Method for intelligent operation of industry based on knowledge evolution of large model and related equipment

Abstract

This application provides a method and related equipment for industrial intelligence operation driven by a large-scale model based on knowledge evolution. In this method, the intelligent processing system first uses a large language model to perform semantic parsing of the target question, extracting object entities, question intent, and question type, and establishing a connection between the question and the domain knowledge system. Then, the intelligent processing system calculates the similarity between the semantic feature vector and historical question features. When the similarity does not reach a threshold, it retrieves multi-dimensional related knowledge elements from the knowledge graph, and then obtains the knowledge completeness based on these elements, clarifying the current knowledge's support for the question answer. Next, based on the knowledge completeness, it calls the large language model to generate the response text. Finally, the intelligent processing system updates the knowledge graph and the knowledge completeness result database, and synchronizes them to multiple system nodes to achieve collaborative knowledge sharing. This method alleviates the technical problem of one-sided and distorted content in industrial intelligent responses and improves the accuracy of industrial intelligent question answering.

Description

Technical Field

[0001] This application relates to the field of knowledge graph technology, and in particular to a method and related equipment for driving intelligent industrial operation based on a large model of knowledge evolution. Background Technology

[0002] With the rapid development of artificial intelligence technology, the intelligent transformation of industries has become a key engine for promoting high-quality economic development. In this process, how to efficiently utilize massive amounts of multi-source heterogeneous data (such as enterprise operational data, industry reports, and scientific research results) to assist decision-making has become a focus for major enterprises and research institutions. In practical industrial application scenarios, users often need to obtain accurate and in-depth answers to specific problems. This requires the system not only to understand the problem but also to provide reliable support based on a vast industry knowledge system.

[0003] In related technologies, industrial intelligent question-answering systems primarily rely on traditional retrieval-based question-answering architectures or retrieval-enhanced generation techniques. In practice, these systems first slice and index massive amounts of industry documents, manuals, and reports to build a large-scale vector database. When a user asks a question, the system calculates the similarity between the question and document slices, retrieving several text paragraphs that most closely match the literal meaning of the question. Subsequently, the system directly uses these retrieved text paragraphs as background knowledge, concatenating them into the input prompt box of the large model, instructing the model to generate the final answer based on these fragments.

[0004] However, when dealing with highly specialized and logically interconnected industry-level problems, the system often fails to grasp the deep logical connections and overall coherence between multiple data sources because it only extracts information fragments based on literal or semantic similarity. This leads to situations where a user's question requires comprehensive information from multiple dimensions, or where the data itself is fragmented. The system often resorts to mechanically piecing together fragments of information, resulting in responses that, while grammatically correct, are often disjointed, one-sided, or even taken out of context, leading to a lack of rigor in the user's answers. Summary of the Invention

[0005] This application provides a method and related equipment for driving industrial intelligence operation based on a large model of knowledge evolution, which can alleviate the technical problem of one-sided and distorted response content in existing industrial intelligence question answering systems and improve the accuracy of industrial intelligence question answering.

[0006] Firstly, this application provides a method for driving industrial intelligence operation based on a large model of knowledge evolution, comprising: receiving a target problem submitted by a user; performing semantic parsing on the target problem using a large language model to obtain the object entities, problem intent, and problem type involved in the target problem; mapping the object entities to corresponding ontology nodes in a domain ontology library to generate a semantic feature vector containing a combination of ontology nodes, problem intent, and problem type; calculating the feature similarity between the semantic feature vector and the features of historical problems in a knowledge completeness result library, wherein the knowledge completeness result library is used to store the semantic feature vectors of historical problems and their corresponding knowledge completeness; when the feature similarity does not reach a preset similarity threshold, retrieving multidimensional related knowledge elements associated with the target problem from a knowledge graph, wherein the multidimensional related knowledge elements are... The knowledge elements include at least triples, attributes, and historical question-and-answer records. Based on the multi-dimensional related knowledge elements associated with the target question, a pre-defined knowledge completeness measurement model is used to calculate the data coverage of the target question in multiple dimensions to obtain the current knowledge completeness level, which includes incomplete, partially complete, and highly complete levels. Based on the knowledge completeness level, a large language model is invoked to generate response text corresponding to the target question. Based on the response text, the knowledge graph is updated, and the knowledge completeness result database is updated based on the semantic feature vector of the target question, the knowledge completeness level, and the response text. Data changes in the knowledge completeness result database are monitored, and newly added or changed knowledge completeness result data is sent to multiple system nodes within the communication range, including enterprise nodes and research institution nodes.

[0007] By adopting the above technical solution, the intelligent processing system first uses a large language model to perform semantic parsing on the target question, extracting object entities, question intent, and question type. Then, it maps the object entities to a domain ontology library to generate semantic feature vectors, establishing a connection between the question and the domain knowledge system. Next, the intelligent processing system calculates the similarity between the semantic feature vectors and historical question features to determine whether historical knowledge resources can be reused. When the similarity does not reach a threshold, the intelligent processing system retrieves multi-dimensional related knowledge elements from the knowledge graph, achieving the convergence of multi-source knowledge. Based on these multi-dimensional related knowledge elements, it calculates multi-dimensional data coverage to obtain the knowledge completeness, clarifying the current knowledge's support capability for question answering. Then, based on the knowledge completeness, it calls the large language model to generate the response text. Finally, the intelligent processing system updates the knowledge graph and knowledge completeness result library based on the response text, synchronizing newly added or changed knowledge completeness result data to multiple system nodes to achieve collaborative knowledge sharing and improve the resource utilization rate of the entire industrial intelligent processing system. This method alleviates the technical problem of one-sided and distorted content in industrial intelligent responses and improves the accuracy of industrial intelligent question answering. In conjunction with some embodiments of the first aspect, in some embodiments, after calculating the feature similarity between the semantic feature vector and the historical records in the knowledge completeness result library, the method further includes: when the feature similarity reaches a preset similarity threshold, determining that there is a historical problem in the knowledge completeness result library that belongs to the same knowledge category as the target problem, and obtaining the knowledge completeness of the historical problem; if the knowledge completeness of the historical problem is highly complete, then calling the existing knowledge in the knowledge graph, using the large language model to generate the response text, and performing an incremental update operation; if the knowledge completeness of the historical problem is partially complete or incomplete, then generating a first prompt message, the first prompt message being used to indicate that there is a gap in the current domain knowledge.

[0008] By adopting the above technical solution, the intelligent processing system first determines the existence of historical questions belonging to the same knowledge category when the feature similarity reaches a preset threshold. Then, it adopts differentiated processing based on the different levels of knowledge completeness of the historical questions. When the historical knowledge is highly complete, the intelligent processing system calls upon existing knowledge to generate a response and performs incremental updates, reducing repetitive knowledge retrieval and completeness calculation processes, improving processing efficiency while ensuring response reliability. When the historical knowledge is partially complete or incomplete, it generates knowledge gap prompts to avoid generating one-sided responses based on insufficient knowledge. This method mitigates the response risk when historical knowledge is insufficient and further improves the logic of knowledge reuse and problem handling.

[0009] In conjunction with some embodiments of the first aspect, in some embodiments, based on the multi-dimensional related knowledge elements associated with the target problem, a preset knowledge completeness measurement model is used to calculate the data coverage of the target problem in multiple dimensions to obtain the current knowledge completeness. Specifically, this includes: based on the multi-dimensional related knowledge elements associated with the target problem, using a preset knowledge completeness measurement model, calculating the data coverage of the target problem in multiple dimensions, where the multiple dimensions include at least performance dimension, operating condition dimension, mechanism dimension, and solution dimension. The performance dimension and operating condition dimension are basic dimensions, while the mechanism dimension and solution dimension are extended dimensions. If the coverage of the basic dimensions does not reach a first preset coverage threshold, the knowledge completeness of the target problem is determined to be incomplete. If the coverage of the basic dimensions reaches the first preset coverage threshold and the coverage of the extended dimensions does not reach a second preset coverage threshold, the knowledge completeness of the target problem is determined to be partially complete. If the coverage of the basic dimensions reaches the first preset coverage threshold and the coverage of the extended dimensions reaches the second preset coverage threshold, the knowledge completeness of the target problem is determined to be highly complete.

[0010] By adopting the above technical solution, the intelligent processing system first divides the dimensions of knowledge completeness measurement into basic dimensions and extended dimensions. The basic dimensions ensure the basic validity of the solution, while the extended dimensions enhance the depth and comprehensiveness of the solution. Then, the intelligent processing system uses a preset knowledge completeness measurement model to calculate the data coverage of each dimension and determines the degree of knowledge completeness (incomplete, partially complete, highly complete) through a dual-threshold classification, achieving a quantitative assessment of the knowledge completeness status. This method, by combining different dimensional divisions with classification criteria, enables the intelligent processing system to clearly define the support boundaries of current knowledge resources for the target problem, improving the reliability of knowledge completeness assessment.

[0011] In conjunction with some embodiments of the first aspect, in some embodiments, based on the knowledge completeness, a large language model is invoked to generate a response text corresponding to the target question. Specifically, this includes: when the knowledge completeness is incomplete, the large language model is invoked to generate an initial response text containing second prompt information based on existing knowledge. The second prompt information is used to indicate that the data coverage of the basic dimension has not reached a first preset coverage threshold; a knowledge gap task is generated based on the dimensions contained in the second prompt information; when the knowledge completeness is partially complete, the large language model is invoked to generate an inference chain and answer content based on known relationships in the knowledge graph, and the confidence of each part in the inference chain is calculated; confidence scores below a preset confidence threshold are marked at the corresponding positions in the inference chain, and combined with the answer content, an initial response text is generated; a supplementary data task is generated based on the extended dimensions whose data coverage has not reached a second preset coverage threshold; when the knowledge completeness is incomplete or partially complete, the basic... Based on the question intent of the target problem and the corresponding initial response text, a knowledge construction task following the DIKW principle is generated. An automated workflow construction suggestion based on the construction task results is generated using a large language model. This automated workflow construction suggestion is added to the corresponding initial response text to obtain the final response text. When the knowledge completeness is highly complete, the large language model is invoked, and based on the knowledge graph, an initial response text is generated. This initial response text includes the reasoning path and text, with the text including tags indicating the knowledge source. The reasoning path is stored as a complete knowledge path in the knowledge graph. Simultaneously, based on the question intent of the target problem and the initial response text, a first automated workflow matching the question intent and initial response text is retrieved from the workflow repository, or a second automated workflow is generated using the large language model according to the DIKW principle. The first or second automated workflow is added to the initial response text to obtain the final response text.

