Energy archive classified retrieval method and system fusing risk early warning atlas

Abstract

The application discloses a kind of energy archives classified retrieval method and system of fusion risk early warning atlas, it is related to digital archives management technical field.Uniform node relationship table is constructed by reading energy archives and historical risk event, the atlas node similarity of archive anchor point and risk event anchor point is calculated;Based on archive anchor point, text evidence fragment and image evidence area are executed bidirectional alignment, generate dynamic classification representation and form classification candidate set jointly static original label;According to classification candidate set, keyword, full-text and vector recall are executed, risk perception score is carried out in combination with risk event anchor point set, and retrieval and association result is output based on classification conflict correction and hierarchical ordering, to improve risk-related archives recall accuracy, ordering stability and retrieval efficiency.

Description

Technical Field

[0001] This invention belongs to the field of digital archives management technology, specifically relating to an energy archives classification and retrieval method and system that integrates risk warning maps. Background Technology

[0002] As the digital transformation of the energy industry continues, a large amount of energy records have accumulated across various business processes, including production operations, equipment maintenance, safety supervision, project construction, dispatch instructions, defect rectification, accident handling, and compliance auditing. These records typically contain heterogeneous data such as structured fields, text content, and image attachments. For energy record retrieval, practical engineering requirements go beyond simply locating the record itself; they also involve simultaneously determining the record's business category, the relationship between the record and risk events, and whether a traceable early warning chain exists during the retrieval phase.

[0003] In existing technologies, traditional keyword matching, full-text retrieval, and manual tag indexing mainly rely on inverted term indexing, static field filtering, and local relevance ranking, which are insufficient to express cross-document relationships between equipment, components, defects, operating conditions, and handling actions. When the same risk fact is expressed using aliases, abbreviations, professional abbreviations, or different languages ​​in different documents, relying solely on literal terms or static tags for retrieval can easily lead to missed or poorly ranked risk-related documents. This, in turn, results in untimely location of key evidence during safety inspections and incident debriefings, increases the cost of manual review, and affects the reliability of retrieval decisions.

[0004] For example, Chinese patent CN202510938561.4 discloses an intelligent risk review and early warning method and system that integrates knowledge graphs. Its focus is on identifying abnormal path structures from historical risk events and triggering linked early warnings. The knowledge graphs used in this type of method primarily serve event review and early warning analysis, without directly entering the main link of candidate recall and ranking for energy archives. Therefore, when the overlap between query terms and the text content is low but highly relevant in the risk propagation path, it is difficult to reliably recall the corresponding archives in advance. This can lead to the risk of missing key evidence in accident tracing scenarios.

[0005] For example, Chinese patent CN202510578193.7 discloses an intelligent classification and retrieval method and system for archives, which achieves archive classification and semantic ranking through multi-source feature extraction, adaptive classification, and semantic association graphs. Chinese patent CN202510993213.7 discloses an intelligent database storage system and method for archive data based on dynamic sharding and adaptive indexing, which focuses on optimizing storage sharding and index access efficiency. The aforementioned solutions improve local processes from the semantic or storage layers, but they typically assume that classification results, graph relationships, and retrieval indexes can be decoupled, and a technical link has not yet been formed that allows risk warning graphs to directly participate in the joint solution of classification and retrieval. Especially in the energy archive scenario, when there is a lack of bidirectional constraints around the same risk anchor point between equipment tag numbers, component names, and defect descriptions in text and abnormal areas, appearance labels, and instrument readings in images, archives that appear similar but have different business facts are easily aggregated incorrectly, leading to mis-retrieval, mis-ranking, and an increased burden of subsequent manual verification.

[0006] Furthermore, existing manual classification and static labeling systems typically treat archive categories as fixed attributes at the time of archiving, making it difficult to respond to the actual situation where the semantics of energy archive risks shift with the operation, maintenance, and accident review phases. This leads to engineering problems where the management classification is correct, but the retrieval and sorting do not conform to the current risk perspective, making it difficult for the platform to simultaneously address both archive management and risk tracing business objectives.

[0007] Therefore, how to move the risk warning map forward into the main link of energy archive retrieval, how to establish a two-way alignment mechanism between textual evidence and image evidence around the facts of equipment defects, and how to generate dynamic classification results that can respond to risk semantic shifts while retaining static original labels have become technical problems that urgently need to be solved in this field. Summary of the Invention

[0008] The purpose of this invention is to provide an energy archive classification and retrieval method and system that integrates a risk warning map, thereby solving the problems in existing technologies such as the separation of the risk map from the main retrieval link, the lack of bidirectional constraints on risk anchor points in textual and image evidence, and the inability of static classification labels to respond to risk semantic shifts, leading to missed recall of key archives, inaccurate ranking, and increased manual review costs. To achieve the above objective, this invention provides an energy archive classification and retrieval method that integrates a risk warning map, comprising: Step S1: Read energy archives and historical risk events, construct a unified node relationship table corresponding to the risk warning map, and calculate the map node similarity between archive anchor points and risk event anchor points based on the unified node relationship table; Step S2: Perform bidirectional alignment on text evidence fragments and image evidence regions in the energy archives based on the archive anchor points, and generate dynamic classification representations based on the similarity of the map nodes and the bidirectional alignment results, and form a classification candidate set by combining the static original labels; Step S3: Receive the query statement, perform keyword recall, full-text recall and vector recall based on the classification candidate set to obtain an initial candidate file set, and calculate the risk perception candidate score of the initial candidate file set by combining the risk event anchor point set corresponding to the query and the similarity of the graph nodes. Step S4: Based on the conflict relationship between the static original labels and the dynamic classification representation, perform classification conflict correction and hierarchical sorting on the initial candidate file set to obtain the classification retrieval results; Step S5: Write the classification search results into the result cache area and generate a result page. When the sorting results meet the early warning joint investigation conditions, expand the joint investigation results of related risk events and related files.

[0009] Preferably, constructing a unified node relationship table corresponding to the risk warning map includes: reading the structured fields, text, and image attachments of energy archives; extracting equipment codes, component names, defect types, operating condition terms, handling actions, and time markers from the text; and extracting component area labels, abnormal appearance labels, and instrument value labels from the image attachments to obtain archive anchor field; reading event numbers, path levels, responsible units, and event time information from historical risk events to obtain risk event anchor field; generating archive anchor primary keys based on the archive anchor field and risk event anchor primary keys based on the risk event anchor field, and writing the archive anchor primary key and risk event anchor primary key into the unified node relationship table; obtaining a pre-stored equipment alias mapping table; when the text lacks standard equipment codes but has equipment aliases, restoring the standard equipment codes according to the equipment alias mapping table, and organizing the node primary key field, node type field, standard equipment code field, component name field, defect type field, time slice field, path level field, edge type field, and edge weight field into the unified node relationship table.

[0010] Preferably, calculating the graph node similarity between the archive anchor point and the risk event anchor point based on the unified node relationship table includes: establishing node connections based on the composition relationship from equipment to component, the state relationship from component to defect, the handling relationship from defect to handling action, and the evidence relationship from the archive anchor point to the risk event anchor point, to obtain type edges and evidence edges in the risk warning graph; for the evidence edges, calculating the edge weights based on entity consistency, temporal proximity, and defect fact consistency, and filtering high-weight evidence edges according to preset retention rules; generating entity consistency scores based on the consistency of equipment codes, component names, and defect types, generating temporal proximity scores based on the temporal proximity of archive time slices and event time slices, and generating path adjacency scores based on path hierarchy proximity and graph hop count; fusing the entity consistency scores, temporal proximity scores, and path adjacency scores according to preset graph weights to obtain the graph node similarity, and writing the graph node similarity into the graph index.

[0011] Preferably, performing bidirectional alignment of text evidence fragments and image evidence regions in an energy archive based on archive anchors includes: segmenting the text into text evidence fragments, and labeling each text evidence fragment with equipment code, component name, defect type, operating condition term, and handling action to obtain a set of text evidence fragments; performing region detection or region proposal based on image attachments to extract image evidence regions containing component regions, anomaly labels, and confidence levels to obtain a set of image evidence regions; filtering fragments and regions related to the same equipment defect fact from the set of text evidence fragments and the set of image evidence regions according to the archive anchor primary key; obtaining a pre-stored business mapping table, and calculating a bidirectional alignment score between text evidence fragments and image evidence regions based on the anchor field matching degree, business context consistency degree, and the joint reliability of image region confidence and text completeness, wherein the text completeness is determined by the completeness of the identification of equipment code, component name, defect type, and handling action; when the energy archive does not contain image attachments, using the text evidence fragments as unilateral evidence in subsequent dynamic classification representation generation; and performing a write-back update on the evidence edge weights based on the bidirectional alignment results.

[0012] Preferably, generating a dynamic classification representation based on the graph node similarity and bidirectional alignment results includes: generating a text main class score based on the text evidence fragment, generating an equipment main class score based on the equipment professional category rules, and generating a defect main class score based on the defect classification rules; generating a risk association main class score by fusing the graph node similarity and the bidirectional alignment score according to the set of risk event anchor points related to the current archive's professional category; constructing a dynamic classification representation for vector retrieval and main class determination based on the text main class score, equipment main class score, defect main class score, and risk association main class score; jointly retaining the dynamic main class in the dynamic classification representation with the static original label, and recording a conflict marker when the two are inconsistent, to obtain the classification candidate set; the dynamic classification representation includes text evidence features, image evidence features, graph adjacency features, and classification rule statistical features.

[0013] Preferably, the process of performing keyword retrieval, full-text retrieval, and vector retrieval based on the dynamic classification representation, according to the classification candidate set, includes: parsing the query statement to obtain a keyword set, a device code set, a time range, file category limiting words, and risk limiting words, and generating a retrieval constraint table; obtaining a pre-stored category mapping table, determining the query limiting category according to the category mapping table, and locating the corresponding risk event anchor point set when the risk limiting words exist; combining the time range and file category limiting words, performing keyword matching retrieval based on the retrieval constraint table to obtain a keyword candidate set; performing full-text index retrieval based on the retrieval constraint table to obtain a full-text candidate set; performing vector retrieval retrieval based on the dynamic classification representation to obtain a vector candidate set; and performing a union of the keyword candidate set, the full-text candidate set, and the vector candidate set to obtain the initial candidate file set.

