belonging to, located at, causing, detected, affecting, and composed of, generating sets of triples. Entities are used as nodes, and triples are used as edges to establish a graph structure. Different representations of the same entity are normalized to ensure that each entity is unique in the graph. The graph structure is stored in a graph database to complete the initialization of the knowledge graph and maintain historical event frequency records for each entity in the knowledge graph. The absolute value of the difference between the frequency of events occurring in the current time window and the historical average is taken. If the absolute value is greater than the product of the historical standard deviation and the fixed sensitivity coefficient, the update mechanism is triggered. The newly occurring abnormal event sequence is scanned, and entity combinations not recorded in the graph are identified. If the same entity combination appears three times in a row, the corresponding relation edge is added. All relations in the knowledge graph are traversed, and for each relation, its current co-occurrence frequency is compared with the historical highest frequency. If the current frequency is higher than the historical highest frequency, its weight value is increased by one unit. If the associated path forms the same sequence three times in a row in time sequence, the path record is added, the attribute information of the affected entity is updated, and the changes are written to the graph database to complete the incremental update of the knowledge graph. By extracting entities and relationships from maintenance records, process specifications, and FMEA documents, unstructured knowledge is transformed into structured knowledge. Named entity recognition, dependency parsing, and entity normalization ensure the accuracy and uniqueness of knowledge representation. Graph database storage supports efficient querying and relational reasoning. Frequency change detection and continuous pattern recognition trigger incremental updates. Adding relation edges, increasing weights, and recording high-frequency paths continuously evolve the knowledge graph. Avoiding full reconstruction improves update efficiency, achieving a fusion of static knowledge construction and dynamic experience accumulation, and enhancing adaptability and application value in industrial scenarios.