[0012] By adopting the above technical solution, when knowledge is incomplete, the intelligent processing system generates an initial response with basic dimension gap hints and creates a knowledge gap task to clarify the direction of knowledge supplementation. When knowledge is partially complete, the intelligent processing system generates an inference chain and response content with low confidence labels to indicate the uncertainty of the response, and simultaneously creates an extended dimension data supplementation task. Then, in scenarios with incomplete or partially complete knowledge, knowledge construction tasks and automated workflow construction suggestions are generated based on the DIKW principle and integrated into the final response. When knowledge is highly complete, the intelligent processing system generates an initial response with inference paths and knowledge source tags to ensure the traceability and credibility of the response, and simultaneously retrieves or generates automated workflows to integrate into the response. This method ensures the relevance and rationality of the response content under different knowledge completeness states, and improves the practicality and guidance of the response text.

[0013] In conjunction with some embodiments of the first aspect, in some embodiments, calculating the feature similarity between the semantic feature vector and the features of historical questions in the knowledge completeness result base specifically includes: obtaining the historical ontology node combinations, historical question intents, and historical question types of historical questions in the knowledge completeness result base; calculating the overlap between the ontology node combinations in the semantic feature vector and the historical ontology node combinations, the semantic similarity between the question intent and the historical question intent, and the consistency between the question type and the historical question type, respectively; and performing a weighted calculation on the overlap, semantic similarity, and consistency to obtain the feature similarity between the semantic feature vector and the features of historical questions.

[0014] By adopting the above technical solution, the intelligent processing system first decomposes the core components of the semantic feature vector (ontology node combination, question intent, and question type), and calculates the matching indicators (overlap, semantic similarity, and consistency) between each component and the corresponding components of historical questions. The overlap of ontology node combinations reflects the degree of association with the domain knowledge involved in the question; the semantic similarity of the question intent reflects the degree of alignment with the user's core needs; and the consistency of the question type reflects the degree of matching of the question form. Then, the intelligent processing system integrates the matching indicators of each dimension through weighted calculation to obtain a comprehensive feature similarity. This method alleviates the one-sidedness of single-dimensional similarity evaluation and improves the accuracy of historical question matching.

[0015] In conjunction with some embodiments of the first aspect, in some embodiments, the knowledge graph is updated based on the response text, and the knowledge completeness result library is updated based on the semantic feature vector of the target question, the knowledge completeness, and the response text. Specifically, this includes: extracting newly generated triples, relational structures, and explanatory structures from the response text, and adding the triples, relational structures, and explanatory structures to the knowledge graph; establishing a mapping relationship between the semantic feature vector of the target question, the knowledge completeness, the gap dimension description information, and the timestamp; and storing the mapping relationship as a new knowledge completeness result in the knowledge completeness result library.

[0016] By adopting the above technical solution, the intelligent processing system first extracts newly generated triples, relational structures, and explanatory structures from the response text to update the knowledge graph, enriching its content dimensions and association strength. Then, the intelligent processing system establishes a mapping relationship between the semantic feature vector of the target question and the knowledge completeness, gap dimension description information, and timestamps, storing this as a new result in the knowledge completeness result repository. This method enables the knowledge graph and the knowledge completeness result repository to be dynamically updated as the problem is processed, enhancing the continuous reuse value of knowledge resources.

[0017] In conjunction with some embodiments of the first aspect, in some embodiments, monitoring data changes in the knowledge completeness result database and sending newly added or modified knowledge completeness result data to multiple system nodes within the communication range specifically includes: determining whether the knowledge completeness result database meets preset synchronization conditions, the preset synchronization conditions including the number of newly added highly complete record results exceeding a preset threshold or the time since the last synchronization exceeding a preset duration; if the knowledge completeness result database meets the preset synchronization conditions, extracting newly added and modified knowledge completeness result data; sending the newly added and modified knowledge completeness result data to the system nodes within the communication range, and receiving external completeness result data sent by the system nodes; detecting the external completeness result data, and when it is detected that the same problem is marked as highly complete in the external completeness result data but is recorded locally as incomplete or partially complete, updating the local knowledge completeness result database using the external completeness result data.

[0018] By adopting the above technical solution, the intelligent processing system first determines whether to trigger data synchronization based on preset synchronization conditions (threshold for the number of newly added highly complete records and synchronization time interval), avoiding resource consumption caused by frequent synchronization. Then, when the synchronization conditions are met, the intelligent processing system extracts the newly added and changed data and sends it to multiple system nodes, realizing cross-node sharing of knowledge resources. It then receives complete result data from external nodes, detects differences in the completeness of knowledge for the same problem, and updates local records using external highly complete data, achieving cross-node complementarity and improvement of knowledge resources. This method realizes the collaborative evolution of knowledge resources among multiple system nodes, improves the completeness and consistency of knowledge resources in the entire system cluster, and enhances the overall collaborative capability of the industrial intelligent processing system.

[0019] In a second aspect, this application provides an intelligent processing system, including one or more processors and a memory; the memory is coupled to the one or more processors, and the memory is used to store computer program code, the computer program code including computer instructions, which the one or more processors call to cause the intelligent processing system to perform the methods described in the first aspect and any possible implementation thereof.

[0020] Thirdly, this application provides a computer-readable storage medium including computer instructions that, when executed on an intelligent processing system, cause the intelligent processing system to perform the method described in the first aspect and any possible implementation thereof.

[0021] Fourthly, this application provides a computer program product, including a computer program / instruction that, when run on an intelligent processing system, causes the intelligent processing system to perform the method described in the first aspect and any possible implementation thereof.

[0022] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: 1. By employing techniques such as semantic parsing using a large language model, mapping domain ontology to generate semantic feature vectors, multi-dimensional knowledge retrieval from knowledge graphs, calculating multi-dimensional data coverage using a preset model to obtain knowledge completeness, generating responses based on the completeness and dynamically updating the knowledge graph and complete result database, the technical problem of one-sided and distorted content in industrial intelligent responses is alleviated, and the accuracy of industrial intelligent question answering is improved.

[0023] 2. By adopting technical means that divides the knowledge completeness measurement dimensions into basic dimensions and extended dimensions, calculates the data coverage of each dimension using a preset model, and judges the knowledge completeness level by dual thresholds, the technical problem of lack of basis for response generation is alleviated, and an accurate assessment of the knowledge completeness level is achieved.

[0024] 3. By adopting technical means such as generating responses based on the level of knowledge completeness, incorporating the DIKW principle to generate knowledge building tasks and automated workflow suggestions, the technical problem of lack of specificity in response generation under different knowledge completeness states is alleviated, and the response content is accurately adapted to different knowledge states. At the same time, it provides feasible workflow solutions for knowledge building and business implementation, and improves the practicality of the response content. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating a method for driving intelligent industrial operation based on a large model of knowledge evolution in an embodiment of this application. Figure 2 This is another flowchart illustrating the large-scale model-driven intelligent operation method based on knowledge evolution in the embodiments of this application; Figure 3 This is a logical schematic diagram of the knowledge evolution processing method in the embodiments of this application; Figure 4 This is a schematic diagram of the hardware structure of the intelligent processing system in the embodiments of this application. Detailed Implementation

[0026] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification and appended claims of this application, the singular expressions “a,” “an,” “the,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to any or all possible combinations including one or more of the listed items.

[0027] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0028] For ease of understanding, the method provided in this implementation is described in process below. Please refer to [link / reference]. Figure 1 This is a flowchart illustrating a method for driving industrial intelligent operation based on a large model of knowledge evolution in an embodiment of this application.

[0029] 101. Receive the target question submitted by the user, use the large language model to perform semantic parsing on the target question, and obtain the object entities involved in the target question, the question intent and the question type.

[0030] The target question refers to the query content submitted by the user to the intelligent processing system through text, voice, or other means; the large language model refers to a large-scale pre-trained language model trained based on deep learning technology, such as the GPT series, BERT series, or domain-customized industry large models; the object entity refers to the specific things, concepts, or technical objects involved in the target question; the question intent is used to indicate the user's fundamental purpose in raising the target question or the type of information they expect to obtain; the question type indicates the classification label of the target question based on the question's expression and answering method, with typical types including fault analysis, optimization suggestions, and mechanism explanation.

[0031] Specifically, after receiving the target question submitted by the user, the intelligent processing system first performs preprocessing operations on the input raw text. This includes removing special characters and noise data, standardizing the text encoding format, performing word segmentation and part-of-speech tagging, and recognizing technical terms and abbreviations, ensuring the standardization and consistency of the input data. Subsequently, the intelligent processing system inputs the preprocessed text data into a large language model. The large language model uses its internal multi-layer Transformer neural network structure to perform deep encoding of the text, utilizing a self-attention mechanism to capture semantic associations and dependencies between words, generating vector representations containing rich contextual information. During the encoding process, the intelligent processing system calls the named entity recognition function of the large language model. By analyzing the context, semantic features, and syntactic roles of each word, it identifies key object entities in the target question. These object entities constitute the core elements of the question, determining the scope and direction of subsequent knowledge retrieval. For example, from "how to optimize the performance of new energy vehicle power batteries under extreme temperatures," the key entities "new energy vehicle" and "power battery" are identified. Meanwhile, the intelligent processing system leverages the intent classification capabilities of large language models. By analyzing multi-dimensional information such as the semantic structure of the question, interrogative word types, and verb phrase features, combined with a predefined intent classification system and a large number of training samples, it determines the user's fundamental purpose and information need type. For example, it determines whether the user wants to obtain technical improvement solutions, understand the cause of a fault, compare different technical routes, or predict development trends. Furthermore, the intelligent processing system also analyzes the overall sentence structure, language expression patterns, and logical relationships of the question using large language models to determine the question's category.

[0032] 102. Map the object entity to the corresponding ontology node in the domain ontology library to generate a semantic feature vector containing the ontology node combination, question intent, and question type.

