An industrial knowledge graph self-evolution method and system based on physical topology constraints

Abstract

The application relates to the technical field of industrial data processing, and discloses an industrial knowledge graph self-evolution method and system based on physical topology constraints, which comprises the following steps: extracting drawing knowledge and text knowledge from engineering drawings and technical manuals and recording source positions; performing consistency comparison, checking conflict relationships according to industrial physical topology rules, and forming to-be-stored knowledge; comparing the to-be-stored knowledge with existing knowledge, generating a conflict node when there is an attribute conflict, determining a target attribute according to external evidence, and updating the industrial knowledge graph; updating equipment relationship weights according to industrial business feedback, knowledge source time difference and fault severity, deactivating equipment relationships below a safety threshold and having source positions, and reextracting drawing knowledge from a target area of the engineering drawings. The application realizes multi-source industrial knowledge fusion, conflict correction and abnormal relationship directional updating, and improves the graph updating accuracy.

Description

Technical Field

[0001] This invention relates to the field of industrial data processing technology, and more specifically, to a method and system for the self-evolution of industrial knowledge graphs based on physical topological constraints. Background Technology

[0002] With the development of industrial digitalization technology, manufacturing enterprises generate various types of data during product design, equipment operation, and production maintenance, including database files, 3D computer-aided design models, engineering drawings, technical documents, equipment ledgers, and maintenance records. Knowledge graphs can unify the organization of equipment, parameters, and relationships from different sources through entity recognition, relation extraction, entity alignment, and knowledge fusion, providing a data foundation for equipment information retrieval, fault analysis, and production management.

[0003] For example, Chinese Patent Publication No. CN117236432A discloses a method and system for constructing a manufacturing process knowledge graph for multimodal data. This method extracts entities, attributes, and relationships from database files, 3D models, drawings, and text, and uses semantic similarity and entity matching to complete data alignment and fusion. However, with continuous updates to different versions of engineering drawings, technical documents, and equipment records, the graph may contain relationships inconsistent with the actual connection conditions of the equipment, outdated parameters, or contradictory data. Existing methods struggle to promptly determine whether existing knowledge remains reliable. Furthermore, when conflicting data cannot be immediately determined, directly overwriting, discarding, or stopping the writing process may result in the loss of valid information or obstruction of the update process.

[0004] Therefore, it is necessary to design a self-evolution method and system for industrial knowledge graphs based on physical topological constraints to solve the problems existing in the current technology. Summary of the Invention

[0005] In view of this, the present invention proposes an industrial knowledge graph self-evolution method and system based on physical topology constraints, which aims to solve the problems that the knowledge graph is unable to identify relationships and parameters that are inconsistent with the actual connection status of the equipment or have failed in a timely manner during the continuous updating of industrial multimodal data, and is unable to maintain the continuity of knowledge updates when new data conflicts with existing knowledge.

[0006] This invention proposes a self-evolution method for industrial knowledge graphs based on physical topological constraints, comprising: Extract drawing knowledge and textual knowledge from engineering drawings and technical manuals, and record the source location of the drawing knowledge; The consistency of the drawing knowledge and text knowledge is compared, and the physical feasibility of conflicting relationships is verified according to the preset industrial physical topology rules. The processed drawing knowledge and text knowledge are then merged to form knowledge to be added to the database. The knowledge to be added to the database is compared with the existing knowledge in the industrial knowledge graph; if there is no attribute conflict, it is written into the industrial knowledge graph; if there is an attribute conflict, a conflict node is generated that retains mutually exclusive candidate attributes, external evidence is obtained, the target attribute is determined from the mutually exclusive candidate attributes based on the external evidence, the remaining candidate attributes and their corresponding relationships are deleted, and the industrial knowledge graph is updated. After the industrial knowledge graph is updated, the weights of equipment relationships in the industrial knowledge graph are updated according to industrial business feedback, time difference of knowledge source, and severity of fault. When the weight of the equipment relationship is lower than the safety threshold and the corresponding equipment relationship has the source location, the corresponding equipment relationship is deactivated, drawing knowledge is re-extracted from the target area of ​​the engineering drawing according to the source location, and the industrial knowledge graph is updated according to the re-extracted drawing knowledge.

[0007] Furthermore, the extraction of the drawing knowledge and text knowledge includes: Equipment identifiers, equipment attributes, and equipment relationships are extracted from the engineering drawings to form the drawing knowledge; equipment identifiers, equipment attributes, and equipment relationships are extracted from the technical manual to form the text knowledge; the knowledge source and extraction confidence level of each equipment attribute and equipment relationship are recorded.

[0008] Furthermore, the consistency comparison includes: Associating the drawing knowledge and text knowledge with the device identifier, the connection objects and connection directions of the mutually associated device relationships are compared; if both the connection objects and connection directions are consistent, the corresponding device relationship is retained; if at least one is inconsistent, the corresponding device relationship is determined to be the conflict relationship.

[0009] Furthermore, the physical feasibility verification includes: A structural feasibility tensor is constructed based on the industrial physical topology rules. Verification results are obtained based on the device identifier, connection object, and connection direction of the conflict relationship. Conflict relationships that are not feasible are deleted. If multiple feasible conflict relationships remain, the conflict relationship with the highest confidence is retained. The device attributes and device relationships that have not had any conflicts, as well as the conflict relationships retained after verification, are merged to form the knowledge to be added to the database.

[0010] Furthermore, generating the conflicting nodes includes: Device attributes with the same device identifier and attribute type but different attribute values ​​in the knowledge to be added to the database and the existing knowledge are identified as mutually exclusive candidate attributes; candidate confidence is determined based on the extraction confidence and knowledge source of each mutually exclusive candidate attribute, and a probabilistic hybrid node containing device identifier, attribute type, mutually exclusive candidate attribute, knowledge source and candidate confidence is generated, and the probabilistic hybrid node is used as the conflict node.

[0011] Furthermore, obtaining the external evidence includes: Retrieve the equipment maintenance records and equipment ledgers corresponding to the equipment identifier; generate confirmation content based on the mutually exclusive candidate attributes and knowledge sources and receive manual confirmation results; use the retrieval results and manual confirmation results as the external evidence.

[0012] Further, determining the target attribute includes: When the manual confirmation result points to a mutually exclusive candidate attribute, the corresponding mutually exclusive candidate attribute is determined as the target attribute; when it does not point to one, the candidate confidence is updated based on the support results of the equipment maintenance records and equipment ledger for each mutually exclusive candidate attribute, and the mutually exclusive candidate attribute that reaches the confirmation threshold and is higher than the confidence of the other candidates is determined as the target attribute; when the target attribute is not determined, the conflict node is retained.

[0013] Furthermore, updating the device relationship weights includes: The feedback adjustment amount is determined based on the industrial business feedback, the timeliness decay amount is determined based on the knowledge source time difference, the fault penalty amount is determined based on the fault severity, and the equipment relationship weight is updated based on the feedback adjustment amount, timeliness decay amount, and fault penalty amount; when the fault severity reaches a high-risk level, a discrete fault penalty amount corresponding to the high-risk level is adopted.

[0014] Furthermore, when re-extracting the drawing knowledge, it includes: When the weight of the device relationship with the source location is lower than the security threshold, the corresponding device relationship is deactivated. The target area is extracted based on the page number and bounding box coordinates in the source location, and the drawing knowledge is re-extracted. The re-extracted drawing knowledge is compared with the text knowledge for consistency. For conflicting relationships, the physical feasibility is checked to form new knowledge to be added to the database. The new knowledge to be added to the database is compared with the existing knowledge. If there is no attribute conflict, it is written into the industrial knowledge graph. If there is an attribute conflict, the conflict node is generated. The external evidence is obtained. The target attribute is determined from the mutually exclusive candidate attributes based on the external evidence. The remaining candidate attributes and their corresponding relationships are deleted, and the industrial knowledge graph is updated.