[0014] Preferably, calculating the risk perception candidate score of the initial candidate file set includes: calculating the keyword matching score, full-text search relevance score, and vector search score for each candidate file in the initial candidate file set; generating a risk graph perception score based on the similarity of graph nodes between the file anchor points of the candidate file and the risk event anchor point set corresponding to the query, combined with the risk association main class score; fusing the keyword matching score, full-text search relevance score, vector search score, and risk graph perception score according to a preset fusion weight to obtain the risk perception candidate score; and using the risk association main class score as the risk graph perception score when no risk limiting words are parsed from the query.

[0015] Preferably, performing classification conflict correction and hierarchical sorting on the initial candidate file set includes: checking the conflict markers between the static original tags and dynamic main categories of each candidate file, and determining the classification conflict relationship in conjunction with the query-defined category; when the dynamic main category is consistent with the query-defined category but the static original tags are inconsistent, the corresponding candidate file is sorted upwards; when the static original tags are consistent with the query-defined category but the dynamic main category is inconsistent, the corresponding candidate file is sorted downwards, obtaining a classification conflict correction coefficient; based on the evidence edge weights between the candidate file anchor points and the query-related risk event anchor points, the maximum edge weight is extracted as the warning intensity; a final ranking score is generated according to the risk perception candidate score, the classification conflict correction coefficient, and the warning intensity, and the initial candidate file set is hierarchically sorted according to the final ranking score to obtain the classification retrieval result; when the query does not contain category-defined terms, the classification conflict correction coefficient is set to a preset benchmark value.

[0016] Preferably, the classification retrieval results are written to the result cache and a result page is generated. When the sorting results meet the early warning joint investigation conditions, the joint investigation results of related risk events and related files are expanded, including: writing the sorted candidate files to the result cache and organizing the result page according to the file main category, device object, risk event, and time range; synchronously outputting the file number, static original label, dynamic main category, related risk event, evidence fragment summary, and evidence image area number in each result; determining whether the early warning intensity of the first file on the result page reaches the preset early warning threshold, whether there is a related risk event anchor point, and whether there is an evidence edge; when the early warning joint investigation conditions are met, calling the graph database interface to query the adjacent nodes of the first file anchor point, extracting the direct risk event, upstream handling files, and downstream review files, generating the joint investigation results, and outputting them together with the result page.

[0017] This invention also provides an energy archive classification and retrieval system integrating a risk warning graph, comprising: a graph mapping module for reading energy archives and historical risk events, constructing a unified node relationship table corresponding to the risk warning graph, and calculating the graph node similarity between archive anchors and risk event anchors based on the unified node relationship table; an alignment representation module for performing bidirectional alignment of text evidence fragments and image evidence regions in energy archives based on archive anchors, and generating dynamic classification representations based on the graph node similarity and bidirectional alignment results, and forming a classification candidate set by combining static original labels; and a hierarchical recall module for receiving query statements and performing keyword recall, full-text recall, and hierarchical recall based on the classification candidate set. The method employs a vector recall based on the dynamic classification representation to obtain an initial candidate file set. It then calculates the risk perception candidate score of the initial candidate file set by combining the corresponding risk event anchor point set and the similarity of the graph nodes. A conflict correction and sorting module performs classification conflict correction and hierarchical sorting on the initial candidate file set based on the conflict relationship between the static original labels and the dynamic classification representation, obtaining classification retrieval results. An output query module writes the classification retrieval results into a result cache and generates a result page. When the sorting results meet the early warning query conditions, it expands the query results for associated risk events and associated files. Each module implements the steps of an energy file classification retrieval method that integrates a risk early warning graph.

[0018] The present invention has the following beneficial effects:

[0019] Compared with existing technologies, this invention moves the node similarity and evidence edge strength in the risk warning graph to the main link of candidate recall and ranking. By performing risk perception scoring on candidate files after keyword recall, full-text recall and vector recall, it solves the problem that it is difficult to reliably discover files related to risk propagation paths by relying solely on word-meaning hits or general semantic distance. This enables risk-related files to be prioritized for identification in the early stages of retrieval, reducing the burden of manual review caused by the omission of key evidence in accident tracing and defect investigation, and improving the risk orientation and ranking stability of energy file retrieval results.

[0020] This invention establishes a bidirectional alignment mechanism between textual evidence fragments and image evidence regions around archival anchor points. This mechanism creates fact-level constraints between equipment codes, component names, defect types, operating condition terms, and handling actions in the text and component regions, anomaly labels, and instrument readings in the images. This alleviates the problem of mis-clustering of seemingly similar but factually different archives based solely on a unified representation. As a result, it reduces multimodal mismatches, enhances the interpretability of classification results and evidence chains, and improves retrieval accuracy and robustness in complex energy archive scenarios.

[0021] This invention solves the problem of difficulty in balancing management standards and retrieval effectiveness when there is inconsistency between the static classification at the time of archiving and the current risk perspective classification by simultaneously retaining the static original label and the dynamic main class in the dynamic classification representation and performing classification conflict correction based on the conflict relationship between the two during the sorting stage. This allows historical archive categories and current risk semantics to participate in sorting in parallel, improving the ability to move key archives forward in scenarios such as accident review, equipment hazard tracing, and safety auditing, and enhancing the system's adaptability to the time-varying nature of risk semantics.

[0022] This invention employs a hierarchical processing approach of recall, risk scoring, and conflict correction, which limits graph fusion and evidence fusion to the initial candidate archives. This avoids high-cost rearrangement calculations across the entire database, ensuring the depth of risk correlation modeling while suppressing overall retrieval overhead. This helps maintain millisecond- to second-level engineering response capabilities and improves the continuous operation capability and resource utilization efficiency of large-scale energy archive platforms. Attached Figure Description

[0023] Figure 1 This is an overall flowchart of an energy archive classification and retrieval method that integrates risk warning maps in an embodiment of the present invention.

[0024] Figure 2 This is a structural diagram of an energy archive classification and retrieval system that integrates risk warning maps, as described in an embodiment of the present invention.

[0025] Figure 3 This is a flowchart of the risk warning map mapping and node similarity calculation unit in an embodiment of the present invention.

[0026] Figure 4 This is a flowchart of the bidirectional alignment and classification candidate set generation unit for textual and image evidence in an embodiment of the present invention.

[0027] Figure 5 This is a flowchart of the risk perception candidate score calculation and classification retrieval unit in an embodiment of the present invention.

[0028] Figure 6 This is a flowchart of the classification retrieval output and risk linkage unit in an embodiment of the present invention. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will now be described in further detail with reference to the accompanying drawings.

[0030] 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.

[0031] Example 1: Overall Flow of an Energy Archive Classification and Retrieval Method Integrating Risk Early Warning Maps

[0032] like Figure 1 As shown, the energy archive classification and retrieval method integrating risk warning maps provided in this embodiment is executed by an energy archive retrieval device. The energy archive retrieval device can be a single server or a processing cluster consisting of a retrieval server, a graph database server, and index service nodes. The execution entity includes a processor, a memory, a graph database interface, a full-text index interface, and a vector index interface. The processor is responsible for executing steps S1 to S5, and the graph database, full-text index library, and vector index library serve as data resources accessed by the processor.

[0033] In this embodiment, the deployment environment is an energy archive retrieval server configured with a 32-core processor, 128GB of memory, and a 24GB graphics card. The graph database contains 1.2 million nodes and 4.8 million edges; the total number of energy archives is 1 million. The retrieval processing cycle is a single query processing cycle, meaning that a complete retrieval process is executed once for each received query request. The graph update cycle is preferably 1 day. This graph update cycle, in engineering terms, refers to the offline refresh interval of the unified node relationship table and the graph index, which can be obtained through offline calibration based on the incremental archive entry rhythm and the update frequency of historical risk events. In practical engineering applications, its typical value is 1 day, and it can usually be reasonably configured between 12 hours and 3 days depending on the archive update density. The full-text index and vector index are preferably updated offline once a day to balance index freshness and index construction overhead.

[0034] like Figure 3 As shown, in step S1, energy archives and historical risk events are read, a unified node relationship table corresponding to the risk warning map is constructed, and the similarity of map nodes between archive anchor points and risk event anchor points is calculated based on the unified node relationship table.

[0035] In this embodiment, the energy archive includes structured fields, text content, and image attachments; historical risk events include at least event number, path level, responsible unit, and event time information. The processor first reads the energy archive to be added to the database and the historical risk event records, and then maps them to the node and edge relationships in the risk warning graph, thereby providing a graph index that can be directly used in the calculation for subsequent retrieval and sorting.

[0036] Step S101: Constructing a unified node relationship table

[0037] The construction of a unified node relationship table corresponding to the risk warning map includes: reading the structured fields, text, and image attachments of energy archives; extracting equipment codes, component names, defect types, operating condition terms, handling actions, and time markers from the text; and extracting component area labels, abnormal appearance labels, and instrument value labels from the image attachments to obtain archive anchor field; reading event numbers, path levels, responsible units, and event time information from historical risk events to obtain risk event anchor field; generating archive anchor primary keys and risk event anchor primary keys based on the archive anchor field and risk event anchor primary keys, and writing the archive anchor primary keys and risk event anchor primary keys into the unified node relationship table; obtaining a pre-stored equipment alias mapping table; when the text lacks standard equipment codes but has equipment aliases, restoring the standard equipment codes according to the equipment alias mapping table, and organizing the node primary key field, node type field, standard equipment code field, component name field, defect type field, time slice field, path level field, edge type field, and edge weight field into the unified node relationship table.