[0033] A domain ontology is a conceptual hierarchy library built according to the knowledge system of a specific industry domain. It defines the classification, inheritance, and association relationships between concepts within the domain, forming a standardized knowledge representation framework. For example, in the ontology of the new energy domain, "lithium battery" belongs to the "chemical energy storage" subclass under the "energy storage equipment" category. Ontology nodes are used to represent standardized conceptual units in the domain ontology. Each node represents a clearly defined domain concept and includes the concept's attributes, constraints, and relationships with other concepts. For example, the "lithium-ion battery" node includes attribute information such as energy density, cycle life, and operating temperature range. Ontology node combinations represent a set structure formed by multiple related ontology nodes, reflecting the knowledge scope and conceptual associations involved in the target problem. For example, a combination consisting of {lithium-ion battery node, temperature control node, performance optimization node}. Semantic feature vectors are multi-dimensional vector representations formed by numerically encoding ontology node combinations, problem intent, and problem type.

[0034] Specifically, the intelligent processing system first accesses a pre-built domain ontology library, which contains the conceptual hierarchy, attribute definitions, and relational networks of the industry domain. For each object entity extracted from the target problem, the intelligent processing system employs multiple strategies for ontology mapping: first, it performs exact matching to check if there is a node name in the ontology library that is completely identical to the object entity; if exact matching fails, a similarity matching mechanism is initiated, calculating the edit distance and semantic similarity between the object entity and the ontology node name to find the closest ontology node; for abbreviations or colloquialisms of technical terms, the intelligent processing system queries a thesaurus and a domain terminology lookup table to convert them into standard ontology node representations. After completing the mapping of a single entity, the intelligent processing system organizes all the mapped ontology nodes into an ordered node combination. This combination not only includes the nodes themselves but also preserves the potential relationships between the nodes. Next, the intelligent processing system integrates the ontology node combination with the problem intent and problem type to construct a semantic feature vector. This vector comprehensively characterizes the positioning and features of the target problem within the domain knowledge system.

[0035] 103. Calculate the feature similarity between the semantic feature vector and the features of historical problems in the knowledge completeness result library. The knowledge completeness result library is used to store the semantic feature vectors of historical problems and the corresponding knowledge completeness.

[0036] Historical problem features are used to represent the set of semantic feature vectors of previously processed problems; feature similarity represents the numerical similarity between the current semantic feature vector and historical problem features, usually ranging from 0 to 1, with larger values ​​indicating greater semantic similarity between the two problems; historical problems refer to user queries previously received and processed by the intelligent processing system, and these problems and their processing results are stored in the knowledge completeness result base as experiential knowledge; knowledge completeness represents the assessment level of the sufficiency of knowledge resources related to a specific problem, usually divided into three levels: incomplete, partially complete, and highly complete.

[0037] Specifically, the intelligent processing system first acquires the historical ontology node combinations, historical problem intents, and historical problem types for all historical problems in the knowledge completeness database. This information constitutes the complete feature representation of the historical problems. For the extraction of historical ontology node combinations, the intelligent processing system reads the node identifier list corresponding to each historical problem from the database and restores it to a specific ontology node object; the historical problem intents and types are directly obtained from the corresponding fields in the database. Next, the intelligent processing system calculates the overlap between the ontology node combinations in the semantic feature vector and the historical ontology node combinations. This calculation process first counts the intersection size of the two node combinations, i.e., the number of nodes they jointly contain, and then divides it by the union size of the two combinations to obtain the Jaccard similarity coefficient as the overlap index. For example, if the current problem involves nodes {A, B, C}, and the historical problem involves nodes {B, C, D}, the intersection is {B, C}, and the union is {A, B, C, D}, then the overlap is 2 ÷ 4 = 0.5. For semantic similarity calculation of question intent, the intelligent processing system converts the question intent into a semantic vector representation. A dedicated intent encoder can be used, and then the cosine similarity between two intent vectors is calculated. This value reflects the closeness of the intents in the semantic space; for example, the semantic similarity between "performance optimization" and "efficiency improvement" may reach 0.8 or higher. For question type consistency, the intelligent processing system uses discrete matching. If two questions are completely identical, the consistency is 1; if they belong to similar types (e.g., different subclasses of methodological discussion), the consistency is 0.5; and if they are completely different, the consistency is 0. Finally, the intelligent processing system performs a weighted calculation of overlap, semantic similarity, and consistency. The weight coefficients are set based on the characteristics of the industry and historical experience. For example, the weight for ontology node overlap can be set to 0.5, the weight for question intent similarity to 0.3, and the weight for question type consistency to 0.2. The final feature similarity score is obtained by weighted summation, which comprehensively reflects the overall similarity between the current question and historical questions.

[0038] 104. When the feature similarity does not reach the preset similarity threshold, retrieve multidimensional related knowledge elements associated with the target question in the knowledge graph. Multidimensional related knowledge elements include at least triples, attributes, and historical question and answer records.

[0039] A knowledge graph is used to represent a domain knowledge base organized in a graph structure, where nodes represent entities or concepts, and edges represent semantic relationships between entities, forming a large-scale knowledge network. The target question refers to the query submitted by the user that needs to be answered. Multidimensional related knowledge elements represent a set of information describing knowledge related to the target question from different angles and levels. Triples refer to the basic knowledge units in the knowledge graph, which are represented in the form of (subject, predicate, object) to represent factual knowledge, such as (lithium battery, operating temperature range, -20℃ to 60℃). Attributes are used to represent the characteristic description information of entities, including numerical attributes and descriptive attributes, such as the energy density attribute value of a lithium battery being 150Wh / kg. Historical question and answer records represent related questions previously processed by the system and their answers.

[0040] Specifically, when the intelligent processing system detects that the feature similarity has not reached a preset threshold, it indicates that there are no highly similar historical cases in the knowledge completeness database that can be directly reused. It is necessary to retrieve relevant knowledge from the knowledge graph to support the question answer. The intelligent processing system first determines the retrieval starting points in the knowledge graph based on the combination of ontology nodes in the semantic feature vector of the target question. These starting point nodes correspond to the core conceptual entities involved in the question. Starting from each starting point node, the intelligent processing system uses a multi-hop graph traversal algorithm to expand the search outward along the relational edges in the knowledge graph, extracting knowledge elements related to the target question. Regarding triple retrieval, the intelligent processing system not only obtains one-hop triples that directly contain the starting point node, but also obtains multi-hop associated triples through relational path reasoning. For example, for the question "lithium battery charging efficiency," the system will retrieve multi-layered relational chains such as "lithium battery - has - charging characteristics" and "charging characteristics - influencing factors - temperature." Regarding attribute retrieval, the intelligent processing system extracts attribute information from all relevant entity nodes, including static attributes (such as material composition and structural parameters) and dynamic attributes (such as performance indicators and state changes), and sorts and filters them according to the importance and relevance of the attributes. For retrieving historical question-and-answer records, the intelligent processing system uses semantic matching algorithms to search the question-and-answer record database for historical dialogues related to the current target question in terms of topic, intent, or keywords. These records not only contain the final answer but also retain the complete reasoning chain and knowledge references. Through multi-dimensional and multi-level knowledge retrieval, the intelligent processing system constructs a multi-dimensional set of related knowledge elements surrounding the target question.

[0041] 105. Based on the multidimensional related knowledge elements associated with the target problem, use the preset knowledge completeness measurement model to calculate the data coverage of the target problem in multiple dimensions, and obtain the current knowledge completeness level. The knowledge completeness level includes incomplete, partially complete and highly complete.

[0042] The pre-defined knowledge completeness measurement model is a quantitative analysis model specifically designed to evaluate the sufficiency of knowledge resources in supporting a specific problem. This model defines the calculation rules and judgment criteria for completeness, such as rule-based scoring models or machine learning-trained classification models. Data coverage represents the completeness of relevant knowledge elements in a specific dimension, usually expressed as a percentage. For example, an 80% coverage rate for the performance dimension indicates that 80% of the necessary knowledge for that dimension has been acquired. Multiple dimensions refer to different angles for evaluating knowledge completeness, including performance, operating condition, mechanism, and solution dimensions. The performance dimension represents knowledge dimensions related to system or product performance indicators, covering performance parameters such as efficiency, accuracy, and stability, such as battery charge / discharge efficiency and cycle life. The operating condition dimension describes application scenarios and operating conditions, including environmental parameters, load conditions, and usage patterns, such as temperature range, humidity conditions, and vibration environment. The mechanism dimension represents knowledge dimensions that explain underlying principles and causal relationships. Standard mechanism construction trees (SMTs) for entities in different domains are pre-stored in the database. Each SMT defines the set of explanatory knowledge nodes that an entity should ideally possess. For example, for the entity of "power battery", its SMT includes four primary nodes such as "thermodynamic principles", "electrochemical reaction kinetics", "material lattice structure evolution" and "side reaction decay mechanism" as well as 20 secondary key explanation nodes.

[0043] The intelligent processing system first invokes the corresponding standard mechanism construction tree based on the object entity in the target problem (e.g., "ternary lithium battery"). Then, the system maps the triples in the retrieved multidimensional related knowledge elements to construct the current knowledge subgraph. The system uses graph matching algorithms (such as VF2 or semantic vector-based node alignment) to calculate the node mapping ratio between the current knowledge subgraph and the standard mechanism construction tree. The calculation formula is: Mechanism dimension coverage = (Σ(number of successfully matched SMT nodes × node weight)) / (total number of SMT nodes × maximum weight). A successful match means that there is a node in the current knowledge subgraph with a semantic similarity exceeding 0.85 to the SMT node. For the solution dimension, the system invokes the Standard Solution Element Template (SST), which defines the structured fields that a complete technical solution must include, including: implementation step sequence, key process parameters, required resource list, expected output indicators, and risk contingency plan—five major categories of elements. The system uses entity recognition and slot filling technology to detect whether the multidimensional related knowledge elements contain the specific values ​​or descriptions corresponding to the above five categories of elements. Solution dimension coverage is the ratio of the number of filled slots to the total number of slots in the SST template. Only when the coverage calculation is based on a comparison with the aforementioned standard template can the control signal of "incomplete," "partially complete," or "highly complete" be output. Basic dimensions refer to the core knowledge dimensions essential for answering the question; performance and operating condition dimensions fall into this category. Extended dimensions represent supplementary knowledge dimensions that enhance the depth and comprehensiveness of the answer; mechanism and solution dimensions fall into this category. Knowledge completeness represents the level of knowledge sufficiency after comprehensive evaluation, divided into three levels: incomplete (severely insufficient knowledge), partially complete (basically sufficient knowledge), and highly complete (fully complete knowledge).

