A multimodal operation and maintenance document driven knowledge graph construction method

Abstract

The application discloses a kind of multi-modal operation and maintenance document driven knowledge graph construction method, including steps S100 to S700, and optional atlas evolution management steps S810 to S830.The application belongs to the field of computer data processing and information system operation and maintenance technology, specifically relates to multi-modal document layout analysis and knowledge graph construction, more specifically refers to a kind of vector coordinate system of drawing as geometric reference, through layout anchor group semantic alignment and running state data one-way closed loop check, realize the method, device, electronic equipment and computer readable storage medium of operation and maintenance knowledge graph construction.Effectively solve the problem that topological geometric relationship precision is insufficient, the natural corresponding relationship of text and image is invalid, atlas has no closed loop check mechanism, and the credibility of triple cannot be traced and cannot be arbitrated.

Description

Technical Field

[0001] This invention belongs to the field of computer data processing and information system operation and maintenance technology, specifically involving multimodal document layout parsing and knowledge graph construction. More specifically, it refers to a method, device, electronic device, and computer-readable storage medium for constructing an operation and maintenance knowledge graph by using the vector coordinate system of the image as a geometric reference and through semantic alignment of layout anchor groups and one-way closed-loop verification of runtime data. Background Technology

[0002] Knowledge graphs are the core foundation for operational data platforms to achieve root cause analysis of faults, impact analysis of changes, and capacity prediction. Knowledge sources in the field of information system operations and maintenance can be categorized into three types: structured data, semi-structured data, and unstructured documents. Among these, unstructured documents, which contain device connection relationships, port attributes, and dependency topologies, are difficult to fully cover using existing structured data collection methods and are the core processing object of this invention.

[0003] Current industry technology primarily focuses on improving extraction accuracy, but it still lacks traceable, arbitrable, and self-correcting mechanisms for determining the reliability of topological relation triples. Existing representative solutions have the following shortcomings: I. Knowledge Graph Construction Scheme for Wiring Diagrams: This scheme employs a top-down four-level analytical framework to extract topological relationships, combining improved target detection algorithms and OCR technology to identify graphic elements and device numbers, and constructs a hierarchical knowledge graph based on semantic rules. Its drawbacks are: port connection relationships rely on probabilistic inference from the identification results rather than geometric inverse solutions; the graph construction process does not retain triple evidence chains, making it impossible to trace topological deviations; and it lacks access to operational data to form a closed-loop verification mechanism, making the knowledge graph prone to drift with document version iterations.

[0004] II. General Multimodal Fusion Scheme: Semantic features of multimodal data are extracted separately, and different modal data are projected onto the same vector space. Entity mapping relationships are learned through prior alignment data to achieve fusion and disambiguation. Its drawbacks are: reliance on large-scale annotated corpora, which lack corresponding annotation resources in the operations and maintenance field; the use of independent processing links for text and images, resulting in the loss of the natural anchoring relationships between figures, captions, and text citations within the document, making it impossible to automatically identify text-image contradictions; and general visual inspection does not explicitly process the precise connection points between ports and polylines.

[0005] III. Adaptive Operation and Maintenance Graph Generation Scheme: This scheme extracts candidate knowledge units through semantic analysis and machine learning, completes knowledge fusion based on relevance and confidence, and achieves adaptive graph updates by combining new data sources. Its drawbacks are: entities and relationships are derived from semantic inference, which cannot adapt to the strict geometric determinism of port connection relationships in operation and maintenance topology documents; the confidence level is a model self-evaluation result, lacking an arbitration mechanism for comparison with live network operational data, making it difficult to locate and self-correct topology deviations.

[0006] In summary, existing technologies for constructing knowledge graphs for operations and maintenance generally suffer from four problems: insufficient accuracy of geometric relationships, failure of natural correspondence between graphs and texts, lack of closed-loop verification mechanisms, and untraceable and non-arbitrable credibility of triples. Summary of the Invention

[0007] To address the above issues and overcome the shortcomings of existing technologies, this invention proposes a multimodal operation and maintenance document-driven knowledge graph construction method, which effectively solves the problems of insufficient accuracy of topological geometric relationships, failure of natural correspondence between graphs and texts, lack of closed-loop verification mechanism for the graph, and untraceable and unarbitrable credibility of triples.

[0008] The technical solution adopted by this invention is as follows: a multimodal operation and maintenance document-driven knowledge graph construction method, including steps S100 to S700, and optional graph evolution management steps S810 to S830: S100: Obtain the multimodal operation and maintenance document set to be processed. The multimodal operation and maintenance document set includes equipment topology diagrams, data center cabling diagrams, equipment list tables, operation and maintenance manuals, fault handling cases, and shift handover records. Perform layout analysis on the multimodal operation and maintenance document set under a unified coordinate system, extract the figure annotation area, text area, table area, and table title area, and retain the relative position and size of each area in the original coordinate system, and output the document layer with layout annotations.

[0009] S200: Perform a three-segment vector analysis of each drawing, including the equipment topology drawing and the computer room wiring diagram. Using the vector coordinate system carried by the drawing as the geometric reference, separate geometric elements and text elements from the vector description of the drawing. Register the text elements as device name tags and port name tags, and extract the geometric center and circumscribed dimensions of the geometric elements. Label the device type according to the shape characteristics of the geometric elements and the label text of the text elements. Snap the endpoints of the connecting lines to the nearest element boundary point and perform polyline vertex merging and orthogonalization processing. Define the port by the intersection of the orthogonal polyline and the element boundary. Obtain the connection triplet and attribute triplet by the analytical intersection geometry solution of the geometric element boundary and the orthogonal polyline within the vector coordinate system.

[0010] S300: Construct a page anchoring group based on the aforementioned image: Use the image caption text and the citation sentences in the text that cite the image as anchoring corpus, together with the page area coordinates, to form a page anchoring group. Perform normalization matching between the device name and port name in the image element and the normalized name in the anchoring corpus to form an image element-anchor text identity evidence pair, which serves as the unique evidence basis for each triple in the knowledge graph.

[0011] S400: For each device connection, the four signals—normalized name consistency signal, semantic similarity signal, layout context matching signal, and connection line topology neighborhood matching signal—are weighted and fused into an identity determination value. .

[0012] S500: Perform four-level processing on the connection triplet and the attribute triplet according to the identity determination value: If the value is greater than or equal to the first threshold, it will be incorporated into the main map. When the threshold is greater than or equal to the second threshold and less than the first threshold, automatic evidence enhancement review is triggered. If the value is greater than or equal to the third threshold and less than the second threshold, the verification will be suspended. If the value is less than the third threshold, the candidate value is rejected; all four processing steps retain all candidate values ​​and their corresponding evidence chains.

[0013] S600: Unify the operation and maintenance state data into a five-tuple of entity, attribute, value, time, and source, and perform equivalence comparison and conflict detection with the attribute triples in the knowledge graph. For conflicting items, adjust the identity judgment value of the connection relationship of the device to which the conflicting item belongs by the following formula: ,in The identity determination value obtained by weighted fusion in step S400 is... The conflict reduction factor and Repeat the downward adjustment until... If it falls within the verification range of [0.30, 0.50), then... As a valid identity determination value for the connection relationship, the confidence marker of the connection relationship in the main graph is changed from verified to pending verification and an automatic verification task is generated. The adoption value of the conflicting attribute is automatically selected according to the preset value source credibility sorting and the time newer priority rule.

[0014] S700: The weights of the four signals are iteratively updated using the verification results of the automatic verification task as feedback signals. After each round of updates, the weights are renormalized so that their sum is 1.

[0015] The operation and maintenance data is not included in the identity determination value. The calculation is only used as input for conflict detection, value arbitration and weight feedback; and the operation and maintenance status data only has falsification effect, that is, it can only trigger the reduction or maintenance of the identity judgment value or the review judgment value, and cannot be used to increase the identity judgment value or the review judgment value.

[0016] Among them, the It is only used for confidence flag determination and automatic verification task triggering, and is neither written back to the weighted fusion calculation in step S400 nor used for the four-level processing in step S500; the downgrade only changes the confidence flag and does not trigger the removal of the connection relationship from the main map. The removal is only triggered by independent removal conditions. The independent removal conditions are that the verification result of the automatic verification task is a recurrence or rollback and has not been resolved after 3 consecutive verification cycles. The verification cycle is counted from the generation of one verification result to the generation of the next verification result.

[0017] It should be noted that in this invention, "the running data only has falsification effect" means that when the running data is consistent with the document extraction conclusion, it does not constitute a basis for adding points to or incorporating the triple; only when the running data shows that the connection relationship or attribute value is not valid will its judgment value be reduced by a multiplicative reduction coefficient.

[0018] It should be further clarified that in this invention, "falsification" and "value arbitration" belong to two independent semantic levels: at the credibility judgment level, the identity judgment value or review judgment value of the running data is only adjusted down by a multiplicative reduction coefficient. When the running data is consistent with the document extraction conclusion, the reduction coefficient is 1 and does not constitute a bonus. At the value arbitration level, when the running data conflicts with the map value, the adopted value is independently selected according to the credibility of the value source from high to low: change record, configuration management database, monitoring platform time series, and document extraction results. The former determines "whether the triple is credible," and the latter determines "which value to take for the attribute," and the two do not overlap.

[0019] It should also be noted that the "identity determination value" mentioned in this invention "The 'confidence marker' of the triple in the spectrum is one of two independent variables:" The confidence level is a numerical value that dynamically changes with evidence enhancement, runtime conflicts, and weight iterations. The confidence flag is the state identifier of the triple in the main graph, with values ​​of verified, pending verification, or rejected. Step S500's four-level processing determines whether candidate triples not yet included in the graph should be initially merged. Step S600's downgrading applies to triples already included in the main graph, changing only their confidence flag without altering their state. These two processes follow different state transition paths, thus avoiding state machine conflicts caused by the same numerical value simultaneously carrying both confidence and state semantics. In this invention, triples not included in the main graph but suspended in the pending verification buffer have a confidence flag value of pending verification. This confidence flag is maintained independently of the main graph and is not marked as rejected simply because the triple has not been included in the main graph.

