Industrial Agent integrated platform

Abstract

The invention relates to the technical field of industrial intellectualization, in particular to an industrial Agent integrated platform which comprises a knowledge graph construction module, a model training module, an intelligent reasoning module, a user-defined development module and a task output module. According to the method, a unified enterprise knowledge graph set covering multiple dimensions of business, operation, scenes, systems and the like is constructed by automatically collecting and fusing structured, unstructured and even multi-modal industrial data, comprehensive and deep cognition on the industrial environment can be formed, simple keyword matching is surpassed, and the method has the advantages of being high in practicability and easy to popularize. According to the method, accurate understanding of complex intentions behind natural language interaction of a user is achieved, the system can automatically generate a one-step deployment instruction set or construct a complete AI Agent task including data query, code logic and flow arrangement based on the deep understanding, and further, the system can be used for realizing the complete deployment of the user. The core language model can dynamically eliminate redundant information based on the semantic association strength and the actual use frequency of the fields.

Description

Technical Field

[0001] This invention relates to the field of industrial intelligent technology, and in particular to an integrated industrial agent platform. Background Technology

[0002] The field of industrial intelligent technology mainly involves the deep integration of information technology and manufacturing, and the use of artificial intelligence, big data, Internet of Things and other means to improve the intelligence level of industrial production systems, so as to realize the automation of production processes, intelligent decision-making and system collaborative optimization.

[0003] Traditional enterprise-level intelligent agent platforms refer to rule engines or script-driven systems deployed in industrial scenarios for specific task processing. They mainly achieve functions such as equipment control and data question answering through manually configured knowledge bases and fixed process input / output interfaces. They typically use knowledge rule table design, equipment operating parameter mapping, custom query statement construction, and dictionary-based intent recognition methods to solve task response problems for industrial data.

[0004] Existing technologies, relying on manually configured knowledge bases and fixed rule engines, reveal shortcomings in integration and adaptability when facing industrial scenarios with continuously increasing data source types or frequent changes in data formats. Each new type of data access requires extensive manual rule mapping and process reconstruction, easily leading to fragmentation and inconsistency of the knowledge system. Furthermore, the dictionary-based intent recognition method used struggles to perform effective semantic association and reasoning when dealing with user queries that are colloquial, ambiguous, or contain out-of-domain terms. Users must strictly adhere to preset query templates, significantly limiting the naturalness and flexibility of interaction. For example, when the production line updates equipment models or introduces new parameter terms in the process flow documents, the entire system must be manually maintained and updated by professionals, otherwise, commands may not be recognized or incorrect data may be returned. This rigid architecture makes it difficult to adapt to the dynamic and ever-changing needs of modern industrial environments. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies and propose an integrated industrial agent platform.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: The integrated industrial agent platform includes: The knowledge graph construction module collects enterprise industrial data from the enterprise production and operation information system, divides the enterprise industrial data into structured and unstructured data, constructs multiple types of enterprise knowledge graphs, and integrates them into an enterprise knowledge graph set. The model training module obtains the field information and the graph identifier corresponding to each type of enterprise knowledge graph in the enterprise knowledge graph set, and inputs it as a training sample set into the BERT pre-trained language model for training and parameter learning, and generates the trained BERT language model. The intelligent reasoning module receives interactive information input by enterprise users, inputs it into the trained BERT language model for semantic representation and context encoding, and outputs a one-click deployment instruction set. The custom development module receives the task construction target fields, behavioral structure expressions and result field requirements input by enterprise users, combines them with the field information of each type of enterprise knowledge graph in the enterprise knowledge graph set, and outputs a set of AI Agent task statement structures. The task output module connects the one-click deployment instruction set and the AI ​​Agent task statement structure set to the platform instruction execution interface in parallel to generate platform output results.

[0007] As a further aspect of the present invention, the enterprise knowledge graph set includes a business domain knowledge graph, an AGI (Area-of-Knowledge Index) business indicator knowledge graph, a scenario SOP (Standard Operating Procedure) knowledge graph, a system domain knowledge graph, and a custom knowledge graph. The trained BERT language model specifically includes field context representation vectors, field frequency parameters, and pruning and retention tags. The one-click deployment instruction set includes inference path configuration parameters, deployment execution trigger instructions, and learning rate adjustment instructions. The AI ​​Agent task statement structure set includes syntax structure generation statements, job flow logic structures, and query statement mapping results. The platform output results specifically refer to the model deployment structure set and the Agent behavior output structure set.

[0008] As a further aspect of the present invention, the knowledge graph construction module includes: The industrial data acquisition submodule collects field structures and data table relationships from MySQL, PgSQL, Hadoop, and Odps in the enterprise's production and operation information system, and obtains character sequences and multimodal content features from doc, pdf, txt, png, and mp4 files. Based on the differences in data formats, it divides and forms structured and unstructured industrial data. The multi-graph construction submodule, based on the structured and unstructured industrial data, processes the structured data using field boundary recognition and the unstructured data using semantic content segmentation. Based on the content attribution logic, it constructs a business domain knowledge graph, an AGI business indicator knowledge graph, a scenario SOP knowledge graph, a system domain knowledge graph, and a custom knowledge graph, integrating them into a multi-type enterprise knowledge graph. The knowledge graph integration submodule compares the entities and relationships between the various types of enterprise knowledge graphs, calculates the semantic overlap between entities and the consistency between relationships, merges entities with an overlap exceeding a preset threshold, adjusts conflicting relationships, and establishes an enterprise knowledge graph set.

[0009] As a further aspect of the present invention, the model training module includes: The training sample construction submodule obtains the field nodes, connection paths between field nodes, and graph identifiers corresponding to each type of enterprise knowledge graph in the enterprise knowledge graph set, and constructs records with field context relationships and graph source types as training sample sets; The model vector learning submodule inputs the training sample set into the BERT pre-trained language model, extracts the vector representation information of each field node in the context and the field node position embedding vector during the runtime of the BERT pre-trained language model, monitors the frequency of each field occurrence and calculates the frequency distribution interval, evaluates the embedding space distance between fields, and obtains the field vector and space distance parameters. The language model generation submodule combines the embedded spatial distance and frequency distribution interval as the pruning criteria based on the field vector and spatial distance parameter, removes redundant fields with a frequency lower than the frequency threshold, and generates the trained BERT language model.

[0010] As a further aspect of the present invention, in the business domain knowledge graph, the field nodes correspond to the information of enterprise production elements and business entities, including equipment name, equipment model, material code, product specifications, process flow parameters, and supplier name. In the AGI (Area-of-Knowledge Graph) of business indicators, the field nodes correspond to key performance indicators at the enterprise operation and management level, including cost indicators, production data, overall equipment efficiency, inventory turnover rate, and on-time order delivery rate. In the scenario SOP knowledge graph, the field nodes correspond to the steps and constraints in the standard operating procedure, including operation step descriptions, process execution sequence, fault handling plans, safety specifications, and standard operating time. In the system domain knowledge graph, the field nodes correspond to the technical architecture and data structure of the enterprise information system itself, including database table names, interface call addresses, service module names, log event IDs, and data field definitions; In the Custom knowledge graph, the field nodes correspond to non-standardized information defined by enterprise users according to specific analysis or management objectives, including project milestones, task owners, and budget utilization rates defined for a specific project.