[0015] Compared with existing technologies, the beneficial effects of this invention are as follows: By comparing the consistency of drawing knowledge and textual knowledge extracted from engineering drawings and technical manuals, and by using industrial physical topology rules to verify the physical feasibility of conflicting relationships, content that does not conform to the actual connection relationship of equipment can be screened out before knowledge is entered into the database; when there is an attribute conflict, mutually exclusive candidate attributes are retained, and the target attribute is determined by combining external evidence, reducing erroneous updates caused by directly overwriting existing knowledge; at the same time, the weight of equipment relationships is dynamically adjusted according to industrial business feedback, the time difference of knowledge sources, and the severity of faults. When the weight of equipment relationships is lower than the safety threshold, the drawing knowledge is re-extracted from the target area of ​​the engineering drawings based on the source location, realizing the targeted correction of abnormal equipment relationships, thereby improving the accuracy, traceability, and reliability of the continuous updating process of the industrial knowledge graph.

[0016] On the other hand, this application also provides a self-evolving industrial knowledge graph system based on physical topology constraints, used to apply the above-mentioned self-evolving industrial knowledge graph method based on physical topology constraints, including: The knowledge extraction unit is used to extract drawing knowledge and text knowledge from engineering drawings and technical manuals, and record the source location of the drawing knowledge; The topology verification unit is used to compare the consistency of the drawing knowledge and text knowledge, perform physical feasibility verification on conflicting relationships according to preset industrial physical topology rules, and merge the processed drawing knowledge and text knowledge to form knowledge to be added to the database. The conflict healing unit is used to compare the knowledge to be added to the database with the existing knowledge in the industrial knowledge graph; when there is no attribute conflict, it is written into the industrial knowledge graph; when there is an attribute conflict, a conflict node is generated that retains mutually exclusive candidate attributes, external evidence is obtained, the target attribute is determined from the mutually exclusive candidate attributes based on the external evidence, the remaining candidate attributes and their corresponding relationships are deleted, and the industrial knowledge graph is updated. The weight evolution unit is used to update the weight of equipment relationships in the industrial knowledge graph according to industrial business feedback, knowledge source time difference and fault severity after the industrial knowledge graph is updated; when the weight of the equipment relationship is lower than the safety threshold and the corresponding equipment relationship has the source position, the corresponding equipment relationship is deactivated, drawing knowledge is re-extracted from the target area of ​​the engineering drawing according to the source position, and the industrial knowledge graph is updated according to the re-extracted drawing knowledge.

[0017] It is understandable that the above-mentioned industrial knowledge graph self-evolution method and system based on physical topology constraints have the same beneficial effects, and will not be elaborated further here. Attached Figure Description

[0018] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart of an industrial knowledge graph self-evolution method based on physical topological constraints provided in an embodiment of the present invention; Figure 2 A flowchart for determining target attributes provided in an embodiment of the present invention; Figure 3 This is a functional block diagram of an industrial knowledge graph self-evolution system based on physical topological constraints, provided in an embodiment of the present invention. Detailed Implementation

[0019] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] In some embodiments of this application, see Figures 1-2 As shown, this application proposes a self-evolution method for industrial knowledge graphs based on physical topological constraints, including: S100: Extract drawing knowledge and text knowledge from engineering drawings and technical manuals, and record the source location of the drawing knowledge; S200: Perform a consistency comparison between the drawing knowledge and text knowledge, and perform a physical feasibility verification on conflicting relationships according to preset industrial physical topology rules. Then, merge the processed drawing knowledge and text knowledge to form knowledge to be added to the database. S300: Compare the knowledge to be added to the database with the existing knowledge in the industrial knowledge graph; write it into the industrial knowledge graph if there is no attribute conflict; if there is an attribute conflict, generate conflict nodes that retain mutually exclusive candidate attributes, obtain external evidence, determine the target attribute from the mutually exclusive candidate attributes based on the external evidence, delete the remaining candidate attributes and their corresponding relationships, and update the industrial knowledge graph. S400: After the industrial knowledge graph is updated, the weights of the equipment relationships in the industrial knowledge graph are updated according to industrial business feedback, the time difference of knowledge sources, and the severity of the fault. When the weight of the equipment relationship is lower than the safety threshold and the corresponding equipment relationship has the source location, the corresponding equipment relationship is deactivated, the drawing knowledge is re-extracted from the target area of ​​the engineering drawing according to the source location, and the industrial knowledge graph is updated according to the re-extracted drawing knowledge.

[0021] Specifically, in one implementation, the method is executed by an industrial data processing device. The industrial data processing device is an industrial server, industrial computer, edge computing device, or cloud computing node. It is communicatively connected to an engineering drawing library, a technical manual library, an industrial physical topology rule library, an industrial knowledge graph, an equipment maintenance record library, an equipment ledger, and an industrial business feedback database. This connection is used to perform drawing knowledge extraction, text knowledge extraction, consistency comparison, physical feasibility verification, knowledge storage, attribute conflict handling, equipment relationship weight updating, and drawing knowledge re-extraction.

[0022] Engineering drawings include piping and instrumentation diagrams, electrical schematics, equipment assembly drawings, equipment layout diagrams, and two-dimensional drawings derived from three-dimensional computer-aided design models. Technical manuals include equipment instruction manuals, operation manuals, maintenance manuals, equipment parameter tables, and technical modification documents. Engineering drawings and technical manuals respectively record document identifiers, version identifiers, release dates, and validity status. Document identifiers are used to determine the source document, version identifiers are used to distinguish different versions of the same document, and validity status is used to identify the current valid version and previous versions.

[0023] Industrial knowledge in an industrial knowledge graph includes equipment identification, equipment attributes, and equipment relationships. Equipment identification includes equipment tag number, equipment code, nameplate number, or equipment ledger number. Equipment attributes include equipment model, rated voltage, rated pressure, rated flow rate, installation location, and operating status. Equipment relationships are used to represent the piping connections, electrical connections, drive connections, control connections, or installation coordination between two pieces of equipment.

[0024] For each equipment relationship, at least the starting equipment identifier, the connected object, and the connection direction should be recorded. The connection direction of fluid pipelines should be recorded according to the direction of medium flow; the connection direction of electrical circuits should be recorded according to the direction from the power supply equipment to the power receiving equipment; the drive relationship should be recorded according to the direction from the driving equipment to the driven equipment; the connection direction of bidirectional installation and mating relationships should be recorded as bidirectional.

[0025] When extracting drawing knowledge from engineering drawings, the engineering drawings are converted into page images according to the page numbers, and the engineering drawing file identifier, version identifier, page number, page width, and page height corresponding to each page image are recorded.

[0026] This implementation uses the YOLOv11-obb model, which has the capability to detect rotating targets, to identify device symbols, interface symbols, parameter tables, and orientation bounding boxes in the page image. The orientation bounding box records the center point coordinates, width, height, and rotation angle of the device symbol, which is used to determine the position and orientation of the device symbol in the page image.

[0027] Input the image region containing device symbols, interface symbols, and adjacent connection lines into the Florence-2 visual language model to extract device identifiers, device types, interface locations, connection objects, and connection directions. Input the image region containing the parameter table into the Donut visual document parsing model to extract device identifiers, attribute types, and attribute values ​​from the parameter table.

[0028] The YOLOv11-obb model is fine-tuned using engineering drawing images labeled with device symbol categories and directional bounding boxes. The Florence-2 model is fine-tuned using engineering drawing regions labeled with device identifiers, device types, interface locations, connected objects, and connection directions. The Donut model is fine-tuned using engineering drawing parameter tables labeled with device identifiers, attribute types, and attribute values. After model training, the device symbol detection results and recognition scores for each structured field are output.