[0044] Specifically, the intelligent processing system, based on multi-dimensional related knowledge elements associated with the target problem, utilizes a pre-defined knowledge completeness measurement model to first calculate the data coverage rate of the target problem in the performance dimension. The system analyzes the performance-related triples and attributes contained in the knowledge elements, and statistically analyzes the types and quantities of covered performance indicators. For example, for battery performance issues, it checks whether data on key indicators such as energy density, power density, charging speed, and depth of discharge are included. The number of acquired performance indicators is divided by the total number of standard indicators typically required for this type of problem to obtain the performance dimension coverage rate. For the operating condition dimension, the intelligent processing system extracts environmental conditions and application scenario information from the knowledge elements, assessing whether it covers various operating conditions involved in the problem, such as whether the temperature range is complete, whether humidity conditions are clearly defined, and whether the load mode is clear. The operating condition dimension coverage rate is calculated by comparing with a standard operating condition list. When calculating the mechanistic dimension coverage rate, the intelligent processing system analyzes the causal relationships, principle explanations, and mechanism descriptions in the knowledge elements to determine whether they can deeply explain the underlying causes of phenomena. For example, it considers whether it includes deep knowledge such as electrochemical reaction mechanisms, thermodynamic processes, and kinetic models, and provides a coverage rate score based on the completeness and depth of the mechanistic knowledge. For the solution dimension, the intelligent processing system evaluates the number of solutions contained in the knowledge elements, the specificity of the solutions, and their operability. It checks whether there are complete implementation steps, parameter settings, and expected effect descriptions, and calculates the coverage rate of the solution dimension accordingly. After obtaining the coverage rates of the four dimensions, the intelligent processing system performs a weighted average of the performance dimension and the operating condition dimension as basic dimensions to obtain the comprehensive coverage rate of the basic dimensions. If this coverage rate does not reach the first preset coverage rate threshold (the first preset coverage rate threshold is the standard value for judging whether the basic dimensions meet the minimum requirements, usually set to 60%-70%), then the knowledge completeness of the target problem is determined to be incomplete, indicating a lack of necessary basic knowledge to answer the question. If the coverage rate of the basic dimensions reaches the first preset coverage threshold, the intelligent processing system continues to evaluate the extended dimensions, performing a weighted average of the coverage rates of the mechanism dimension and the solution dimension. If the coverage rate of the extended dimensions does not reach the second preset coverage threshold (the second preset coverage threshold is the standard value for judging whether the extended dimensions have reached a higher level, usually set to 75%-85%), then the knowledge completeness is determined to be partially complete, indicating the possession of basic knowledge but a lack of depth. If the coverage of both the basic and extended dimensions reaches the corresponding threshold, the knowledge completeness is determined to be highly complete, indicating that the knowledge resources are sufficient and can provide comprehensive and in-depth answers.

[0045] In some embodiments, the preset knowledge completeness measurement model employs a multi-dimensional weighted quantitative scoring algorithm to calculate the degree of knowledge completeness. The model first defines an overall evaluation function, calculated as follows:

[0046] Where KCS(q): represents the total knowledge completeness score for the target problem q; clip [0,1] The truncation function ensures the final score is between 0 and 1; i represents the set of dimensions involved in the calculation, i∈{E, R, S, T, Q}, corresponding to entity, relation, evidence, timeliness, and semantic dimensions, respectively; T represents the type of the target question (e.g., factual, diagnostic, predictive, etc.); w i (T) represents the weight coefficient of the i-th dimension when the problem type is T; C i (q) represents the coverage score of the target question q on the i-th dimension; w F (T) represents the weight coefficient of the penalty term when the problem type is T; P F (q) represents the conflict and noise penalty score in the knowledge element corresponding to the target question q.

[0047] This formula is used to calculate the final knowledge completeness score KCS(q) for the target question q, and the result is restricted to the interval [0, 1]. It obtains the comprehensive score by weighting and summing the coverage of multiple dimensions such as entities, relations, evidence, timeliness, and semantics, and subtracting the penalty terms for conflicts and noise.

[0048] To achieve adaptive adjustment of the weights, the weight coefficient w i (T) is dynamically calculated based on the problem type and risk coefficient, and the calculation formula is as follows:

[0049] in, ρ(T) represents the basic preset weight of the i-th dimension; ρ(T) represents the risk coefficient of the problem type T, with a value range of [0, 1]; ai represents the sensitivity coefficient of the i-th dimension to risk; aF represents the sensitivity coefficient of the penalty term (F: refers to conflict or noise in knowledge) to the problem risk level.

[0050] This formula normalizes and dynamically adjusts the weights. Based on the risk level of the problem, the system automatically adjusts the importance of each dimension (for example, high-risk problems have higher requirements for evidence and timeliness).

[0051] When calculating the specific scores for each dimension, the final entity coverage calculation formula is as follows:

[0052] Among them, C EThis represents the basic entity coverage score. The higher the value, the more comprehensive the retrieved context coverage of the question; Eq represents the set of all entities extracted from the user query; e represents a single entity in the set Eq; α(e) represents the importance weight of entity e; G∪D indicates that the entity can be matched in the knowledge graph or in the document, and the union of the two is taken; G represents the structured knowledge graph; D represents the unstructured document library; match(e, G∪D) represents the matching detection function used to determine whether entity e exists in the knowledge source. g E The entity gating factor can only be 0 or 1; CriticalSlots(q) represents the set of critical slot entities in query q. This represents the final corrected entity coverage score. This is the system's final criterion for determining whether the current search results are usable.

[0053] Calculate C E The formula calculates a "weighted hit rate." It examines each entity in the user's question; if a match is found, a score (weight α) is added to that entity; otherwise, 0 is added. This reflects the degree of flexibility in how the search content matches the question.

[0054] Calculate g E The formula performs a hard constraint check. If any entity in the critical slot set fails to match (value is 0), the product of these values ​​will result in g. E That is, it is 0; only when all key entities are successfully matched is g E The value is 1. This constitutes the system's "circuit breaker mechanism," preventing illusions caused by the lack of crucial information.

[0055] Regarding the sufficiency of evidence (C) S The model considers both the authority and quantity of sources, and the calculation formula is as follows: First, perform an "independent evidence deduplication count" for each source:

[0056] Strength of evidence:

[0057] Sufficiency normalization:

[0058] Where Auth(src) represents the authority weight of the source src; Count(src) represents the number of unique pieces of evidence provided by the source src after deduplication. The formula uses the natural logarithm ln to handle the number of unique pieces of evidence, reflecting the diminishing marginal utility of evidence from the same source; θ S(T(q)) represents the threshold value for the sufficiency of evidence set for question type T(q); min(1,...): ensures the final sufficiency of evidence score C. S No more than 1.

[0059] This formula assesses the sufficiency of evidence supporting the knowledge. Using a logarithmic function to handle the amount of evidence implies diminishing marginal gains as the amount of evidence from the same source increases. It also incorporates weighting based on source authority and normalizes the data according to a question type threshold.

[0060] Regarding timeliness consistency (C) T The model incorporates a time decay mechanism, and the calculation formula is as follows:

[0061] Among them, C T Represents the timeliness consistency score; conf(x) represents the original confidence level of knowledge element x; age(x) represents the time difference (e.g., days) between knowledge element x and the current system time; λ(r) represents the time difference between knowledge element x and the current system time. x ) represents the relation type r x The time sensitivity decay coefficient (e.g., a large coefficient for "real-time status" and a small coefficient for "historical principles"); τ(x) represents the timestamp of knowledge element x; Δ(q) represents the effective time window set for the target problem q.

[0062] This formula is used to assess the timeliness of knowledge. The older the knowledge, the more its contribution to the score decays exponentially, and the rate of decay depends on how sensitive the relation type is to time.

[0063] 106. Based on the completeness of knowledge, call the large language model to generate the response text corresponding to the target question.

[0064] The response text refers to the answer content generated by the intelligent processing system for the target question. The content depth and expression style of this text vary depending on the completeness of the knowledge.

[0065] Specifically, the intelligent processing system first determines the response generation strategy based on the level of knowledge completeness. When the knowledge completeness is highly complete, the system invokes a large language model to organize rich, multi-dimensional, related knowledge elements into structured input, including relevant triplet facts, detailed attribute data, and valuable historical question-and-answer experience. This guides the model to generate a comprehensive and in-depth response text, which not only directly answers the user's question but also provides in-depth explanations of principles, comparative analyses of multiple solutions, and implementation suggestions. The system also annotates the knowledge sources and reasoning basis in the text to enhance the credibility of the response. When the knowledge completeness is partially complete, the system first generates core response content based on existing basic dimension knowledge to ensure that the basic questions are answered. Then, it clearly points out the knowledge limitations in extended dimensions such as mechanistic explanations or solution details, reminding the user that some analyses may not be in-depth enough or that some suggestions need further verification. Simultaneously, the system invokes the reasoning capabilities of the large language model to make reasonable inferences based on existing knowledge, but clearly marks which content is inferred rather than established facts, and suggests that the user consult specific professional materials or domain experts to obtain more complete information. When the knowledge completeness level is incomplete, the intelligent processing system first generates an honest explanation of the knowledge deficiency through a large language model, clearly informing the user that the current system lacks the key information needed to answer the question, specifically pointing out the missing items in basic dimensions such as performance data or operating condition information. Then, based on limited knowledge, it provides preliminary directional suggestions or explanations of relevant concepts to avoid giving potentially misleading specific conclusions.

[0066] 107. Based on the response text, update the knowledge graph, and based on the semantic feature vector of the target question, the knowledge completeness, and the response text, update the knowledge completeness result database.

[0067] A knowledge-complete results repository is used to store a complete record of historical problem handling.