[0038] In this embodiment, the processor reads a file D1 from memory. Its structured fields include the equipment code "TR220-A-01", the professional category "Transformer", the archiving label "Daily Inspection", and the date "2025-03-10". Key information such as "casing", "crack", "leakage", and "replacement arrangement" are identified from the text. The casing area and crack area are identified from the image attachments. Subsequently, the processor concatenates the equipment code, component name, defect type, and time slice into the file anchor key. In this embodiment, the preset time window, in engineering terms, refers to the time slice division criteria that maps discrete dates to a unified business time scale. It can be obtained through offline calibration by statistically analyzing the duration of risk events in historical accident traceability samples. Its typical value is 7 days, but in actual engineering applications, it can usually be configured between 1 and 14 days, by day, week, or maintenance cycle. Based on this setting, the file anchor key generated for file D1 is "TR220-A-01—Casing—Crack—2025W10".

[0039] Correspondingly, the processor reads event anchor point E7 from historical risk events, with event number "RISK-781", path level "Level 2", and responsible unit "Operations and Maintenance Team 1", and maps it to the same business time slice "2025W10". This generates the risk event anchor point primary key "RISK-781—Level 2—Operations and Maintenance Team 1".

[0040] The unified node relationship table preferably includes the following fields: node primary key field, node type field, standard equipment code field, equipment alias field, component name field, defect type field, handling action field, time slice field, path hierarchy field, responsibility unit field, source file number field, source event number field, edge type field, edge weight field, and update time field. Among these, the node type field must at least distinguish between file anchor nodes, risk event anchor nodes, equipment nodes, component nodes, and defect nodes; the edge type field must at least distinguish between component edges, status edges, handling edges, and evidence edges.

[0041] In cases where a device code is missing but an alias exists, the processor invokes the device alias mapping table to perform standardized restoration. To avoid incorrect edge creation due to alias ambiguity, the device alias mapping table in this embodiment is defined as a table showing the correspondence between standard device codes and their official names, common abbreviations, historical aliases, and professional aliases. Its initial data can be generated from the device ledger and historical manually confirmed files. In practical engineering applications, the mapping table is preferably updated offline weekly, with this update cycle typically set between 3 and 14 days depending on the frequency of professional changes. When the same alias maps to multiple standard device codes, the processor prioritizes comparing whether the responsible unit, professional category, and time slice match simultaneously. If a unique match still cannot be determined, the file is marked as a node to be disambiguated, and only the basic node record is retained without establishing a high-confidence evidence edge.

[0042] Step S102: Node connection and evidence edge screening

[0043] The calculation of the similarity of graph nodes between the archive anchor point and the risk event anchor point based on the unified node relationship table includes: establishing node connections based on the composition relationship from equipment to component, the state relationship from component to defect, the handling relationship from defect to handling action, and the evidence relationship from the archive anchor point to the risk event anchor point, to obtain type edges and evidence edges in the risk warning graph; for the evidence edges, the edge weight of the evidence edge is calculated based on entity consistency, time proximity, and defect fact consistency, and high edge weight evidence edges are selected according to the preset retention rules; in this embodiment, the preset retention rules are used to control the sparsity of the graph and limit the scale of subsequent sorting calculations. Preferably, risk event anchor points of the same profession are first selected by professional category, and then the top 5 are retained by sorting the evidence edge weights from high to low; if there are more than 5 evidence edges with an edge weight of not less than 0.75, all evidence edges with an edge weight of not less than 0.75 are retained, but the total number does not exceed 8; if no evidence edge has an edge weight of 0.35, only the type edge is retained for the archive anchor point, and no high confidence evidence edge is established.

[0044] In this embodiment, the processor establishes candidate evidence edges based on D1 and E7. Since the device codes are identical, the defect types are identical, and the time difference is 0 days, the edge establishment conditions are met. To quantify whether the current file constitutes direct evidence of the corresponding risk event, the processor calculates the edge weights of the evidence edges in this step. The evidence edge fusion coefficients used in this step are obtained through offline calibration using historical risk evidence chain samples.

[0045] Specifically, to address the issue of varying contributions of different sources of consistency to the credibility of evidence, the processor calculates the edge weight of the evidence edge according to the following formula when establishing the evidence edge between the archive anchor point and the risk event anchor point:

[0046] ,

[0047] in, Indicates the archive anchor point Anchor point of risk events Evidence between the edges; Represents the entity consistency score; Indicates fractions that are close in time; Indicates the consistency score of the defective facts; The score represents the consistency between textual and image evidence. , , , The four factors represent the evidence edge fusion coefficients, and their sum is 1.

[0048] In this embodiment, the engineering meaning of the preset evidence edge fusion coefficient is the contribution ratio of entity, time, defective facts, and cross-modal consistency to the evidence strength. It can be obtained by jointly calibrating the evidence edge accuracy retention rate and false connection rate on the confirmed risk evidence chain sample. Its typical values ​​are 0.30, 0.20, 0.30, and 0.20, respectively. In actual engineering, it can be determined by offline scanning in the range of 0.20 to 0.40, 0.10 to 0.30, 0.20 to 0.35, and 0.10 to 0.25.

[0049] To further calculate the consistency of defect facts, the processor uses the following formula:

[0050] ,

[0051] in, This indicates the consistency between text-side defect labels and event defect labels; it is 1 when they are completely consistent, 0.6 when they belong to the same defect family, and 0 otherwise. This indicates the consistency between image-side anomaly labels and event defect labels, and the value selection rules are the same. , These represent the source credibility coefficients on the text side and the image side, respectively. This is a smoothing term used to avoid a zero denominator when both the text and image sides are missing.

[0052] In this embodiment, the text-side source credibility coefficient and the image-side source credibility coefficient are used to measure the basic credibility ratio of textual evidence and image evidence in confirming defective facts. They can be obtained through offline statistical analysis of manually reviewed samples. Their typical values ​​are 0.55 and 0.45, respectively, and can usually be adjusted between 0.40 and 0.70 and 0.30 and 0.60 according to business quality. If the file lacks image attachments, the image-side consistency indicator is set to 0 and the image-side source credibility coefficient is set to 0, and the calculation is completed only by the text side.

[0053] Before the subsequent two-way alignment results are generated, the processor can first use the prior values ​​of text integrity and image detection confidence to form a temporary consistency score; after step S202 is completed, the consistency score of text and image evidence is written back and updated. This consistency score is the maximum two-way alignment score among the alignment candidates to reflect the strongest evidence correspondence in the same file.

[0054] In the evidence edge screening stage, the preset retention rules in this embodiment are used to control the sparsity of the graph and limit the scale of subsequent sorting calculations. These rules can be obtained through joint calibration of the number of graph nodes, average number of edges, and recall delay constraints. A typical strategy is as follows: first, anchor points for risk events within the same professional category are screened; then, evidence edges are sorted from highest to lowest weight, retaining the top 5. If there are more than 5 evidence edges with a weight not lower than 0.75, all evidence edges with a weight not lower than 0.75 are retained, but the total number does not exceed 8. If no evidence edge has a weight reaching 0.35, only the type edge is retained for that anchor point, without establishing a high-confidence evidence edge. Here, the preset high edge weight threshold of 0.75 and the preset minimum retention threshold of 0.35 have engineering meanings of being the high-confidence evidence judgment threshold and the evidence edge establishment threshold, respectively. These can be determined by balancing the false connection rate and the missed connection rate on the confirmed risk-related samples. In practical applications, the former is typically set between 0.65 and 0.85, and the latter is typically set between 0.25 and 0.45.

[0055] For document D1 and risk event E7, combined with the subsequent bidirectional alignment results, we obtain an entity consistency score of 1.00, a time proximity score of 1.00, a defective fact consistency score of 1.00, and a text and image evidence consistency score of 0.99. Therefore, the edge weight of the evidence edge is 0.998, and it is retained as one of the high-confidence evidence edges corresponding to D1. The above edge weight construction and screening process constitutes the aforementioned logic for establishing type edges and evidence edges.

[0056] Step S103: Calculate the similarity of graph nodes and write the index

[0057] Entity consistency scores are generated based on the consistency of device codes, component names, and defect types, respectively. Time proximity scores are generated based on the time proximity of archive time slices and event time slices, and path adjacency scores are generated based on path hierarchy proximity and graph hop count. The entity consistency scores, time proximity scores, and path adjacency scores are then fused according to preset graph weights to obtain the graph node similarity, and the graph node similarity is written into the graph index.

[0058] The similarity of graph nodes reflects the comprehensive degree of object consistency, temporal proximity, and proximity of risk path locations. Together with the aforementioned evidence edge weights, it forms the graph basis for risk perception ranking.

[0059] Specifically, the processor calculates the similarity of graph nodes using the following formula: ,in, Indicates the archive anchor point Anchor point of risk events Graph node similarity; Represents the entity consistency score; Indicates fractions that are close in time; Indicates the path adjacency score; , , This represents the similarity fusion weight of the graph nodes, and the sum of the three is 1.

[0060] In this embodiment, the engineering meaning of the preset graph weight is the proportion of object consistency, time proximity and path proximity in the graph similarity. It can be obtained by jointly optimizing the proportion of the top 10 risk-related factors, the hit rate of the first key file and the average response time on the accident tracing query sample in the past 3 years. Its typical values ​​are 0.45, 0.25 and 0.30, respectively, and can usually be determined by scanning between 0.30 and 0.60, 0.10 and 0.35 and 0.15 and 0.40.

[0061] The entity consistency score is determined by the consistency of three object fields: equipment code, component name, and defect type, and is calculated using the following formula:

[0062] ,

[0063] in, This indicates the device coding consistency indicator; a value of 1 indicates consistency, and a value of 0 indicates inconsistency. This indicates the consistency indicator for component names, with 1 for consistency, 0.5 for mapping to adjacent higher-level components, and 0 for the rest. This indicates the consistency index for defect types, with 1 for consistency, 0.6 for the same defect family, and 0 for the rest. , , This represents the importance coefficient of the equipment code, component name, and defect type. This is a smoothing term.