[0029] Equipment identification is determined based on the equipment tag number recorded in the engineering drawings. The format of the equipment tag number is obtained from the standard equipment identification field in the equipment ledger, including the equipment type letter field, connector, and number field. For example, the standard equipment identification for a pump is P-B01, the standard equipment identification for a valve is V-A02, and the standard equipment identification for a motor is M-03.

[0030] When the identified text content matches the equipment identifier format in the equipment ledger, the text content is identified as the equipment identifier. Text content containing numerical values ​​and units of measurement is identified as a candidate equipment attribute. Text content containing terms such as inlet, outlet, in, out, upstream, or downstream direction is used to determine the connection direction.

[0031] The engineering drawing symbol library records interface numbers, interface locations, and interface directions according to equipment symbol categories. Interface locations are represented by coordinate offsets relative to the center point of the equipment symbol's bounding box. Based on the detected equipment symbol category, the engineering drawing symbol library is queried, and the retrieved interface locations are converted to the page image coordinate system to obtain the interface area of ​​the corresponding equipment symbol.

[0032] When a pipeline intersects the interface area of ​​a device symbol, the corresponding pipeline is identified as a device connection line. Tracing the device connection line outward from the starting device symbol until it intersects the interface area of ​​another device symbol, the device identifier corresponding to that other device symbol is identified as the connection object.

[0033] When a connecting line has an arrow, the connection direction is determined according to the arrow's direction. When a connecting line has both inlet and outlet markings, the connection direction is determined according to the direction from the outlet device to the inlet device. When the equipment symbols corresponding to pumps, compressors, fans, check valves, and one-way valves have a fixed flow direction, the fixed flow direction field for the corresponding equipment type is read from the industrial physical topology rule base, and the connection direction is determined according to the fixed flow direction field.

[0034] For example, if a pump symbol with equipment identification P-B01 is identified in the engineering drawings, and the outlet interface of pump P-B01 is connected to a valve symbol with equipment identification V-A02 via a connecting line, and the arrow on the connecting line points from P-B01 to V-A02, then a device relationship is formed where the starting equipment identification is P-B01, the connected object is V-A02, and the connection direction is from P-B01 to V-A02.

[0035] Drawing knowledge includes equipment identifiers, equipment attributes, and equipment relationships extracted from engineering drawings. Each piece of drawing knowledge synchronously records its source and location. The source includes the engineering drawing's file identifier, version identifier, release date, and validity status.

[0036] The source location includes the document identifier, version identifier, page number, page width, page height, and bounding box coordinates of the engineering drawing. The bounding box coordinates are represented by the coordinates of the top-left corner and bottom-right corner of the page, and these two coordinates are divided by the page width and page height respectively to form normalized bounding box coordinates.

[0037] The bounding box of a device attribute is the smallest bounding rectangle covering the device identifier, attribute type, and attribute value. The bounding box of a device relationship is the smallest bounding rectangle covering the starting device symbol, device connection lines, and the symbols of the connected object devices. The source location is written into the drawing knowledge along with the corresponding device attribute or device relationship.

[0038] When extracting textual knowledge from technical manuals, the manuals are segmented according to headings, paragraphs, and tables. Equipment identifiers, equipment names, equipment models, attribute types, and attribute values ​​in the technical manuals are extracted using the BERT-CRF entity recognition model. The BERT-CRF entity recognition model labels each character in the text with an entity category, grouping consecutive characters labeled with the same entity category into equipment entities or attribute entities.

[0039] The equipment relationships in the technical manual are extracted using the CasRel relationship extraction model. The CasRel model first identifies the starting equipment entity in the statement, then determines the connection objects and relationship types that exist with the starting equipment entity. Relationship types include piping connections, electrical connections, drive connections, control connections, inlet connections, and outlet connections.

[0040] For example, the technical manual states that "valve V-A02 is installed on the outlet pipeline of pump P-B01". The BERT-CRF entity recognition model extracts valve V-A02 and pump P-B01, and the CasRel relationship extraction model extracts the outlet connection relationship between the two, forming a device relationship with the starting device identifier P-B01, the connection object V-A02, and the connection direction P-B01 pointing to V-A02.

[0041] For parameter tables in technical manuals, the cell containing the equipment tag number is designated as the equipment identifier, the table header or row title is designated as the attribute type, and the cell content corresponding to the attribute type is designated as the attribute value. For example, if the equipment tag number is M-03, the table header is rated voltage, and the corresponding cell is 380V, then the equipment attribute is defined as equipment identifier M-03, attribute type is rated voltage, and attribute value is 380V.

[0042] When equipment is referred to by its name in the technical manual, the equipment ledger is consulted based on the equipment name, equipment model, and the process unit it belongs to. If the query result uniquely corresponds to a single equipment identifier, the retrieved equipment identifier is used as the standard equipment identifier. If the query result corresponds to multiple equipment identifiers, an equipment identifier verification record is generated, which includes the equipment name, equipment model, process unit, and candidate equipment identifiers. Engineering personnel then confirm the standard equipment identifier from the candidate equipment identifiers to form textual knowledge.

[0043] Textual knowledge includes device identifiers, device attributes, and device relationships extracted from technical manuals. Each piece of textual knowledge records its corresponding knowledge source, which includes the technical manual's document identifier, version identifier, publication date, validity status, and the corresponding paragraph identifier or table location.

[0044] When extracting knowledge from drawings and text, record the extraction confidence level for each equipment attribute and equipment relationship. The extraction confidence level is represented by a value between 0 and 1.

[0045] The YOLOv11-obb model outputs a score for device symbol detection. The Florence-2 model outputs scores for device identifier, connection object, and connection direction. The Donut model outputs scores for device identifier, attribute type, and attribute value. When a field consists of multiple text tags, the arithmetic mean of the probabilities of each text tag constituting the corresponding field is taken as the field recognition score.

[0046] The equipment attributes in the drawing knowledge consist of equipment identifier, attribute type, and attribute value. The lowest value in the field identification score corresponding to the equipment identifier, attribute type, and attribute value is determined as the extraction confidence of the corresponding equipment attribute.

[0047] The equipment relationships in the drawing knowledge are composed of the starting equipment identifier, the connected object, and the connection direction. The lowest value among the starting equipment symbol detection score, the connected object equipment symbol detection score, the starting equipment identifier field recognition score, the connected object field recognition score, and the connection direction field recognition score is determined as the extraction confidence of the corresponding equipment relationship.

[0048] The BERT-CRF entity recognition model outputs entity annotation scores based on device identifier, attribute type, and attribute value. The CasRel relation extraction model outputs relation recognition scores based on the starting device, connected object, and relation type. For device attributes in textual knowledge, the lowest score among the corresponding device identifier, attribute type, and attribute value is used as the extraction confidence score. For device relations in textual knowledge, the lowest score among the corresponding starting device identifier, connected object, and connection direction is used as the extraction confidence score.

[0049] Once a device attribute has its device identifier, attribute type, and attribute value, it proceeds to subsequent processing. Similarly, once a device relationship has its starting device identifier, connected object, and connection direction, it proceeds to subsequent processing. Extraction results with incomplete fields are recorded as pending supplementary knowledge; engineers will supplement the corresponding fields before proceeding to subsequent processing.

[0050] After acquiring knowledge of the drawings and text, the equipment identification is standardized. Standardization includes removing spaces, converting letters to uniform uppercase and lowercase, converting underscores to horizontal lines, and converting equipment aliases to standard equipment identifiers according to the equipment alias mapping table in the equipment ledger. For example, "PB-01," "P_B01," and "Pump No. 1" are unified into the standard equipment identifier "P-B01."

[0051] The equipment identification is associated with drawing knowledge and text knowledge. Drawing knowledge and text knowledge with the same equipment identification are grouped into the same equipment knowledge group. For equipment relationships within the same equipment knowledge group, the connection objects and connection directions are compared separately.