[0068] Specifically, the intelligent processing system first extracts newly generated triples, relational structures (relational structures represent complex association patterns between entities, including multi-hop relational chains, conditional relations, and causal relations, such as the causal chain of "temperature increase → electrolyte viscosity decrease → ion conductivity increase"), and explanatory structures (explanatory structures are used to represent a systematic exposition framework of concepts, principles, or methods, including elements such as definitions, principles, application conditions, and precautions, such as a complete explanatory system for "solid electrolytes") from the reply text. It then performs deep analysis of the reply text using natural language processing technology, identifying factual statements and converting them into triples. For example, from the sentence "Research shows that solid batteries can still maintain 90% of their capacity at 85℃," it extracts the triple (solid battery, high-temperature capacity retention rate, 90%@85℃). Simultaneously, it analyzes the logical relationships and reasoning processes in the text, extracting multi-step relational structures, such as identifying the progressive relational chain of "material modification → performance improvement → application expansion." For the conceptual explanations and principle descriptions in the text, the intelligent processing system constructs a complete explanatory structure, including information from multiple dimensions such as concept definitions, mechanisms of action, applicable conditions, and comparisons of advantages and disadvantages. After extraction, the intelligent processing system adds these triples, relational structures, and explanatory structures to the knowledge graph. Triples are directly added as new edges between corresponding nodes, relational structures represent complex association paths by creating intermediate nodes and multiple edges, and explanatory structures are stored as rich attributes of nodes. Next, the intelligent processing system establishes a mapping relationship between the semantic feature vector of the target question, the knowledge completeness level, and the response text, creating a structured record with multiple fields. The semantic feature vector field stores a standardized representation of the question for future similarity matching. The knowledge completeness field records the evaluation result (incomplete, partially complete, or highly complete) and the specific coverage values ​​for each dimension. The response text field stores the complete generated content and its metadata (such as generation time, model version used, and cited knowledge sources), while also adding auxiliary information such as timestamps, user identifiers, and question domain classifications. Finally, the intelligent processing system stores this complete mapping relationship as a new knowledge completeness result in the knowledge completeness result library. During storage, the system optimizes the index and establishes a retrieval index based on multiple dimensions such as semantic features, completeness, and time to ensure that relevant records can be quickly found in the future. At the same time, it updates statistical information, such as the number of questions at various completeness levels.

[0069] 108. Monitor data changes in the knowledge completeness result database and send newly added or changed knowledge completeness result data to multiple system nodes within the communication range, including enterprise nodes and research institution nodes.

[0070] Communication range refers to the network coverage area where data transmission can occur between system nodes, which may be a local area network, a wide area network, or a dedicated industrial internet; multiple system nodes represent instances of intelligent processing systems deployed in different locations or institutions, which work together to form a distributed knowledge network; enterprise nodes refer to system nodes deployed within various enterprises; research institution nodes refer to system nodes deployed in universities, research institutes, and other research units.

[0071] Specifically, the intelligent processing system first determines whether the knowledge completeness database meets the preset synchronization conditions (the preset synchronization conditions include the number of newly added highly complete records exceeding a preset threshold or the time since the last synchronization exceeding a preset duration). The intelligent processing system maintains a synchronization status manager, which counts the number of newly added highly complete records in real time. When the number of newly added highly complete records exceeds the preset threshold (e.g., 10 records), a synchronization flag is triggered. Alternatively, when the system clock shows that the time since the last synchronization exceeds a preset duration (e.g., 24 hours), a synchronization operation is also triggered. This dual triggering mechanism ensures the timely sharing of important knowledge and the regular updating of routine knowledge. If the knowledge completeness database meets the preset synchronization conditions, the intelligent processing system begins the data extraction process. The system accurately identifies all newly added knowledge completeness data since the last synchronization through timestamp comparison and version number tracking. This data includes newly processed problem cases, generated response texts, and assessed completeness levels. Simultaneously, it extracts changed knowledge completeness data—records where the completeness level has changed, such as cases that have been upgraded from incomplete to partially complete after knowledge supplementation. During extraction, the system compresses and encrypts the data to ensure transmission efficiency and security. Next, the intelligent processing system sends the newly added and changed knowledge completeness data to system nodes within the communication range, using a publish-subscribe model or a point-to-point transmission protocol. While sending data, the intelligent processing system also receives external completeness data from other system nodes, achieving bidirectional knowledge exchange. Finally, the intelligent processing system detects and analyzes the received external complete result data, focusing on the differences in knowledge completeness. When it detects that the same problem (identified by semantic feature vector matching) is marked as highly complete in the external complete result data but recorded locally as incomplete or partially complete, it indicates that other nodes have obtained more complete knowledge resources. At this time, the intelligent processing system will use the external complete result data to update the local knowledge complete result database.

[0072] This application employs a knowledge evolution-based large-scale model-driven industrial intelligence operation method. The intelligent processing system first uses a large language model to semantically parse the target question, extracting object entities, question intent, and question type. The object entities are then mapped to a domain ontology library to generate semantic feature vectors, establishing a connection between the question and the domain knowledge system. Next, the intelligent processing system calculates the similarity between this semantic feature vector and historical question features in the knowledge completeness database. If the similarity does not reach a preset threshold, multi-dimensional related knowledge elements are retrieved from the knowledge graph. A preset knowledge completeness measurement model is used to calculate multi-dimensional data coverage, clarifying the degree of knowledge completeness. Then, based on the knowledge completeness, the intelligent processing system calls the large language model to generate targeted response text. Finally, based on the response text, the knowledge graph and knowledge completeness database are updated, data changes are monitored and synchronized to multiple system nodes, achieving knowledge collaboration and sharing. This optimizes the entire process of industrial intelligent question answering, from knowledge association and completeness assessment to response generation and knowledge evolution. This method alleviates the technical problem of one-sided and distorted content in industrial intelligent responses and improves the accuracy of industrial intelligent question answering.

[0073] Based on the above, the following is a more detailed description of the process provided in this implementation. Please refer to [link / reference]. Figure 2 This is another flowchart illustrating the large-scale model-driven intelligent operation method based on knowledge evolution in this application.

[0074] 201. Receive the target question submitted by the user, and use a large language model to perform semantic parsing on the target question to obtain the object entities involved in the target question, the question intent, and the question type. (This step has been explained in 101.) 202. Map the object entity to the corresponding ontology node in the domain ontology library, generating a semantic feature vector containing the ontology node combination, question intent, and question type. (This step has been explained in 102.) 203. Calculate the feature similarity between the semantic feature vector and the features of historical questions in the knowledge completeness database. The knowledge completeness database is used to store the semantic feature vectors of historical questions and their corresponding knowledge completeness levels. (This step has been explained in 103.) 204. When the feature similarity reaches the preset similarity threshold, it is determined that there is a historical problem in the knowledge completeness result library that belongs to the same knowledge category as the target problem, and the knowledge completeness of the historical problem is obtained.

[0075] Feature similarity refers to the numerical matching degree between the semantic feature vector of the target problem currently being processed by the intelligent processing system and the features of historical problems stored in the knowledge completeness result library; the preset similarity threshold refers to the critical value set in advance by the intelligent processing system to determine whether two problems belong to the same knowledge category. This threshold is set according to the characteristics of the industry field and historical experience, and is usually between 0.7 and 0.9.

[0076] Specifically, when the feature similarity calculated by the intelligent processing system reaches or exceeds a preset similarity threshold, the system first triggers a historical knowledge reuse judgment mechanism. This indicates that there are likely historical processing cases in the knowledge completeness database that are highly relevant to the current target problem, and historical experience can be directly utilized to accelerate problem processing. The intelligent processing system uses a similarity ranking algorithm to select several candidate problems with the highest similarity from all historical problems that meet the threshold condition. Typically, the top three most similar historical problems are selected as reference cases. These candidate problems are then subjected to in-depth analysis, comparing multiple dimensions such as the core ontology nodes, problem intent, and problem type to confirm that they indeed belong to the same knowledge category as the target problem. After confirming the knowledge category match, the intelligent processing system extracts the complete processing record of the historical problem from the knowledge completeness database and obtains its knowledge completeness assessment result (incomplete, partially complete, or highly complete).

[0077] 205. If the knowledge completeness of the historical question is highly complete, then call the existing knowledge in the knowledge graph, use the large language model to generate the response text, and perform incremental update operations.

[0078] Existing knowledge refers to all knowledge elements that have been stored and organized in the knowledge graph of an intelligent processing system, including structured knowledge such as entity nodes, relation edges, triples, attribute data, and reasoning paths.

[0079] Specifically, when the intelligent processing system confirms that the knowledge completeness of a historical problem is highly complete, it first locates and extracts all existing knowledge resources associated with that historical problem from the knowledge graph. This extraction process starts from the core entity nodes corresponding to the historical problem, and uses a graph traversal algorithm to obtain relevant knowledge within a multi-hop range. This includes directly related triples (such as "lithium iron phosphate battery - operating temperature range -20℃ to 60℃"), indirectly related reasoning chains (such as "lithium iron phosphate battery → cathode material → olivine structure → ion transport channel → low temperature performance"), detailed attribute sets of entities (such as energy density, cycle life, cost, and dozens of other attribute values), as well as reasoning paths and solution templates generated when processing the problem in the past. The intelligent processing system organizes this existing knowledge into a structured knowledge package, sorts and filters it according to importance and relevance, ensuring that the knowledge input into the large language model is both comprehensive and concise, avoiding information overload that could affect the quality of the generated data. Subsequently, the intelligent processing system constructs a prompt template, integrating the target question, existing knowledge, and generation instructions into a complete input. It then invokes a large language model to perform knowledge fusion and text generation tasks. The large language model, based on existing knowledge, performs reasoning and organization to generate a logically clear and detailed response text. This response text not only directly answers the user's core question but also provides in-depth explanations of principles, multi-faceted analysis and comparisons, specific implementation suggestions, and potential risk warnings. Furthermore, it embeds knowledge source annotations within the text to enhance the credibility and traceability of the answer. After generating the response text, the intelligent processing system performs an incremental update operation, analyzing whether new knowledge associations or reasoning paths have been generated in the response text. If so, these are added as new knowledge edges to the knowledge graph. Finally, the relevant records in the knowledge completeness result database are updated, storing the target question and its processing result as a new highly complete case.

[0080] 206. If the knowledge completeness of a historical issue is partially complete or incomplete, a first prompt message is generated. The first prompt message is used to indicate that there is a gap in the knowledge of the current domain.