[0011] As a further aspect of the present invention, the process of eliminating redundant fields whose occurrence frequency is lower than the frequency threshold is specifically as follows: Based on the frequency of occurrence and the embedding space distance of the fields, the redundancy metric score of each field is calculated, and each field is divided into core fields, regular fields and rare fields according to the redundancy metric score. For the rare field, when the redundancy metric score exceeds the first preset redundancy threshold, the rare field is marked as a field to be removed. For the regular field, when the redundancy metric score exceeds the second preset redundancy threshold, the regular field is marked as a field to be removed, wherein the second preset redundancy threshold is greater than the first preset redundancy threshold; No removal operation is performed on the core fields, and all rare fields and regular fields marked as fields to be removed are aggregated.

[0012] As a further aspect of the present invention, the intelligent reasoning module includes: The reasoning input construction submodule receives interactive information input by enterprise users, including natural language task descriptions, target field combinations, and action keywords. It matches and filters the field nodes of each type of enterprise knowledge graph in the enterprise knowledge graph set to obtain the matched field information and construct a reasoning input field sequence. The stride adjustment instruction submodule, based on the inference input field sequence, combines the context vector and frequency features in the field vector and spatial distance parameter to calculate the switching interval period value between fields. When the frequency of field switching interval changes is higher than the set period threshold, a stride adjustment instruction is constructed. The instruction set output submodule combines the inference input field sequence with the stride adjustment instruction, inputs it into the inference path of the trained BERT language model for path generation and response judgment, performs semantic representation and context encoding, and outputs a one-click deployment instruction set.

[0013] As a further aspect of the present invention, the custom development module includes: The custom field input submodule receives the task construction target fields, behavioral structure expressions and result field requirements input by enterprise users, matches the field path combinations related to the target task for each type of enterprise knowledge graph in the enterprise knowledge graph set, decomposes the semantic association relationship between fields and identifies semantic roles, and establishes a custom field input set. The task logic generation submodule inputs the custom field input set into the AICoding engine to generate function statement structure, into the SOP path engine to construct task step logic, and into the text2sql conversion engine to generate SQL query logic expression, thereby obtaining multi-dimensional task logic statements; The statement structure integration submodule, based on the multi-dimensional task logic statement, integrates the function statement structure, task step logic, and SQL query logic expression, and associates and arranges the statements generated by each engine to match the user's input task construction target fields, behavioral structure expressions, and result field requirements, and outputs a set of AI Agent task statement structures.

[0014] As a further aspect of the present invention, the task output module includes: The deployment instruction processing submodule connects the one-click deployment instruction set to the platform instruction execution interface in parallel to generate a model deployment structure set. The Agent statement parsing submodule connects the AI ​​Agent task statement structure set to the platform instruction execution interface in parallel to obtain the Agent behavior output structure set. The platform instruction execution submodule, based on the model deployment structure set and the agent behavior output structure set, executes the model deployment inference process and the agent development task process respectively through the platform instruction execution interface, and generates platform output results.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, by automatically collecting and fusing structured, unstructured, and even multimodal industrial data, a unified enterprise knowledge graph collection covering multiple dimensions such as business, operations, scenarios, and systems is constructed. This enables a comprehensive and in-depth understanding of the industrial environment, going beyond simple keyword matching to achieve a precise understanding of the complex intentions behind users' natural language interactions. Based on this deep understanding, the system can automatically generate a one-step deployment instruction set or construct a complete AI Agent task that includes data querying, code logic, and process orchestration. Furthermore, its core language model can dynamically eliminate redundant information based on the semantic association strength and actual usage frequency of fields, ensuring that the model can continuously focus on the core business of the enterprise while maintaining high efficiency and accuracy, thus realizing a transformation from passive response to proactive adaptation. Attached Figure Description

[0016] Figure 1 This is a platform flowchart of the present invention; Figure 2 This is a flowchart of the knowledge graph construction module of the present invention; Figure 3 This is a flowchart of the model training module of the present invention; Figure 4 This is a flowchart of the intelligent reasoning module of the present invention; Figure 5 The flowchart is for the custom development module of this invention; Figure 6 This is a flowchart of the task output module of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0018] Please see Figure 1 The integrated industrial agent platform includes: The knowledge graph construction module collects enterprise industrial data from the enterprise production and operation information system, divides the enterprise industrial data into structured and unstructured data, constructs multiple types of enterprise knowledge graphs, and integrates them into an enterprise knowledge graph set. The model training module obtains the field information and the corresponding graph identifier of each type of enterprise knowledge graph in the enterprise knowledge graph set, and inputs them as training sample set into the BERT pre-trained language model for training and parameter learning, and generates the trained BERT language model. The intelligent inference module receives interactive information input by enterprise users, inputs it into the trained BERT language model for semantic representation and context encoding, and outputs a one-click deployment instruction set. The custom development module receives the task construction target fields, behavioral structure expressions and result field requirements input by enterprise users, combines them with the field information of each type of enterprise knowledge graph in the enterprise knowledge graph set, and outputs a set of AI Agent task statement structures. The task output module connects the one-click deployment instruction set and the AI ​​Agent task statement structure set to the platform instruction execution interface in parallel to generate the platform output results; The enterprise knowledge graph set includes business domain knowledge graph, AGI (Area-of-Knowledge Index) business indicator knowledge graph, scenario SOP (Standard Operating Procedure) knowledge graph, system domain knowledge graph, and custom knowledge graph. The trained BERT language model specifically includes field context representation vectors, field frequency parameters, and pruning and retention tags. The one-click deployment instruction set includes inference path configuration parameters, deployment execution trigger instructions, and learning rate adjustment instructions. The AI ​​Agent task statement structure set includes syntax structure generation statements, job flow logic structures, and query statement mapping results. The platform output results specifically refer to the model deployment structure set and the Agent behavior output structure set.

[0019] Please see Figure 2 The knowledge graph construction module includes: The industrial data acquisition submodule collects field structures and data table relationships from MySQL, PgSQL, Hadoop, and Odps in the enterprise's production and operation information system, and obtains character sequences and multimodal content features from doc, pdf, txt, png, and mp4 files. Based on the differences in data formats, it divides and forms structured and unstructured industrial data. Information is collected from the enterprise's production and operation information system. Specifically, the process first connects to MySQL and PostgreSQL relational databases, executes pre-defined SQL queries, and iterates through the "Equipment Information Table," "Bill of Materials Table," and "Production Work Order Table" within the system. It extracts the field structures defined in the table headers, such as "Equipment ID," "Equipment Name," and "Model" fields in the "Equipment Information Table." These fields represent the basic identity information and technical parameters of the equipment in the industrial data. Explicit foreign key relationships between fields are also captured to form a structured data source. Simultaneously, the system connects to the Hadoop Distributed File System and ODPS big data computing service, calling their data access interfaces to scan stored log files and production reports. It parses the column definitions within the files and their mapping relationships to the data tables. For example, it parses the field structures of "Timestamp," "Equipment ID," and "Temperature Reading" fields recording equipment operating status in the log files. For unstructured data, a file system crawler is launched to access specified doc, pdf, and txt format files for "Equipment Operation Manuals," "Fault Repair Reports," and "Process Instructions." A text extraction library is used to read the file content one by one, converting continuous text information into a character sequence stream. For PNG format "equipment structure diagrams" and "product defect diagrams," and MP4 format "equipment operation videos" and "worker operation videos," a multimodal content analysis interface is used to perform object recognition on the image data, extracting the contour features and spatial location information of key components in the images. For the video data, frame-by-frame processing is performed to identify operational actions and state changes within the video frames, encoding this visual information into feature vectors. After data acquisition, the data is categorized based on its source and inherent format. Data with clearly defined fields and table structures obtained from MySQL, PostgreSQL, Hadoop, and Odps is classified as structured industrial data, while character sequences and multimodal content features extracted from doc, pdf, txt, png, and mp4 files are classified as unstructured industrial data, ultimately forming two main categories: structured and unstructured industrial data.