[0052] If the connection objects and connection directions are the same in both the drawing knowledge and the text knowledge, the corresponding device relationships are determined to be consistent and retained. If the connection objects or connection directions are different in the drawing knowledge and the text knowledge, the inconsistent device relationships are determined to be conflicting relationships.

[0053] For example, the drawing information states that pump P-B01 is connected to valve V-A02 at the outlet, while the text information states that pump P-B01 is connected to valve V-A02 at the inlet. Since the starting device identifier and the connected object are the same for both devices, but the connection directions are different, the two devices are identified as a conflicting relationship.

[0054] Before performing a physical feasibility check on conflict relationships, a structural feasibility tensor is constructed based on industrial physical topology rules.

[0055] Industrial physical topology rules are derived from currently valid and approved equipment connection lists, equipment interface compatibility tables, piping and instrumentation connection tables, electrical wiring tables, and equipment installation mutual exclusion tables.

[0056] The equipment connection list records the equipment identifiers and corresponding interface numbers that allow direct connection, categorized by equipment identification. The equipment interface compatibility table records the allowed interface types, categorized by interface type. The piping and instrumentation diagram connection table records the piping connection direction, categorized by starting equipment identifier, connection object, and media flow direction. The electrical wiring table records the electrical connection direction, categorized by power supply equipment identifier, power receiving equipment identifier, and terminal block number. The equipment installation mutual exclusion table records combinations of equipment types that prohibit direct connection, categorized by equipment type combination.

[0057] Each rule record in the industrial physical topology rule base includes the starting device identifier, the connection object, the connection direction, and the physical feasibility value.

[0058] When generating a rule record, the device connection list is queried based on the starting device identifier and the connected object. If the device connection list shows that two devices are allowed to connect directly, the connection list is approved; if the device connection list does not show a direct connection between the two devices, the connection list is rejected.

[0059] The device interface compatibility table is consulted based on the interface numbers of the two devices. If the interface types of the two devices belong to an allowed connection combination, the interface compatibility result is "interface compatibility passed"; if the interface types of the two devices do not belong to an allowed connection combination, the interface compatibility result is "interface compatibility failed".

[0060] The connection direction of fluid pipelines is determined based on the connection table of the piping and instrumentation diagram. When the determined connection direction is consistent with the medium flow direction in the corresponding record, the flow direction is considered to be through; when the two directions are opposite, the flow direction is considered to be not through.

[0061] The electrical connection relationship is determined based on the electrical wiring diagram to determine the connection direction. When the starting device is a power supply device, the connection object is a power receiving device, and the wiring terminals correspond, the electrical direction is considered to be correct; if the power supply device, power receiving device, or wiring terminals do not correspond, the electrical direction is considered to be incorrect.

[0062] The device installation mutual exclusion table is queried based on the device type of the starting device and the connected object. If the corresponding device type combination does not exist in the device installation mutual exclusion table, the mutual exclusion check passes; if the device installation mutual exclusion table records that the corresponding device type combination prohibits direct connection, the mutual exclusion check fails.

[0063] For fluid pipeline relationships, if the connection list passes, the interface compatibility passes, the flow direction passes, and the mutual exclusion check passes, the physical feasibility value of the corresponding rule record is set to 1; if any result fails, the physical feasibility value is set to 0.

[0064] For electrical connection relationships, if the connection list passes, the interface compatibility passes, the electrical direction passes, and the mutual exclusion check passes, the physical feasibility value of the corresponding rule record is set to 1; if any result fails, the physical feasibility value is set to 0.

[0065] For drive connections, control connections, and installation compatibility, the corresponding device connection list, interface compatibility table, and device installation mutual exclusion table are used for judgment. If all judgment results pass, the physical feasibility value is set to 1; if any judgment result fails, the physical feasibility value is set to 0.

[0066] Before processing a batch of engineering drawings and technical manuals, the industrial data processing equipment reads all equipment identifiers, equipment types, and interface numbers involved in the batch, and checks whether corresponding rule records exist in the industrial physical topology rule base. If rule records are missing, the equipment supplements the records based on the currently valid and approved equipment connection list, equipment interface compatibility table, piping and instrumentation flow diagram connection table, electrical wiring table, and equipment installation mutual exclusion table. After the supplementation is completed, the knowledge processing for this batch is executed.

[0067] A structural feasibility tensor is constructed based on an industrial physical topology rule base. The structural feasibility tensor uses the starting device identifier, the connected object, and the connection direction as three index dimensions. When the physical feasibility value of the corresponding combination in the industrial physical topology rule is 1, a 1 is written to the corresponding index position of the structural feasibility tensor; when the physical feasibility value is 0, a 0 is written to the corresponding index position.

[0068] When performing physical feasibility verification on a conflict relationship, the corresponding index position in the structural feasibility tensor is determined based on the starting device identifier, connection object, and connection direction of the conflict relationship. A tensor to be verified with the same index structure as the structural feasibility tensor is created, setting the index position corresponding to the conflict relationship to 1 and the remaining index positions to 0.

[0069] The tensor to be verified is multiplied element-wise with the structural feasibility tensor, and the product result at the index of the conflict relationship is read. When the product result is 1, the corresponding conflict relationship is determined to be physically feasible and retained; when the product result is 0, the corresponding conflict relationship is determined to be physically infeasible, the corresponding conflict relationship is deleted, and the extraction confidence of the corresponding conflict relationship is set to 0.

[0070] For example, the drawing knowledge indicates that valve V-A02 is directly connected to the outlet of high-pressure centrifugal pump P-B01. However, the equipment interface compatibility table states that the interface type of valve V-A02 and the outlet interface type of pump P-B01 are not allowed connection combinations. Therefore, the interface compatibility result recorded in the corresponding rule fails, and the physical feasibility value is 0. The product result of the corresponding index position in the structural feasibility tensor is 0, and the corresponding equipment relationship is deleted.

[0071] If, after deleting physically infeasible conflict relationships, only one physically feasible conflict relationship remains, that conflict relationship is retained. If multiple physically feasible conflict relationships remain, the extraction confidence of each conflict relationship is compared, and the conflict relationship with the highest extraction confidence is retained.

[0072] When multiple physically feasible conflict relationships have the same extraction confidence level, the validity status of the knowledge sources for each conflict relationship is compared, and conflict relationships originating from currently valid documents are retained. If all sources originate from currently valid documents, the publication time of each knowledge source is compared, and conflict relationships with the more recent publication time are retained. When publication times are the same, a relationship verification record is created, and engineers determine the retained device relationships based on the current device connection status.

[0073] Merge the equipment attributes and relationships that do not conflict in the drawing knowledge and text knowledge with the conflicting relationships retained after the physical feasibility verification.

[0074] Device relationships with identical device identifiers, connection objects, and connection directions are merged into a single device relationship, and the merged device relationship stores the knowledge source, source location, and extraction confidence score. Device attributes with identical device identifiers, attribute types, and attribute values ​​are merged into a single device attribute, and the knowledge source, source location, and extraction confidence score are stored. After merging, the knowledge to be added to the database is formed.

[0075] The knowledge to be added to the database is compared with the existing knowledge in the industrial knowledge graph. The device node in the industrial knowledge graph is queried based on the device identifier, and then the existing device attributes under the corresponding device node are queried based on the attribute type.

[0076] If there is no existing device attribute matching the corresponding attribute type in the knowledge to be added to the database, it is determined that there is no attribute conflict, and the corresponding device attribute is written into the industrial knowledge graph.

[0077] If the device identifier, attribute type, and attribute value of the knowledge to be added to the database are the same as those of the existing knowledge, it is determined that there is no attribute conflict. The existing device attributes are retained, and the knowledge source, source location, and extraction confidence of the knowledge to be added to the database are written into the source record of the existing device attributes.