[0081] The first prompt message is a warning text generated by the intelligent processing system to inform users or system administrators of the current knowledge status. This message clearly points out the deficiencies in domain knowledge. Current domain knowledge refers to the existing knowledge resources in the knowledge graph and knowledge base of the intelligent processing system for the domain to which the target question belongs. Knowledge gaps refer to the knowledge content that is needed to fully answer the target question but is not yet available in the current system.

[0082] Specifically, when the intelligent processing system detects that the knowledge completeness of a historical problem is partially complete or incomplete, it first initiates a knowledge gap analysis process. For partially complete cases, the intelligent processing system will analyze the missing details in the extended dimensions in detail, identifying which key theoretical explanations, physical models, or chemical mechanism descriptions are missing in the mechanism dimension, and which specific technical routes, implementation steps, or parameter configuration suggestions are missing in the solution dimension. For example, when dealing with the problem of "fatigue failure mechanism of carbon fiber composite materials," the system may find that although there is basic material performance data and testing conditions, there is a lack of explanation of the microscopic fracture mechanism and fatigue life prediction models. For incomplete cases, the intelligent processing system will analyze the serious deficiencies in the basic dimensions in more detail, clearly pointing out which core indicator data, testing standards, or evaluation systems are missing in the performance dimension, and which key application scenario descriptions, environmental parameter ranges, or boundary condition definitions are missing in the operating condition dimension. Based on the knowledge gap analysis results, the intelligent processing system generates a structured initial prompt message. This message contains multiple layers of content: first, a clear warning statement; second, a detailed description of the gap, listing the specific missing knowledge dimensions, data types, and information categories; third, an impact assessment explanation, clarifying the potential impact of the knowledge gap on the question's answer; fourth, knowledge acquisition suggestions, providing possible ways to supplement knowledge, such as "recommending consulting recently published academic papers"; and finally, alternative solutions, offering possible workarounds in cases of insufficient knowledge. The initial prompt message is presented in various ways, including displaying a prominent warning icon on the user interface, adding a disclaimer at the beginning of the response text, and sending a knowledge gap record notification to the system administrator.

[0083] 207. When the feature similarity does not reach the preset similarity threshold, retrieve multidimensional related knowledge elements associated with the target question from the knowledge graph. These multidimensional related knowledge elements must include at least triples, attributes, and historical question-and-answer records. (This step has been explained in section 104.) 208. Based on the multidimensional related knowledge elements associated with the target problem, use a preset knowledge completeness measurement model to calculate the data coverage of the target problem in multiple dimensions. The multiple dimensions include at least performance dimension, operating condition dimension, mechanism dimension and solution dimension. The performance dimension and operating condition dimension are basic dimensions, while the mechanism dimension and solution dimension are extended dimensions.

[0084] The pre-defined knowledge completeness measurement model represents a mathematical model or algorithmic framework pre-built by the intelligent processing system for quantitatively evaluating the sufficiency of knowledge resources. This model defines the evaluation indicators, calculation formulas, and weight coefficients for each dimension, such as a rule-based weighted scoring model or a machine learning-based classification model.

[0085] Specifically, the intelligent processing system first performs structured preprocessing on multi-dimensional related knowledge elements, classifying and standardizing fragmented knowledge into four dimensions: performance, operating conditions, mechanism, and solution. This ensures that each knowledge element is correctly categorized into its corresponding dimension. After independently evaluating the four dimensions, the intelligent processing system focuses on the performance and operating condition dimensions as foundational dimensions, as these directly determine whether a basic, usable answer can be provided. The mechanism and solution dimensions are used as extended dimensions to assess the depth and value of the response.

[0086] The pre-defined knowledge completeness measurement model is an evaluation model built by the intelligent processing system using machine learning methods. Its construction process includes: first, collecting historical problem-solving cases and labeling the knowledge coverage and final completeness level of each problem across four dimensions: performance, operating conditions, mechanism, and solution; then, extracting the semantic features of the problem, the structural features of the knowledge graph, and content features as training samples; next, employing a deep neural network architecture, designing a multi-layered network structure including a feature extraction layer, a dimension evaluation layer, and a comprehensive judgment layer. The dimension evaluation layer calculates the coverage rate for each of the four knowledge dimensions, and the comprehensive judgment layer outputs the final completeness level based on the coverage rate of each dimension and pre-defined threshold rules; finally, model training is completed through cross-validation and parameter optimization. The model's input is the semantic feature vector of the target problem (including ontology node combinations, problem intent type, etc.) and the subgraph structure related to the problem in the knowledge graph (including knowledge elements such as entities, relationships, and attributes). The output is the classification result of the knowledge completeness level (incomplete, partially complete, or highly complete) and the specific coverage rate values ​​(0-100%) for the four dimensions.

[0087] 209. If the coverage of the basic dimensions does not reach the first preset coverage threshold, then the knowledge completeness of the target problem is determined to be incomplete. (This step has been explained in 105.) 210. If the coverage of the basic dimensions reaches the first preset coverage threshold and the coverage of the extended dimensions does not reach the second preset coverage threshold, then the knowledge completeness of the target problem is determined to be partially complete. (This step has been explained in 105.) 211. If the coverage of the basic dimensions reaches the first preset coverage threshold and the coverage of the extended dimensions reaches the second preset coverage threshold, then the knowledge completeness of the target problem is determined to be highly complete. (This step has been explained in section 105) 212. Based on the completeness of knowledge, call the large language model to generate the response text corresponding to the target question.

[0088] DIKW represents the cognitive model of Data-Information-Knowledge-Wisdom, including four levels, namely the data layer (determining the data source address to be collected, setting the sampling frequency, defining the data communication protocol, and configuring the denoising filter parameters for the original data. The output of this step is a standardized original data set), the information layer (extracting key fields using regular expressions, removing abnormal outliers according to the preset threshold, interpolating and complementing missing values, and converting unstructured text into structured key-value pairs), the knowledge layer (specifically calling the entity recognition model and the relationship extraction model to extract (entity-relationship-entity) triples from the structured information table and performing entity alignment operations to link the newly extracted triples to the existing knowledge graph to form a new topological structure), and the wisdom layer (indicating that the system calls the preset decision solver (Solver) to construct constraint equations and objective functions based on the expanded knowledge graph (for example: maximizing efficiency Y under the constraint of cost < X), and outputting the optimal control parameter group or the specific operation path through iterative calculations).

[0089] Specifically, when the knowledge completeness level is incomplete, the intelligent processing system first invokes the large language model to organize the currently acquired limited knowledge elements into an input context. Simultaneously, it explicitly informs the model of the current severe knowledge deficiency, guiding it to generate an initial response text containing secondary prompts based on existing knowledge. During this generation process, the large language model first analyzes which basic dimensions have insufficient coverage. For example, if it finds that the performance dimension has only 45% coverage, specifically missing key performance indicators, the model will organize these specific missing items into structured secondary prompts. Based on the specific missing dimensions contained in the secondary prompts, the intelligent processing system automatically generates knowledge gap tasks. Each task includes a clear knowledge acquisition objective, suggested data sources, and estimated completion time and resource requirements. Next, based on the problem intent and initial response text, the intelligent processing system generates knowledge building tasks according to the DIKW principle. This process includes four levels of planning: at the data layer, it determines the raw test data, experimental records, and observation results to be collected; at the information layer, it plans how to clean, standardize, and structure the raw data; at the knowledge layer, it designs how to extract patterns, build models, and form theories from the information; and at the wisdom layer, it plans how to transform knowledge into decision support and problem solutions. For example, it generates tasks such as: "Data layer task: collect complete lifecycle test data from 100 samples → Information layer task: extract performance degradation curves and key inflection points → Knowledge layer task: build a performance degradation prediction model → Wisdom layer task: formulate maintenance strategies to extend service life." Subsequently, the intelligent processing system uses a large language model to generate automated workflow construction suggestions based on the knowledge building tasks. The model analyzes the logical relationships and dependencies of the tasks, designs parallel executable task branches and sequential task chains, and generates a complete workflow plan that includes specific execution steps, required tools, personnel division of labor, time nodes, and quality checkpoints. Finally, the intelligent processing system integrates the suggestions for building automated workflows into the initial response text to form the final response text. This final text not only includes preliminary answers based on existing knowledge and clear hints of knowledge gaps, but also provides systematic knowledge supplementation solutions and actionable workflow suggestions, enabling users to understand the current limitations and obtain a clear path to solve the problem.

[0090] Specifically, when the knowledge completeness is partially complete, the intelligent processing system first invokes a large language model to construct a reasoning chain based on known relationships in the knowledge graph. This process begins with the core entity nodes involved in the target question and proceeds along the relationship edges of the knowledge graph through multi-hop reasoning. Each reasoning step is based on explicitly defined relational triples in the knowledge graph. The large language model not only generates reasoning chains but also organizes the answer content based on these reasoning relationships, integrating technical principles, influencing factors, and improvement directions into coherent text. While generating the reasoning chains, the intelligent processing system calculates the confidence score of each reasoning step. This calculation comprehensively considers factors such as the authority of the knowledge source, the timeliness of the data, and the size of the reasoning span. For example, the confidence score of a reasoning step directly based on experimental data is 0.9, the confidence score based on theoretical derivation is 0.75, and the confidence score based on empirical rules is 0.6. When a reasoning step involves missing knowledge in an extended dimension, the confidence score will decrease accordingly; for example, the confidence score of a reasoning step lacking a complete mechanistic explanation may only be 0.5. The intelligent processing system explicitly marks inference steps below a preset reliability threshold (e.g., 0.65) in the inference chain. Combining the marked inference chain with the response content, it generates an initial response text. Next, based on the expanded dimensions where data coverage does not reach a second preset coverage threshold, the system generates supplementary data tasks, specifically analyzing missing details in the mechanism and solution dimensions, creating specific supplementary tasks for each missing item. Then, based on the question intent of the target problem and the initial response text, the system generates knowledge building tasks according to the DIKW principle. The system uses a large language model to generate automated workflow building suggestions based on the building tasks. Finally, the system integrates the automated workflow building suggestions into the initial response text to form the final response text.