[0020] The multi-graph construction submodule is based on structured and unstructured industrial data. It processes structured data using field boundary recognition and unstructured data using semantic content segmentation. Based on the content ownership logic, it constructs business domain knowledge graphs, AGI business indicator knowledge graphs, scenario SOP knowledge graphs, system domain knowledge graphs, and custom knowledge graphs, integrating them into a multi-type enterprise knowledge graph. After receiving structured and unstructured industrial data, a knowledge graph is constructed. For structured data, a field boundary identification method is used. Specifically, the "Equipment ID" and "Equipment Name" fields in the "Equipment Information Table" of the enterprise's production and operation information system are read, and each independent field name is identified as the boundary of an entity or attribute. For example, the value "Mixer A" under the "Equipment Name" field is identified as an entity. For unstructured data, a semantic content segmentation method is used. When processing a "Fault Repair Report" txt file, natural language processing technology is used to identify noun phrases and verb phrases in the text. For example, the sentence "The technician replaced the motor of Mixer A" is segmented into semantic units such as "technician," "replace," "mixer A," and "motor." After data processing is completed, five types of knowledge graphs are constructed based on content ownership logic. When constructing the business domain knowledge graph, the identified "Mixer A" entity and "Motor" entity, as well as the "Motor Model" attribute extracted from the "Bill of Materials" table, are connected through the relationship of "containing components." The fields in the business domain knowledge graph represent information about enterprise production factors and business entities in industrial data, including equipment name, equipment model, material code, product specifications, process parameters, and supplier name. When constructing the AGI (Area-of-Growth Index) business indicator knowledge graph, the "daily output" and "running time" data for "Mixer A" are extracted from production reports. This data is used as input for the "Equipment Overall Efficiency" indicator and associated with the "Mixer A" entity. The fields in the AGI business indicator knowledge graph represent key performance indicators (KPIs) at the enterprise operation and management level in industrial data, including cost indicators, output data, equipment overall efficiency, inventory turnover rate, and on-time order delivery rate. When constructing the scenario SOP (Standard Operating Procedure) knowledge graph, the description of "Mixer A Start-up Steps" in the "Equipment Operation Manual" is parsed, connecting operational steps such as "Step 1: Check Power Supply" and "Step 2: Press Start Button" as ordered nodes to form a standard operating procedure. The fields in the scenario SOP knowledge graph represent the steps and constraints in the standard operating procedures of industrial data, including operational step descriptions, process execution sequence, fault handling plans, safety regulations, and standard operating time. For the system domain knowledge graph, information such as the table name of the "Equipment Information Table" and the data type "Integer" for the "Equipment ID" field are extracted from the database metadata to construct a technical graph describing the structure of the data source itself. The fields in the system domain knowledge graph represent the technical architecture and data structure of the enterprise information system itself within the industrial data, including database table names, interface call addresses, service module names, log event IDs, and data field definitions. Finally, based on a "Progress Report for Specific Project A" document, a Custom knowledge graph is constructed, extracting predefined information such as "Project Milestone: Completed Equipment Debugging" and "Task Leader: Engineer Zhang" to establish a knowledge graph related to project management.Fields in the Custom knowledge graph refer to non-standardized information defined by enterprise users in industrial data according to specific analysis or management objectives. This includes project milestones, task owners, and budget utilization rates defined for a specific project. This process is repeated until all structured and unstructured industrial data is processed, ultimately integrating them into a multi-category enterprise knowledge graph that includes five categories: business, operations, scenarios, systems, and custom data.

[0021] The knowledge graph integration submodule compares the entities and relationships between multiple types of enterprise knowledge graphs, calculates the semantic overlap between entities and the consistency between relationships, merges entities with overlap exceeding a preset threshold, adjusts conflicting relationships, and establishes a set of enterprise knowledge graphs. Based on multiple enterprise knowledge graphs, entity and relation alignment and merging are performed. First, entities are compared: the entity "Equipment: Mixer A" from the business domain knowledge graph and the entity "Assessment Unit: Mixer A" from the AGI (Action Group Index) knowledge graph are selected, and their descriptive attributes are extracted. For example, the former's "Model: JB-001" and "Installation Location: Workshop 1" are used, while the latter's "Asset Number: ZC-9527" and "Production Line: Production Line 1" are used. Then, the semantic overlap between these two entities is calculated. This calculation process converts all attribute text of the entities into high-dimensional vectors, calculates their dot product, and then normalizes them. The calculation formula is: ; in, Representing entities and The semantic overlap between them ranges from 0 to 1. and These are two entities that need to be compared, such as "Equipment: Mixer A" and "Assessment Unit: Mixer A". and They are entities and The vector representation of attribute text in the word vector model. It is the dot product of two vectors. It is the product of the magnitudes of two vectors, with the first term calculating the cosine similarity of the entity description text. It is a physical entity and The exact match rate of key attributes between two entities is calculated by counting the proportion of key fields such as "ID" and "number" shared by the two entities out of the total number of key fields. The value ranges from 0 to 1. This is the attribute matching weight, used to balance the impact of semantic similarity and the precise matching of key attributes, and it is a dimensionless coefficient value. This weight value is set based on the fitting optimization results of one hundred manually labeled entity pairs to maximize classification accuracy. In specific settings, if the key attributes of entities in the business scenario (such as "asset number") are highly unique and the data quality is high, then it should be appropriately increased. The value should be adjusted to increase the impact of exact matches; conversely, if key attributes are missing or inconsistent, and the entity's descriptive text is more reliable, the value should be lowered. value.

[0022] 1. Parameter settings: The entity “Equipment: Mixer A” has the attribute text “Model JB-001 Installation location: Workshop No. 1”. The entity "Assessment Unit: Mixer A" has the attribute text "Asset No. ZC-9527 belongs to Production Line 1". Using the BERT model, the cosine similarity of the attribute text vectors of the two entities is calculated. The value is The key attribute list is defined as: "ID", "Number", "Model", and "Asset Number". Entity Entity with the key attribute "model" Both have the key attribute "Asset Number". They do not share any key attributes, therefore the shared attribute count is 0. The total number of key attributes is 2. Therefore... Based on business experience, set attribute matching weights. for .

[0023] 2. Substitute into the formula to calculate: .