[0019] Step 3: Based on user questions and multi-turn dialogue context, integrate real-time operating data and structured abnormal events, combine entities and relationships in the knowledge graph to perform semantic understanding and intent recognition, complete and reconstruct ambiguous or incomplete questions, and output structured question representations; Step three involves fusing user questions and multi-turn dialogue contexts, as well as integrating real-time operational data and structured anomaly events. Obtain the semantic representation of the user's current question, obtain the semantic representation of each statement in the multi-turn dialogue history, and calculate the median value of the association strength between the question and the historical statements using a formula: ,in Score the attention paid to the historical statements in the j-th round by the i-th round of questions. Let be the query vector for the i-th round of questions. Let j be the key vector of the historical statements in the j-th round. Let be the dimension of the key vector. This represents the difference between the current round and the previous rounds. For relative position embedding functions, output a vector of preset dimensions. As a learnable weight vector, the calculation result reflects the combined strength of semantic matching and positional relationship. The calculation result is multiplied by the time decay factor, which decays exponentially according to the dialogue round interval. The product result is normalized to generate the weighting coefficient of the historical statements. The semantic representation of each historical statement is multiplied by the corresponding weighting coefficient and summed to obtain the historical context representation. The semantic representation of the current question is merged with the historical context representation to generate the dialogue context fusion result. Semantic representations of real-time operating conditions and structured abnormal events are obtained. The dialogue context fusion results, real-time operating condition semantic representations, and structured abnormal event semantic representations are mapped to a unified semantic space to generate modal gating vectors. The modal gating vectors are obtained by nonlinear transformation of the joint features of the three inputs. The cross-modal fusion representation is generated by the following formula: ,in Let be the gate vector for the m-th mode. Values ​​can be text, sensor data, or events. These are the embedding vectors for the three modalities, The weight matrix is ​​a learnable matrix. For bias vectors, For the Sigmoid function, The fused vector A projection matrix specifically for modal analysis. This is element-wise multiplication; Using the dialogue context fusion result as the query, and the real-time working condition semantic representation and the structured abnormal event semantic representation as the key and value, a cross-modal attention operation is performed. The cross-modal attention output is added to the dialogue context fusion result to generate an enhanced context representation. The enhanced context representation is then output as the input for subsequent reasoning tasks. Step three, which combines entities and relationships in the knowledge graph for semantic understanding and intent recognition, includes: Get the text content of the user's current question, match semantically related candidate entities in the knowledge graph based on the question content, obtain the contextual semantic representation of the multi-turn dialogue, align the contextual semantic representation with the candidate entities, confirm the entity association, obtain the entity relationship structure in the knowledge graph, and extract the directly associated adjacency subgraph with the candidate entity as the center. Identify multi-hop connection paths between entities in the subgraph to form a set of path patterns. Fuse the path patterns with the contextual semantics to generate a joint semantic representation. Determine the intent category corresponding to the question based on the matching strength between the path structure and entity relationships in the joint semantic representation. The system retrieves the time information mentioned in the question and associates it with the time entity in the knowledge graph; retrieves the geographical location information mentioned in the question and associates it with the location entity in the knowledge graph; retrieves the device identification information mentioned in the question and associates it with the device entity in the knowledge graph; and generates a structured semantic understanding result containing intent category, time information, geographical location information, device identification information, and associated path. Step three involves completing and reconstructing the fuzzy or incomplete problem and outputting a structured problem representation. The system obtains the text content of the user's current question, the contextual semantic representation of the multi-turn dialogue, the entity set and relation set in the knowledge graph, and the initial structured representation corresponding to the user's question. The initial structured representation includes the fields of subject, action, object, condition, and time. Based on the contextual semantic representation and the entity relations in the knowledge graph, the system identifies the missing semantic components in the question and uses the association information in the knowledge graph to complete the missing fields in the initial structured representation. The completed structured representation is matched against preset syntax and semantic rules to determine whether it satisfies all constraints. For structured representations that do not meet the constraints, a repair operation is performed. The repair process aims to minimize the amount of modification and constraint conflicts. The repair objective is defined by the following formula: ,in This is the initial structured representation. This is the restored representation. To be a set of legal representations that satisfy syntactic and semantic constraints, For distance L1, for The number of constraints violated To constrain the penalty coefficient, a preset constant is used. Repair operations include entity replacement, relation adjustment, and slot completion. The repaired representation must belong to the set. The output is a repaired structured problem representation, which includes the fields of subject, action, object, condition, and time. By integrating user questions with multi-turn dialogue context, dynamic tracking of interaction history is achieved; by introducing real-time operating data and structured abnormal events, the on-site perception capability of semantic understanding is enhanced; through cross-modal alignment and attention mechanisms, the semantic space of text, sensors, and events is unified, improving the accuracy of context representation; by matching candidate entities with knowledge graphs and extracting adjacency subgraphs, and using multi-hop paths to identify potential intentions, deep semantic parsing based on domain knowledge is achieved; by associating information such as time, location, and equipment with graph entities, the spatiotemporal consistency of semantic understanding results is ensured; by identifying semantic gaps and combining them with knowledge graphs to complete the question structure, ambiguous or incomplete inputs are effectively addressed; through rule constraints and repair mechanisms, the grammatical validity and logical integrity of the output structured question representation are guaranteed; and overall, a reliable conversion from natural language to machine-executable queries is achieved, significantly improving the robustness and accuracy of intelligent question answering in industrial scenarios.