[0064] In this embodiment, the importance coefficients of equipment code, component name, and defect type are used to reflect the distinguishing contribution of the three types of object fields in accident review. They can be obtained through offline calibration of historical manual review samples. Their typical values ​​are 0.40, 0.25, and 0.35, respectively, and can be adjusted between 0.20 and 0.50 in actual engineering.

[0065] Time proximity score is used to measure the business proximity between archive time slices and event time slices, and is calculated using the following formula:

[0066] ,

[0067] in, This indicates the time difference between the archive anchor time slice and the risk event anchor time slice, in days. This represents the time decay constant.

[0068] To distinguish between short-term correlated events and long-term weakly correlated events, this step introduces a preset time decay constant, which represents the rate at which the correlation decays as the time difference increases. It is usually calibrated offline based on the optimal recall window of historical accident traceability samples, with a typical value of 7 days. In practical engineering applications, it can usually be configured between 3 and 14 days.

[0069] Path adjacency score is used to describe the hierarchical proximity between file anchors and risk event anchors in the risk propagation path, and is calculated using the following formula:

[0070] ,

[0071] in, This indicates the path hierarchy in the graph that maps archive anchor points to their corresponding file anchor points. The path hierarchy representing the anchor point of a risk event; Represents a node With nodes The number of hops on the shortest path in the graph; This represents the penalty coefficient for the number of jumps.

[0072] The preset hop count penalty coefficient is used to measure the attenuation effect of the path jump length on risk relevance. It can be obtained through offline testing of historical associated path samples. Its typical value is 0.20, and it can usually be set between 0.05 and 0.30 based on the average path length of the map.

[0073] In this embodiment, for D1 and E7, the device code, component name, and defect type are completely identical, therefore the entity consistency score is 1.00; the time difference is 0 days, therefore the time proximity score is 1.00; file anchor level Event anchor level Shortest number of hops Therefore, the path adjacency score is approximately 0.83, resulting in a final graph node similarity of 0.95. The processor writes this score into the graph index table for direct use in subsequent risk perception ranking.

[0074] When an archive lacks image attachments, only the device, component, and defect tags extracted from the text are used to construct the node primary key. When the text lacks an explicit device code but has a device alias, the standard device code is first restored through the device alias mapping table. When any node cannot locate the path level, the path adjacency score is set to 0. When both archive and event time information are missing, the time proximity score is set to 0. This boundary processing ensures that the similarity of graph nodes can always be calculated, preventing graph indexing failure due to missing items. The above calculation process implements the aforementioned writing of graph node similarity into the graph index.

[0075] like Figure 4 As shown, in step S2, bidirectional alignment is performed on the text evidence fragments and image evidence regions in the energy archive based on the archive anchor points, and a dynamic classification representation is generated according to the similarity of the map nodes and the bidirectional alignment results, and a classification candidate set is formed by combining the static original labels.

[0076] In this embodiment, the goal of step S2 is to constrain and fuse textual and image evidence in the archive around the same device defect fact, so that subsequent classification and vector retrieval no longer rely solely on general representations, but explicitly embed the device, component, defect, and risk adjacency relationships.

[0077] Step S201: Extraction of text evidence fragments and image evidence regions

[0078] Performing bidirectional alignment of text evidence fragments and image evidence regions in energy archives based on archive anchor points includes: segmenting the text into text evidence fragments, and labeling each text evidence fragment with equipment code, component name, defect type, operating condition term and handling action to obtain a set of text evidence fragments; performing region detection or region proposal based on image attachments to extract image evidence regions containing component regions, anomaly labels and confidence levels to obtain a set of image evidence regions.

[0079] In this embodiment, the processor segments the main text of file D1 into four text evidence fragments using periods, semicolons, and entry marks. Fragment T2 contains the message "The crack on the casing surface continues to expand, accompanied by slight water seepage; power outage and replacement recommended." Simultaneously, the processor performs region detection or region proposal on the image attachments, retaining output fields related to component regions, anomaly labels, and confidence levels. For D1, three image evidence regions are ultimately retained: region R1 is labeled "Casing Crack" with a confidence level of 0.92; region R2 is labeled "Nameplate" with a confidence level of 0.88; and region R3 is labeled "Background Equipment" with a confidence level of 0.40.

[0080] When the image detection module outputs multiple overlapping regions, the processor preferably uses non-maximum suppression or an overlap rate threshold to ensure the overlap rate is not lower than a preset confidence threshold of 0.50. In this embodiment, the overlap rate threshold is defined as the threshold for determining overlap between two candidate regions as duplicate detections, and can be obtained through offline calibration using image annotation samples; its typical value can be set to 0.50, and in practical engineering, it can usually be reasonably configured between 0.40 and 0.70. The aforementioned text fragments and image regions constitute the aforementioned set of text evidence fragments and set of image evidence regions.

[0081] Step S202: Two-way alignment score calculation

[0082] Based on the primary key of the archive anchor point, segments and regions related to the same equipment defect fact are filtered from the set of text evidence segments and the set of image evidence regions; a pre-stored business mapping table is obtained, and a bidirectional alignment score between text evidence segments and image evidence regions is calculated based on the degree of matching of anchor point fields, the degree of consistency of business context, and the joint reliability of image region credibility and text completeness. The text completeness is determined by the completeness of identification of equipment code, component name, defect type, and handling action; when the energy archive does not contain image attachments, the text evidence segments are used as unilateral evidence in the subsequent dynamic classification and representation generation.

[0083] For file D1, only fragment T2 and region R1 are retained for alignment calculation.

[0084] To measure whether text fragments and image regions jointly describe the same risk fact, the processor calculates a two-way alignment score using the following formula: ,in, Indicates the first The text evidence fragment and the first Bidirectional alignment score between image evidence regions; This indicates that the anchor field matches the score; Indicates the context consistency score; Indicates the regional confidence score; , , This represents the two-way alignment weight, and the sum of the three is 1.

[0085] In this embodiment, a preset bidirectional alignment weight is used to control the contribution ratio of object field matching, business context consistency, and joint reliability in cross-modal fact verification. It can be obtained by offline fitting of manually annotated text fragments and image regions with fact samples. Its typical values ​​are 0.50, 0.30, and 0.20, and can usually be determined by scanning within the range of 0.35 to 0.60, 0.20 to 0.40, and 0.10 to 0.30.

[0086] The anchor field matching score is calculated using the following formula:

[0087] ,

[0088] in, Indicates the device code matching indicator quantity; Indicates the component name matching indicator quantity; Indicates the defect type matching indicator quantity; , , This represents the importance coefficient of the three types of matching items; This is a smoothing term.

[0089] In this embodiment, the device code matching weight, component matching weight, and defect matching weight are used to represent the distinguishing ability of each field in the judgment of the consistency between text and image facts. They can be obtained through offline statistics of manually reviewed samples; their typical values ​​are 0.35, 0.25, and 0.40, respectively.

[0090] Contextual consistency score reflects whether there is a business-related co-occurrence between condition words, actions, and image anomaly labels in the text, and is calculated using the following formula:

[0091] ,

[0092] in, This indicates the number of co-occurrence entries formed in the business mapping table between the condition words and handling actions in the text evidence fragments and the image anomaly labels; Indicates the first The total number of work-related terms and actions identified in each text evidence fragment; This is a smoothing term.

[0093] If no work condition words and handling actions are detected in a certain text segment, that is... If so, the context consistency score is set to 0 to prevent abnormal amplification when there is no context information.

[0094] The region confidence score is used to combine image detection confidence with text fragment completeness, and is calculated using the following formula: ,in, Indicates the first Detection confidence of each image evidence region; Indicates the first The completeness score of each text evidence fragment.

[0095] The text completeness score is determined by the completeness of identification of four fields: equipment code, component name, defect type, and handling action, and is calculated using the following formula: ,in, This indicates whether the device code has been identified; 1 indicates identification, and 0 indicates otherwise. This indicates whether the component name has been identified; 1 indicates identification, and 0 indicates otherwise. This indicates whether a defect type has been identified; 1 indicates identification, and 0 indicates otherwise. This indicates whether a action was identified; a score of 1 indicates identification, and a score of 0 indicates otherwise. Therefore, the completeness score is 1 when all four items are present, 0.75 when one item is missing, 0.50 when two items are missing, 0.25 when three items are missing, and 0 when all items are missing.

[0096] The aforementioned preset text completeness rule is used to measure whether a text fragment has the completeness of fields that can independently support the determination of risk facts. It can be obtained by manually verifying the information completeness rate statistics in the sample. In this embodiment, the rule is implemented by a four-item equal distribution method, which is applicable to energy files such as equipment operation and maintenance, defect investigation and accident review.

[0097] For fragment T2 and region R1 of file D1, the equipment code, component name, and defect type match perfectly, so the anchor field matching score is 1.00; there are two working condition words and handling actions in the text, namely "continue to expand" and "suggest replacement", and these two, together with the image anomaly label, form two business co-occurrence entries, so the context consistency score is 1.00; fragment T2 has all four fields, so the text integrity score is 1.00; the detection confidence of region R1 is 0.92, so the region confidence score is 0.96; finally, the bidirectional alignment score is 0.99.

[0098] When the document has no image attachments, the image-side region is not calculated, and the text evidence fragment is directly used as unilateral evidence in the subsequent dynamic classification representation. When the image detection retains only one region, it still participates in alignment as a single region without cross-region competition. The above calculation process realizes the aforementioned bilateral alignment score generation.

[0099] Step S203: Dynamic classification representation generation

[0100] The dynamic classification representation generated based on the graph node similarity and bidirectional alignment results includes: generating a text main class score based on the text evidence fragment, generating an equipment main class score based on the equipment professional category rules, and generating a defect main class score based on the defect classification rules; generating a risk association main class score by fusing the graph node similarity and the bidirectional alignment score according to the set of risk event anchor points related to the professional to which the current file belongs; constructing a dynamic classification representation for vector retrieval and main class determination based on the text main class score, equipment main class score, defect main class score, and risk association main class score; jointly retaining the dynamic main class in the dynamic classification representation with the static original label, and recording a conflict marker when the two are inconsistent, to obtain the classification candidate set; the dynamic classification representation includes text evidence features, image evidence features, graph adjacency features, and classification rule statistical features.