[0024] In this step, the preset threshold is set based on the distribution of semantic overlap scores calculated from two hundred pairs of samples known to be the same or different entities. Experimental data shows that the overlap scores for different entities range from 0.1 to 0.5, while the overlap scores for the same entities range from 0.7 to 0.95. To ensure the accuracy of merging, the preset threshold is set to 0.8, which is the critical point between the two distribution ranges and can effectively distinguish between matching and non-matching entity pairs. Since the calculated overlap of 0.75 does not exceed the preset threshold of 0.8, the entities "Equipment: Mixer A" and "Assessment Unit: Mixer A" are considered different entities and are not merged. The advantage of this method is that it accurately matches by introducing key attributes. This approach compensates for potential biases caused by the diversity of description methods in pure semantic matching, thus improving the accuracy of entity merging. Next, relationship consistency is compared and adjusted to check for conflicts in the relationships of the merged new entities. For example, the business domain knowledge graph contains the relationship "Mixer A - Located in - Workshop 1," while the system domain knowledge graph, obtained through log analysis, contains the relationship "Mixer A - Reported Data Source - Workshop 2 Data Acquisition Terminal." These two relationships point to different geographical locations, constituting a conflict. The adjustment process involves querying the timestamp of the original data corresponding to each relationship. It was found that the data for "Workshop 2 Data Acquisition Terminal" is the most recent record, indicating a change in equipment location. Therefore, the relationship "Mixer A - Located in - Workshop 2" is retained, and the relationship "Mixer A - Located in - Workshop 1" is marked as a "historical location." By repeatedly performing the complete process of comparison, calculation, merging, and adjustment on all entities and relationships in the entire graph, a unified enterprise knowledge graph set without redundant entities and conflicting relationships is ultimately established.

[0025] Please see Figure 3 The model training module includes: The training sample construction submodule obtains the field nodes, connection paths between field nodes, and graph identifiers corresponding to each type of enterprise knowledge graph in the enterprise knowledge graph set, and constructs records with field context relationships and graph source types as training sample sets; In the business domain knowledge graph, the field nodes correspond to the information of enterprise production elements and business entities, including equipment name, equipment model, material code, product specifications, process parameters, and supplier name. In the AGI (Area-of-Knowledge Graph), the field nodes correspond to key performance indicators at the enterprise operation and management level, including cost indicators, production data, overall equipment efficiency, inventory turnover rate, and on-time order delivery rate. In the scenario SOP knowledge graph, field nodes correspond to the steps and constraints in the standard operating procedure, including operation step descriptions, process execution sequence, fault handling plans, safety specifications, and standard operating time. In the system domain knowledge graph, field nodes correspond to the technical architecture and data structure of the enterprise information system itself, including database table names, interface call addresses, service module names, log event IDs, and data field definitions; In the Custom knowledge graph, field nodes correspond to non-standardized information defined by enterprise users according to specific analysis or management objectives, including project milestones, task owners, and budget utilization rates defined for a specific project. After acquiring the enterprise knowledge graph set, the field information of each knowledge graph type in the set is extracted. Specifically, the business domain knowledge graph is traversed to extract field nodes, such as "Equipment Name" with the value "Mixer A"; field nodes represent the smallest unit of information or entity / attribute label in the industrial data. Next, the paths directly or indirectly connected to "Mixer A" are traced to obtain the inter-field connection paths, such as "Mixer A - Included Components - Motor" and "Motor - Own Parameters - Rated Power"; these inter-field connection paths represent the semantic association chain between entities, relationships, and entities or attributes in the industrial data. Finally, a graph identifier, "Business Domain," is appended to all fields along this path. The graph identifier corresponding to a field indicates the type of enterprise knowledge graph from which the field in the industrial data originates. Through this processing, a record is constructed: "[Equipment Name: Mixer A, Included Components: Motor, Own Parameters: Rated Power]; Graph Source: Business Domain". This process is performed on all fields and their connection paths in the enterprise knowledge graph set. For example, in the AGI (Aggregate Girdling Indicator) knowledge graph, the field "Equipment Overall Efficiency" is extracted, with the connection path "Mixer A - Calculated Indicator - Equipment Overall Efficiency," and the graph is identified as "AGI Business Indicator," thus constructing the record: "[Equipment Name: Mixer A, Calculated Indicator: Equipment Overall Efficiency]; Graph Source: AGI Business Indicator." Similarly, "Operation Step Description" is extracted from the Scenario SOP (Standard Operating Procedure) knowledge graph, constructing the record: "[Operation Name: Power On, Execution Sequence: Step 1, Operation Step Description: Check Power Supply]; Graph Source: Scenario SOP." "Database Table Name" is extracted from the System Domain knowledge graph, constructing the record: "[System Name: MES, Includes: Database Table Name: Equipment Information Table]; Graph Source: System Domain." "Project Milestone" is extracted from the Custom knowledge graph, constructing the record: "[Project Name: Production Line Upgrade, Includes: Project Milestone: Complete Equipment Debugging]; Graph Source: Custom." Repeat the above extraction and labeling process for all field nodes and connection paths in the entire graph to finally generate a set of records containing a large number of field context relationships and graph source types. This set is the training sample set used for subsequent model training.

[0026] The model vector learning submodule inputs the training sample set into the BERT pre-trained language model, extracts the vector representation information of each field node in the context and the field node position embedding vector during the runtime of the BERT pre-trained language model, monitors the frequency of each field occurrence and calculates the frequency distribution interval, evaluates the embedding space distance between fields, and obtains the field vector and spatial distance parameters. The training sample set is input into the BERT pre-trained language model. The BERT model adopts the Transformer Encoder structure, consisting of a multi-layer bidirectional self-attention mechanism and a feedforward neural network. Internally, each input record, such as "Device Name: Blender A, Components: Motor," is converted into a series of input tokens. The attention mechanism layer of the BERT model calculates the mutual influence weights between each field node (e.g., "Blender A") and all other field nodes (e.g., "Device Name," "Components," "Motor"). Through multi-layered attention calculations, the model generates a context-sensitive vector representation for the "Blender A" field node, capturing the semantics of "Blender A" in the specific context of "as a device name, and containing a motor." Simultaneously, the model generates a positional embedding vector for "Blender A" at its second position in the input sequence, representing its order information within the path. During the processing of the entire training sample set, the submodule continuously monitors the total occurrence count of each field (e.g., "Device Name," "Motor") and calculates the frequency distribution range based on the overall frequency of all fields. For example, statistics show that "device name" appears 10,000 times, "motor" appears 8,000 times, "rated power" appears 6,000 times, while a less frequently used "historical fault code" appears only 10 times. Next, the submodule evaluates the embedding space distance between fields. Specifically, it extracts the context vectors generated for "mixer A" and "motor" in the model, and then calculates the Euclidean distance between these two vectors in a high-dimensional space. A small distance value, such as 0.2, indicates that "mixer A" and "motor" are semantically closely related. By performing vector extraction and distance calculation on all field pairs in the sample set, a complete set of parameters is obtained. This set of parameters includes the final vector representation of each field and the spatial distance values ​​between all field pairs, together constituting the field vectors and spatial distance parameters.