[0078] When the device identifier and attribute type are the same as those in the existing knowledge but the attribute values ​​are different, it is determined that there is an attribute conflict, and the device attributes with different attribute values ​​are identified as mutually exclusive candidate attributes.

[0079] For example, the existing knowledge record in the industrial knowledge graph states that the rated voltage of motor M-03 is 380V, while the knowledge to be added to the database states that the rated voltage of motor M-03 is 220V. The two device attributes have the same device identifier and attribute type, but different attribute values; therefore, rated voltage 380V and rated voltage 220V are identified as mutually exclusive candidate attributes.

[0080] When attribute conflicts exist, a probabilistic hybrid node is generated and designated as the conflicting node. A temporary attribute relationship is established between the probabilistic hybrid node and the corresponding device node. The probabilistic hybrid node records the device identifier, attribute type, each mutually exclusive candidate attribute, extraction confidence score of each mutually exclusive candidate attribute, knowledge source, candidate score, and candidate confidence score.

[0081] Knowledge sources are grouped according to source type, validity status, and approval status. Currently valid and approved engineering drawings, historical engineering drawings, currently valid and approved technical manuals, and historical technical manuals form separate knowledge source groups.

[0082] The knowledge source reliability coefficient for each knowledge source group is determined based on the historical device attributes that have been verified. All device attributes with final verification results are read from the corresponding knowledge source group. The number of verification results that match the original device attributes is counted, and this number is divided by the total number of device attributes with final verification results. The resulting ratio is used as the knowledge source reliability coefficient for the corresponding knowledge source group.

[0083] The final verification results are derived from the confirmation results of engineering personnel, post-maintenance parameters in equipment maintenance records, or registered parameters in the current valid equipment ledger. When the knowledge source group has not yet formed historical verification results, the reliability coefficient of the corresponding knowledge source is set to 1 as a neutral coefficient that does not change the extraction confidence level, and conflict nodes generated by the corresponding knowledge source are marked as requiring manual confirmation.

[0084] For each new final verification result obtained, the corresponding verification result is added to the knowledge source group to which it belongs, the number of consistent results and the total number of verifications are recalculated, and the knowledge source reliability coefficient is updated.

[0085] When determining candidate confidence levels, the extraction confidence level of each mutually exclusive candidate attribute is multiplied by the corresponding knowledge source reliability coefficient to obtain the candidate score for the corresponding mutually exclusive candidate attribute. Each candidate score is then divided by the sum of all candidate scores to obtain the candidate confidence level for each mutually exclusive candidate attribute, and the sum of all candidate confidence levels is 1.

[0086] For example, in historical technical manuals, the extraction confidence score for the 380V attribute is 0.80, corresponding to a knowledge source reliability coefficient of 0.75 and a candidate score of 0.60; in currently valid engineering drawings, the extraction confidence score for the 220V attribute is 0.90, corresponding to a knowledge source reliability coefficient of 1.00 and a candidate score of 0.90. The sum of the two candidate scores is 1.50, the candidate confidence score for the 380V attribute is 0.40, and the candidate confidence score for the 220V attribute is 0.60. The probabilistic hybrid node simultaneously retains the two mutually exclusive candidate attributes, 380V and 220V, along with their knowledge sources and candidate confidence scores.

[0087] After generating probabilistic hybrid nodes, equipment maintenance records and equipment ledgers are retrieved based on equipment identifier and attribute type, respectively.

[0088] Equipment maintenance records should include at least the equipment identifier, maintenance time, maintenance items, pre-maintenance parameters, post-maintenance parameters, and review status. Filter equipment maintenance records by equipment identifier, and then filter those records by attribute type that relate to the corresponding equipment attributes.

[0089] If the parameter after maintenance is the same as a mutually exclusive candidate attribute, and the maintenance time is later than the publication time of the corresponding knowledge source, the corresponding device maintenance record will be determined as a search result supporting that mutually exclusive candidate attribute. If the device maintenance record explicitly states that the device changed from one attribute value to another, the mutually exclusive candidate attribute corresponding to the parameter before maintenance will be determined as a rejected candidate attribute, and the mutually exclusive candidate attribute corresponding to the parameter after maintenance will be determined as a supported candidate attribute.

[0090] The equipment ledger should include at least the equipment identifier, current registration parameters, ledger version, update time, and review status. Search for currently valid and reviewed equipment ledger versions by equipment identifier and attribute type. If the current registration parameters match a mutually exclusive candidate attribute, the equipment ledger will be selected as a search result supporting that mutually exclusive candidate attribute.

[0091] Confirmation content is generated based on mutually exclusive candidate attributes and knowledge sources. The confirmation content includes equipment identifier, attribute type, all mutually exclusive candidate attributes, knowledge source for each mutually exclusive candidate attribute, extraction confidence level, candidate confidence level, equipment maintenance record retrieval results, and equipment ledger retrieval results.

[0092] When mutually exclusive candidate attributes originate from engineering drawings, the corresponding drawing area is extracted based on the file identifier, version identifier, page number, and bounding box coordinates in the source location, and this drawing area is added to the confirmation content. When mutually exclusive candidate attributes originate from technical manuals, the corresponding paragraph or table row is added to the confirmation content. The confirmation content is sent to the engineer's terminal, and manual confirmation is received.

[0093] Manual confirmation results include confirming one mutually exclusive candidate attribute, confirming that none of the mutually exclusive candidate attributes match the current device status, and temporarily being unable to determine the result.

[0094] When the manual confirmation result points to a mutually exclusive candidate attribute, the corresponding mutually exclusive candidate attribute is determined as the target attribute.

[0095] If the manual verification results show that none of the mutually exclusive candidate attributes match the current device status, the conflicting nodes are retained, and a knowledge re-collection task containing the device identifier, attribute type, and all mutually exclusive candidate attributes is generated.

[0096] When the manual confirmation result is temporarily uncertain, the candidate confidence is updated based on the support results of each mutually exclusive candidate attribute in the equipment maintenance records and equipment ledger.

[0097] When updating candidate confidence levels, equipment maintenance records and equipment ledgers are treated as separate external evidence categories. If an equipment maintenance record supports a mutually exclusive candidate attribute, the corresponding external evidence category is marked as supported; if an equipment maintenance record negates a mutually exclusive candidate attribute, the corresponding external evidence category is marked as negated. Equipment ledgers are processed in the same way to generate either supporting or negating results.

[0098] When both equipment maintenance records and equipment ledgers have currently valid and audited search results, the number of external evidence categories is 2. For each mutually exclusive candidate attribute, the number of external evidence categories supporting that mutually exclusive candidate attribute is divided by 2 to obtain the updated candidate confidence level.

[0099] When both equipment maintenance records and equipment ledgers support the same mutually exclusive candidate attribute, the updated candidate confidence level for the corresponding mutually exclusive candidate attribute is 1, while the updated candidate confidence level for all other mutually exclusive candidate attributes is 0. When equipment maintenance records and equipment ledgers support different mutually exclusive candidate attributes, the updated candidate confidence level for both mutually exclusive candidate attributes is 0.5. When one external evidence category supports a certain mutually exclusive candidate attribute, but another external evidence category does not provide corresponding attribute information, the updated candidate confidence level for the corresponding mutually exclusive candidate attribute is 0.5.

[0100] This implementation sets the confirmation threshold to 1. A confirmation threshold of 1 indicates that both the equipment maintenance record and the equipment ledger, two independent external evidence categories, support the same mutually exclusive candidate attribute. When automatically determining the target attribute using external evidence, it is also required that both the equipment maintenance record and the equipment ledger have currently valid and audited search results.

[0101] When the confidence level of a mutually exclusive candidate attribute after its update reaches the confirmation threshold 1 and is higher than the candidate confidence levels of the other mutually exclusive candidate attributes, the corresponding mutually exclusive candidate attribute is determined as the target attribute.