[0101] In this embodiment, the processor calculates a text main category score, an equipment main category score, a defect main category score, and a risk-related main category score for each file. The text main category score is obtained by weighting the degree of support each text evidence fragment provides for the candidate main category; the equipment main category score is obtained by mapping the rules between equipment code prefixes, professional categories, responsible units, and typical component lists; and the defect main category score is obtained by mapping the rules between defect names, defect families, handling actions, and risk levels. These three types of scores can be generated by combining rule engines and statistical features, and are not limited to a single model.

[0102] The risk-related primary class score measures which primary class the current file is closer to from a risk perspective, and is calculated using the following formula:

[0103] ,

[0104] in, This indicates the score of the main category associated with the current file's risk classification. This represents the set of risk event anchors related to the profession to which the file belongs; Indicates the current file anchor point; Indicates the similarity of graph nodes; Indicates two-way aligned scores; This represents the fusion coefficient of the risk-related main class.

[0105] In this embodiment, the preset fusion coefficient is used to balance the role of the proximity relationship of the graph and the consistency of the text image facts in the determination of the risk main class. It can be obtained by jointly optimizing the classification consistency rate and risk recall rate on the validation set. Its typical value is 0.55, and it can usually be configured between 0.40 and 0.70.

[0106] For file D1, since the highest similarity to the graph node of E7 is 0.95 and the maximum bidirectional alignment score is 0.99, the risk association main class score is 0.97. Combining the text main class score, equipment main class score, and defect main class score, the processor determines the dynamic main class of D1 as "casing crack risk traceability file", while retaining the static original label "daily inspection" and recording conflict markers, rather than directly overwriting the original archive label.

[0107] In this embodiment, the dynamic classification representation preferably adopts a 256-dimensional representation. The engineering meaning of this dimension is the length of the classification retrieval features used for vector retrieval and main class determination, which can be determined offline by comparing index usage, single re-ranking time, the proportion of the top 10 risk-related features, and the dynamic main class consistency rate; its typical value is 256 dimensions, but in actual engineering, it can be selected from 128, 192, 256, and 320 dimensions according to computing power and accuracy requirements. Specifically, this 256-dimensional representation is formed by concatenating 64 dimensions of textual evidence features, 64 dimensions of image evidence features, 64 dimensions of graph adjacency features, and 64 dimensions of classification rule statistical features.

[0108] The text evidence features include at least the main text class score, keyword coverage statistics, equipment code hit statistics, defect word frequency, operating condition word frequency, handling action word frequency, time slice location code, and fragment completeness aggregation value, which are then linearly projected to form a 64-dimensional representation. The image evidence features include at least the regional anomaly label statistics, main region confidence, regional location distribution, equipment component category distribution in the image, and high-confidence region count after alignment with the text, which are then linearly projected to form a 64-dimensional representation. The graph adjacency features include at least the edge weight statistics of the first few evidence edges, path hierarchy distribution, shortest hop count statistics, number of related risk events, and graph adjacency density, which are then linearly projected to form a 64-dimensional representation. The classification rule statistical features include at least the main text class score, equipment class score, defect class score, risk association class score, static label one-hot representation, conflict marker, and query history hit statistics, which are then linearly projected to form a 64-dimensional representation. The process of combining these four types of features constitutes the aforementioned dynamic classification representation.

[0109] When the set of risk event anchor points is empty, the risk association main class score is directly set to the maximum two-way alignment score; when there is neither image alignment result nor any mapped risk event, the static original label is retained and the risk association main class score is set to 0 to ensure that weak information files can still enter the classification link.

[0110] like Figure 5As shown, in step S3, a query statement is received, and keyword recall, full-text recall, and vector recall based on the dynamic classification representation are performed according to the classification candidate set to obtain an initial candidate file set. The risk perception candidate score of the initial candidate file set is calculated by combining the risk event anchor point set corresponding to the query and the similarity of the graph nodes.

[0111] Step S301: Retrieve constraint table generation

[0112] The process of performing keyword recall, full-text recall, and vector recall based on the dynamic classification representation based on the classification candidate set includes: parsing the query statement to obtain a keyword set, a device code set, a time range, file category limiting words, and risk limiting words, and generating a retrieval constraint table; obtaining a pre-stored category mapping table, determining the query limiting category based on the category mapping table, and locating the corresponding risk event anchor point set when the risk limiting words exist.

[0113] In this embodiment, the user inputs the query "Retrieve historical files and risk records related to the crack in the A-phase bushing of the 220kV main transformer". The processor parses the query to obtain the keyword set {220kV, main transformer, A-phase, bushing, crack}, the equipment code set {TR220-A-01}, and the risk qualifier set {crack, risk record}, ​​and writes them into the retrieval constraint table.

[0114] The category mapping table includes at least five fields: standard category name, synonym, superordinate category, subordinate category, and professional alias. The processor preferably determines the query-limited category in the order of exact match, synonym match, and superordinate / subordinate match. In this embodiment, the category mapping table matches "risk record" to the "risk traceability file" category. Here, the engineering meaning of the preset category matching rule is the semantic hierarchy mapping basis between different category terms, which can be constructed offline through historical query logs and manual classification results; its typical matching strength can be set to 1 for exact match, 0.9 for synonym match, and 0.7 for superordinate / subordinate match, and can be adjusted between 0.6 and 1 in actual engineering according to the complexity of the category system.

[0115] If multiple professional categories appear in the query and conflict with each other, the processor determines the primary query category according to the rule of prioritizing equipment codes, followed by risk qualifiers, and then text occurrence positions. If it still cannot be determined, the category qualifiers are left blank, retaining only the risk qualifiers and equipment codes. The results of the above processing constitute the aforementioned search constraint table.

[0116] Step S302: Layered Recall Execution

[0117] Combining the time range and file category qualifiers, keyword matching retrieval is performed based on the retrieval constraint table to obtain a keyword candidate set; full-text index retrieval is performed based on the retrieval constraint table to obtain a full-text candidate set; vector retrieval retrieval is performed based on the dynamic classification representation to obtain a vector candidate set; and a union is performed on the keyword candidate set, the full-text candidate set, and the vector candidate set to obtain the initial candidate file set.

[0118] In this embodiment, the processor first performs keyword matching recall, then full-text index recall, and finally vector retrieval on the dynamic classification representation table. To control the scale of each type of recall and suppress subsequent reordering overhead, the keyword recall upper limit in this embodiment is used to limit the number of term-driven candidates, typically set at 50, and can be configured between 20 and 100 depending on the index density; the full-text recall upper limit is used to limit the number of full-text related candidates, typically set at 100, and can be configured between 50 and 200; the vector recall upper limit is used to limit the number of vector nearest neighbor candidates, typically set at 80, and can be configured between 30 and 150. These upper limits can be obtained through offline calibration using average query latency and the top 10 hit rate.

[0119] For the current query, keyword recall returned 45 files, full-text recall returned 96 files, and vector recall returned 72 files, which, after being combined, form 184 initial candidate files. File D1 hit all three recall sets simultaneously. By prioritizing recall followed by scoring, the complex graph fusion can be limited to the candidate range, reducing overall computational overhead. The above process achieves the generation of the aforementioned initial candidate file set.

[0120] Step S303: Calculation of candidate scores for risk perception

[0121] Calculating the risk perception candidate score of the initial candidate file set includes: calculating the keyword matching score, full-text search relevance score, and vector search score for each candidate file in the initial candidate file set; generating a risk graph perception score based on the similarity of graph nodes between the file anchor points of the candidate file and the risk event anchor point set corresponding to the query, combined with the risk association main class score; fusing the keyword matching score, full-text search relevance score, vector search score, and risk graph perception score according to a preset fusion weight to obtain the risk perception candidate score; and using the risk association main class score as the risk graph perception score when no risk limiting words are parsed from the query.

[0122] In this embodiment, the keyword matching score can be calculated by combining the number of hit keywords and the hit rate of key fields; the full-text search relevance score can be obtained by mapping the original full-text index score to the quantiles in the current query candidate set; and the vector retrieval score can be calculated by converting the vector distance into relative proximity in the current recall batch. In order to move the graph factors forward to the main link, the processor further calculates the risk graph perception score.

[0123] Specifically, the risk perception candidate score of the candidate profile is calculated using the following formula:

[0124] ,

[0125] in, Indicates candidate files Risk perception candidate scores; Indicates the keyword matching score; This indicates the score related to full-text search. Indicates the vector retrieval score; This represents the risk map perception score; , , , This represents the risk perception fusion weight, and the sum of the four is 1.

[0126] In this embodiment, the engineering meaning of the preset fusion weight is the contribution ratio of word literal hit, full-text relevance, vector proximity, and graph association to the total relevance of candidate files. It can be determined by joint optimization of the manual relevance annotation set and the accident tracing retrieval set. Its typical values ​​are 0.20, 0.20, 0.25, and 0.35, respectively. In actual engineering, it can be determined by scanning within the range of 0.10 to 0.30, 0.10 to 0.30, 0.15 to 0.35, and 0.20 to 0.45.

[0127] Risk perception score is calculated using the following formula:

[0128] ,

[0129] in, This represents the set of risk event anchor points obtained based on the query statement; Indicates candidate files Archive anchor points; Indicates candidate files Risk-related main class score; This represents the spectral fusion coefficient.

[0130] To enhance the risk association constraints in the query context, this step introduces a preset graph fusion coefficient, which represents the relative weight of the query-corresponding risk event anchor point and the inherent risk category of the file in the risk score. It is usually based on the joint calibration of the top 10 risk-related proportions, the top 5 hit rates and false positive rates, with a typical value of 0.60. In actual engineering, it can be set between 0.40 and 0.75.