[0020] It should also be noted that the "automation" referred to in this invention means that the entire process from document input to graph output does not require manual judgment and entry of each item: merging, review, suspension, rejection, arbitration and weight update are all automatically executed by the system according to the rules; among them, the suspended triples to be verified are processed in a closed loop by subsequent automatic verification tasks.

[0021] Preferably, step S200 specifically includes: S210: Read the vector description carried by the graphic, which includes primitive geometric objects, text objects and polyline objects, and the coordinates of the three are all represented by the vector coordinate system; directly read the geometric boundary of the primitive geometric object and the path vertex sequence of the polyline object from the vector description to obtain the primitive boundary and polyline path without rasterization and image segmentation; S215: For the map that does not carry a text object or is missing a text object in the vector description, the map is rasterized and rendered at a preset resolution. A text detection network is used to perform text region detection on the rendered map, and the text box coordinates and text content are output. The text box coordinates are converted back to the vector coordinate system according to the ratio between the preset resolution and the vector coordinate system, and then filled back as text primitives. S220: Separate geometric primitives and text primitives according to the vector coordinate system: Register text primitives located inside the geometric object of the primitive and matching the device naming convention regular template as the device name label of the primitive; register text primitives located within a preset proximity range of the boundary of the geometric object of the primitive and matching the port number regular template set as the port name label of the primitive; the device name label and the port name label constitute the device name and port name of the primitive; the geometric primitive is retained as the primitive outline; S230: Calculate the geometric center coordinates and the size of the bounding rectangle of the primitive geometric object from the geometric boundary, and calculate the aspect ratio and the number of skeleton branches of the primitive from the geometric boundary as shape features; S240: The device type is determined by combining the shape feature and the device name label. The candidate device type set includes server, switch, router, firewall, load balancer, storage device, and terminal. The device type and confidence level of each graphic element are output. S250: The nearest point algorithm is used to snap the endpoints of the connection line to the nearest primitive boundary point. Then, the vertex merging and orthogonalization of the polyline are performed to obtain the endpoint port identifier and connection type. The connection type includes direct connection, uplink, downlink, and relay link. When there are multiple candidate endpoints for the same connection line, the geometric distance and port number lexicographical order are used to disambiguate and select a unique endpoint. S260: Define the intersection of orthogonal polylines and primitive boundaries as ports, assembling them into connection triples: source device: source port, connection type, destination device: destination port, and attribute triples: device name, attribute, and value. Attribute values ​​cover four mandatory categories: network segment, VLAN, IP address, and service affiliation.

[0022] The connection triplet is obtained by the inverse geometric solution of the intersection of the primitive boundary and the orthogonal polygonal line in the vector coordinate system, and is not determined by probabilistic inference of text recognition results or visual detection results, thereby ensuring the geometric determinism of the port connection relationship.

[0023] The device name label and the port name label are distinguished according to the following rules: text elements located inside the geometric object and matching the device naming convention regular expression template are registered as device name labels; text elements located within a preset proximity range of the boundary of the geometric object and matching the port number regular expression template set are registered as port name labels; the preset proximity range is 3% to 8% of the short side of the outer rectangle of the element.

[0024] The upward and downward directions are determined by the vertical relative positions of the graphic elements on the graphic layout: the graphic element with the smaller ordinate of its geometric center is located at the top of the layout, and the graphic element with the larger ordinate is located at the bottom of the layout. A connecting line pointing from a graphic element located at the bottom of the layout to a graphic element located at the top of the layout is considered upward, and a connecting line pointing from a graphic element located at the top of the layout to a graphic element located at the bottom of the layout is considered downward. A direct connection is determined by the orthogonal broken line not passing through other graphic elements, and a relay link is determined by the orthogonal broken line passing through multiple graphic elements carrying the same aggregation group identifier. When the difference in the ordinates of the geometric centers of the graphic elements is less than a preset ordinate difference threshold, it is determined to be a direct connection within the same layer. In step S250, the endpoint adsorption radius is taken as 2% to 5% of the shorter side of the outer rectangle of the graphic element, and the orthogonalization tolerance is taken as 10° to 20°.

[0025] When the map does not carry a vector description and is provided only as a raster image, as an alternative branch to step S210 in step S200 where the vector description carried by the map is read, the raster image is first calibrated to the vector coordinate system, and steps S211 to S213 are executed: S211. Use a text detection network to detect the coordinates of the text box, and perform masking and zeroing on the area covered by the coordinates of the text box to obtain a residual image; S212. First, determine the value range of the segmentation threshold for the residual image according to the gray-level histogram. The value range is from 0.5 times the gray-level mean to 0.95 times the gray-level mean. Then, within this value range, use the maximum inter-class variance method to determine the segmentation threshold. S213. Perform morphological thinning and Zhang-Suen thinning on the segmentation results to obtain the primitive central skeleton, and extract the primitive geometric center coordinates and the outer rectangle size from the primitive central skeleton, and convert them back to the vector coordinate system.

[0026] Specifically, steps S211 to S213 and step S210 are executed selectively; that is, step S210 is not executed when steps S211 to S213 are executed, and step S220 and its subsequent steps continue to apply after steps S211 to S213 are completed. At this time, the endpoints of the connecting lines are determined by the endpoints of the central skeleton of the primitive, and the connecting triplets are still obtained by geometric inverse solving of the intersection of the primitive boundary and the orthogonal broken line within the vector coordinate system.

[0027] Preferably, step S300 specifically includes: S310: Based on the document layer with layout annotations, locate the caption area and text reference area corresponding to a single image; S320: Establish a one-to-one correspondence between map identifiers and map caption text, and extract complete sentences in the main text paragraphs that cite the map identifiers as citation sentences; S330: Perform name normalization processing, mapping port numbers and English abbreviations to normalized names in a unified format. The name normalization processing includes unifying full-width and half-width characters, expanding English abbreviations, and normalizing port number formats. S340: A page anchor group is composed of a quadruple of map primitive, normalized name, anchor text and page coordinates, which serves as the unique evidence basis for each triple in the knowledge graph. S350: For primitives for which no anchoring corpus was found in either the caption area or the text area, the normalized name consistency signal and semantic similarity signal are set to zero, and the connection triples and attribute triples corresponding to the primitives are suspended for verification and not incorporated into the main graph. The confidence flag of the suspended triples not incorporated into the main graph is set to verification, and the confidence flag is maintained independently of the main graph.

[0028] The significance of setting S350 is that the anchored group is the unique evidence basis for the triple; without anchored corpus, there is no evidence basis. If such primitives participate in weighted fusion according to the default values ​​of consistency signal and semantic similarity signal, their identity determination value will be determined solely by the prior knowledge of layout position and connection, resulting in inclusion without textual evidence support, which constitutes a missing evidence chain. Suspending them directly as pending verification can prevent the inclusion of primitives without evidence without deleting the evidence itself, and can be automatically re-evaluated after subsequent document versions supplement the captions or quotations.

[0029] In step S330, the expansion of English abbreviations is performed based on the device type-abbreviation mapping table constructed according to the device naming specification document in the operation and maintenance field, and the port number format normalization is performed based on the port number regular expression template set; for names not covered by the mapping table and the regular expression template set, the original string is retained and marked as unnormalized, and the normalized name consistency signal of the unnormalized name is zero in step S400.

[0030] To facilitate the four-signal fusion determination, in step S400, the identity determination value is calculated using the following formula: ;in The normalized name consistency signal is calculated from the matching degree of the device name of the graphic element and the normalized name in the anchor text using the Jaccard string. The semantic similarity signal is calculated from the cosine similarity of the sentence vectors of the anchored text and the primitive-annotated text. The layout context matching degree signal is calculated from the matching degree between the relative position of the graphic element and the prior position of the device of this type; The connection topology neighborhood matching degree signal is calculated from the matching degree between the connection topology neighborhood and the connection prior of this type of device; each signal value satisfies 0 ≤ ≤1.

[0031] The location prior and the connection prior are constructed according to the device type. The construction method is as follows: using historical map instances that have been incorporated into the main map as statistical samples, the relative position distribution of the device elements in the map layout and the connection degree distribution of the device ports are statistically analyzed respectively, and the normalized probability density of the distribution is used as the matching degree value; when the historical samples are insufficient, the rules exported from the equipment networking specification document in the operation and maintenance field are used to determine the matching degree.

[0032] The weight values ​​range are , , , And after selecting from the range of values, normalization is performed to make After each round of update in step S700, if any component of the updated weight exceeds the range of values, the component is clipped to the boundary of the corresponding interval and then renormalized so that the sum is 1. The clipping and normalization are repeated until all four components fall within the range of values, or until the number of iterations reaches the preset upper limit of the number of iterations, where the preset upper limit of the number of iterations is 10.

[0033] Special note: The runtime consistency signal is not involved in the calculation of this judgment value. It is only used as a conflict detection and arbitration trigger signal in the subsequent runtime verification stage and has only the effect of falsification.

[0034] Further, in step S500, the first threshold is 0.75, the second threshold is 0.50, and the third threshold is 0.30. The four-level processing is as follows: When merged into the main atlas; Automatic review of evidence enhancement is triggered at any time; The application is currently suspended pending verification. When merging, the system rejects the entry. All candidate values ​​and evidence chains are retained in all four positions. When merging, existing values ​​are not overwritten, and when rejecting, evidence is not deleted, so that each triple in the graph is always in one of three states: verified, pending verification, or rejected.