[0027] The language model generation submodule combines the embedding spatial distance and frequency distribution interval as the pruning criteria based on the field vector and spatial distance parameter, removes redundant fields with a frequency lower than the frequency threshold, and generates the trained BERT language model. The process of removing redundant fields that occur less frequently than a frequency threshold is as follows: Based on the frequency of occurrence of the fields in each industrial application scenario, the fields are divided into core fields, regular fields, and rare fields; For rare fields, when the redundancy metric score exceeds the first preset redundancy threshold, the rare field is marked as a field to be removed. For regular fields, when the redundancy metric score exceeds the second preset redundancy threshold, the regular field is marked as a field to be removed, wherein the second preset redundancy threshold is greater than the first preset redundancy threshold; Do not perform removal operations on core fields, and summarize all rare fields and regular fields that are marked as fields to be removed; Based on field vectors and spatial distance parameters, language model pruning and optimization are performed. First, the embedding spatial distance of a field is combined with its frequency distribution range to determine if a field is redundant. Specifically, fields are first divided into three categories based on their total frequency of occurrence across all industrial application scenarios. Frequency statistics show that the total frequency of all fields ranges from 1 to 10,000 times. Based on this range, two frequency thresholds are set. Frequency threshold one (lower limit for core fields) is set at 80% of the total frequency, or 8,000 times. This value is determined based on the field hit rate in historical tasks; fields with a frequency exceeding this value are considered high-frequency dependencies in task execution. Frequency threshold two (upper limit for rare fields) is set at 2% of the total frequency, or 200 times; fields below this value are considered low-frequency, low-value fields. Fields appearing in the top 20% (e.g., more than 8,000 times) are defined as core fields, such as "Device Name"; fields appearing in the middle 60% (e.g., between 200 and 8,000 times) are defined as regular fields, such as "Rated Power"; and fields appearing in the bottom 20% (e.g., less than 200 times) are defined as rare fields, such as "Historical Fault Codes". Next, a redundancy metric score is calculated for each field, using the following method: ; in, Indicates the currently evaluated number fields The redundancy metric score has the same dimensions as the Euclidean distance and is a unitless value. It is the first one currently being evaluated Each field. It represents the total number of fields, for example, one thousand fields. It is the first one currently being evaluated fields and any field in the summation loop The Euclidean distance between context vectors is calculated as the square root of the sum of the squares of the differences in each dimension of the two vectors, and is a unitless positive value. Calculated fields The average embedding spatial distance to all other fields. It is a field The frequency of occurrence, for example, the rare field "historical fault code" appears 10 times. It represents the highest frequency of occurrence across all fields, for example, 10,000 times. This is a frequency penalty coefficient used to reduce the redundancy score of high-frequency fields, ensuring the consistency of units between the two terms in the formula. This coefficient is set based on optimization of the removal accuracy of 500 known redundant fields. The principle for setting this coefficient is: when it is necessary to prioritize retaining core fields that frequently appear in multiple business scenarios, it should be appropriately increased. The value can be adjusted to increase the penalty (reduction) for redundancy scores, making them less likely to be removed; conversely, if a greater emphasis is placed on the semantic isolation between fields, the value can be appropriately reduced. value.

[0028] 1. Parameter settings: : Rare field to be evaluated: "Historical Fault Codes". The total number of fields is set to Through calculation With the rest Calculate the Euclidean distances of the fields and average them, assuming the average distance is obtained. The value is . The frequency of occurrence of the field "Historical Fault Codes" is: . The highest frequency of occurrence among all fields is To be conservative, frequently used fields will be retained first, with a frequency penalty coefficient applied. Set as .

[0029] Substitute into the formula to calculate: .

[0030] Next, set the first preset redundancy threshold. Second preset redundancy threshold The setting of these two thresholds is based on the inflection point of the distribution after sorting the redundancy measurement scores of all fields. For example, statistical analysis of the scores of one thousand fields reveals a clear inflection point around 0.9. Therefore, the first preset redundancy threshold... Set to 0.9 to filter weakly related items in rare fields; second preset redundancy threshold. The redundancy score is set to 0.95, which is greater than the first preset redundancy threshold of 0.9, and is used to filter extremely weak correlations in regular fields. For the rare field "Historical Fault Codes," its calculated redundancy score is 1.4999. Because 1.4999 exceeds the first preset redundancy threshold of 0.9, "Historical Fault Codes" is marked as a field to be removed. For the regular field "Rated Power," if its redundancy score is 0.88, no action is taken because 0.88 does not exceed the second preset redundancy threshold of 0.95. Core fields, such as "Equipment Name," are not removed regardless of their redundancy score. The advantage of this method is that by introducing a frequency penalty coefficient... This approach allows the redundancy metric score to consider the actual usage frequency of fields while measuring the weakness of semantic relevance, avoiding the erroneous removal of occasionally occurring but crucial fields, thus achieving precise vocabulary reduction in the model. Finally, all rare and regular fields marked as to be removed are aggregated and their corresponding vectors are removed from the BERT language model's vocabulary, resulting in a more concise and efficient trained BERT language model.

[0031] Please see Figure 4 The intelligent reasoning module includes: The reasoning input construction submodule receives interactive information input by enterprise users, including natural language task descriptions, target field combinations, and action keywords. It matches and filters the field nodes of each type of enterprise knowledge graph in the enterprise knowledge graph set, obtains the matched field information, and constructs the reasoning input field sequence. The module initiates upon receiving interactive information from the enterprise user. Assume the user inputs a natural language task description: "Query the real-time power of the mixer on production line one," specifying the target field combination: "mixer," "power," and providing the action keyword: "query." The submodule first performs word segmentation on the input natural language task description, obtaining tokens such as "query," "production line one," "mixer," "real-time," and "power." Then, it matches these tokens, along with the user-specified target field combination "mixer" and "power," against field nodes of all categories in the enterprise knowledge graph set. The matching process is word-by-word. For example, it compares the string similarity of "mixer" with all entity values ​​under the "equipment name" field in the graph, finding a high match with "mixer A." Similarly, "power" matches field nodes such as "rated power" and "real-time power." The filtering process is based on the tightness of the match and contextual relevance. Finally, the submodule filters out the most relevant field information. For example, it matches the field node "Equipment Name: Mixer A" from the business domain knowledge graph and the field node "Performance Index: Real-time Power" from the AGI business indicator knowledge graph. After obtaining these matched field information, they are arranged according to the logical order of their appearance in the user input or the inherent association order in the graph, constructing an ordered sequence, namely the inference input field sequence: "[Equipment Name: Mixer A, Performance Index: Real-time Power]".