[0102] If the equipment maintenance record and the equipment ledger do not simultaneously form a currently valid and reviewed search result, or if the confidence level of any mutually exclusive candidate attribute after updating is less than 1, the conflicting node is retained, and subsequent equipment maintenance records, equipment ledgers, or manual confirmation results are continued to be received.

[0103] After determining the target attribute, it is written into the industrial knowledge graph and set as the current valid attribute of the corresponding device node. The remaining candidate attributes are deleted from the current valid attribute set, and the current valid associations between the remaining candidate attributes and the device node, process node, or control node are also deleted.

[0104] The attribute value, knowledge source, extraction confidence level, candidate confidence level, deletion time, and deletion basis of the deleted candidate attribute are written into the history record for subsequent auditing and updating of the knowledge source reliability coefficient.

[0105] For example, existing knowledge in the industrial knowledge graph records that the rated voltage of motor M-03 is 380V, while the currently valid engineering drawings record a rated voltage of 220V. Equipment maintenance records indicate that motor M-03 has been changed from a 380V model to a 220V model, and the currently valid and approved equipment ledger also records the rated voltage as 220V. Both the equipment maintenance record and the equipment ledger support the 220V attribute, and the candidate confidence level after the 220V attribute update is 1, reaching the confirmation threshold. Therefore, 220V is identified as the target attribute, the 380V attribute and its currently valid association are deleted, and the rated voltage of motor M-03 is updated to 220V.

[0106] After the industrial knowledge graph completes the writing of knowledge to be added to the database or the handling of attribute conflicts, the weights of device relationships in the industrial knowledge graph are updated.

[0107] When a device relationship is first written into the industrial knowledge graph, the extraction confidence score of the corresponding device relationship is used as the initial device relationship weight. When the same device relationship is supported by both drawing knowledge and text knowledge, the arithmetic mean of the extraction confidence scores of the corresponding device relationship in the drawing knowledge and text knowledge is taken as the initial device relationship weight. Device relationship weights are represented by values ​​between 0 and 1 and recorded with one decimal place.

[0108] Industrial business feedback comes from equipment inspection records, maintenance work orders, operator confirmation results, equipment fault reports, and knowledge query and error correction records. Each piece of industrial business feedback includes equipment identification, corresponding equipment relationship, feedback conclusion, feedback time, and review status.

[0109] Industrial business feedback that confirms the equipment relationship matches the current equipment status is identified as positive feedback, while industrial business feedback that confirms the equipment relationship is incorrect, has failed, or has changed due to technical upgrades is identified as negative feedback.

[0110] When updating the weight of the same device relationship, the number of positive and negative feedback responses that have been reviewed since the last weight update is counted. If the number of positive feedback responses is greater than the number of negative feedback responses, the user feedback value is incremented by 1; if the number of negative feedback responses is greater than the number of positive feedback responses, the user feedback value is decremented by 1; if the number of positive feedback responses is equal to the number of negative feedback responses, the user feedback value is 0. If there is negative feedback corresponding to a high-risk fault, the user feedback value is directly decremented by 1.

[0111] The basic update step size is consistent with the recording precision of the device relationship weights. Device relationship weights are recorded to one decimal place, with a minimum weight change unit of 0.1; therefore, the basic update step size is 0.1. The feedback adjustment amount is the product of the basic update step size and the user feedback value. When the user feedback value is +1, the feedback adjustment amount is +0.1; when the user feedback value is -1, the feedback adjustment amount is -0.1; and when the user feedback value is 0, the feedback adjustment amount is 0.

[0112] The knowledge source time difference is the difference between the current weight update time and the publication time of the corresponding device relationship knowledge source. When the same device relationship has multiple knowledge sources, the knowledge source time difference is calculated using the currently valid and most recently published knowledge source.

[0113] The time-lapse rate corresponding to the time difference in knowledge sources is determined based on the validity period of the knowledge source. The validity period is read from the review period field of the corresponding file type in the enterprise document management system. Engineering drawings use the engineering drawing review period, and technical manuals use the technical manual revision review period.

[0114] When the time difference of the knowledge source does not exceed one effective period, the time decay is 1; when it exceeds one effective period but does not exceed two effective periods, the time decay is 2; when it exceeds two effective periods but does not exceed three effective periods, the time decay is 3; and when it exceeds three effective periods, the time decay is 4.

[0115] When the enterprise document management system does not record the review cycle for the corresponding file type, the timeliness decay is set to 1, and a file review cycle supplementary task is generated. Therefore, knowledge sources lacking review cycles will not receive additional decay due to the time difference; the decay will be calculated according to the aforementioned rules after the review cycle is supplemented.

[0116] The severity of a fault is classified into general, major, serious, and high-risk levels based on the consequences of the fault on equipment operation, safety, and production.

[0117] If the equipment can maintain its original operating state, and the fault manifests as general wear, slight vibration, or localized temperature rise, and there is no need to reduce the load during operation, it is determined to be of the general level, and the fault penalty is set to 1.

[0118] When equipment performance deteriorates, requiring reduced load operation, isolation of partial functions, or scheduled maintenance, it is classified as a major fault, with a fault penalty of 2.

[0119] If the main function of the equipment fails and requires immediate shutdown, or if the protective device is activated or a single production unit stops operating, it is classified as a critical level, and the fault penalty is 3.

[0120] When there is a risk of high-pressure leakage, hazardous medium leakage, protection interlock failure, uncontrolled high-temperature release, or deflagration, or when there is a risk of serious injury or death to personnel or destructive damage to equipment, it is classified as a high-risk level and a discrete fault penalty of 4 is adopted.

[0121] When a fault meets the criteria for multiple severity levels, the severity is determined by the highest level. For example, if a bearing experiences general wear while the equipment remains in its original operating state, it is classified as a general level; if the compressor needs to operate under reduced load, it is classified as a major level; if a motor burns out and causes a single production unit to shut down, it is classified as a critical level; and if the compressor cylinder experiences high-pressure leakage or safety interlock failure, it is classified as a high-risk level.

[0122] The equipment relationship weights are updated based on feedback adjustments, time-related degradation, and fault penalties. The equipment relationship weights are calculated using the following formula: ; in, This indicates the updated device relationship weights; This indicates the device relationship weights before the update; Update the step size based on the baseline; This represents user feedback values; Indicates the amount of penalty for failure; This indicates the amount of decay over time.

[0123] When the user feedback value is +1, both the fault penalty and the time-decrease amount are set to 1, and the device relationship weight is increased according to the basic update step size.

[0124] When the user feedback value is -1 and the feedback content does not involve equipment failure, the failure penalty is 1, and the timeliness decay is determined according to the time difference of the knowledge source.

[0125] When the user feedback value is -1 and the feedback content involves equipment failure, the failure penalty amount is determined according to the severity of the failure, and the timeliness decay amount is determined according to the time difference of the knowledge source.

[0126] When the user feedback value is 0, the device relationship weight remains unchanged. When the calculation result is greater than 1, the updated device relationship weight is set to 1; when the calculation result is less than 0, the updated device relationship weight is set to 0.

[0127] The safety threshold is determined based on confirmed valid and invalid historical device relationships. Historical device relationships are marked as valid or invalid according to the actual device status, and the updated weights of each historical device relationship are calculated according to the aforementioned device relationship weight update method.

[0128] The updated weights of each historical device relationship are arranged from low to high, and the median value between each different weight and two adjacent different weights is used as the candidate security threshold.

[0129] For each candidate security threshold, the number of valid device relationships with weights lower than the candidate security threshold after the update is counted as the number of valid relationships that are falsely stopped; the number of failed device relationships with weights not lower than the candidate security threshold after the update is counted as the number of failed relationships that are retained.