[0035] The enhanced automatic verification of evidence includes: S510, Anchor Corpus Expansion: Retrieve the occurrence positions of the same canonical name in other diagrams, table areas, device lists, and the main text of the maintenance manual within the multimodal operation and maintenance document set, construct an expanded anchor corpus, and recalculate the canonical name consistency signal and semantic similarity signal accordingly; S520, Operational Bypass Verification: Read the operational quintuple of the ports involved in the connection relationship, determine the non-existence of the connection relationship based on the port status, link traffic, and alarm events, and output a bypass reduction coefficient. The bypass reduction coefficient takes a preset value of 0.5 to 0.8 when the detected port is closed, the link traffic is continuously zero, or there is a link interruption alarm and the operational status indicates that the connection relationship does not exist; otherwise, it takes 1; S530, Topology Neighborhood backtracking: Using the set of adjacent edges that the primitive has been incorporated into the main graph as constraints, check whether the connection relationship to be verified constitutes port duplication or link non-pair conflict with the topology already incorporated into the graph, and output the backtracking pass coefficient. The backtracking pass coefficient takes a preset value of 0.5 to 0.8 when a conflict is detected, and takes 1 when no conflict is detected; S540, multiply the four-signal weighted value after the recalculation of the extended anchoring corpus by the bypass reduction coefficient and the backtracking pass coefficient to obtain the verification judgment value; the bypass reduction coefficient and the backtracking pass coefficient are both not greater than 1, so that the running bypass verification and the topology neighborhood backtracking only have the effect of falsification; when the verification judgment value is greater than or equal to the first threshold, it is incorporated into the main graph, otherwise it is suspended as pending verification and all candidate values ​​and evidence chains are retained.

[0036] Since the verification judgment value is only multiplicatively reduced by a coefficient not greater than 1 on the recalculated weighted value of the four signals, the bypass reduction coefficient is 1 when the running data is consistent and does not constitute an addition. The reduction is only made when the running data indicates that the connection is not established. Therefore, the verification process does not constitute a circular argument to prove the connection relationship with the running data.

[0037] To achieve closed-loop self-calibration, step S600 specifically includes: S610: Extract asset attributes from the configuration management database, extract time-series indicators from the monitoring platform, extract alarm events from the alarm platform, and extract change records from the change system, and uniformly format them into a five-tuple of entity, attribute, value, time, and source; S620: Perform an equivalence comparison and conflict detection between the quintuple and the attribute triple in the knowledge graph; S630: Comparison results are divided into three levels: Consistent items are not updated; conflicting items are adjusted down for the connection relationship using the following formula: ,in The identity determination value obtained by weighted fusion in step S400 The conflict reduction factor and Repeat the downward adjustment until... If it falls within the verification range of [0.30, 0.50), then... As the valid identity determination value of this connection relationship, the It is only used for confidence mark determination and automatic verification task triggering, and is neither written back to the weighted fusion calculation in step S400 nor used for the four-level processing in step S500; at the same time, the confidence mark of the connection relationship in the main graph is changed from verified to pending verification, and an automatic verification task is generated; the pending verification item remains unchanged. S640: For conflicting attributes, the change records, configuration management database, monitoring platform time series, and document extraction results are extracted in descending order of the credibility of the value source. Within the same source, the newest value is given priority. The adopted value is automatically selected and written into the graph, while the overwritten value along with its evidence chain is retained.

[0038] Wherein, the conflict reduction coefficient The value is taken as 0.6 to 0.8, preferably 0.7. The downward adjustment does not change the existence state of the connection relationship in the main graph. The removal of the connection relationship from the main graph is triggered only by an independent removal condition. The independent removal condition is: the verification result of the automatic verification task is a recurrence or rollback, and it has not been resolved after three consecutive verification cycles; the verification cycle is counted from the generation of one verification result to the generation of the next verification result. Therefore, the conflict only leads to the downgrading of the confidence flag and the generation of the automatic verification task, and will not cause a conflict with the "..." in step S500 due to the judgment value falling into the verification interval. The rule of "suspending for verification and not merging for the time being" creates a state semantic conflict.

[0039] The runtime data comes from within the operation and maintenance system. The extraction process involves de-identification and field-level anonymization, retaining only the device and port dimension attributes required for graph construction.

[0040] To achieve weight adaptation, in step S700, the weight update is performed according to the following formula: ,in Let be the weight vector updated in round t, and let the learning rate α be between 0.10 and 0.20, preferably 0.15. Let t be the feedback gradient. The normalized vector.

[0041] When there are samples with multiple verification results in round t. The arithmetic mean of the feedback gradients of all verified samples in this round is then normalized to obtain the result. The weight update is performed in batches: each round consists of no less than 600 verified triples. The feedback gradient of each sample is determined according to the sample category, and then the weight update is performed after taking the arithmetic mean.

[0042] The rule for constructing the feedback gradient is as follows: Let the mean vector of the resolved class samples in round t be denoted as . The mean vector of the relapsed samples on the four signals is: Then take the already resolved samples. For relapsed samples That is, complementary components are selected one by one; for rollback samples, complementary components are selected... This causes the weights of the four signals to decay towards a uniform distribution in order to suppress the dominance of a single signal.

[0043] The verification results are automatically determined by the runtime data without external intervention: if the conflict of the same entity attribute does not reappear within K consecutive collection cycles from the generation of the automatic verification task and the map value is consistent with the runtime value, it is determined to be resolved, and K is 3; if the same conflict reappears within K collection cycles, it is determined to be recurred; if the system records a rollback event associated with the conflict attribute, it is determined to be rolled back.

[0044] The updated samples are verified triples with a size of no less than 600. The convergence criterion is that the change in the weight vector is less than 0.005 for 5 consecutive rounds. After each round of update, the weight vector is renormalized so that the sum of its components is 1. If any component exceeds the range of values ​​after normalization, it is clipped to the corresponding interval boundary and normalized again. This process is iterated until all components fall within the range of values ​​or the preset number of iterations is reached.

[0045] To ensure the stability of the graph during document version iterations, this method also includes a graph evolution management step: S810: When a new schema layer is generated during an upgrade of the operation and maintenance documentation version, instance migration mapping is performed on the shadow graph: device instances are matched by normalized name, and port instances are matched by the composite key of device normalized name and port normalized name; when multiple candidate instances are matched for the same normalized name, they are sorted from high to low by normalized name similarity, and if the similarity is the same, they are sorted from near to far by page coordinate distance, and one is selected for matching; the attribute values ​​and connection edges of successfully matched instances remain unchanged during migration, and unmatched instances are incorporated as new instances in the new graph, which is not considered a violation of conservation; the number and attribute values ​​of existing device instances and port instances with the same name remain unchanged during migration. S820: Perform three checks on the migration results: instance conservation check, i.e., the number of existing device instances and port instances with the same name does not increase or decrease before and after the migration; endpoint connectivity check, i.e., each port instance has at least one connection edge and does not exceed the maximum number of connections, the maximum number of connections is preset according to the port type, with 1 for ordinary ports and the upper limit of the number of members configured in the aggregation group for trunk aggregation ports; and required attribute non-empty rate check, i.e., the proportion of non-empty required attributes is not lower than that of the previous version. S830: If all three checks pass, the mode layer is switched atomically; if any one check fails, the original mode layer is retained and an evolution check report is generated. The migration is then re-executed after correction in a subsequent version.

[0046] Corresponding to the above method, the present invention also provides a knowledge graph construction device driven by multimodal operation and maintenance documents, including a layout analysis module, a vector parsing module, an anchoring group construction module, an identity determination module, an evidence chain retention and processing module, a runtime verification module, and a weight feedback module, and also includes an optional graph evolution management module; the layout analysis module is used to perform unified coordinate system layout analysis of the multimodal operation and maintenance document set and output a document layer with layout annotations; the vector parsing module is used to perform three-segment vector parsing of primitive-label-connection using the graph vector coordinate system as the geometric reference, and obtain connection triples and attribute triples through geometric inverse solution; the anchoring group construction module is used to construct graph-annotation-reference layout anchoring groups, forming primitive-anchor text identity evidence pairs as unique triples. The evidence base; the identity determination module is used to weight and fuse four signals—normalized name consistency, semantic similarity, layout context matching degree, and connector topology neighborhood matching degree—into an identity determination value; the evidence chain retention processing module is used to perform four levels of processing—inclusion, automatic review, suspension, and rejection—based on the identity determination value, and retain all candidate values ​​and evidence chains; the runtime verification module is used to unify the operation and maintenance runtime data into five-tuples, perform conflict detection, value arbitration, and confidence flag downgrading, and generate automatic verification tasks; the weight feedback module is used to iteratively update the weights of the four signals and re-normalize them based on the verification results; the graph evolution management module is used to perform instance migration mapping and three verifications on the shadow graph when the operation and maintenance document version is upgraded, and complete the atomic switching of the mode layer.

[0047] The present invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described method for constructing a knowledge graph in the field of information system operation and maintenance driven by multimodal operation and maintenance documents.

[0048] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described method for constructing a knowledge graph in the field of information system operation and maintenance driven by multimodal operation and maintenance documents; the computer-readable storage medium is a non-transitory tangible storage medium.