[0032] The stride adjustment instruction submodule, based on the inference input field sequence, combines the context vector and frequency features in the field vector and spatial distance parameter to calculate the switching interval period value between fields. When the frequency of field switching interval changes is higher than the set period threshold, a stride adjustment instruction is constructed. After obtaining the inference input field sequence, such as "[Device Name: Blender A, Performance Indicator: Real-time Power]", the operation begins. The submodule first extracts the context vector and frequency features of each field in the sequence from the field vectors and spatial distance parameters obtained during the previous model training phase. Specifically, it extracts the context vector of "Blender A" and its frequency features as a core field, as well as the context vector of "Real-time Power" and its frequency features as a regular field. Next, it calculates the switching interval period value between fields. This calculation is based on the Euclidean distance between the context vectors of the two fields in the embedding space and their normalized frequency difference. The calculation formula is: add the Euclidean distance between the two fields to the maximum Euclidean distance of all field pairs to obtain a distance summation term. Divide this distance summation term by the product of the Euclidean distance, the smoothing ratio of the normalized frequency difference, and the baseline period coefficient to obtain the switching interval period value. The smoothing ratio of the normalized frequency difference is calculated by adding a frequency difference smoothing term to the absolute difference of the normalized frequencies of the two fields, and then dividing by the frequency difference smoothing term. The larger the Euclidean distance, or the greater the difference in normalized frequencies, the smaller the calculated switching interval period value, indicating that the model needs a finer, smaller step size when processing these two fields. For example, the Euclidean distance between the fields "Blender A" (core field) and "Real-time Power" (regular field) is 0.3. The maximum Euclidean distance for all field pairs is set to 3.0. The normalized frequency difference smoothing term is set based on testing with 300 field pairs to ensure the smoothness and stability of the calculation results. It is set to a small, unitless value, such as 0.1. The principle for setting this smoothing term is: when the field frequency distribution is relatively concentrated and the differences are generally small, a smaller value should be set to amplify the impact of these subtle differences; conversely, when the frequency distribution is very sparse and the differences are significant, a slightly larger value can be set to avoid extreme frequency differences causing the period value to be too small. The baseline period coefficient is set based on the analysis of 500 historical inference tasks, statistically determining the optimal average switching interval. It is set to a unitless value, such as 20. The principle for setting this coefficient is: this coefficient directly affects the overall range of the output period value. If a larger default step size is desired for the model's inference and lower sensitivity to field switching, this coefficient should be increased; if a smaller default step size is desired and the model is more sensitive to semantic and frequency changes between fields, this coefficient should be decreased. Assuming the normalized frequency of "Blender A" is 0.9 and the normalized frequency of "Real-time Power" is 0.6, the absolute difference in normalized frequencies is 0.3. The smoothing ratio for the difference in normalized frequencies is calculated as follows: The calculated switching interval period is: Then, the calculated switching interval period value is compared with a predefined period threshold. The period threshold is determined by analyzing 500 historical inference tasks and statistically analyzing the optimal average switching interval value; for example, it is set to 40. In the current example, the calculated switching interval period value of 55 is higher than the set period threshold of 40, indicating a strong semantic association between "mixer A" and "real-time power," and the frequency of field switching interval changes is lower than the set period threshold. When the frequency of field switching interval changes is higher than the set period threshold (i.e., the calculated period value is less than the period threshold), a step adjustment instruction is constructed. For example, if the period value of another sequence "[Device Name: Mixer A, Historical Fault Code]" is calculated to be 10, which is lower than the period threshold of 40, a step adjustment instruction is constructed, with the instruction content being "reduce inference step size, coefficient 0.5".

[0033] The instruction set output submodule combines the inference input field sequence with stride adjustment instructions, inputs them into the inference path of the trained BERT language model for path generation and response judgment, performs semantic representation and context encoding, and outputs a one-click deployment instruction set. The inference input field sequence "[Device Name: Blender A, Performance Metric: Real-time Power]" is combined with a stride adjustment instruction. If no stride adjustment instruction was generated in the previous step, an empty instruction is generated. This combination process appends the stride adjustment instruction as metadata to a specific node transformation of the field sequence, forming a structured input object. This complete input object is then fed into the inference path of the trained BERT language model. Upon receiving the input, the model first performs semantic representation and context encoding on the field sequence, converting "Blender A" and "Real-time Power" into vector representations in a high-dimensional space. When the model's inference path moves from "Blender A" to "Real-time Power," the attached stride adjustment instruction is detected. If the instruction is "reduce inference stride, coefficient 0.5," the model dynamically adjusts the computational scope of its internal attention mechanism, focusing more on "Blender A," which is directly related to "Real-time Power," rather than other more distant fields in the sequence when calculating the context representation of "Real-time Power." Based on this, the model performs path generation and response judgment, predicting the action or code snippet most likely to be connected between these two fields. Based on the action keyword "query" and the meaning of the fields themselves, the model determines that this is a data query task. Ultimately, based on this determination, the model generates a series of standardized, directly executable instructions. For example, it generates a JSON-formatted instruction: "{'action': 'query', 'target': {'device': 'blender A', 'metric': 'real-time power'}, 'condition': 'latest'}". This JSON object is the final output one-click deployment instruction set.

[0034] Please see Figure 5 Custom development modules include: The custom field input submodule receives the task construction target fields, behavioral structure expressions and result field requirements input by enterprise users, matches the field path combinations related to the target task for each type of enterprise knowledge graph in the enterprise knowledge graph set, decomposes the semantic relationship between fields and identifies semantic roles, and establishes a custom field input set. The submodule receives task construction information input by enterprise users through the interface, including the target field "monthly production report", the behavioral structure expression "statistics on the total production of all mixers and categorizes them by model", and the result field requirements "model, total production". The submodule first matches the enterprise knowledge graph set based on the target field "monthly production report" and keywords such as "mixer", "production", and "model" in the behavioral structure. It traverses the business domain knowledge graph and the AGI (Area-of-Knowledge) business indicator knowledge graph, searching for field path combinations related to these keywords. For example, it finds the field connection path "equipment-ownership attribute-model" in the business domain knowledge graph and "equipment-generated data-production" in the AGI business indicator knowledge graph. The submodule combines these paths to form a preliminary field network. Next, it decomposes the semantic relationships between fields and identifies semantic roles. In the behavioral structure expression "statistics on the total production of all mixers and categorizes them by model", it identifies "statistics" as the main action, "total production" as the operation object, "mixer" as the limitation of the operation scope, and "categorize by model" as the grouping basis. In this way, the natural language commands input by the user are broken down into structured semantic roles. Finally, the identified field path combinations (such as "equipment.model" and "equipment.output") are integrated with the identified semantic roles (such as action: statistics, grouping basis: model) to establish a set of custom field inputs that contains all the necessary information and has a clear structure, thus preparing for the subsequent generation of task logic.

[0035] The task logic generation submodule takes the custom field input set and inputs it into the AICoding engine to generate function statement structure, into the SOP path engine to build task step logic, and into the text2sql conversion engine to generate SQL query logic expression, thus obtaining multi-dimensional task logic statement; After receiving the custom field input set, its contents are distributed in parallel to three different engines for processing. First, the logic related to data processing and calculation, such as "total production" and "classification by model," is input to the AICoding engine. The AICoding engine contains a code generation model that generates specific function statement structures based on the semantic description of the input. For example, for "total production and classification by model," it generates a Python pseudocode: "function calculate_production(data): group_by(data, 'model').sum('production')". Second, the part of the set describing the operation sequence and process is input to the SOP path engine. If the user's requirements involve a series of steps, such as "export data first, then perform statistics, and finally generate charts," the SOP path engine constructs a task step logic, defining the input, output, and execution order of each step, forming a process definition file. For the simple task of "statistical production," the step logic may only involve single-step calculations. Finally, the part of the set involving data querying, i.e., which data source and which fields need to be retrieved, is input to the text2sql conversion engine. Based on the requirement of "obtaining the production and model data of all mixers", the text2sql transformation engine analyzes its relationship with the database tables "Production Record Table" and "Equipment Information Table" in the knowledge graph, and generates the corresponding SQL query logic expression: "SELECT T1.Model, SUM(T2.Production) FROM Equipment Information Table AST1 JOIN Production Record Table AST2 ON T1.Equipment ID = T2.Equipment ID WHERET1.Equipment Type = 'Mixer' GROUP BY T1.Model". Through the parallel processing of these three engines, a set of multi-dimensional task logic statements containing function definitions, step flow, and data query statements is finally obtained.