[0130] For ordinary equipment relationships, select the candidate safety threshold with the minimum sum of the number of valid relationships that fail to stop and the number of failed relationships that are retained. If multiple candidate safety thresholds correspond to the same minimum sum, select the candidate safety threshold with the lower value.

[0131] For equipment relationships involving high-voltage equipment, hazardous media, or safety interlocks, first select candidate safety thresholds with a retention count of 0 for failed relationships, and then select the candidate safety threshold with the fewest false shutdowns for valid relationships from among them. If multiple candidate safety thresholds correspond to the same number of false shutdowns for valid relationships, select the candidate safety threshold with the higher value.

[0132] This implementation obtains a safety threshold of 0.4 through historical device relationship calibration. The historical device relationships, valid or invalid flags, weights before and after the update, and the final safety threshold used in the safety threshold calibration are all stored in the threshold calibration record.

[0133] For example, the original equipment relationship weight for compressor C is 0.9. When a maintenance technician submits negative feedback regarding high-pressure leakage in the compressor C cylinder, the user feedback value is -1. High-pressure leakage in the cylinder is considered high-risk, with a penalty of 4. The knowledge source time difference in the corresponding technical manual exceeds one valid period but does not exceed two valid periods, resulting in a time decay of 2. The basic update step size is 0.1. The updated equipment relationship weight is: 0.9 + 0.1 × ( 1)×4×2=0.1.

[0134] The updated device relationship weight of 0.1 is lower than the safety threshold of 0.4, so the corresponding device relationship is disabled.

[0135] When a device relationship is deactivated, its valid status field is changed to "deactivated," preventing the corresponding device relationship from participating in industrial knowledge queries, association reasoning, and business judgments. The starting device identifier, connection object, connection direction, knowledge source, source location, historical weight, deactivation time, and deactivation reason of the device relationship are retained.

[0136] When the weight of a device relationship is lower than the safety threshold and the corresponding device relationship has a source location, the corresponding page image is read based on the engineering drawing file identifier, version identifier, and page number in the source location.

[0137] Based on the normalized bounding box coordinates from the source location and the page width and page height of the current page image, restore the pixel bounding box coordinates of the corresponding device relationship in the current page image.

[0138] The recovered pixel bounding box is used as the initial target region. If the initial target region already covers the starting device symbol, device connection lines, and connected object device symbols, the initial target region is determined as the target region.

[0139] If the initial target area does not completely cover the device connection line, pixel connectivity tracing is performed along the connection line starting from the initial device interface within the initial target area until the connection line intersects with the interface area of ​​the connected object device symbol. The minimum bounding rectangle of the initial device symbol, the complete device connection line, and the connected object device symbol is determined as the target area.

[0140] The target area is scaled proportionally to the input size specified in the YOLOv11-obb model configuration file, with empty areas padded using page background pixels. The scaled target area is then input into the YOLOv11-obb model. The areas containing the detected device symbols, interface symbols, device connection lines, and parameter tables are then input into the Florence-2 model and the Donut model, respectively, to re-extract device identifiers, device attributes, and device relationships.

[0141] When re-extracting drawing knowledge, the extraction confidence of each equipment attribute and equipment relationship is re-determined according to the aforementioned method, and the source position is re-recorded based on the position of the re-extraction result in the original page image.

[0142] The re-extracted drawing knowledge and text knowledge are associated with standard equipment identifiers, and the connection objects and directions of the associated equipment relationships are compared. When the connection objects or directions are inconsistent, the corresponding equipment relationship is identified as a conflict relationship, and physical feasibility is verified according to the aforementioned structural feasibility tensor.

[0143] Conflicting relationships that are deemed infeasible after verification are deleted. If multiple feasible conflicting relationships remain, the conflicting relationship with the highest confidence level is retained for extraction. Device attributes and relationships that do not have conflicts, along with the conflicting relationships retained after physical feasibility verification, are merged to form new knowledge to be added to the database.

[0144] The new knowledge to be added to the database is compared with the existing knowledge in the industrial knowledge graph. If there are no attribute conflicts, the new knowledge to be added to the industrial knowledge graph is written.

[0145] When attribute conflicts exist, mutually exclusive candidate attributes are determined in the aforementioned manner, probabilistic hybrid nodes are generated, equipment maintenance records and equipment ledgers are retrieved based on equipment identifiers and attribute types, confirmation content is generated and manual confirmation results are received, the target attribute is determined based on the candidate confidence level updated according to the manual confirmation results or external evidence, the remaining candidate attributes and their corresponding currently valid associations are deleted, and the industrial knowledge graph is updated.

[0146] For example, in the industrial knowledge graph, the original equipment relationship record shows that compressor C is connected to safety valve D. After the weight of this equipment relationship is updated, it is lower than the safety threshold. The industrial data processing equipment reads the corresponding engineering drawing page based on the source location, restores the original bounding box coordinates, and traces the connection object along the interface of compressor C and the equipment connection line, forming a target area containing compressor C, the complete connection line, and safety valve D.

[0147] After re-extracting the target area, the drawing knowledge of the actual connection between compressor C and safety valve E is obtained. The re-extracted drawing knowledge is compared with the textual knowledge in the technical manual, and the structural feasibility tensor is used to verify compressor C, safety valve E, and the connection direction. If the verification result is feasible, the equipment relationship between compressor C and safety valve E is formed into new knowledge to be added to the database. After comparing the new knowledge with existing knowledge, it is written into the industrial knowledge graph, and the equipment relationship between compressor C and safety valve E is set to the current valid state. This completes the updating of equipment relationship weights, the deactivation of equipment relationships, the location of the source, the re-extraction of the target area, and the updating of the industrial knowledge graph.

[0148] In summary, by comparing the consistency of drawing knowledge and textual knowledge extracted from engineering drawings and technical manuals, and by using industrial physical topology rules to verify the physical feasibility of conflicting relationships, content that does not conform to the actual connection relationships of equipment can be filtered out before knowledge is added to the database. When attribute conflicts occur, mutually exclusive candidate attributes are retained, and the target attribute is determined by combining external evidence, reducing erroneous updates caused by directly overwriting existing knowledge. At the same time, the weight of equipment relationships is dynamically adjusted based on industrial business feedback, the time difference of knowledge sources, and the severity of faults. When the weight of equipment relationships is lower than the safety threshold, the drawing knowledge is re-extracted from the target area of ​​the engineering drawings based on the source location, realizing the targeted correction of abnormal equipment relationships. This improves the accuracy, traceability, and reliability of the continuous updating process of the industrial knowledge graph.

[0149] Based on another preferred embodiment described above, see [link to preferred embodiment]. Figure 3 As shown, this embodiment provides a self-evolving industrial knowledge graph system based on physical topology constraints, used to apply the above-described self-evolving industrial knowledge graph method based on physical topology constraints, including: The knowledge extraction unit is used to extract drawing knowledge and text knowledge from engineering drawings and technical manuals, and record the source location of the drawing knowledge; The topology verification unit is used to compare the consistency of the drawing knowledge and text knowledge, perform physical feasibility verification on conflicting relationships according to preset industrial physical topology rules, and merge the processed drawing knowledge and text knowledge to form knowledge to be added to the database. The conflict healing unit is used to compare the knowledge to be added to the database with the existing knowledge in the industrial knowledge graph; when there is no attribute conflict, it is written into the industrial knowledge graph; when there is an attribute conflict, a conflict node is generated that retains mutually exclusive candidate attributes, external evidence is obtained, the target attribute is determined from the mutually exclusive candidate attributes based on the external evidence, the remaining candidate attributes and their corresponding relationships are deleted, and the industrial knowledge graph is updated. The weight evolution unit is used to update the weight of equipment relationships in the industrial knowledge graph according to industrial business feedback, knowledge source time difference and fault severity after the industrial knowledge graph is updated; when the weight of the equipment relationship is lower than the safety threshold and the corresponding equipment relationship has the source position, the corresponding equipment relationship is deactivated, drawing knowledge is re-extracted from the target area of ​​the engineering drawing according to the source position, and the industrial knowledge graph is updated according to the re-extracted drawing knowledge.