[0049] The beneficial effects achieved by this invention using the above structure stem from the integrated synergy of three elements: geometric inverse solution of the vector coordinate system, preservation of the unique evidence basis and evidence chain of the page anchoring group, and the one-way falsification channel in the running state. This is not a simple superposition of the effects of each step, but rather manifested in the following ways: 1. Instead of relying on image recognition and probabilistic inference to determine connection relationships, it directly uses the vector coordinates inherent in the drawing to accurately calculate the port docking relationship through the intersection of the element border and the connecting line, resulting in significantly less geometric precision loss and more accurate extraction of port attributes; 2. By utilizing the inherent correspondence between figures, captions, and text references within the document, it automatically compares and identifies discrepancies between the figures and text. Inconsistent points are directly marked as pending verification, significantly reducing the workload of manual verification. Third, each connection retains a complete chain of evidence and is managed uniformly according to three states: "verified," "pending verification," and "rejected." Erroneous relationships can be accurately located and will not spread disorderly within the graph, ensuring full traceability of all relationships. Fourth, operational data is used only for "error detection," not for "score enhancement," avoiding the logical problem of self-verification. The graph can self-correct based on actual operational conditions, preventing it from becoming increasingly skewed over time. Furthermore, correction only addresses problematic localities, avoiding complete recalculation each time and saving computational resources. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the overall process of a multimodal operation and maintenance document-driven knowledge graph construction method proposed in this invention; Figure 2 This is a schematic diagram of the primitive-tag-connection vector analysis of a multimodal operation and maintenance document-driven knowledge graph construction method proposed in this invention; Figure 3 This is a schematic diagram of the diagram-annotation-index page anchoring group structure for a multimodal operation and maintenance document-driven knowledge graph construction method proposed in this invention; Figure 4 This is a schematic diagram illustrating the four-signal identity determination and four-level evidence chain retention process of a multimodal operation and maintenance document-driven knowledge graph construction method proposed in this invention. Figure 5 This is a schematic diagram illustrating the runtime verification and conflict arbitration of a multimodal operation and maintenance document-driven knowledge graph construction method proposed in this invention. Figure 6 This is a schematic diagram of the weight feedback learning process of a multimodal operation and maintenance document-driven knowledge graph construction method proposed in this invention. Detailed Implementation

[0051] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0052] like Figure 1 As shown, this invention proposes a multimodal operation and maintenance document-driven knowledge graph construction method. The overall process starts with acquiring a multimodal operation and maintenance document set, which includes equipment topology diagrams, data center cabling diagrams, equipment lists, operation and maintenance manuals, fault handling cases, and shift handover records. The process outputs a document layer with layout annotations from the layout analysis layer, and then sequentially enters six processing stages: primitive vector parsing, layout anchoring group construction, four-signal identity determination, evidence chain retention in four levels, runtime verification and conflict arbitration, and weight feedback learning, finally outputting an incremental knowledge graph. The document layer with layout annotations output by the layout analysis module is transmitted to the vector parsing module and the anchoring group construction module, respectively. The vector parsing module outputs a continuous... The triplet, attribute triplet, and geometric coordinate data are transmitted to the identity determination module. The anchor group output by the anchor group construction module and the identity evidence pair are synchronously transmitted to the identity determination module. After calculating the identity determination value, the identity determination module transmits it to the evidence chain retention processing module, which outputs candidate triplets and corresponding evidence chains. The evidence chain retention processing module synchronizes the triplets that have been incorporated into the main graph to the runtime verification module. The runtime verification module connects to the external configuration management database, monitoring platform, alarm platform, and change system, and outputs conflict handling results and automatic verification tasks to the weight feedback module. The weight feedback module outputs the updated weight vector and sends it back to the identity determination module, forming a self-correcting closed loop.

[0053] like Figure 1 As shown, step S100 uses a layout analysis model to perform document layout parsing. The input is an RGB image with a tensor of 3×H×W, where H and W are the image height and width, respectively. The output is a region mask obtained by performing Softmax and argmax operations on the probability map of each pixel category. The document is divided into six regions: captions, body text, tables, table titles, headers, and footers. The coordinates and dimensions of each region in the original pixel coordinate system are preserved to provide a region boundary reference for subsequent anchor group construction.

[0054] For the table area, locate the table semantics according to the table title, restore the cell structure of the equipment list table, and output the records of table identifier, row identifier, column name, cell value, and cell page coordinates; for the operation and maintenance manual, fault handling cases and shift handover records in the main text area, retain the paragraph page coordinates after splitting by paragraph and chapter.

[0055] like Figure 1 and Figure 2 As shown, step S200, the vector parsing module, performs the primitive-label-connection vector parsing stage, using the vector coordinate system carried by the map as the geometric reference. The map includes equipment topology maps and server room wiring diagrams. The connection relationships are obtained by the analytical intersection geometry inverse solution of the primitive boundary and orthogonal polygonal lines within the coordinate system, without rasterization processing or image segmentation, and without probabilistic inference from text recognition or visual detection results. Specific execution steps: S210: Read the vector description carried by the drawing, which includes primitive geometric objects, text objects, and polyline objects, all of which are represented in the vector coordinate system of the drawing; directly read the geometric boundary of the primitive geometric objects and the path vertex sequence of the polyline objects from the vector description to obtain the primitive boundary and the polyline path. Typical forms of the vector description include SVG path and text elements, Visio ShapeSheet geometric and text fields, DXF primitive and polyline records, and path and text objects in PDF drawing instructions.

[0056] S215: For maps whose vector descriptions do not carry text objects or whose text objects are missing, the maps are rasterized and rendered at a preset resolution. A text detection network is used to perform text region detection on the rendered map, outputting the coordinates of the text boxes and the text content. The coordinates of the text boxes are then converted back to the vector coordinate system according to the ratio between the preset resolution and the vector coordinate system, and then filled back as text primitives.

[0057] S220: Separate geometric primitives and text primitives according to the vector coordinate system: Register text primitives located inside the geometric object of the primitive and matching the device naming convention regular template as the device name label of the primitive; register text primitives located within a preset proximity range of the boundary of the geometric object of the primitive and matching the port number regular template set as the port name label of the primitive; the device name label and the port name label constitute the device name and port name of the primitive; the geometric primitive is retained as the primitive outline. In this embodiment, the preset proximity range is 5% of the short side of the bounding rectangle of the primitive, the device naming convention regular template is in the form of "computer room identifier - rack identifier - serial number", and the port number regular template set covers prefixes such as GE, XGE, 10GE, Eth-Trunk, etc.

[0058] S230: Calculate the geometric center coordinates and circumscribed rectangle dimensions of the primitive geometric object from the geometric boundary, and calculate the aspect ratio and number of skeleton branches of the primitive from the geometric boundary as shape features.

[0059] S240: The device type is determined by combining the shape features and the device name label. The candidate device type set is set to seven categories: server, switch, router, firewall, load balancer, storage device, and terminal. The device type and confidence level corresponding to each graphic element are output.

[0060] S250: The nearest point algorithm is used to snap the endpoints of the connection line to the nearest primitive boundary point. The snapping radius is 2% to 5% of the short side of the primitive's circumscribed rectangle. Then, the vertex merging and orthogonalization of the polyline are performed. The orthogonalization tolerance is 10° to 20°. After processing, the endpoint port identifier and connection type are obtained. The connection type includes direct connection, uplink, downlink, and trunk link. When there are multiple candidate endpoints for the same connection line, disambiguation is performed according to the geometric distance and the lexicographical order of the port number to select a unique endpoint.

[0061] In this system, upward and downward connections are determined by the vertical relative positions of the elements on the page: the element with the smaller ordinate of its geometric center is located at the top of the page, and the element with the larger ordinate is located at the bottom. A connection line pointing from an element at the bottom of the page to an element at the top is considered upward, and vice versa. When the difference between the ordinates of the geometric centers of the two elements is less than a preset threshold, it is considered a direct connection within the same layer. Direct connections and relay links are determined by whether an orthogonal polyline passes through other elements: an orthogonal polyline that does not pass through other elements is considered a direct connection, and one that passes through multiple elements carrying the same aggregation group identifier is considered a relay link. This determination only requires the ordinates of the geometric centers of the elements and the vertex sequence of the polyline path, without introducing prior knowledge of the hierarchical topology, and can therefore be directly derived from geometric quantities within a vector coordinate system.

[0062] S260: Define the intersection of the orthogonal polyline and the boundary of the primitive as a port, and assemble it into a connection triple (source device: source port, connection type, destination device: destination port) and an attribute triple (device name, attribute, value). The attribute value covers four types of required attributes: network segment, VLAN, IP address, and service affiliation.

[0063] Cathyide branch: When the map does not carry a vector description and is provided only as a raster image, as an alternative branch to step S210 in step S200 which reads the vector description carried by the map, the raster image is first calibrated to the vector coordinate system, and steps S211 to S213 are executed: S211. Use a text detection network to detect the coordinates of text boxes, and then mask and zero out the area covered by the text box coordinates to obtain the residual image. S212. First, determine the range of the segmentation threshold for the residual image based on the gray-level histogram. The range is from 0.5 times the gray-level mean to 0.95 times the gray-level mean. Then, within this range, use the maximum inter-class variance method to determine the segmentation threshold. S213. Then, morphological thinning and Zhang-Suen thinning are performed to obtain the primitive central skeleton. The geometric center coordinates and the size of the circumscribed rectangle of the primitive are extracted from the primitive central skeleton and converted back to the vector coordinate system.

[0064] Specifically, steps S211 to S213 and step S210 are executed selectively; that is, step S210 is not executed when steps S211 to S213 are executed, and step S220 and its subsequent steps continue to apply after steps S211 to S213 are completed. At this time, the endpoints of the connecting lines are determined by the endpoints of the central skeleton of the primitive, and the connecting triplets are still obtained by geometric inverse solving of the intersection of the primitive boundary and the orthogonal broken line within the vector coordinate system.

[0065] like Figure 1 and Figure 3 As shown, step S300 uses the anchor group construction module to perform the page anchor group construction stage, constructing a page anchor group for each single image, establishing a unique evidence association between the image and text. The specific execution flow is as follows: S310: Based on the layout analysis results, locate the caption area and text reference area corresponding to a single figure; S320: Establish a one-to-one correspondence between map identifiers and map caption texts, and extract complete sentences in the main text paragraphs that cite the map identifier as citation sentences; S330: Perform name normalization processing, such as normalizing GE0 / 1 into a unified port number format and mapping the English abbreviation Core-SW1 into a normalized name; the expansion of the English abbreviation is performed according to the device type-abbreviation mapping table constructed by the device naming specification document in the operation and maintenance field, and the port number format normalization is performed according to the port number regular expression template set; for names not covered by the mapping table and the regular expression template set, the original string is retained and marked as unnormalized, and the normalized name consistency signal of the unnormalized name is zero in step S400; S340: A page anchor group is composed of a quadruple of map primitive, normalized name, anchor text and page coordinates, which serves as the unique evidence basis for each triple in the knowledge graph. S350: For primitives for which no anchored corpus was found in either the caption area or the text area, the normalized name consistency signal and semantic similarity signal are set to zero. The connection triples and attribute triples corresponding to the primitive are suspended for verification and are not incorporated into the main graph. The confidence flag of the suspended triples is set to be verified and is maintained independently of the main graph.