[0020] Example 2: The technical solution of this embodiment of the invention differs from that of Example 1 in that: like Figure 1 As shown, step four: Based on the structured question representation generated by semantic understanding, combined with knowledge graph and real-time working condition data, an initial reasoning strategy is generated, multi-task parallel reasoning is executed, the reasoning path and task priority are dynamically adjusted, and the reasoning model parameters and strategy rules are updated using user feedback. Step four involves generating an initial reasoning strategy from a structured question representation based on semantic understanding, combined with a knowledge graph and real-time operational data. The process involves obtaining a structured question representation from the semantic understanding stage, acquiring device functions and operating specifications from the knowledge graph, determining operational dependencies between entities based on the entities and operating specifications in the structured question representation, acquiring sensor readings and system status flags from real-time operating data, and assigning dynamic confidence weights to each entity in the structured question representation using the real-time operating data. The assignment process is defined by the following formula: ,in Let represent the confidence level of the i-th entity at the current moment. It is the knowledge graph entity identifier corresponding to the i-th entity. It is a knowledge source reliability score, calculated by weighting the source's authority, update frequency, and historical accuracy with fixed weights. It is the real-time observation associated with the i-th entity. It is a data source quality score, calculated by weighting sensor health status, signal-to-noise ratio, sampling integrity, and timestamp validity with fixed weights. These represent the weight coefficients of the knowledge source and the data source, respectively. It is the Sigmoid function, with an output range of (0,1). The confidence score is used for information weighting in subsequent inference processes. Based on operational dependencies and dynamic confidence weights, a set of candidate inference strategies is constructed. Each strategy defines a mapping from the current system state to an action sequence. The expected benefit and execution cost of each candidate strategy under the current state distribution are evaluated to generate an initial set of inference strategies. The generation process is defined by the following optimization objective: ,in This represents the initial set of inference strategies generated. Represents the set of all candidate strategies. It is a subset selected from it. It is a strategy In state The following revenue estimates It is a strategy Cost estimation, expected value Indicates distribution in the current context The average value below, The upper limit of the total system resources consists of the number of parallel slots, the maximum computation time, and the maximum memory quota. The output generated result is used as the initial inference strategy. Step four involves performing multi-task parallel inference, dynamically adjusting the inference path and task priority, and updating the inference model parameters and policy rules based on user feedback. The system obtains the execution status of the current inference task, the latest confidence level of each entity and the urgency information of newly arrived tasks, and recalculates the priority of each inference task based on the changes in confidence level, task progress and urgency. Based on the updated priority, the system adjusts the task execution sequence to ensure that the dependencies between tasks remain unchanged. The system also obtains the available parallel slots and hardware resource status of the task scheduling system, and generates an updated task scheduling plan based on the new execution sequence and resource status. The execution engine dynamically adjusts task execution according to the updated plan. Obtain user feedback data, update inference model parameters based on user feedback, improve the selection tendency of high-return actions, update relevant rules in the policy rule base, perform rule addition, replacement or revocation operations, perform consistency verification on the updated rule set to ensure that the syntax is legal, the logic is consistent and there are no conflicts, write the verified model parameters and policy rules to persistent storage, and output the updated inference model parameters and policy rules. By combining structured problem representation, knowledge graphs, and real-time operational data, multi-source fusion and context awareness of inference input are achieved. Operational dependencies are constructed to clarify execution order constraints between entities, ensuring that inference logic conforms to industrial process specifications. Dynamic confidence weight allocation integrates the reliability of knowledge sources and data quality, enabling quantifiable evaluation of information source credibility and improving inference robustness. Candidate strategy set generation and benefit-cost evaluation generate an optimal strategy subset under resource constraints, ensuring the feasibility and efficiency of the inference scheme. A multi-task parallel execution mechanism enhances inference throughput. Real-time monitoring of task status, confidence changes, and task urgency dynamically adjusts task priorities and scheduling plans, achieving optimal resource utilization while ensuring dependencies. User feedback drives model parameter updates and rule base evolution, supporting continuous strategy optimization. Rule addition, replacement, revocation, and consistency verification ensure logical self-consistency and conflict-free knowledge base logic. Overall, this achieves end-to-end adaptive inference from static strategy generation to dynamic execution control and feedback loop optimization, significantly enhancing the agent's decision-making ability and response flexibility in complex industrial scenarios.

[0021] Step 5: Respond to the user's natural language question, maintain the multi-turn dialogue context, initiate the semantic understanding and reasoning process, return the result and record the user feedback information; Step five, which involves responding to user natural language questions, maintaining the multi-turn dialogue context, initiating the semantic understanding and reasoning process, returning results, and recording user feedback, includes: The system receives natural language questions from users, obtains the existing dialogue state, including the current intent, the current slot, and the historical interaction sequence, merges the semantic parsing results of the current question into the current slot, updates the dialogue state, transforms the updated dialogue state into a formal query expression, obtains entity attributes, domain rules, and system dynamic states from the knowledge base, performs knowledge retrieval based on the formal query, generates a matching fact set, applies reasoning rules to deduce conclusions from the premises, and continues to expand until no new propositions are generated, generating a reasoning result set. The reasoning results are organized into structured conclusions, the structured conclusions are converted into natural language response text, response information is returned to the user, the user's feedback behavior after this interaction is obtained, including operation continuity, rating and comment content, the feedback information is written into the user behavior log, and the interaction completion indicator is output. By responding to user questions and updating the dialogue status in real time, the system maintains the coherence of multiple rounds of context. By generating formal queries and combining the knowledge base and system status to perform retrieval and rule-based reasoning, the system ensures accurate and complete conclusions. By organizing structured results and generating natural language, the system achieves readable response output. By collecting and persisting user feedback, the system supports subsequent optimization. The overall process is complete and logically clear, ensuring the continuity, accuracy, and adaptability of industrial intelligent question answering.

[0022] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and its improved concept, should be covered within the scope of protection of the present invention.