[0091] Specifically, when the knowledge completeness level is highly complete, the intelligent processing system first invokes the large language model to organize the rich multi-dimensional knowledge elements in the knowledge graph into a complete input context. This includes a comprehensive performance dataset, a complete operating condition matrix, an in-depth mechanism explanation system, and a mature solution library. Based on this high-quality knowledge, the large language model generates initial response text. During the generation process, the model not only provides direct answers to the questions but also constructs clear reasoning paths, with each reasoning node supported by explicit knowledge. In the text portion, the intelligent processing system requires the large language model to insert knowledge source markers at key argument positions, such as "According to the test results of [Lithium Battery Database: LB2023-0456], the capacity retention rate is 85.3% after 1000 cycles at 25℃." These markers include not only source identifiers but also access links or retrieval codes to ensure the traceability and verifiability of the knowledge. After generating the initial response text, the intelligent processing system stores the reasoning path as a complete knowledge path in the knowledge graph. Simultaneously, based on the question intent and the initial response text, the system performs an intelligent retrieval in the workflow repository. The system uses a semantic matching algorithm to search for historical workflows highly relevant to the current question intent. For example, inputting the intent "improve battery cycle life" retrieves candidate solutions such as "Lithium-ion battery life optimization workflow v2.3" and "Standard process for improving energy storage system performance." Through similarity calculation and suitability evaluation, the system selects the most suitable first automated workflow. This workflow contains a complete process with specific operation guidelines for each step. If no highly matching existing solution is found in the workflow repository, the intelligent processing system uses a large language model to generate a second automated workflow according to the DIKW principle. The model analyzes the problem characteristics and knowledge structure to design a customized workflow solution, complete with execution scripts, monitoring indicators, and contingency plans for handling anomalies. Finally, the intelligent processing system integrates the selected first automated workflow or the generated second automated workflow into the initial response text to form the final response text. This final text not only provides an authoritative answer based on sufficient knowledge and a clear reasoning process, but also includes a directly executable workflow solution, so that users can obtain both theoretical answers to their questions and specific guidance for practical operation.

[0092] 213. Based on the response text, update the knowledge graph, and based on the semantic feature vector of the target question, the knowledge completeness, and the response text, update the knowledge completeness result database.

[0093] Specifically, the intelligent processing system first parses the response text, extracting newly generated triples, relational structures, and explanatory structures. It then uses natural language processing (NLP) technology to identify entities, relationships, and knowledge frameworks. After standardizing the extracted new knowledge elements, it adds them to the knowledge graph, completing the incremental update of the knowledge graph. Next, the intelligent processing system focuses on establishing mapping relationships. The system uses the semantic feature vector of the target question as the primary key. This vector contains numerical representations of core features such as the combination of ontology nodes and the question's intent type. It uses the knowledge completeness assessment result as the core attribute, recording the incomplete, partially complete, or highly complete judgment results, as well as the specific coverage values ​​for each dimension. It records detailed descriptions of the missing dimensions, clearly indicating which dimensions (performance, operating conditions, mechanism, and solution) have insufficient knowledge, what specific key information is missing, and a quantitative assessment of the degree of deficiency. It records the processing timestamp precisely to the millisecond level for knowledge timeliness management and historical traceability. The system associates and binds these four types of core information through data structures to form a complete mapping record. This record also includes auxiliary fields such as a summary of the response text, a list of cited knowledge sources, and metadata about the processing process. Finally, the intelligent processing system stores the constructed mapping relationship as a new knowledge completeness result in the knowledge completeness result library. During storage, a unique identifier is generated, a multi-dimensional index structure is established to support fast retrieval, and the statistical metadata in the library is updated.

[0094] In some embodiments, the evolution update process of the knowledge completeness result can also be performed based on the response text to form a traceable knowledge evolution state, thereby constraining the processing path of subsequent issues.

[0095] Please see Figure 3 This is a logical schematic diagram of the knowledge evolution processing method in the embodiments of this application. The following is a combination of... Figure 3 This paper further explains the logical flow of the knowledge evolution processing method.

[0096] Specifically, after generating a response text corresponding to the target question, the intelligent processing system performs structured parsing on the response text, extracts the knowledge elements it contains, and analyzes the completeness of the knowledge content covered by the response text from four preset analysis dimensions: performance, operating conditions, mechanism, and solution. For each analysis dimension, the system clearly identifies whether there are any knowledge deficiencies and further determines the specific missing key information items and their corresponding degree of deficiency. The degree of deficiency is calculated using preset quantitative evaluation rules to form the knowledge completeness evaluation result for the corresponding dimension.

[0097] The intelligent processing system associates the knowledge completeness assessment results, corresponding missing key information identifiers, and quantitative assessment values ​​under the above four analytical dimensions with the semantic feature vector of the target problem, and performs consistency verification and difference comparison with the existing knowledge in the current knowledge graph. When a new knowledge relationship is detected, a change in the original knowledge attribute, or an expansion of the knowledge coverage, the evolution and update processing flow of the knowledge completeness results is triggered.

[0098] During the evolutionary update process, the intelligent processing system maintains a corresponding evolutionary state for each knowledge completeness result. This state characterizes the stage position of the knowledge completeness result in the knowledge evolution process. The evolutionary state includes at least an initial generation state, a supplementary evolutionary state, and a stable evolutionary state. The initial generation state indicates that the knowledge completeness result is generated for the first time by large-scale model reasoning and has not yet undergone multiple rounds of verification or supplementation. The supplementary evolutionary state indicates that the knowledge completeness result is corrected or expanded in subsequent problem processing through new data, external system synchronization, or repeated verification. The stable evolutionary state indicates that the knowledge completeness result remains consistent across multiple target problem processing scenarios, and its knowledge coverage across the four analytical dimensions of performance, operating conditions, mechanism, and solution all meet preset stability conditions, making it a reusable stable knowledge unit.

[0099] The change in the evolutionary state does not occur automatically based on time sequence, but is driven by at least one evolutionary triggering condition. These triggering conditions include: obtaining consistent knowledge completeness assessment results across different processing batches for the same semantic feature vector matching problem; filling knowledge gaps in one or more historically identified analysis dimensions, with the corresponding quantitative assessment value of the missing degree being lower than a preset threshold; or synchronously obtaining knowledge results with a higher completeness level in the corresponding analysis dimension from external system nodes. When the evolutionary triggering condition is met, the intelligent processing system updates the evolutionary state of the corresponding knowledge completeness result.

[0100] When a new target question is received, the intelligent processing system not only matches historical questions based on the semantic feature vector of the target question, but also dynamically selects the problem processing path based on the evolution state of the matched knowledge completeness results and the missing information in each analysis dimension. When the matched knowledge completeness results are in a stable evolution state, the system directly reuses the corresponding knowledge completeness results, skipping the complete multi-dimensional related knowledge element retrieval process; when the matched knowledge completeness results are in a supplementary evolution state, the system reuses existing results and triggers targeted knowledge supplementation processing only for analysis dimensions where knowledge is still insufficient; when the matched knowledge completeness results are in an initial generation state, the system executes the complete multi-dimensional related knowledge element retrieval and knowledge completeness reassessment process.

[0101] To ensure the traceability of knowledge evolution, the intelligent processing system records the processing timestamp to the millisecond level when each complete knowledge result is generated. This is used for knowledge timeliness management and historical tracing, and the system maintains evolution trajectory information for the corresponding complete knowledge result. The evolution trajectory information includes at least an evolution version identifier, an evolution source marker, a parent version reference relationship, and an explanation of the evolution trigger reason. The system associates and binds the completeness evaluation results of the four analytical dimensions (performance, operating condition, mechanism, and solution), the identifiers of missing key information, and the quantitative evaluation values ​​with the timestamp information through a data structure, forming a complete mapping record.

[0102] In addition, the mapping record also includes auxiliary fields such as a summary of the response text, a list of cited knowledge sources, and metadata of the processing process, to support the auditing, backtracking, and discrepancy analysis of subsequent knowledge evolution results. When new evidence is detected that conflicts with the existing stable evolution state, the system reverts to the corresponding historical version based on the parent version reference relationship in the mapping record, and re-triggers the knowledge completeness assessment and evolution update process.

[0103] Finally, the intelligent processing system stores the updated knowledge completeness results, corresponding evolution states, evolution trajectory information, and mapping records as an integrated data structure in the knowledge completeness result library. It also establishes an index relationship based on semantic feature vectors, evolution states, completeness levels, and processing timestamps to support the rapid location, state determination, and processing path selection of knowledge evolution results in subsequent problem processing.

[0104] 214. Monitor data changes in the knowledge completeness result database and send newly added or changed knowledge completeness result data to multiple system nodes within the communication range. These system nodes include enterprise nodes and research institution nodes. (This step has been explained in section 108.) The knowledge evolution-based large-scale model-driven industrial intelligent operation method in this application adopts the following approach: First, the intelligent processing system uses a large language model to semantically analyze the target problem, extracting object entities, problem intent, and problem type. The object entities are then mapped to a domain ontology library to generate semantic feature vectors, establishing a connection between the problem and the domain knowledge system. Next, the intelligent processing system calculates the feature similarity between this semantic feature vector and historical problem features in the knowledge completeness database. When a preset threshold is reached, the system differentiates processing based on the knowledge completeness of the historical problem. For highly complete problems, existing knowledge is reused to generate responses and incremental updates are performed. For partially complete or incomplete problems, knowledge gaps are indicated. If the threshold is not reached, multi-dimensional related knowledge elements are retrieved, and the completeness is determined by calculating multi-dimensional data coverage using a preset knowledge completeness measurement model. Finally, the intelligent processing system generates targeted responses based on the knowledge completeness. For incomplete responses, gaps are indicated and a knowledge construction task is generated. For partially complete responses, the confidence level of the inference chain is marked and a data supplementation task is performed. For highly complete responses, an inference path and automated workflow are provided. Finally, the intelligent processing system updates the knowledge graph and the knowledge completeness database. When synchronization conditions are met, it sends data to multiple system nodes and receives external data to update the local database, achieving collaborative knowledge evolution. This method alleviates the technical problem of incomplete and distorted content in industrial intelligent responses, improves the accuracy of industrial intelligent question answering, optimizes the logic for reusing historical knowledge, and enhances the dynamic improvement capability of knowledge reserves and the level of cross-node collaborative processing.