[0066] In this embodiment, the determination in S350 is completed using the map identifier as the retrieval key: if a map element does not have a corresponding map caption text in the caption area, and there is no sentence in the text area referencing the map identifier, then the anchoring corpus for that map element is empty. The triples corresponding to such map elements, along with their page coordinates, are written into the verification cache area. After subsequent document versions supplement the map captions or citation sentences, S310 to S340 will be automatically re-executed.

[0067] The significance of setting S350 is that the anchor group is the sole evidence basis for the triple; without anchored corpus, there is no evidence basis. If such primitives participate in weighted fusion according to the default values ​​of consistency signal and semantic similarity signal, their identity determination value will be determined solely by the prior of layout position and connection, resulting in inclusion without textual evidence support, which constitutes a missing evidence chain. Suspending them directly as pending verification can prevent the inclusion of primitives without evidence without deleting the evidence itself, and can be automatically re-evaluated after subsequent document versions supplement the captions or quotations.

[0068] like Figure 1 and Figure 4 As shown, step S400 employs a four-signal identity determination module to perform the identity determination stage for each device connection relationship, fusing the four types of signals to calculate the identity determination value. The calculation formula is as follows: .

[0069] in: The normalized name consistency signal is calculated from the Jaccard string matching degree between the primitive device name and the normalized name in the anchor text, with corresponding weights. Take 0.25; The semantic similarity signal is calculated from the cosine similarity of the BERT sentence vectors of the anchor text and the primitive-annotated text, with corresponding weights. Take 0.20; The layout context matching degree signal is calculated from the matching degree between the relative position of the primitive and its prior type, with corresponding weights. Take 0.15; The connection line topology neighborhood matching degree signal is calculated from the matching degree between the connection line topology neighborhood and the type prior, with corresponding weights. We set the weights to 0.40. The four weights satisfy... .

[0070] The location prior and connection prior are constructed separately according to the device type. The construction method is as follows: using historical map instances that have been incorporated into the main map as statistical samples, the relative position distribution of the device elements in the map layout and the connection degree distribution of the device ports are statistically analyzed respectively, and the normalized probability density of the distribution is used as the matching degree value; when the historical samples are insufficient, the rules exported from the equipment networking specification document in the operation and maintenance field are used to determine the matching degree.

[0071] The basis for determining the range of weight values ​​is as follows: The geometric and topological determinism that directly reflects the connection relationship is consistent with the technical line of the geometric inverse solution of this invention, and therefore is given the highest weight; Strong matching evidence at the name level, with secondary weight; This is a weak semantic-level matching evidence, easily affected by devices with the same name, and therefore has a lower weight. As statistical evidence at the location level, it has the weakest discriminative power and the lowest weight. Based on this, it is determined that... , , , And after selecting from the range of values, normalization is performed to make .

[0072] Special note: The runtime consistency signal is not involved in the calculation of this judgment value and only has falsification effect, serving only as a conflict detection and arbitration trigger signal in the subsequent runtime verification stage. This design allows the same data to be processed separately at the two semantic levels of credibility judgment and conflict arbitration, avoiding the circular reasoning defect of using the model's self-scoring as both the judgment basis and the verification basis.

[0073] like Figure 1 and Figure 4 As shown, step S500 employs the evidence chain preservation processing module to perform a four-tiered processing stage for evidence chain preservation, applying a four-tiered hierarchical processing method to the triples based on the identity determination value: when When the condition is met, it is determined to be merged and directly merged into the main graph, while retaining all candidate values ​​and the complete evidence chain, which includes the original anchor text, page coordinates, and the values ​​of each signal component. when When the evidence is enhanced, an automatic review is triggered. The system automatically expands the evidence and recalculates the review judgment value. If the review judgment value is not less than 0.75, it is incorporated into the main graph; otherwise, it is suspended for verification. when When the condition is met, it is determined to be pending verification and is suspended in the pending verification cache area, and is not merged into the main graph for the time being; when If the value is rejected, it will not be included in the main graph, but all candidate values ​​and evidence chains will be retained for backtracking.

[0074] The enhanced evidence automatic review is performed according to S510 to S540: First, the document set is searched for the occurrence of the same normalized name in other maps, equipment lists, and operation and maintenance manuals to construct an extended anchoring corpus, and s1 and s2 are recalculated; then, the running state quintuple of the ports involved in the connection relationship is read, and the non-existence of the connection relationship is determined based on the port status, link traffic, and alarm events, and a bypass reduction coefficient is output. This coefficient takes a preset value of 0.5 to 0.8 when the port is detected to be closed, the link traffic is continuously zero, or there is a link interruption alarm and the running state indicates that the connection relationship does not exist, and takes 1 in other cases; then, the set of adjacent edges that the primitive has been incorporated into the main graph is used as a constraint to check for repeated port occupation and unpaired link conflicts, and a backtracking pass coefficient is output. This coefficient takes a preset value of 0.5 to 0.8 when a conflict is detected, and takes 1 when no conflict is detected; finally, the weighted value of the four signals after the extended anchoring corpus is recalculated is multiplied by the bypass reduction coefficient and the backtracking pass coefficient to obtain the review judgment value.

[0075] Since both coefficients are not greater than 1, the verification judgment value is not higher than the recalculated weighted value of the four signals. The running data and topological neighborhood constraints can only be falsified, not verified. When the verification judgment value is not less than 0.75, it is merged into the main graph; otherwise, it is suspended for verification and all candidate values ​​and evidence chains are retained.

[0076] When merging, existing values ​​are not overwritten; when rejecting, evidence is not deleted. This ensures that each triple in the graph is always in one of three states: verified, pending verification, or rejected. When an erroneous relationship is disproven, it is only downgraded rather than silently deleted, thus ensuring traceability across the entire chain.

[0077] like Figure 1 and Figure 5 As shown, step S600 uses the runtime verification module to perform runtime verification and conflict arbitration, accesses multi-source operation and maintenance runtime data to perform closed-loop verification, and realizes the self-correction capability of the map. The specific process is as follows: S610: Extract asset attributes from the configuration management database, extract time-series indicators from the monitoring platform, extract alarm events from the alarm platform, and extract change records from the change system, and uniformly format them into a five-tuple of entity, attribute, value, time, and source; S620: Perform equivalence comparison and conflict detection between the quintuple and the attribute triple in the graph; S630: The comparison results are divided into three categories: consistent items are not updated; conflicting items are adjusted according to the following formula: ,in The identity determination value obtained by weighted fusion in step S400, where δ is the conflict reduction coefficient and Repeat the downward adjustment until... If it falls within the verification range of [0.30, 0.50), then... As the valid identity determination value of this connection relationship, the It is only used for confidence mark determination and automatic verification task triggering, and is neither written back to the weighted fusion calculation in step S400 nor used for the four-level processing in step S500; at the same time, the confidence mark of the connection relationship in the main graph is changed from verified to pending verification, and an automatic verification task is generated; the pending verification item remains unchanged. S640: For conflicting attributes, the change records, configuration management database, monitoring platform time series, and document extraction results are extracted in descending order of the credibility of the value source. Within the same source, the newest value is given priority. The adopted value is automatically selected and written into the graph, while the overwritten value along with its evidence chain is retained.

[0078] In this embodiment, Take 0.7. Based on a certain connection relationship... For example: after the first conflict It still doesn't fall within the [0.30, 0.50) range, so we repeat the reduction once more to get... If the value falls within the verification interval, the reduction stops, and 0.402 is used as the valid identity determination value. Simultaneously, the confidence flag for this connection is changed from verified to pending verification, and an automatic verification task is generated. The weighted fusion result, serving as evidence on the document side, is fully preserved and is not overwritten or deleted due to conflicts.

[0079] in, The confidence level is numerical, while the confidence label is a state-based identifier; the two are independent of each other. for Derivatives in conflict states, and Stored separately. Downgrading due to a conflict only changes the confidence flag and generates an automatic verification task; it does not trigger the removal of the autonomous graph of the connection relationship. Removal is only triggered by independent removal conditions, i.e., the verification result of the automatic verification task is a recurrence or rollback that has not been resolved after three consecutive verification cycles. A verification cycle is defined as the period from the generation of one verification result to the generation of the next verification result. This eliminates the semantic conflict between the downgraded judgment value falling into the verification interval and the "verification file not merged" rule in step S500.

[0080] The runtime data comes from within the operation and maintenance system. The extraction process involves de-identification and field-level anonymization, retaining only the device and port dimension attributes required for graph construction.

[0081] It should be noted that the "automation" referred to in this invention is manifested in the following ways at this stage: conflict detection, reduction, confidence tag downgrading, automatic verification task generation and value arbitration are all automatically executed by the system according to the rules, without the need for manual judgment of each item; only when the automatic verification task fails to converge after multiple iterations is the batch of triples confirmed at once on the manual assistance interface, which does not constitute manual graph construction of individual triples.

[0082] like Figure 1 and Figure 6 As shown, step S700 employs a weight feedback module to perform the weight feedback learning phase, iteratively updating the weights of the four types of signals based on the verification results of the automatic verification task. The update formula is as follows: ,in For the first The updated weight vector and learning rate in each round The value is typically between 0.10 and 0.20; in this embodiment, it is 0.15. For the first Round feedback gradient The normalized vector.

[0083] In this embodiment, each round consists of a batch of no less than 600 verified triples: first, the feedback gradient of each sample is determined according to the category of the verification result; then, the arithmetic mean of the feedback gradients of all samples in the batch is taken and normalized to obtain the result. And perform a weight update.

[0084] The rule for constructing the feedback gradient is as follows: Let the mean vector of the resolved class samples in round t be denoted as . The mean vector of the relapsed samples on the four signals is: ,but: Take samples that have been dissolved That is, to increase the weight of signals that have higher values ​​on resolved samples and lower values ​​on relapsed samples; Samples of relapse That is, to take complementary values ​​for each component and reduce the weight of the signal with higher values ​​in the relapse sample; Take rollback samples This causes the weights of the four signals to decay towards a uniform distribution in order to suppress the dominance of a single signal.

[0085] The technical meaning is: the positive discriminative power of each signal is measured by the difference in the values ​​of each signal between the resolved sample and the relapsed sample. The direction of weight iteration is determined by the difference in signal distribution between the two types of samples, rather than by the absolute mean level of the signal.