Claims

1. An adaptive question-answering method for knowledge base construction and intelligent agents in industrial manufacturing, characterized in that, Includes the following steps: Step 1: Collect multi-source observation data during equipment operation, process execution, environmental monitoring, quality inspection, and production management. Clean the collected data, handle missing values, filter noise, standardize, and align the data over time. Generate structured abnormal events based on preset thresholds. Step 2: Based on maintenance records, process specifications, and FMEA documents, extract entities and relationships of equipment, processes, and failure modes through natural language processing, construct a knowledge graph, and perform incremental updates based on the frequency and pattern of abnormal events; Step 3: Based on user questions and multi-turn dialogue context, integrate real-time operating data and structured abnormal events, combine entities and relationships in the knowledge graph to perform semantic understanding and intent recognition, complete and reconstruct ambiguous or incomplete questions, and output structured question representations; Step 4: Based on the structured question representation generated by semantic understanding, combined with knowledge graph and real-time working condition data, generate an initial reasoning strategy, execute multi-task parallel reasoning, dynamically adjust the reasoning path and task priority, and update the reasoning model parameters and strategy rules using user feedback. Step 5: Respond to the user's natural language question, maintain the multi-turn dialogue context, initiate the semantic understanding and reasoning process, return the results, and record the user feedback information.

2. The adaptive question-answering method for knowledge base construction and intelligent agents in industrial manufacturing according to claim 1, characterized in that, Step two, based on maintenance records, process specifications, and FMEA documents, involves extracting entities and relationships related to equipment, processes, and failure modes through natural language processing to construct a knowledge graph. The process of incrementally updating this knowledge graph based on the frequency and patterns of abnormal events includes: Entities and relationships of equipment, processes, and failure modes are extracted from maintenance records, process specifications, and FMEA documents to construct a knowledge graph. Incremental updates are performed based on the frequency and pattern changes of abnormal events. Named entity recognition is performed to extract entities of equipment, components, processes, failure modes, detection indicators, and operators, generating entity sets. Dependency parsing is performed to identify semantic relationships between entities, generating sets of triples. Entities are used as nodes and triples as edges to establish a graph structure. Different representations of the same entity are normalized. The graph structure is stored in a graph database, completing the initialization of the knowledge graph. The absolute value of the difference between the frequency of events occurring in the current time window and the historical average is taken. If the absolute value is greater than the product of the historical standard deviation and the fixed sensitivity coefficient, the update mechanism is triggered. The newly occurring abnormal event sequence is scanned, entity combinations not recorded in the graph are identified, the attribute information of the affected entities is updated, and the changes are written to the graph database to complete the incremental update of the knowledge graph.

3. The adaptive question-answering method for knowledge base construction and intelligent agents in industrial manufacturing according to claim 1, characterized in that, The process of fusing user questions and multi-turn dialogue context, and integrating real-time operating data with structured abnormal events in step three includes: The system obtains the semantic representation of the user's current question and the semantic representation of each statement in the multi-turn dialogue history. It calculates the median value of the association strength between the question and the historical statements using a formula. The calculation result reflects the comprehensive strength of semantic matching and positional relationship. The calculation result is multiplied by the time decay factor, and the product result is normalized to generate the weighting coefficient of the historical statements. The semantic representation of each historical statement is multiplied by the corresponding weighting coefficient and summed to obtain the historical context representation. The semantic representation of the current question is merged with the historical context representation to generate the dialogue context fusion result. Obtain semantic representations of real-time operating condition data and structured abnormal events. Map the dialogue context fusion results, real-time operating condition semantic representations, and structured abnormal event semantic representations to a unified semantic space to generate modal gating vectors. Using the dialogue context fusion result as the query, and the real-time working condition semantic representation and the structured abnormal event semantic representation as the key and value, perform cross-modal attention operation, add the cross-modal attention output to the dialogue context fusion result, generate an enhanced context representation, and output the enhanced context representation.