[0131] For file D1, the keyword matching score is 0.88, the full-text search relevance score is 0.81, and the vector search score is 0.93. Since the query risk event anchor point set includes E7, and the graph node similarity is 0.95 and the risk association main class score is 0.97, the risk graph perception score is 0.96, and the final risk perception candidate score is 0.91.

[0132] When the query fails to parse out risk qualifiers, resulting in an empty set of anchor points for the query risk events, the risk graph perception score is directly set to the risk association main category score. When a candidate file does not enter the vector recall set, the vector retrieval score is set to 0. When a candidate file does not hit the keyword recall set, the keyword matching score is set to 0. The above processing constitutes the aforementioned risk perception candidate score.

[0133] like Figure 5 As shown, in step S4, based on the conflict relationship between the static original labels and the dynamic classification representation, classification conflict correction and hierarchical sorting are performed on the initial candidate file set to obtain the classification retrieval results.

[0134] Step S401: Classification conflict correction

[0135] Performing classification conflict correction and hierarchical sorting on the initial candidate file set includes: checking the conflict markers between the static original labels and dynamic main classes of each candidate file, and determining the classification conflict relationship in conjunction with the query-defined category; when the dynamic main class is consistent with the query-defined category but the static original labels are inconsistent, the corresponding candidate file is sorted upwards; when the static original labels are consistent with the query-defined category but the dynamic main class are inconsistent, the corresponding candidate file is sorted downwards, thus obtaining the classification conflict correction coefficient.

[0136] In this embodiment, the processor checks for a conflict between the static original label "Daily Inspection" and the dynamic main class "Casing Crack Risk Tracing File" of file D1. Since the dynamic main class matches the query-defined category, D1 is ranked higher. To move more relevant files from a risk perspective forward without excessively overriding the original archiving specifications, the processor calculates a classification conflict correction coefficient according to rules.

[0137] In this embodiment, the preset upward adjustment coefficient measures the ranking improvement when the dynamic main category matches the query-defined category but the static tags lag behind. This coefficient can be obtained offline through calibration of the key file forward movement effect in historical review queries; its typical value is 0.30, and in practice, it is usually configured between 0.10 and 0.40. The preset downward adjustment coefficient measures the ranking reduction when the static tags appear relevant but dynamic evidence is insufficient. This coefficient can be obtained through joint calibration of the erroneous file extraction ratio and the key file forward movement capability; its typical value is 0.20, and in practice, it is usually configured between 0.10 and 0.30. If the category relationship is only weakly consistent (upper and lower levels), it is preferable to multiply by a weak consistency correction factor of 0.85 to reduce ranking fluctuations caused by hierarchical differences.

[0138] When the query does not contain category qualifiers, the classification conflict correction coefficient is set to a preset baseline value. This preset baseline value, in engineering terms, is a default coefficient that does not impose additional bias on static labels and dynamic main class conflicts when there are no category constraints. It can be directly set according to the ranking stability requirements; its typical value is 1.00, and it can usually be configured between 0.95 and 1.05. The above processing achieves the aforementioned classification conflict correction.

[0139] Step S402: Output of final sorting and classification retrieval results

[0140] Based on the evidence edge weights between candidate file anchors and query-related risk event anchors, the maximum edge weight is extracted as the warning intensity; a final ranking score is generated according to the risk perception candidate score, classification conflict correction coefficient, and warning intensity, and the initial candidate file set is hierarchically sorted according to the final ranking score to obtain the classification retrieval results; when the query does not contain category qualifiers, the classification conflict correction coefficient is set to a preset benchmark value.

[0141] In this embodiment, the processor extracts the maximum edge weight from the evidence edge between the candidate file anchor point and the query-related risk event anchor point as the warning intensity. For D1, its maximum evidence edge weight is 0.998, therefore the warning intensity is 0.998.

[0142] To ensure that high-credibility risk evidence is appropriately moved forward in the final ranking, the processor introduces a warning amplification factor. In this embodiment, the preset warning amplification factor controls the degree to which the warning intensity enhances the final ranking score. It can be obtained by jointly calibrating the erroneous inclusion rate and the ability to move key documents forward in historical manual review results; its typical value is 0.25, and in actual engineering, it can usually be configured between 0.10 and 0.35.

[0143] Based on this, the final ranking score is generated by fusing the risk perception candidate score, the classification conflict correction coefficient, and the warning intensity. For D1, the risk perception candidate score is 0.91, the classification conflict correction coefficient is 1.29, the warning intensity is 0.998, and the final ranking score is 1.20, placing it in the first position.

[0144] The processor then outputs a list of results in descending order of the final sorting score, returning the file number, static original label, dynamic main category, associated risk event, evidence fragment summary, and evidence image region number for each result. For D1, the result simultaneously returns the file number, the static original label "routine inspection," the dynamic main category "casing crack risk traceability file," the associated risk event "RISK-781," the evidence fragment T2 summary, and the image region R1. This output constitutes the aforementioned classification retrieval results.

[0145] When multiple candidate files have the same final ranking score, priority is given to those with higher time proximity, and then to those whose static original tags are consistent with the dynamic main class, in order to enhance the stability of the results.

[0146] like Figure 6 As shown, in step S5, the classification retrieval results are written into the result cache area and a result page is generated. When the sorting results meet the early warning joint investigation conditions, the joint investigation results of related risk events and related files are expanded.

[0147] Step S501: Result caching and result page generation

[0148] The classification retrieval results are written to the result cache and a result page is generated. When the sorting results meet the early warning joint investigation conditions, the joint investigation results of related risk events and related files are expanded, including: writing the sorted candidate files to the result cache and organizing the result page according to the file main category, device object, risk event and time range; and synchronously outputting the file number, static original label, dynamic main category, related risk event, evidence fragment summary and evidence image area number in each result.

[0149] In this embodiment, the processor writes the sorting results to the result cache. The result cache, in engineering terms, is a temporary storage area that stores the current query result page and its context, and can be implemented through memory caching or high-speed key-value storage. Its cache retention period can be configured based on user page-turning behavior and the frequency of repeated accesses; in practical engineering applications, a typical value is 10 minutes, and it can usually be set between 1 minute and 30 minutes. The result page is organized hierarchically according to "File Main Category—Device Object—Risk Event—Time Range," allowing users to view results from multiple dimensions, including category, object, and risk chain.

[0150] Step S502: Triggering of Early Warning Joint Investigation and Deployment of Related Results

[0151] Determine whether the warning intensity of the first file on the results page reaches the preset warning threshold, whether there is an anchor point for a related risk event, and whether there is an evidence edge; when the warning joint investigation conditions are met, call the graph database interface to query the adjacent nodes of the anchor point of the first file, extract the direct risk event, upstream handling file, and downstream review file, generate the joint investigation results, and output them together with the results page.

[0152] In this embodiment, the preset warning threshold is used to determine whether the first file in the current result is sufficient to trigger a risk joint investigation. Its engineering meaning is the minimum credible threshold for a file to serve as direct evidence of a risk fact. It can be obtained by offline calibration of the false trigger rate and the missed trigger rate in historical joint investigation samples. Its typical value is 0.65, and it can usually be configured between 0.50 and 0.80.

[0153] When the warning intensity of the first file on the results page is higher than 0.65, and there is at least one associated risk event anchor point and at least one evidence edge, the processor determines that the warning linkage conditions are met. Subsequently, the processor calls the graph database interface to query the one-hop and two-hop adjacent nodes of the first file anchor point, reads the associated file ID, associated event ID, and edge type corresponding to the evidence edge, and generates the linkage results page in the order of "direct risk event - upstream handling file - downstream review file".

[0154] Through this step, the system ultimately outputs not only a sorted list of files, but also the risk chain lookup results centered around the top high-risk file. This allows users to directly access the associated chain of risk events, handling files, and review files from a single file, forming a closed loop from retrieval to tracing. The above processing achieves the aforementioned results page output and risk chain lookup.

[0155] In this embodiment, the graph weight, bidirectional alignment weight, evidence edge fusion coefficient, and risk perception fusion weight are all obtained through offline grid scanning of historical samples. The graph weight is based on 2800 accident tracing query samples from the past three years; the bidirectional alignment weight comes from 8600 pairs of manually annotated text fragments and image regions; the evidence edge parameters come from 4200 sets of confirmed risk evidence chains; and the dynamic classification representation dimensions are determined by comparing 128, 192, 256, and 320 dimensions. Offline replay results show that, under the test conditions of 1 million energy files, 30 consecutive days, and 3000 tracing queries, the top 10 risk-related files in this embodiment account for 90%, while the corresponding scheme using keyword plus full-text search accounts for 60%; the hit rate of the first key file reaches 89%, while the corresponding scheme using keyword, full-text, and general vector search accounts for 68%; the average response time is 428ms, while the corresponding scheme is 517ms. The results demonstrate that by introducing graph node similarity, bidirectional alignment score, and conflict correction mechanism into the same main retrieval link, it is possible to improve the ability to move key risk files forward while controlling retrieval latency.

[0156] Example 2: Mismatch Suppression in Scenes with Mixed Images of Corrosion and Flashover on Transmission Line Fittings

[0157] This embodiment illustrates how to suppress mismatches using the aforementioned bidirectional alignment and graph edge weighting mechanism in scenarios where images appear similar but have different risk facts. The scenario involves 200,000 files, 350,000 image attachments, a graph update cycle of one day, and a single query processing cycle.

[0158] In this embodiment, the engineering meaning of the preset time window is to constrain the time difference threshold for comparability between archives and risk events within the same business time scale, which can be statistically calibrated through the defect evolution cycle in the line operation and maintenance scenario; its typical value is 7 days, and it can usually be configured between 1 day and 14 days. The specific typical values ​​of the preset bidirectional alignment weight are still 0.50, 0.30, and 0.20; the typical value of the preset fusion coefficient is 0.55; the typical value of the preset early warning threshold is 0.65. The engineering definition and acquisition method of the above parameters are consistent with those in Embodiment 1.