[0086] For rollback samples, the uniform vector, after normalization, needs to be clipped according to its value range: Trim the component to the maximum value of 0.22. The component was clipped to the upper limit of 0.18, and the w4 component was clipped to the lower limit of 0.32. After renormalization, The proportion decreased and The increase in the proportion actually converges in the direction of reducing the proportion of the dominant signal and relatively increasing the proportion of other signals within the allowable range, which is consistent with the design intention of suppressing the dominance of a single signal. Figure 1 The value range here acts as a limiter to ensure that a single rollback does not violate the prior ordering constraints of the weights. The pruning and normalization iteratively executes until all components fall within the value range or the upper limit of 10 iterations is reached: for example, (0.18, 0.13, 0.11, 0.58) obtained in a certain update is pruned to (0.20, 0.15, 0.12, 0.45) and then normalized to (0.2174, 0.1630, 0.1304, 0.4891). If the boundary is crossed again, the next round of trimming and normalization is required until convergence.

[0087] Regarding the combination of learning rate and convergence criterion: the change in the weight vector in a single round is approximately .Pick A value of 0.10 to 0.20 can make the single-round change significantly higher than the convergence criterion of 0.005, and the weights gradually converge after the substantial adjustment is completed within a finite number of rounds.

[0088] In this implementation environment, the weights converge to (0.24, 0.21, 0.16, 0.39) after 21 iterations starting from the initial value (0.25, 0.20, 0.15, 0.40). The changes in the weight vector in the last 5 iterations are 0.0044, 0.0039, 0.0033, 0.0028, and 0.0022, respectively, all of which are less than the convergence criterion of 0.005.

[0089] The verification results are automatically determined by the runtime data: if the conflict of the same entity attribute does not reappear within K consecutive collection cycles since the automatic verification task is generated and the map value is consistent with the runtime value, it is determined to be resolved, and K is 3; if the same conflict reappears within K collection cycles, it is determined to be recurred; if the system records a rollback event associated with the conflict attribute, it is determined to be rolled back.

[0090] The updated samples are verified triples with a size of no less than 600. The convergence criterion is that the change in the weight vector is less than 0.005 for 5 consecutive rounds. After each round of update, the weight vector is renormalized so that the sum of its components is 1. If any component exceeds the value range described in step S400 after normalization, the component is clipped to the corresponding interval boundary and normalized again. The process is iterated until convergence or the upper limit of the number of iterations is reached.

[0091] like Figure 1As shown, steps S810-S830 employ the graph evolution management module to perform graph evolution management. When the maintenance document version is upgraded, instance migration mapping is performed on the shadow graph. Specifically, device instances are matched by normalized names, and port instances are matched by the composite key of device normalized name and port normalized name. When multiple candidate instances are matched for the same normalized name, they are sorted from high to low by normalized name similarity, and if the similarity is the same, they are sorted from near to far by page coordinate distance, and one is selected for matching. Successfully matched instances retain their attribute values ​​and connection edges during migration, while unmatched instances are incorporated as new instances in the newly added graph and are not considered as conservation violations. During migration, the number and attribute values ​​of existing device instances and port instances with the same name remain unchanged.

[0092] Then, three checks are performed: Instance conservation check: The number of existing device instances and port instances with the same name does not increase or decrease before and after the migration; Endpoint connectivity verification: Each port instance has at least one connection edge and does not exceed the maximum number of connections. The maximum number of connections is preset according to the port type. For ordinary ports, it is 1, and for trunk aggregation ports, it is the upper limit of the number of members configured in its aggregation group. Required attribute non-empty rate validation: The proportion of non-empty required attributes is not lower than that of the previous version.

[0093] If all three checks pass, the mode layer is switched atomically. If any one check fails, the original mode layer is retained and an evolution check report is generated. The migration is then re-executed after correction in subsequent versions to ensure the stability of the graph during document version iterations.

[0094] like Figure 1 , Figure 2 , Figure 3 and Figure 5 As shown, corresponding to the above method embodiments, the present invention also provides a multimodal operation and maintenance document-driven knowledge graph construction device, including a layout analysis module, a vector parsing module, an anchoring group construction module, an identity determination module, an evidence chain preservation and processing module, a runtime verification module, and a weight feedback module. Each module is used to execute the corresponding steps in the above method embodiments. The runtime verification module does not output a value for calculating the identity determination value to the identity determination module, and its output bypass signal only has falsification effect, and is only used to lower or maintain the identity determination value and the verification determination value, and must not be used to raise it. It neither writes back to the identity determination module nor is it used for the fourth-level processing of the evidence chain preservation module; the vector parsing module includes a geometric reading unit, a tag registration unit, an endpoint processing unit, and a triplet assembly unit, and may further include a grid catch-all unit, which are used to execute steps S210, S220, S250, S260, and the catch-all branch when no vector description is carried; the anchor group construction module also includes an anchor corpus missing processing unit, which is used to execute step S350; the running state verification module includes a quintuple collection unit, a conflict detection unit, a reduction unit, and an arbitration unit, which are used to execute the reduction part of steps S610, S620, S630, and S640, respectively; the device includes a graph evolution management module, which is used to execute steps S810 to S830.

[0095] The present invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method embodiments.

[0096] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described method embodiments.

[0097] Experimental results To verify the effectiveness of this invention, ablation experiments were performed on a self-built operation and maintenance documentation test suite. The test suite contains 217 device topology diagrams and data center cabling diagrams for 1260 devices, along with accompanying device lists, operation and maintenance manuals, fault handling cases, and shift handover records. The annotation method used the asset and connection records in the current network configuration management database as a benchmark, with two annotators cross-validating to form a gold standard. The gold standard connection triplets totaled 3842. Evaluation metrics included: precision (P), the ratio of correctly resolved connection triplets to the total number of resolved connection triplets; recall (R), the ratio of correctly resolved connection triplets to the total number of gold standard connection triplets; and the F1 score, the harmonic mean of precision and recall. The number of triplets requiring external intervention and the non-null rate of required attributes were also statistically analyzed.

[0098] Ablation test results

[0099] As shown in the ablation experiment results table: After removing one-way falsification in variant A, the judgment value will rise when the running data is consistent with the document extraction conclusion, causing some contradictory relationships that should have been suspended to be directly incorporated. The recall rate decreased by 4.7 percentage points, and the number of triples requiring external intervention increased by about 1.3 times, indicating that one-way falsification makes a substantial contribution to suppressing false incorporation. After removing the page anchoring group in variant B, the contradiction between the image and text cannot be found at the source, and the recall rate decreased by 6.6 percentage points. After removing the geometric inverse solution in variant C, the port connection relationship degenerates into the probability inference of the recognition result, and the recall rate decreased by 8.1 percentage points, which is the single item with the greatest impact, which is consistent with the technical main line of this invention, which uses the vector coordinate system as the geometric reference. After removing the weight feedback in variant D, the judgment rule cannot adapt to the verification result, and the recall rate decreased by 2.2 percentage points, which is consistent with the design that the weight can be substantially adjusted within a limited number of rounds when the learning rate is 0.10 to 0.20. The above results indicate that the contributions of each technical means are not independent and additive. Removing any one of them leads to a simultaneous deterioration in recognition performance and intervention cost. Moreover, the deterioration caused by removing geometric inverse solution in the vector analysis stage is the most significant, which confirms the synergistic relationship between geometric inverse solution, one-way falsification, and layout anchoring.

[0100] Regarding the prior statistics and test set division: The location prior and connectivity prior described in step S400 use historical map instances that have been incorporated into the main map atlas as statistical samples. These statistical samples strictly do not overlap with the test set of this section. Specifically, the division is based on document version and archiving time. The test set of this section uses 217 maps from the latest version, while the prior statistical samples use historical map instances from previous versions. Simultaneously, 5-fold cross-validation is used, and the prior statistics for each fold do not include the test maps for that fold. This eliminates data leakage introduced by the prior statistics and ensures that the above indicators reflect the generalization performance of the method.

[0101] It should be noted that the values ​​listed in this section are exemplary results under one implementation environment of the present invention. The specific values ​​vary with the composition of the test set, the labeling caliber, and the completeness of the runtime data, and do not constitute a limitation on the scope of protection of the present invention.