[0036] The statement structure integration submodule integrates function statement structure, task step logic and SQL query logic expression based on multi-dimensional task logic statements. It associates and arranges the statements generated by each engine to match the user's input task construction target fields, behavior structure expression and result field requirements, and outputs a set of AI Agent task statement structures. After acquiring the multi-dimensional task logic statements, integration and orchestration begin. The submodule first examines the function statement structure generated by the AICoding engine, the task step logic constructed by the SOP path engine, and the SQL query logic expression generated by the text2sql conversion engine. The goal of integration is to establish the relationship between these three to fully match the user's initial input requirements for task construction target fields, behavioral structure expressions, and result fields. The orchestration process is as follows: First, the SQL query logic is used as the source of data input, and its execution is the first step. Then, the result of the SQL query (a dataset containing "model" and "total output") is used as the input to the "data" parameter in the function statement structure generated by the AICoding engine. In this way, the function statement structure is associated with the SQL query logic expression. The entire task execution flow is orchestrated as follows: first, the SQL query is executed to retrieve raw data from the database; then, the query result is passed to the generated function for calculation and processing. Finally, the integrated logic is encapsulated into a complete, executable task unit. The output of this unit (i.e., the function's return value) is matched and validated against the user's required result fields, "model" and "total output." Once confirmed, a structured set of AI Agent task statements is output. This set clearly defines where the data comes from (SQL), how it is processed (function logic), and the task execution order (SOP logic), forming a complete and automated task solution.

[0037] Please see Figure 6 The task output module includes: The deployment instruction processing submodule connects the one-click deployment instruction set to the platform instruction execution interface in parallel to generate a model deployment structure set; The submodule receives a one-click deployment instruction set output by the intelligent inference module, such as a JSON object: "{'action': 'query', 'target': {'device': 'blender A', 'metric': 'real-time power'}, 'condition': 'latest'}". The submodule then connects this instruction set to the platform instruction execution interface in parallel. "Parallel connection" means that if multiple inference tasks generate instruction sets simultaneously, the submodule can process these instructions concurrently, generating a task deployment structure independently for each instruction. Upon receiving the JSON object, the platform instruction execution interface parses it. First, it identifies the top-level action "action" as "query," i.e., data query. Next, it parses the "target" field to determine that the query target is the "real-time power" of "blender A". Finally, it parses the "condition" field, explicitly requesting the latest data. Based on this parsed information, the platform instruction execution interface matches and instantiates a task template for real-time data monitoring from a pre-defined template library. This instantiation process generates a model deployment structure set, which is specifically represented as a configured, executable data stream task. This task defines the data source (the sensor data interface connected to "Blender A"), the data processing node (extracting the "Real-time Power" field), and the data output endpoint (returning the data to the user interface). The final generated model deployment structure set is then submitted to the platform's execution engine, awaiting scheduling and execution.

[0038] The Agent statement parsing submodule connects the AI ​​Agent task statement structure set to the platform instruction execution interface in parallel to obtain the Agent behavior output structure set. The system receives a set of AI Agent task statement structures output by the custom development module. This set contains a series of integrated instructions, such as SQL query statements, Python function code, and task execution flow definitions. The submodule also connects this set of structures to the platform instruction execution interface in parallel. Upon receiving this set, the platform instruction execution interface begins to parse its contents one by one. First, the interface identifies the SQL query statement and understands that it is a data extraction request. Next, it parses the Python function code and analyzes that it is a logic of data aggregation and computation. Finally, it parses the task execution flow definition, clarifying that the SQL query must be executed first, and then the query result is used as input to execute the Python function. Based on the parsing results, the platform instruction execution interface dynamically constructs a multi-step Agent behavior. This construction process is the core of obtaining the Agent behavior output structure set; it combines the originally separate statements into an executable workflow with internal logic and data flow. For example, the interface might create a workflow with two steps: first, configure a database connector and execute the specified SQL; second, configure a code executor, take the output of the first step as input, and execute the specified Python function. This fully defined workflow is the final set of Agent behavior output structures obtained.

[0039] The platform instruction execution submodule, based on the model deployment structure set and the agent behavior output structure set, executes the model deployment inference process and the agent development task process respectively through the platform instruction execution interface, and generates platform output results. Based on the model deployment structure set and the Agent behavior output structure set, the actual execution process is initiated. The platform instruction execution interface first processes the model deployment structure set, submitting the structure describing the real-time data query task to the platform's model deployment inference process. This process mobilizes underlying computing resources to establish a data pipeline from the sensor to the user end, continuously acquiring the "real-time power" data of "mixer A". Based on the known power range of industrial mixers... arrive Between these points, assuming the real-time power value is... This value As a platform output, this is the real-time monitoring data that the user expects to obtain through the instruction set automatically generated by the intelligent inference module. Simultaneously, the platform instruction execution interface also processes the Agent behavior output structure set, submitting the workflow describing the custom statistical task to the Agent development task flow. This flow executes the steps in the workflow sequentially: First, it initiates an SQL query to the production database to obtain the model and output data of all mixers; then, it sends the returned data table to the code executor, running a Python function to perform grouping and summation calculations. After the calculation, the result is, for example, a dataset containing "{Model: JB-001, Total Output: 5000}, {Model: JB-002, Total Output: 7200}". This dataset is returned as another platform output, representing the business statistical report that the user expects to obtain through the custom development module. Finally, the platform presents the results corresponding to the instructions generated by the user through different modules on the user's monitoring dashboard and report pages, respectively.

[0040] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. An industrial agent integration platform, characterized in that, The platform comprises: A knowledge graph construction module, which collects enterprise industrial data from an enterprise production and operation information system, divides structured and unstructured enterprise industrial data, constructs multiple types of enterprise knowledge graphs, and integrates them into an enterprise knowledge graph set; A model training module, which obtains field information and graph identifiers corresponding to the fields of each type of enterprise knowledge graph in the enterprise knowledge graph set as a training sample set, inputs them into a BERT pre-training language model for training and parameter learning, and generates a trained BERT language model; An intelligent reasoning module, which receives interactive information input by an enterprise user, inputs it into the trained BERT language model for semantic representation and context encoding, and outputs a one-key deployment instruction set; A self-defined development module, which receives a task construction target field, behavior structure expression and result field demand input by an enterprise user, combines them with the field information of each type of enterprise knowledge graph in the enterprise knowledge graph set, and outputs an AI Agent task sentence structure set; A task output module, which accesses a platform instruction execution interface in parallel with the one-key deployment instruction set and the AI Agent task sentence structure set, and generates a platform output result.