[0159] In candidate file D2, the text fragment T5 states, "The corrosion of the crossarm hardware on tower No. 5 has worsened; rust removal and repainting are recommended." The detection label for image region R4 is "insulator flashover," with a confidence level of 0.90. In traditional unified vector retrieval, due to the similar visual patterns of "surface abnormal patches," D2 may be incorrectly grouped into the flashover file cluster.

[0160] In this embodiment, in step S202, the processor first checks whether the anchor field of fragment T5 and region R4 are consistent. Since the component name of T5 is "crossarm hardware" and the component label of R4 is "insulator", the component name matching indicator is 0; the defect types "rust" and "flashover" do not belong to the same defect family, so the defect type matching indicator is 0; even if the equipment codes are on the same tower, the anchor field matching score still drops to 0.37. Furthermore, since the handling action "rust removal and repainting" in the text and the image label "flashover" do not have a co-occurrence relationship in the business mapping table, the context consistency score is 0. Although fragment T5 lacks an explicit equipment code, it has a component name, defect type, and handling action, so the text completeness score is 0.75, and the final bidirectional alignment score is 0.36.

[0161] Furthermore, in step S102, if an attempt is made to connect D2 to the "flashover" risk event anchor point, the resulting evidence edge weight will be lower than 0.35 due to the low consistency of both the defective facts and the text / image evidence, and therefore will not be included in the high-confidence evidence edge set. Here, the aforementioned preset minimum retention threshold of 0.35 represents the minimum credibility requirement for establishing high-confidence evidence edges. This can be controlled by the false connection rate during offline calibration, with a typical value of 0.35, and can usually be configured between 0.25 and 0.45.

[0162] Conversely, for the genuine hardware corrosion file D3, the text fragment T7 contains "horizontal brace hardware corrosion," and the image region R6 is labeled "hardware corrosion," with a confidence level of 0.94. After calculation in step S202, the bidirectional alignment score of D3 is 0.95, and the risk association main class score obtained in step S203 is 0.91. After entering step S303, its risk perception candidate score is significantly higher than that of D2, and therefore it is moved forward in the final result.

[0163] In this embodiment, the business mapping table was manually compiled by line operation and maintenance professionals. It contains 146 sets of "operating condition words or handling actions - image anomaly tags" relationships, such as "rust removal and repainting - rust" and "cleaning treatment - flashover marks". The engineering meaning of this business mapping table is a co-occurrence constraint rule base between text business actions and image anomalies. It can be constructed through manual review of historical archives and professional knowledge compilation, and updated offline monthly or quarterly. Offline replay results show that, under test conditions of 15 days and 1200 mixed text and image queries, compared to the control scheme of unified vector retrieval plus static tag filtering, this embodiment reduced the mismatch archive rate from 18.0% to 7.0%, increased the consistency rate between dynamic main categories and manually reviewed categories from 82.5% to 93.0%, and increased the hit rate of the top 5 risks from 64.0% to 86.0%.

[0164] Subsequently, based on the above calculation and processing results, the processor continues to execute the subsequent steps as described in Example 1, completing operations such as risk perception scoring, classification conflict correction, result page organization, and joint query output. Its specific execution logic is consistent with Example 1, and will not be repeated here.

[0165] Example 3: Dynamic Classification Conflict Correction in Relay Protection Failure to Operate Replay Scenario

[0166] This embodiment illustrates how to move key archives forward in a scenario where static archive tags lag behind current risk semantics, through dynamic classification representation and conflict correction mechanisms. The scenario involves 150,000 archives, 180,000 risk event anchor points, and a query statement of "retrieving historical archives related to the failure of protection on a certain 35kV line at a certain station."

[0167] In this embodiment, the preset upward adjustment coefficient is defined as the ranking increase when the dynamic main category and the query-defined category are consistent, but the static labels are inconsistent. It can be jointly calibrated by the historical sample statistics of the professional review query and the manual relevance score; its typical value is 0.30, and it can usually be configured between 0.10 and 0.40. The preset downward adjustment coefficient is used to suppress the ranking when the static labels are relevant but the dynamic evidence is insufficient; its typical value is 0.20, and it can usually be configured between 0.10 and 0.30. The typical value of the preset warning amplification coefficient is 0.25, and the typical value of the preset warning threshold is 0.65. The engineering definition and acquisition method of the above parameters are consistent with those in Embodiment 1.

[0168] The static original label of a certain file D4 is "experiment record", but in step S103, its similarity with the graph node of the "protection refusal to move" risk event anchor point reaches 0.89; in step S202, the bidirectional alignment score between the text fragment and the test waveform screenshot reaches 0.93; in step S203, the risk association main class score is 0.91, and the dynamic main class is identified as "protection refusal to move risk traceability file".

[0169] The query explicitly contains the category qualifier "reluctance to act," therefore, in step S301, the processor maps it to the standard category "Protection of Reluctance to Act Risk Tracing File" according to the category mapping table. In step S401, D4 satisfies the condition that "the dynamic main class is consistent with the query-qualified category and the static original label is inconsistent," resulting in a classification conflict correction coefficient of 1.27. Combined with its maximum evidence edge weight of 0.72, after calculating the final ranking score in step S402, D4 moves from the 8th position to the 3rd position under the static label ranking scheme.

[0170] Compared to another file, D5, its static original label is "accident analysis," but its graph node similarity is only 0.42, its bidirectional alignment score is only 0.35, and its risk association main class score is 0.39. Although its static label appears to be closer to the query category, it will be slightly downgraded in step S401 due to insufficient dynamic evidence, and therefore will not remain at the top for long.

[0171] In this embodiment, the upward adjustment coefficient of 0.30 and the downward adjustment coefficient of 0.20 are derived from the optimization of manual relevance scores based on 1800 retrospective queries. If the upward adjustment is too high, the dynamic main category will overly cover the static management specifications; if the downward adjustment is too high, it will push back the truly key but lagging files. Therefore, this embodiment selects a parameter combination that balances retrieval advancement capability and management stability. Offline replay results show that, under the test conditions of 10 days and a total of 800 retrospective queries, compared with the control scheme of static tag sorting plus full-text search, this embodiment moved the average ranking position of key files from 6.8th to 2.9th, increased the top 5 hit rate from 52.0% to 84.0%, and increased the satisfaction rate of manual review from 76.0% to 91.0%.

[0172] Subsequently, based on the above calculation and processing results, the processor continues to execute the subsequent steps as described in Example 1, completing operations such as writing the result cache, generating the result page, and outputting the risk chain joint query when the joint query conditions are met. Its specific execution logic is the same as that in Example 1, and will not be repeated here.

[0173] Example 4: Offline Training and Online Inference of Dynamic Classification Representation Generation Model

[0174] This embodiment illustrates the modeling implementation of the aforementioned dynamic classification representation. The model employs an offline training and online inference deployment approach. Training samples are derived from historical archives, manually reviewed annotation results, and confirmed risk events from the atlas.

[0175] In this embodiment, the input of the text evidence recognition model is a sequence of text segments with a maximum length of 512 words, and the output is the device code, component name, defect type, operating condition words, and action label; the input of the image evidence recognition model is an archive image with a resolution of 640×640, and the output is the component region box, anomaly label, and confidence score; the input of the dynamic classification representation model is not a general semantic vector, but rather text main class features, image evidence features, graph adjacency features, and static label conflict features, and the output is a 256-dimensional dynamic classification representation.

[0176] During offline training, the training set, validation set, and test set are divided in an 8:1:1 ratio; the optimal batch size is 32; and the optimal initial learning rate is... The optimal number of training epochs is 30. Training should stop when the overall metric on the validation set shows no improvement for five consecutive epochs. Here, the preset batch size refers to the number of samples used in a single parameter update, which can be set based on memory capacity and training stability; a typical value is 32, and it can usually be configured between 16 and 64. The preset initial learning rate controls the parameter update step size and can be determined by the convergence of the validation set; a typical value is... , usually available to The preset number of stops is used to distinguish between short-term verification fluctuations and continuous no-improvement states. The typical value is 5 consecutive rounds, and it can usually be set between 3 and 8 rounds.

[0177] In the dynamic classification representation model, the input layer is organized into four branches: the text evidence branch receives 64 text statistics and fragment support features, the image evidence branch receives 64 region labels and confidence features, the graph adjacency branch receives 64 edge weights, path levels and hop count statistics features, and the rule statistics branch receives 64 main class scores, static labels and conflict marker features. The four branches are linearly transformed to obtain 64-dimensional intermediate representations, which are then concatenated to form a 256-dimensional classification retrieval feature representation, and finally output as the final representation for vector retrieval and main class determination.

[0178] To ensure that the model output is simultaneously subject to class supervision, cross-carrier fact consistency, and graph proximity constraints, the processor employs a joint loss during the offline training phase: ,in, Indicates loss in the main category classification of archives; This represents the alignment loss between textual evidence fragments and image evidence regions; This represents the graph proximity constraint loss; , This represents the loss weight.

[0179] In this embodiment, the preset loss weight is used to balance the impact of alignment constraints and graph constraints on the main class classification task. It can be obtained by jointly optimizing the dynamic main class consistency rate, the top 10 hit rate of risk relevance and training stability on the validation set. Its typical values ​​are 0.35 and 0.25, respectively, and can usually be configured between 0.10 and 0.50 and 0.10 and 0.40.

[0180] Among them, the cross-entropy loss is preferred for the main category classification loss of archives. Its input is the main category prediction distribution of the dynamic classification representation output, and the label is the manually reviewed or confirmed archive category.

[0181] Alignment loss is defined as:

[0182] ,

[0183] in, This represents the set of positive sample pairs, indicating that textual evidence fragments and image evidence regions describe the same equipment defect fact; This represents a set of negative sample pairs, indicating that the two do not describe the same equipment defect. Indicates the number of positive sample pairs; Indicates the number of negative sample pairs; This indicates the negative sample separation interval.