[0105] The methods provided in the above embodiments can be executed by an intelligent processing system. The intelligent processing system in the embodiments of this invention is described below from a hardware processing perspective; please refer to [link / reference]. Figure 4 This is a schematic diagram of the physical device structure of an intelligent processing system in an embodiment of this application.

[0106] It should be noted that, Figure 4 The structure of the intelligent processing system shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of the present invention.

[0107] like Figure 4As shown, the intelligent processing system includes a Central Processing Unit (CPU) 301, which can perform various appropriate actions and processes based on programs stored in Read-Only Memory (ROM) 302 or programs loaded from storage portion 308 into Random Access Memory (RAM) 303, such as performing the methods described in the above embodiments. The RAM 303 also stores various programs and data required for system operation. The CPU 301, ROM 302, and RAM 303 are interconnected via a bus 304. An Input / Output (I / O) interface 305 is also connected to the bus 304.

[0108] The following components are connected to I / O interface 305: input section 306 including audio input devices, push-button switches, etc.; output section 307 including a liquid crystal display (LCD) and audio output devices, indicator lights, etc.; storage section 308 including a hard disk, etc.; and communication section 309 including a network interface card such as a LAN (Local Area Network) card, modem, etc. Communication section 309 performs communication processing via a network such as the Internet. Drive 310 is also connected to I / O interface 305 as needed. Removable media 311, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 310 as needed so that computer programs read from them can be installed into storage section 308 as needed.

[0109] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing computer programs for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 309, and / or installed from removable medium 311. When the computer program is executed by central processing unit (CPU) 301, it performs the various functions defined in the present invention.

[0110] It should be noted that specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0111] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. Each block in a flowchart or block diagram may represent a module, program segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those shown in the drawings.

[0112] Specifically, the intelligent processing system in this embodiment includes a processor and a memory. The memory stores a computer program. When the computer program is executed by the processor, it implements the knowledge evolution-based large model-driven industrial intelligent operation method provided in the above embodiment.

[0113] In another aspect, the present invention also provides a computer-readable storage medium, which may be included in the intelligent processing system described in the above embodiments; or it may exist independently and not assembled into the intelligent processing system. The storage medium carries one or more computer programs, which, when executed by a processor of the intelligent processing system, enable the intelligent processing system to implement the knowledge evolution-based large-model-driven industrial intelligent operation method provided in the above embodiments.

[0114] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0115] As used in the above embodiments, depending on the context, the term "when..." can be interpreted as meaning "if...", "after...", "in response to determining...", or "in response to detecting...". Similarly, depending on the context, the phrase "when determining..." or "if (the stated condition or event) is interpreted as meaning "if determining...", "in response to determining...", "when (the stated condition or event) is detected", or "in response to detecting (the stated condition or event)".

[0116] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive), etc.

[0117] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM or random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A method for industry intelligent operation driven by a large model based on knowledge evolution, characterized in that, include: The system receives a target question submitted by a user and uses a large language model to perform semantic parsing on the target question to obtain the object entities, question intent, and question type involved in the target question. The object entity is mapped to the corresponding ontology node in the domain ontology library to generate a semantic feature vector containing ontology node combination, question intent and question type; Calculate the feature similarity between the semantic feature vector and the features of historical problems in the knowledge completeness result library, wherein the knowledge completeness result library is used to store the semantic feature vectors of historical problems and the corresponding knowledge completeness. When the feature similarity does not reach the preset similarity threshold, multidimensional related knowledge elements associated with the target question are retrieved from the knowledge graph. The multidimensional related knowledge elements include at least triples, attributes, and historical question and answer records. Based on the multidimensional related knowledge elements associated with the target problem, a preset knowledge completeness measurement model is used to calculate the data coverage of the target problem in multiple dimensions to obtain the current knowledge completeness level, which includes incomplete, partially complete, and highly complete. Based on the aforementioned knowledge completeness, the large language model is invoked to generate a response text corresponding to the target question; Based on the response text, update the knowledge graph, and based on the semantic feature vector of the target question, the knowledge completeness, and the response text, update the knowledge completeness result database. Monitor data changes in the knowledge completeness result database and send newly added or changed knowledge completeness result data to multiple system nodes within the communication range, including enterprise nodes and research institution nodes.

2. The method of claim 1, wherein, After calculating the feature similarity between the semantic feature vector and the historical problem features in the knowledge completeness result base, the method further includes: When the feature similarity reaches the preset similarity threshold, it is determined that there is a historical problem in the knowledge completeness result library that belongs to the same knowledge category as the target problem, and the knowledge completeness of the historical problem is obtained; If the knowledge completeness of the historical issue is highly complete, then the existing knowledge in the knowledge graph is called, the large language model is used to generate the response text, and an incremental update operation is performed. If the knowledge completeness of the historical problem is partially complete or incomplete, a first prompt message is generated, which is used to indicate that there is a gap in the current domain knowledge.

3. The method of claim 1, wherein, Based on the multidimensional associated knowledge elements related to the target problem, a preset knowledge completeness measurement model is used to calculate the data coverage of the target problem in multiple dimensions, thereby obtaining the current knowledge completeness level, specifically including: Based on the multidimensional related knowledge elements associated with the target problem, the data coverage of the target problem in multiple dimensions is calculated using the preset knowledge completeness measurement model. The multiple dimensions include at least performance dimension, operating condition dimension, mechanism dimension and solution dimension. The performance dimension and the operating condition dimension are basic dimensions, and the mechanism dimension and the solution dimension are extended dimensions. If the coverage of the basic dimension does not reach the first preset coverage threshold, then the knowledge completeness of the target problem is determined to be incomplete. If the coverage of the basic dimension reaches the first preset coverage threshold and the coverage of the extended dimension does not reach the second preset coverage threshold, then the knowledge completeness of the target problem is determined to be partially complete. If the coverage of the basic dimension reaches the first preset coverage threshold and the coverage of the extended dimension reaches the second preset coverage threshold, then the knowledge completeness of the target problem is determined to be highly complete.

4. The method of claim 3, wherein, Based on the stated knowledge completeness, the large language model is invoked to generate a response text corresponding to the target question, specifically including: When the knowledge completeness is incomplete, the large language model is invoked to generate an initial response text containing a second prompt message based on the existing knowledge. The second prompt message is used to indicate that the data coverage has not reached the first preset coverage threshold of the basic dimension. Based on the dimensions contained in the second prompt information, a knowledge gap task is generated; When the knowledge completeness is partially complete, the large language model is invoked to generate an inference chain and answer content based on the known relationships in the knowledge graph, and the confidence level of each part in the inference chain is calculated. Confidence levels below a preset confidence threshold are marked at the corresponding positions in the inference chain, and combined with the answer content, an initial response text is generated; Based on the extended dimensions where the data coverage does not reach the second preset coverage threshold, a supplementary data task is generated. When the knowledge completeness level is incomplete or partially complete, based on the question intent of the target question and the corresponding initial response text, a knowledge building task following the DIKW principle is generated, and the large language model is used to generate automated workflow building suggestions based on the results of the building task. The automated workflow construction suggestions are added to the corresponding initial response text to obtain the final response text; When the knowledge completeness is highly complete, the large language model is invoked, and an initial response text is generated based on the knowledge graph. The initial response text includes a reasoning path and text, and the text includes a mark of the knowledge source. The reasoning path is stored as a complete knowledge path in the knowledge graph. Simultaneously, based on the question intent of the target question and the initial response text, a first automated workflow matching the question intent and the initial response text is retrieved from the workflow repository, or a second automated workflow is generated using the large language model according to the DIKW principle. The first automated workflow or the second automated workflow is added to the initial response text to obtain the final response text.

5. The method of claim 1, wherein, Calculating the feature similarity between the semantic feature vector and the historical problem features in the knowledge completeness result base specifically includes: Obtain the historical ontology node combinations, historical problem intents, and historical problem types from the knowledge completeness result database; Calculate the overlap between the ontology node combination and the historical ontology node combination in the semantic feature vector, the semantic similarity between the question intent and the historical question intent, and the consistency between the question type and the historical question type, respectively. The overlap, semantic similarity, and consistency are weighted and calculated to obtain the feature similarity between the semantic feature vector and the historical problem features.

6. The method of claim 4, wherein, Based on the response text, the knowledge graph is updated, and based on the semantic feature vector of the target question, the knowledge completeness, and the response text, the knowledge completeness result database is updated, specifically including: Extract the newly generated triples, relational structures, and explanatory structures from the response text, and add the triples, relational structures, and explanatory structures to the knowledge graph; Establish a mapping relationship between the semantic feature vector of the target problem, the knowledge completeness, the gap dimension description information, and the timestamp; The mapping relationship is stored as a new knowledge completion result in the knowledge completion result library.

7. The method of claim 1, wherein, Monitoring data changes in the knowledge completeness result database and sending newly added or changed knowledge completeness result data to multiple system nodes within the communication range, specifically including: Determine whether the knowledge completeness result library meets the preset synchronization conditions. The preset synchronization conditions include the number of newly added highly complete record results exceeding a preset number threshold or the time since the last synchronization exceeding a preset duration. If the knowledge completeness result database meets the preset synchronization conditions, extract the newly added knowledge completeness result data and the changed knowledge completeness result data; The newly added knowledge completeness result data and the changed knowledge completeness result data are sent to system nodes within the communication range, and external completeness result data sent by the system nodes are received. The external complete result data is inspected. When the same problem is detected as highly complete in the external complete result data but incomplete or partially complete in the local record, the local knowledge complete result database is updated using the external complete result data.

8. An intelligent processing system, characterized in that, Includes one or more processors and memory; The memory is coupled to the one or more processors, the memory being used to store computer program code, the computer program code including computer instructions, the one or more processors invoking the computer instructions to cause the intelligent processing system to perform the method as described in any one of claims 1-7.

9. A computer-readable storage medium comprising computer instructions, characterized in that, When the computer instructions are executed on the intelligent processing system, the intelligent processing system performs the method as described in any one of claims 1-7.

10. A computer program product, comprising a computer program / instructions, characterized in that, When the computer program / instructions are run on the intelligent processing system, the intelligent processing system performs the method as described in any one of claims 1-7.

Images