[0102] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A multimodal operation and maintenance document-driven knowledge graph construction method, characterized in that: The method includes steps S100 to S700: S100: Obtain the multimodal operation and maintenance document set to be processed. The multimodal operation and maintenance document set includes equipment topology diagrams, data center cabling diagrams, equipment list, operation and maintenance manuals, fault handling cases, and shift handover records. Perform layout analysis in a unified coordinate system on the multimodal operation and maintenance document set, extract the figure caption area, text area, table area and table title area, and retain the relative position and size of each area in the original coordinate system, and output the document layer with layout annotations; S200: Perform a three-segment vector analysis of each drawing, including the equipment topology drawing and the computer room wiring diagram. Using the vector coordinate system carried by the drawing as the geometric reference, separate geometric elements and text elements from the vector description of the drawing. Register the text elements as device name tags and port name tags, and extract the geometric center and circumscribed dimensions of the geometric elements. Label the device type according to the shape characteristics of the geometric elements and the label text of the text elements. The endpoints of the connecting lines are snapped to the nearest primitive boundary point and polyline vertex merging and orthogonalization are performed. The port is defined by the intersection of the orthogonal polyline and the primitive boundary. In the vector coordinate system, the connection triplet and attribute triplet are obtained by the analytical intersection geometry inverse solution of the geometric primitive boundary and the orthogonal polyline. S300: Construct a page anchoring group based on the aforementioned image: Use the image caption text and the citation sentences in the text that cite the image as anchoring corpus, together with the page area coordinates, to form a page anchoring group. Perform normalization matching between the device name and port name in the image element and the normalized name in the anchoring corpus to form an image element-anchor text identity evidence pair, which serves as the unique evidence base for each triple in the knowledge graph. S400: For each device connection, the four signals—normalized name consistency signal, semantic similarity signal, layout context matching signal, and connection line topology neighborhood matching signal—are weighted and fused into an identity determination value. ; S500: Perform four-level processing on the connection triplet and the attribute triplet according to the identity determination value: If the value is greater than or equal to the first threshold, it will be incorporated into the main map. When the threshold is greater than or equal to the second threshold and less than the first threshold, automatic evidence enhancement review is triggered. If the value is greater than or equal to the third threshold and less than the second threshold, the verification will be suspended. If the value is less than the third threshold, it is rejected from being included; all four processing steps retain all candidate values ​​and corresponding evidence chains; the evidence enhancement automatic review outputs a review judgment value, which is obtained by multiplying the weighted value of the four recalculated signals by a reduction coefficient not greater than 1; if the review judgment value is greater than or equal to the first threshold, it is included in the main spectrum, otherwise it is suspended for verification. S600: Unify the operation and maintenance status data into a five-tuple of entity, attribute, value, time, and source, and perform equivalence comparison and conflict detection with the attribute triples in the knowledge graph. For conflicting items, adjust the identity judgment value of the connection relationship of the device to which the conflicting item belongs by the following formula: ,in The identity determination value obtained by weighted fusion in step S400 is... The conflict reduction factor and Repeat the downward adjustment until... If it falls within the verification range of [0.30, 0.50), then... As the valid identity determination value of the connection relationship, the confidence flag of the connection relationship in the main graph is changed from verified to pending verification and an automatic verification task is generated. The adoption value of the conflicting attribute is automatically selected according to the preset value source credibility sorting and the time newer priority rule. S700: The weights of the four signals are iteratively updated using the verification results of the automatic verification task as feedback signals. After each round of updates, the weights are renormalized so that their sum is 1. The operation and maintenance data is not included in the identity determination value. The calculation is only used as input for conflict detection, value arbitration and weight feedback; and the operation and maintenance status data only has the effect of falsification, that is, it can only trigger the reduction or maintenance of the identity judgment value or the review judgment value, and cannot be used to increase the identity judgment value or the review judgment value. Among them, the It is only used for confidence flag determination and automatic verification task triggering, and is neither written back to the weighted fusion calculation in step S400 nor used for the four-level processing in step S500; the downgrade only changes the confidence flag and does not trigger the removal of the connection relationship from the main map. The removal is only triggered by independent removal conditions. The independent removal conditions are that the verification result of the automatic verification task is a recurrence or rollback and has not been resolved after 3 consecutive verification cycles. The verification cycle is counted from the generation of one verification result to the generation of the next verification result.

2. The multimodal operation and maintenance document-driven knowledge graph construction method according to claim 1, characterized in that: It also includes map evolution management steps S810 to S830: S810: When a new schema layer is generated during an upgrade of the operation and maintenance documentation version, instance migration mapping is performed on the shadow graph: device instances are matched by normalized name, and port instances are matched by the composite key of device normalized name and port normalized name; when multiple candidate instances are matched for the same normalized name, they are sorted from high to low by normalized name similarity, and if the similarity is the same, they are sorted from near to far by page coordinate distance, and one is selected for matching; the attribute values ​​and connection edges of successfully matched instances remain unchanged during migration, and unmatched instances are incorporated as new instances in the new graph, which is not considered a violation of conservation; the number and attribute values ​​of existing device instances and port instances with the same name remain unchanged during migration. S820: Perform three checks on the migration results: instance conservation check, i.e., the number of existing device instances and port instances with the same name does not increase or decrease before and after the migration; endpoint connectivity check, i.e., each port instance has at least one connection edge and does not exceed the maximum number of connections, the maximum number of connections is preset according to the port type, with 1 for ordinary ports and the upper limit of the number of members configured in the aggregation group for trunk aggregation ports; and required attribute non-empty rate check, i.e., the proportion of non-empty required attributes is not lower than that of the previous version. S830: If all three checks pass, the mode layer is switched atomically; if any one check fails, the original mode layer is retained and an evolution check report is generated. The migration is then re-executed after correction in a subsequent version.

3. The multimodal operation and maintenance document-driven knowledge graph construction method according to claim 1, characterized in that: Step S200 specifically includes: S210: Read the vector description carried by the graphic, which includes primitive geometric objects, text objects and polyline objects, and the coordinates of the three are all represented by the vector coordinate system; directly read the geometric boundary of the primitive geometric object and the path vertex sequence of the polyline object from the vector description to obtain the primitive boundary and polyline path without rasterization and image segmentation; S215: For the map that does not carry a text object or is missing a text object in the vector description, the map is rasterized and rendered at a preset resolution. A text detection network is used to perform text region detection on the rendered map, and the text box coordinates and text content are output. The text box coordinates are converted back to the vector coordinate system according to the ratio between the preset resolution and the vector coordinate system, and then filled back as text primitives. S220: Separate geometric primitives and text primitives according to the vector coordinate system: Register text primitives located inside the geometric object of the primitive and matching the device naming convention regular template as the device name label of the primitive; register text primitives located within a preset proximity range of the boundary of the geometric object of the primitive and matching the port number regular template set as the port name label of the primitive; the device name label and the port name label constitute the device name and port name of the primitive; the geometric primitive is retained as the primitive outline; S230: Calculate the geometric center coordinates and the size of the bounding rectangle of the primitive geometric object from the geometric boundary, and calculate the aspect ratio and the number of skeleton branches of the primitive from the geometric boundary as shape features; S240: The device type is determined by combining the shape feature and the device name label. The candidate device type set includes server, switch, router, firewall, load balancer, storage device, and terminal. The device type and confidence level of each graphic element are output. S250: The nearest point algorithm is used to snap the endpoints of the connection line to the nearest primitive boundary point. Then, the vertex merging and orthogonalization of the polyline are performed to obtain the endpoint port identifier and connection type. The connection type includes direct connection, uplink, downlink, and relay link. When there are multiple candidate endpoints for the same connection line, the geometric distance and port number lexicographical order are used to disambiguate and select a unique endpoint. S260: Define the intersection of the orthogonal polyline and the boundary of the primitive as a port, and assemble it into a connection triple (source device: source port, connection type, destination device: destination port) and an attribute triple (device name, attribute, value). The attribute values ​​cover four types of required attributes: network segment, VLAN, IP address, and service affiliation. The connection triplet is obtained by the inverse geometric solution of the analytical intersection of the primitive boundary and the orthogonal polygonal line in the vector coordinate system, and is not determined by probabilistic inference of text recognition results or visual detection results, thereby ensuring the geometric determinism of port connection relationships; The upward and downward directions are determined by the vertical relative positions of the graphic elements on the graphic layout: the graphic element with the smaller ordinate of its geometric center is located at the top of the layout, and the graphic element with the larger ordinate is located at the bottom of the layout. The connecting line is determined to be upward when it points from a graphic element located at the bottom of the layout to a graphic element located at the top of the layout, and downward when it points from a graphic element located at the top of the layout to a graphic element located at the bottom of the layout. The direct connection is determined by the orthogonal broken line not passing through other graphic elements, and the relay link is determined by the orthogonal broken line passing through multiple graphic elements carrying the same aggregation group identifier. In step S250, the endpoint adsorption radius is taken as 2% to 5% of the short side of the outer rectangle of the graphic element, and the orthogonalization tolerance is taken as 10° to 20°. When the map does not carry a vector description and is provided only as a raster image, as an alternative branch to step S210 in step S200 where the vector description carried by the map is read, the raster image is first calibrated to the vector coordinate system, and steps S211 to S213 are executed: S211. Use a text detection network to detect the coordinates of the text box, and perform masking and zeroing on the area covered by the coordinates of the text box to obtain a residual image; S212. First, determine the value range of the segmentation threshold for the residual image according to the gray-level histogram. The value range is from 0.5 times the gray-level mean to 0.95 times the gray-level mean. Then, within this value range, use the maximum inter-class variance method to determine the segmentation threshold. S213. Perform morphological thinning and Zhang-Suen thinning on the segmentation results to obtain the primitive central skeleton, and extract the primitive geometric center coordinates and the outer rectangle size from the primitive central skeleton, and convert them back to the vector coordinate system. Specifically, steps S211 to S213 and step S210 are executed selectively; that is, step S210 is not executed when steps S211 to S213 are executed, and step S220 and its subsequent steps continue to apply after steps S211 to S213 are completed. At this time, the endpoints of the connecting lines are determined by the endpoints of the central skeleton of the primitive, and the connecting triplets are still obtained by geometric inverse solving of the intersection of the primitive boundary and the orthogonal broken line within the vector coordinate system.

4. The multimodal operation and maintenance document-driven knowledge graph construction method according to claim 1, characterized in that: Step S300 specifically includes: S310: Based on the document layer with layout annotations, locate the caption area and text reference area corresponding to a single image; S320: Establish a one-to-one correspondence between map identifiers and map caption text, and extract complete sentences in the main text paragraphs that cite the map identifiers as citation sentences; S330: Perform name normalization processing, mapping port numbers and English abbreviations to normalized names in a unified format. The name normalization processing includes unifying full-width and half-width characters, expanding English abbreviations, and normalizing port number formats. S340: A page anchor group is composed of a quadruple of map primitive, normalized name, anchor text and page coordinates, which serves as the unique evidence basis for each triple in the knowledge graph. S350: For primitives for which no anchoring corpus was found in either the caption area or the text area, their normalized name consistency signal and semantic similarity signal are set to zero. The connection triples and attribute triples corresponding to the primitives are suspended for verification and not incorporated into the main graph. The confidence flag of the suspended triples not incorporated into the main graph is set to "to be verified" and this confidence flag is maintained independently of the main graph. In step S330, the expansion of English abbreviations is performed based on the device type-abbreviation mapping table constructed according to the device naming specification document in the operation and maintenance field, and the port number format normalization is performed based on the port number regular expression template set; for names not covered by the mapping table and the regular expression template set, the original string is retained and marked as unnormalized, and the normalized name consistency signal of the unnormalized name is zero in step S400.