2. The industrial agent integration platform of claim 1, wherein, The enterprise knowledge graph set comprises a business domain knowledge graph, an AGI operating indicator knowledge graph, a scene SOP knowledge graph, a system domain knowledge graph, and a Custom self-defined knowledge graph, the trained BERT language model specifically comprises a field context representation vector, a field frequency parameter, and a pruning retention mark, the one-key deployment instruction set comprises a reasoning path configuration parameter, a deployment execution trigger instruction, and a learning rate adjustment instruction, and the AI Agent task sentence structure set comprises a syntax structure generation sentence, a job flow logical structure, and a query sentence mapping result, and the platform output result specifically refers to a model deployment structure set and an Agent behavior output structure set.

3. The industrial agent integration platform of claim 1, wherein, The knowledge graph construction module comprises: An industrial data collection submodule, which collects field structures and data table relationships in Mysql, Pgsql, Hadoop and Odps from an enterprise production and operation information system, and obtains character sequences and multi-modal content features in doc, pdf, txt, png and mp4, divides and forms structured and unstructured industrial data according to data format differences; A multi-type graph construction submodule, which processes structured data by field boundary recognition and unstructured data by semantic content segmentation based on the structured and unstructured industrial data, and constructs a business domain knowledge graph, an AGI operating indicator knowledge graph, a scene SOP knowledge graph, a system domain knowledge graph and a Custom self-defined knowledge graph according to content attribution logic, and integrates them into a multi-type enterprise knowledge graph; A knowledge graph integration submodule, which compares entities and relationships between each type of enterprise knowledge graph according to the multi-type enterprise knowledge graph, calculates the semantic overlap degree between entities and the consistency between relationships, merges entities with an overlap degree exceeding a preset threshold, adjusts conflicting relationships, and establishes an enterprise knowledge graph set.

4. The industrial agent integration platform of claim 3, wherein, The model training module comprises: The training sample construction submodule obtains the field nodes, the connection paths between the field nodes, and the graph identifiers corresponding to the fields of each type of enterprise knowledge graph in the set of enterprise knowledge graphs, constructs records with field context relationships and graph source types, and takes the records as a training sample set; The model vector learning submodule inputs the training sample set into a BERT pre-training language model, extracts vector expression information of each field node in a context and a field node position embedding vector when the BERT pre-training language model runs, monitors the frequency of occurrence of each field and calculates a frequency distribution interval, evaluates the embedding space distance between fields, and obtains field vectors and space distance parameters; The language model generation submodule combines the embedding space distance and the frequency distribution interval as a pruning judgment basis according to the field vectors and the space distance parameters, removes redundant fields with a frequency of occurrence lower than a frequency threshold, and generates a trained BERT language model.

5. The industrial agent integration platform of claim 4, wherein, In the business domain knowledge graph, the field nodes correspond to information of enterprise production factors and business entities, including device names, device models, material codes, product specifications, process flow parameters, and supplier names. In the AGI operation index knowledge graph, the field nodes correspond to key performance indicators of enterprise operation and management, including cost indicators, yield data, device comprehensive efficiency, inventory turnover rate, and order on-time delivery rate. In the scenario SOP knowledge graph, the field nodes correspond to links and constraint conditions in the standard operation procedure, including operation step descriptions, process execution sequences, fault handling plans, safety specification requirements, and standard operation time lengths. In the system domain knowledge graph, the field nodes correspond to the technical architecture and data structure of the enterprise information system itself, including database table names, interface call addresses, service module names, log event IDs, and data field definitions. In the Custom custom knowledge graph, the field nodes correspond to non-standardized information defined by enterprise users according to specific analysis or management targets, including project milestones, task responsible persons, and budget usage rates defined for a specific project.

6. The industrial agent integration platform of claim 4, wherein, The process of removing redundant fields with a frequency of occurrence lower than a frequency threshold is as follows: According to the frequency of occurrence of the fields and the embedding space distance, a redundancy measurement score of each field is calculated, and each field is divided into a core field, a regular field, and a rare field according to the redundancy measurement score; For the rare field, when the redundancy measurement score exceeds a first preset redundancy threshold, the rare field is marked as a field to be removed; For the regular field, when the redundancy measurement score exceeds a second preset redundancy threshold, the regular field is marked as a field to be removed, where the second preset redundancy threshold is greater than the first preset redundancy threshold; The core field is not subjected to a removal operation, and all rare fields and regular fields marked as fields to be removed are summarized.

7. The industrial agent integration platform of claim 6, wherein, For calculating the redundancy measurement score of each field, the following formula is used: ； wherein, denotes the redundancy measure score of the current evaluated field , is the current evaluated field , is the total number of fields, is the current evaluated field and the Euclidean distance of the context vector over any one of the fields in the summation loop, the average of the embedding space distance of the field to all other fields is calculated, is the frequency of occurrence of the field , is the maximum frequency of occurrence among all fields, is the frequency penalty coefficient, is the loop index variable.​​ 8. The industrial agent integration platform of claim 4, wherein, The intelligent reasoning module includes: The reasoning input construction submodule receives interactive information input by the enterprise user, including a natural language task description, a target field combination, and an action keyword, performs matching and screening according to field nodes of each type of enterprise knowledge graph in the enterprise knowledge graph set, acquires matched field information, and constructs a reasoning input field sequence; The step adjustment instruction submodule calculates a switching interval period value between fields based on the reasoning input field sequence, in combination with a context vector and a frequency feature in the field vector and the spatial distance parameter, and constructs a step adjustment instruction when a field switching interval change frequency is higher than a set period threshold value; The instruction set output submodule combines the reasoning input field sequence and the step adjustment instruction, inputs them into an inference path of the trained BERT language model for path generation and response judgment, performs semantic representation and context encoding, and outputs a one-key deployment instruction set.

9. The industrial agent integration platform of claim 8, wherein, The self-defined development module includes: The self-defined field input submodule receives a task construction target field, a behavior structure expression, and a result field requirement input by the enterprise user, matches a field path combination related to a target task in each type of enterprise knowledge graph in the enterprise knowledge graph set, splits and performs semantic role recognition on a semantic association relationship between fields, and establishes a self-defined field input set; The task logic generation submodule inputs the self-defined field input set into an AICoding engine generation function sentence structure, inputs it into a SOP path engine to construct a task step logic, and inputs it into a text2sql conversion engine to generate a SQL query logic expression, acquires a multi-dimensional task logic sentence, and The sentence structure integration submodule integrates a function sentence structure, a task step logic, and a SQL query logic expression based on the multi-dimensional task logic sentence, associates and arranges sentences generated by each engine to match the task construction target field, the behavior structure expression, and the result field requirement input by the user, and outputs an AI Agent task sentence structure set.

10. The industrial agent integration platform of claim 9, wherein, The task output module includes: The deployment instruction processing submodule parallelly accesses a platform instruction execution interface with the one-key deployment instruction set, and generates a model deployment structure set; The Agent statement analysis submodule parallelly accesses the platform instruction execution interface with the AI Agent task sentence structure set, and acquires an Agent behavior output structure set; The platform instruction execution submodule executes a model deployment reasoning process and an Agent development task process based on the model deployment structure set and the Agent behavior output structure set, respectively, through the platform instruction execution interface, and generates a platform output result.

Images