4. The adaptive question-answering method for knowledge base construction and intelligent agents in industrial manufacturing according to claim 3, characterized in that, Step three, which involves semantic understanding and intent recognition by combining entities and relationships in the knowledge graph, includes: Get the text content of the user's current question, match semantically related candidate entities in the knowledge graph based on the question content, obtain the contextual semantic representation of the multi-turn dialogue, align the contextual semantic representation with the candidate entities, confirm the entity association, obtain the entity relationship structure in the knowledge graph, and extract the directly associated adjacency subgraph with the candidate entity as the center. Identify multi-hop connection paths between entities in the subgraph to form a set of path patterns. Fuse the path patterns with the contextual semantics to generate a joint semantic representation. Determine the intent category corresponding to the question based on the matching strength between the path structure and entity relationships in the joint semantic representation. The system retrieves the time information mentioned in the question and associates it with the time entity in the knowledge graph; it retrieves the geographical location information mentioned in the question and associates it with the location entity in the knowledge graph; it retrieves the device identification information mentioned in the question and associates it with the device entity in the knowledge graph; and generates a structured semantic understanding result that includes intent category, time information, geographical location information, device identification information, and associated paths.

5. The adaptive question-answering method for knowledge base construction and intelligent agents in industrial manufacturing according to claim 4, characterized in that, The process of completing and reconstructing the fuzzy or incomplete problem and outputting a structured problem representation in step three includes: The system obtains the text content of the user's current question, the contextual semantic representation of the multi-turn dialogue, the entity set and relation set in the knowledge graph, the initial structured representation of the user's question, and based on the contextual semantic representation and the entity relations in the knowledge graph, identifies the missing semantic components in the question and uses the association information in the knowledge graph to complete the missing fields in the initial structured representation. The completed structured representation is matched with the preset syntax and semantic rules to determine whether the completed structured representation meets all constraints. For structured representations that do not meet the constraints, a repair operation is performed. The repair process aims to minimize the scope of modification and constraint conflicts, and outputs the repaired structured problem representation.

6. The adaptive question-answering method for knowledge base construction and intelligent agents in industrial manufacturing according to claim 1, characterized in that, The process of generating an initial reasoning strategy in step four, based on the structured question representation generated by semantic understanding and combined with knowledge graphs and real-time operational data, includes: Obtain the structured question representation output from the semantic understanding stage, obtain the device functions and operation specifications from the knowledge graph, determine the operation dependencies between entities based on the entities and operation specifications in the structured question representation, obtain sensor readings and system status flags from real-time operating data, and assign dynamic confidence weights to each entity in the structured question representation in combination with real-time operating data. Based on operation dependencies and dynamic confidence weights, a set of candidate inference strategies is constructed. Each strategy defines a mapping from the current system state to the action sequence. The expected benefits and execution costs of each candidate strategy under the current state distribution are evaluated, an initial set of inference strategies is generated, and the generated results are output as the initial inference strategy.

7. The adaptive question-answering method for knowledge base construction and intelligent agents in industrial manufacturing according to claim 6, characterized in that, Step four involves performing multi-task parallel inference, dynamically adjusting the inference path and task priority, and updating the inference model parameters and policy rules using user feedback. The system obtains the execution status of the current inference task, the latest confidence level of each entity and the urgency information of newly arrived tasks, and recalculates the priority of each inference task based on the changes in confidence level, task progress and urgency. Based on the updated priority, the system adjusts the task execution sequence, obtains the available parallel slots and hardware resource status of the task scheduling system, and generates an updated task scheduling plan based on the new execution sequence and resource status. The execution engine dynamically adjusts task execution according to the updated plan. Obtain user feedback data, update inference model parameters based on user feedback, update relevant rules in the policy rule base, perform rule addition, replacement or revocation operations, perform consistency verification on the updated rule set, write the verified model parameters and policy rules to persistent storage, and output the updated inference model parameters and policy rules.

8. The adaptive question-answering method for knowledge base construction and intelligent agents in industrial manufacturing according to claim 1, characterized in that, Step five, which involves responding to user natural language questions, maintaining the multi-turn dialogue context, initiating the semantic understanding and reasoning process, returning results, and recording user feedback, includes the following: Receive user natural language questions, obtain the existing dialogue state, merge the semantic parsing results of the current question into the current slot, update the dialogue state, transform the updated dialogue state into a formal query expression, obtain entity attributes, domain rules and system dynamic state from the knowledge base, perform knowledge retrieval based on formal query, generate a matching fact set, apply reasoning rules to deduce conclusions from premises, continuously expand until no new propositions are generated, and generate a reasoning result set. The reasoning results are organized into structured conclusions, the structured conclusions are converted into natural language response text, response information is returned to the user, the user's feedback behavior after this interaction is obtained, the feedback information is written into the user behavior log, and the interaction completion indicator is output.

Images