5. The multimodal operation and maintenance document-driven knowledge graph construction method according to claim 1, characterized in that: In step S400, the identity determination value is calculated using the following formula: ;in The normalized name consistency signal is calculated from the matching degree of the device name of the graphic element and the normalized name in the anchor text using the Jaccard string. The semantic similarity signal is calculated from the cosine similarity of the sentence vectors of the anchored text and the primitive-annotated text. The layout context matching degree signal is calculated from the matching degree between the relative position of the graphic element and the prior position of the device of this type; The connection topology neighborhood matching degree signal is calculated from the matching degree between the connection topology neighborhood and the connection prior of this type of device; each signal value satisfies 0 ≤ ≤1; The location prior and the connection prior are constructed separately according to the device type. The construction method is as follows: using historical map instances that have been incorporated into the main map as statistical samples, the relative position distribution of the device elements in the map layout and the connection degree distribution of the device ports are statistically analyzed respectively, and the normalized probability density of the distribution is used as the matching degree value; when the historical samples are insufficient, the rules exported from the equipment networking specification document in the operation and maintenance field are used to determine the matching degree. The weight values ​​range are , , , And after selecting from the range of values, normalization is performed to make After each round of update in step S700, if any component of the updated weight exceeds the range of values, the component is clipped to the corresponding interval boundary and then re-normalized so that the sum is 1. The clipping and normalization are repeated until all four components fall into the range of values, or until the number of iterations reaches the preset upper limit of the number of iterations, where the preset upper limit of the number of iterations is 10. In step S700, the weight update is performed according to the following formula: ,in Let be the weight vector updated in round t, and let the learning rate α be between 0.10 and 0.20, preferably 0.

15. Let t be the feedback gradient. Normalized vector; The rule for constructing the feedback gradient is as follows: Let the mean vector of the resolved class samples in round t be denoted as . The mean vector of the relapsed samples on the four signals is: Then take the already resolved samples. For relapsed samples That is, complementary components are selected one by one; for rollback samples, complementary components are selected... This causes the weights of the four signals to decay towards a uniform distribution in order to suppress the dominance of a single signal. The verification results are automatically determined by the runtime data without external intervention: if the conflict of the same entity attribute does not reappear within K consecutive collection cycles from the generation of the automatic verification task and the graph value is consistent with the runtime value, it is determined to be resolved, and K is 3; if the same conflict reappears within K collection cycles, it is determined to be recurred; if the change system records a rollback event associated with the conflict attribute, it is determined to be rolled back. The updated samples are verified triples with a size of no less than 600. The convergence criterion is that the change in the weight vector is less than 0.005 for 5 consecutive rounds. After each round of update, the weight vector is renormalized so that the sum of its components is 1.

6. The multimodal operation and maintenance document-driven knowledge graph construction method according to claim 5, characterized in that: In step S500, the first threshold is 0.75, the second threshold is 0.50, and the third threshold is 0.

30. The evidence enhancement automatic review includes: S510, anchoring corpus expansion: retrieving the occurrence positions of the same canonical name in other diagrams, table areas, device lists, and the main text of the operation and maintenance manual within the multimodal operation and maintenance document set, constructing an expanded anchoring corpus, and recalculating the canonical name consistency signal and semantic similarity signal accordingly; S520, runtime bypass verification: reading the runtime quintuple of the ports involved in the connection relationship, determining the non-existence of the connection relationship based on the port status, link traffic, and alarm events, and outputting a bypass reduction coefficient. The bypass reduction coefficient is 0.5 when the detected port is closed, the link traffic is continuously zero, or there is a link interruption alarm and the runtime indicates that the connection relationship does not exist. The preset value is 0.8, and 1 is taken in other cases; S530, Topological Neighborhood Backtracking: Using the set of adjacent edges that the primitive has been incorporated into the main graph as constraints, check whether the connection relationship to be verified constitutes port duplication or link non-pair conflict with the topology already incorporated into the graph, and output the backtracking pass coefficient. The backtracking pass coefficient is a preset value of 0.5 to 0.8 when a conflict is detected, and 1 when no conflict is detected; S540, Multiply the four-signal weighted value after the recalculation of the extended anchoring corpus by the bypass reduction coefficient and the backtracking pass coefficient to obtain the verification judgment value; The bypass reduction coefficient and the backtracking pass coefficient are both not greater than 1, so that the running bypass verification and the topological neighborhood backtracking only have the effect of falsification; When the verification judgment value is greater than or equal to the first threshold, it is incorporated into the main graph, otherwise it is suspended as pending verification and all candidate values ​​and evidence chains are retained.

7. The multimodal operation and maintenance document-driven knowledge graph construction method according to claim 1, characterized in that: Step S600 specifically includes: S610: Extract asset attributes from the configuration management database, extract time-series indicators from the monitoring platform, extract alarm events from the alarm platform, and extract change records from the change system, and uniformly format them into a five-tuple of entity, attribute, value, time, and source; S620: Perform an equivalence comparison and conflict detection between the quintuple and the attribute triple in the knowledge graph; S630: Comparison results are divided into three levels: Consistent items are not updated; conflicting items are adjusted down for the connection relationship using the following formula: ,in The identity determination value obtained by weighted fusion in step S400 is The conflict reduction factor and Repeat the downward adjustment until... If it falls within the verification range of [0.30, 0.50), then... As the valid identity determination value of this connection relationship, the It is only used for confidence mark determination and automatic verification task triggering, and is neither written back to the weighted fusion calculation in step S400 nor used for the four-level processing in step S500; at the same time, the confidence mark of the connection relationship in the main graph is changed from verified to pending verification, and an automatic verification task is generated; the pending verification item remains unchanged. S640: For conflicting attributes, the change records, configuration management database, monitoring platform time series, and document extraction results are extracted in descending order of the credibility of the value source. Within the same source, the newest time is given priority. The adopted value is automatically selected and written into the graph, while the overwritten value along with its evidence chain is retained. Wherein, the conflict reduction coefficient The value is set to 0.6 to 0.8, preferably 0.

7. The reduction does not change the existence of the connection in the main graph. The removal of the connection from the main graph is triggered only by an independent removal condition, which is: the verification result of the automatic verification task is a recurrence or rollback, and it has not been resolved after 3 consecutive verification cycles.

8. A multimodal operation and maintenance document-driven knowledge graph construction device, characterized in that, include: The system includes a layout analysis module, a vector parsing module, an anchoring group construction module, an identity determination module, an evidence chain preservation and processing module, a runtime verification module, and a weight feedback module, as well as a graph evolution management module. The layout analysis module is used to acquire the multimodal operation and maintenance document set to be processed and perform layout analysis under a unified coordinate system, outputting a document layer with layout annotations. The vector parsing module is used to perform a three-segment vector parsing of the drawing elements, labels, and connections, and to obtain connection triples and attribute triples through geometric inverse solution in the vector coordinate system. The vector parsing module includes: a label registration unit, used to register text primitives located inside the primitive geometric object and matching the device naming convention regular template as device name labels, and to register text primitives located within a preset proximity range of the primitive geometric object boundary and matching the port number regular template set as port name labels; an endpoint processing unit, used to snap the endpoints of the connecting line to the nearest primitive boundary point and perform polyline vertex merging and orthogonalization processing, determine uplink and downlink based on the relative size of the ordinate of the primitive geometric center, and determine direct connection and relay link based on whether the orthogonal polyline passes through other primitives; and a raster catch-all unit, used to calibrate the raster image to the vector coordinate system when the primitive does not carry a vector description, obtain the primitive central skeleton through zeroing the text region mask, segmentation using the maximum inter-class variance method within the grayscale histogram value range, and Zhang-Suen thinning processing, and convert it back to the vector coordinate system; The anchor group construction module is used to construct image-annotation-citation layout anchor groups on an image-by-image basis, forming an image element-anchor text identity evidence pair as the unique evidence basis for the triple; the anchor group construction module also includes an anchor corpus missing processing unit, used to set the normalized name consistency signal and semantic similarity signal of images elements for which no anchor corpus was found in either the caption area or the main text area to zero, and to suspend the connection triple and attribute triple corresponding to the image element as pending verification and not to be incorporated into the main graph; the confidence flag of the suspended triple is set to pending verification and is maintained independently of the main graph. The identity determination module is used to weight and fuse the normalized name consistency signal, semantic similarity signal, layout context matching degree signal and connector topology neighborhood matching degree signal into an identity determination value. The evidence chain retention processing module is used to perform four levels of processing based on the identity determination value: incorporation, automatic review of evidence enhancement, suspension pending verification, and rejection, while retaining all candidate values ​​and evidence chains. The runtime verification module is used to unify the operation and maintenance runtime data into five-tuples and compare them with the attribute triples, and to handle conflicting items according to... The effective identity determination value was obtained by lowering the value. The system sets the confidence flag to be verified, generates an automatic verification task, and automatically selects the adoption value according to the source credibility and the rule of prioritizing the newest time. The runtime verification module includes a reduction unit, which is used to... Repeat the downgrade until It falls within the verification interval of [0.30, 0.50), and the stated It is neither written back to the identity determination module nor used for the four-level processing of the evidence chain preservation module; The weight feedback module is used to iteratively update the weights of the four signals and renormalize them using the verification results of the automatic verification task as feedback signals. The graph evolution management module is used to match device instances on the shadow graph by normalized name when a new mode layer is generated by the upgrade of the operation and maintenance document version, match port instances by the composite key of device normalized name and port normalized name and perform instance migration mapping, and perform instance conservation verification, endpoint connectivity verification and required attribute non-empty rate verification on the migration results. If all three verifications pass, the mode layer is switched atomically. If any one of them fails, the original mode layer is retained and an evolution verification report is generated. The running verification module does not output values ​​for calculating the identity determination value to the identity determination module, and its output bypass signal only has the effect of falsification and is only used to lower or maintain the identity determination value and the review determination value, and shall not be used to raise it; the device will only remove the connection relationship from the main map when the verification result of the automatic verification task is a recurrence or rollback and has not been resolved after 3 consecutive verification cycles.

9. An electronic device, characterized in that, It includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method of any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images