[0150] It is understandable that the above-mentioned industrial knowledge graph self-evolution method and system based on physical topology constraints have the same beneficial effects, and will not be elaborated further here.

[0151] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the present invention.

Claims

1. A self-evolution method for industrial knowledge graphs based on physical topological constraints, characterized in that, include: Extract drawing knowledge and textual knowledge from engineering drawings and technical manuals, and record the source location of the drawing knowledge; The consistency of the drawing knowledge and text knowledge is compared, and the physical feasibility of conflicting relationships is verified according to the preset industrial physical topology rules. The processed drawing knowledge and text knowledge are then merged to form knowledge to be added to the database. The knowledge to be added to the database is compared with the existing knowledge in the industrial knowledge graph; if there is no attribute conflict, it is written into the industrial knowledge graph. When attribute conflicts exist, conflict nodes that retain mutually exclusive candidate attributes are generated, external evidence is obtained, the target attribute is determined from the mutually exclusive candidate attributes based on the external evidence, the remaining candidate attributes and their corresponding relationships are deleted, and the industrial knowledge graph is updated. After the industrial knowledge graph is updated, the weights of equipment relationships in the industrial knowledge graph are updated according to industrial business feedback, time difference of knowledge source, and severity of fault. When the weight of the device relationship is lower than the safety threshold and the corresponding device relationship has the source location, the corresponding device relationship is deactivated, drawing knowledge is re-extracted from the target area of ​​the engineering drawing based on the source location, and the industrial knowledge graph is updated based on the re-extracted drawing knowledge.

2. The self-evolution method for industrial knowledge graphs based on physical topological constraints according to claim 1, characterized in that, When extracting the drawing knowledge and text knowledge, the following are included: Equipment identifiers, equipment attributes, and equipment relationships are extracted from the engineering drawings to form the drawing knowledge; equipment identifiers, equipment attributes, and equipment relationships are extracted from the technical manual to form the text knowledge; the knowledge source and extraction confidence level of each equipment attribute and equipment relationship are recorded.

3. The self-evolution method for industrial knowledge graphs based on physical topological constraints according to claim 2, characterized in that, The consistency comparison includes: Associating the drawing knowledge and text knowledge with the device identifier, the connection objects and connection directions of the mutually associated device relationships are compared; if both the connection objects and connection directions are consistent, the corresponding device relationship is retained; if at least one is inconsistent, the corresponding device relationship is determined to be the conflict relationship.

4. The self-evolution method for industrial knowledge graphs based on physical topological constraints according to claim 3, characterized in that, The physical feasibility verification includes: A structural feasibility tensor is constructed based on the industrial physical topology rules. Verification results are obtained based on the device identifier, connection object, and connection direction of the conflict relationship. Conflict relationships that are not feasible are deleted. If multiple feasible conflict relationships remain, the conflict relationship with the highest confidence is retained. The device attributes and device relationships that have not had any conflicts, as well as the conflict relationships retained after verification, are merged to form the knowledge to be added to the database.

5. The self-evolution method for industrial knowledge graphs based on physical topological constraints according to claim 4, characterized in that, When generating the conflicting nodes, the following steps are included: Device attributes with the same device identifier and attribute type but different attribute values ​​in the knowledge to be added to the database and the existing knowledge are identified as mutually exclusive candidate attributes; candidate confidence is determined based on the extraction confidence and knowledge source of each mutually exclusive candidate attribute, and a probabilistic hybrid node containing device identifier, attribute type, mutually exclusive candidate attribute, knowledge source and candidate confidence is generated, and the probabilistic hybrid node is used as the conflict node.

6. The self-evolution method for industrial knowledge graphs based on physical topological constraints according to claim 5, characterized in that, Obtaining the external evidence includes: Retrieve the equipment maintenance records and equipment ledgers corresponding to the equipment identifier; generate confirmation content based on the mutually exclusive candidate attributes and knowledge sources and receive manual confirmation results; use the retrieval results and manual confirmation results as the external evidence.

7. The self-evolution method for industrial knowledge graphs based on physical topological constraints according to claim 6, characterized in that, Determining the target attribute includes: When the manual confirmation result points to a mutually exclusive candidate attribute, the corresponding mutually exclusive candidate attribute is determined as the target attribute; when it does not point to one, the candidate confidence is updated based on the support results of the equipment maintenance records and equipment ledger for each mutually exclusive candidate attribute, and the mutually exclusive candidate attribute that reaches the confirmation threshold and is higher than the confidence of the other candidates is determined as the target attribute; when the target attribute is not determined, the conflict node is retained.

8. The self-evolution method for industrial knowledge graphs based on physical topological constraints according to claim 1, characterized in that, Updating the device relationship weights includes: The feedback adjustment amount is determined based on the industrial business feedback, the timeliness decay amount is determined based on the knowledge source time difference, the fault penalty amount is determined based on the fault severity, and the equipment relationship weight is updated based on the feedback adjustment amount, timeliness decay amount, and fault penalty amount; when the fault severity reaches a high-risk level, a discrete fault penalty amount corresponding to the high-risk level is adopted.

9. The self-evolution method for industrial knowledge graphs based on physical topological constraints according to claim 8, characterized in that, When re-extracting the drawing knowledge, the following is included: When the weight of the device relationship with the source location is lower than the security threshold, the corresponding device relationship is deactivated. The target area is extracted based on the page number and bounding box coordinates in the source location, and the drawing knowledge is re-extracted. The re-extracted drawing knowledge is compared with the text knowledge for consistency. For conflicting relationships, the physical feasibility is checked to form new knowledge to be added to the database. The new knowledge to be added to the database is compared with the existing knowledge. If there is no attribute conflict, it is written into the industrial knowledge graph. If there is an attribute conflict, the conflict node is generated. The external evidence is obtained. The target attribute is determined from the mutually exclusive candidate attributes based on the external evidence. The remaining candidate attributes and their corresponding relationships are deleted, and the industrial knowledge graph is updated.

10. A self-evolving industrial knowledge graph system based on physical topology constraints, used to apply the self-evolving industrial knowledge graph method based on physical topology constraints as described in any one of claims 1-9, characterized in that, include: The knowledge extraction unit is used to extract drawing knowledge and text knowledge from engineering drawings and technical manuals, and record the source location of the drawing knowledge; The topology verification unit is used to compare the consistency of the drawing knowledge and text knowledge, perform physical feasibility verification on conflicting relationships according to preset industrial physical topology rules, and merge the processed drawing knowledge and text knowledge to form knowledge to be added to the database. The conflict healing unit is used to compare the knowledge to be added to the database with the existing knowledge in the industrial knowledge graph; when there is no attribute conflict, it is written into the industrial knowledge graph. When attribute conflicts exist, conflict nodes that retain mutually exclusive candidate attributes are generated, external evidence is obtained, the target attribute is determined from the mutually exclusive candidate attributes based on the external evidence, the remaining candidate attributes and their corresponding relationships are deleted, and the industrial knowledge graph is updated. The weight evolution unit is used to update the equipment relationship weights in the industrial knowledge graph based on industrial business feedback, knowledge source time difference, and fault severity after the industrial knowledge graph is updated. When the weight of the device relationship is lower than the safety threshold and the corresponding device relationship has the source location, the corresponding device relationship is deactivated, drawing knowledge is re-extracted from the target area of ​​the engineering drawing based on the source location, and the industrial knowledge graph is updated based on the re-extracted drawing knowledge.

Images