[0184] In this embodiment, the preset negative sample separation interval is used to control the safety separation margin of negative sample pairs in bidirectional alignment scores. It can be obtained by jointly calibrating the alignment accuracy and mismatch suppression rate on manually labeled samples; its typical value is 0.30, and it can usually be set between 0.20 and 0.50. Positive sample pairs are composed of text fragments and image regions within the same file that have been manually verified and confirmed. Negative sample pairs are composed of combinations of different parts within the same file, combinations of different defects within the same equipment, and random sampling combinations across files.

[0185] The graph proximity constraint loss is defined as:

[0186] ,

[0187] in, This represents a sample set of file anchors and risk event anchors with confirmed risk associations; This represents the predicted correlation between the archive anchor point and the risk event anchor point obtained from the output of the dynamic classification representation; Indicates the number of samples; This represents a set of weakly associated samples that have a graph path but a large number of hops. Indicates the number of weakly associated samples; Indicates the shortest number of hops between nodes; This represents the penalty coefficient for the number of jumps.

[0188] To reflect that weakly associated paths should be subject to additional suppression, this step introduces a preset hop count penalty coefficient, which represents the additional constraint strength of weakly associated samples with long hop counts in the graph. It can be jointly calibrated offline by the risk recall rate and false association rate on the validation set. Its typical value is 0.10, and it can usually be configured between 0.05 and 0.20.

[0189] When the file contains only text and no images, the alignment loss is set to 0; when the file has no adjacent risk events in the graph, the graph proximity constraint loss is set to 0, and only the classification loss is retained for training. After this joint training, the dynamic classification representation can more stably reflect the three types of information—category supervision, text-image fact consistency, and graph proximity—during the online inference stage.

[0190] Example 5: An Energy Archive Classification and Retrieval System Integrating Risk Early Warning Maps

[0191] like Figure 2 As shown, when implemented in a system form, the system in this embodiment includes a map mapping module, an alignment representation module, a hierarchical recall module, a conflict correction and sorting module, and an output query module.

[0192] The system comprises the following modules: a graph mapping module for reading energy archives and historical risk events, constructing a unified node relationship table corresponding to the risk warning graph, and calculating the graph node similarity between archive anchors and risk event anchors based on the unified node relationship table; an alignment representation module for performing bidirectional alignment of text evidence fragments and image evidence regions in energy archives based on archive anchors, generating dynamic classification representations based on the graph node similarity and bidirectional alignment results, and forming a classification candidate set by combining static original labels; a hierarchical recall module for receiving query statements, performing keyword recall, full-text recall, and vector recall based on the dynamic classification representations to obtain an initial candidate archive set, and calculating the risk perception candidate score of the initial candidate archive set by combining the risk event anchor set corresponding to the query and the graph node similarity; a conflict correction and sorting module for performing classification conflict correction and hierarchical sorting on the initial candidate archive set based on the conflict relationship between static original labels and the dynamic classification representations to obtain classification retrieval results; and an output query module for writing the classification retrieval results into the result cache and generating a result page, and expanding the query results of associated risk events and associated archives when the sorting results meet the warning query conditions.

[0193] The data flow between modules is as follows: the graph mapping module outputs a unified node relationship table, evidence edge weights, and graph node similarity index to the alignment representation module and the hierarchical recall module; the alignment representation module outputs a dynamic classification representation and a classification candidate set to the hierarchical recall module and the conflict correction and ranking module; the hierarchical recall module outputs risk-aware candidate scores to the conflict correction and ranking module; the conflict correction and ranking module outputs the final ranking results to the output query module; and the output query module finally outputs the classification retrieval result page and the query result page when the conditions are met. This system thus implements all the functions of the aforementioned method steps.

[0194] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for classifying and retrieving energy archives by integrating risk early warning maps, characterized in that, include: Step S1: Read energy archives and historical risk events, construct a unified node relationship table corresponding to the risk warning map, and calculate the map node similarity between archive anchor points and risk event anchor points based on the unified node relationship table; Step S2: Perform bidirectional alignment on text evidence fragments and image evidence regions in the energy archives based on the archive anchor points, and generate dynamic classification representations based on the similarity of the map nodes and the bidirectional alignment results, and form a classification candidate set by combining the static original labels; Step S3: Receive the query statement, perform keyword recall, full-text recall and vector recall based on the classification candidate set to obtain an initial candidate file set, and calculate the risk perception candidate score of the initial candidate file set by combining the risk event anchor point set corresponding to the query and the similarity of the graph nodes. Step S4: Based on the conflict relationship between the static original labels and the dynamic classification representation, perform classification conflict correction and hierarchical sorting on the initial candidate file set to obtain the classification retrieval results; Step S5: Write the classification search results into the result cache area and generate a result page. When the sorting results meet the early warning joint investigation conditions, expand the joint investigation results of related risk events and related files.

2. The energy archive classification and retrieval method integrating risk early warning maps according to claim 1, characterized in that, The unified node relationship table corresponding to the risk warning map includes: Read the structured fields, text, and image attachments of the energy archive, extract the equipment code, component name, defect type, operating condition words, handling action, and time stamp from the text, and extract the component area label, abnormal appearance label, and instrument value label from the image attachments to obtain the archive anchor field; Read the event number, path level, responsible unit, and event time information from historical risk events to obtain the risk event anchor point field; Generate an archive anchor key based on the archive anchor key field, generate a risk event anchor key based on the risk event anchor key field, and write the archive anchor key and the risk event anchor key into the unified node relationship table; Obtain the pre-stored device alias mapping table. When the text lacks a standard device code but has a device alias, restore the standard device code according to the device alias mapping table, and organize the node primary key field, node type field, standard device code field, component name field, defect type field, time slice field, path level field, edge type field, and edge weight field into the unified node relationship table.

3. The energy archive classification and retrieval method integrating risk early warning maps according to claim 2, characterized in that, The calculation of graph node similarity between archive anchors and risk event anchors based on the unified node relationship table includes: Based on the compositional relationship between equipment and components, the state relationship between components and defects, the handling relationship between defects and handling actions, and the evidentiary relationship between file anchor points and risk event anchor points, node connections are established to obtain the type edges and evidence edges in the risk warning map; For the aforementioned evidence edges, the edge weights are calculated based on entity consistency, temporal proximity, and consistency of defective facts, and evidence edges with high edge weights are selected according to preset retention rules. Entity consistency scores are generated based on the consistency of equipment codes, component names, and defect types, respectively. Time proximity scores are generated based on the time proximity of archive time slices and event time slices. Path adjacency scores are generated based on the path hierarchy proximity and graph hop count. The entity consistency score, temporal proximity score, and path adjacency score are fused according to a preset graph weight to obtain the graph node similarity, and the graph node similarity is written into the graph index.

4. The energy archive classification and retrieval method integrating risk early warning maps according to claim 1, characterized in that, Performing bidirectional alignment of textual and image evidence regions in energy archives based on archival anchors includes: The main text is segmented into text evidence fragments, and each text evidence fragment is labeled with the equipment code, component name, defect type, operating condition word and handling action to obtain a set of text evidence fragments; Based on image attachments, perform region detection or region proposal to extract image evidence regions containing component regions, anomaly labels, and confidence levels to obtain an image evidence region set; based on the archive anchor key, filter fragments and regions related to the same equipment defect facts from the text evidence fragment set and the image evidence region set; Obtain a pre-stored business mapping table, and calculate a bidirectional alignment score between text evidence fragments and image evidence regions based on the degree of matching of anchor field, the degree of consistency of business context, and the joint reliability of image region credibility and text integrity. The text integrity is determined by the completeness of identification of device code, component name, defect type, and handling action. When the energy archive does not contain image attachments, the text evidence fragments will be used as unilateral evidence in the subsequent dynamic classification and representation generation. Based on the bidirectional alignment results, the edge weights of the evidence are updated by writing back.

5. The energy archive classification and retrieval method integrating risk early warning maps according to claim 4, characterized in that, The dynamic classification representation generated based on the graph node similarity and bidirectional alignment results includes: A text main category score is generated based on the text evidence fragments; an equipment main category score is generated based on the equipment professional category rules; and a defect main category score is generated based on the defect classification rules. Based on the set of risk event anchor points related to the current archive's specialty, the similarity of the graph nodes and the bidirectional alignment score are fused to generate a risk association main category score. Based on the main text score, main device score, defect score, and risk association score, a dynamic classification representation for vector retrieval and main class determination is constructed. The dynamic main class and static original label in the dynamic classification representation are jointly retained, and a conflict flag is recorded when the two are inconsistent, thus obtaining the classification candidate set; The dynamic classification representation includes textual evidence features, image evidence features, graph adjacency features, and classification rule statistical features.

6. The energy archive classification and retrieval method integrating risk early warning maps according to claim 1, characterized in that, Performing keyword recall, full-text recall, and vector recall based on the dynamic classification representation based on the aforementioned candidate classification set includes: Parse the query statement to obtain the keyword set, device code set, time range, file category limiting words and risk limiting words, and generate a search constraint table; Obtain a pre-stored category mapping table, determine the query limitation category based on the category mapping table, and locate the corresponding risk event anchor point set when the risk limitation term exists; Combining the time range and file category limiting words, keyword matching and recall are performed based on the retrieval constraint table to obtain a keyword candidate set; Based on the retrieval constraint table, a full-text index recall is performed to obtain a full-text candidate set; Based on the dynamic classification representation, vector retrieval and recall are performed to obtain a vector candidate set; Perform a union operation on the keyword candidate set, the full-text candidate set, and the vector candidate set to obtain the initial candidate file set.

7. The energy archive classification and retrieval method integrating risk early warning maps according to claim 6, characterized in that, Calculating the risk perception candidate score of the initial candidate file set includes: For each candidate file in the initial candidate file set, the keyword matching score, full-text search relevance score, and vector search score are calculated respectively. Based on the similarity of graph nodes between the candidate archive anchor points and the set of risk event anchor points corresponding to the query, and combined with the risk association main class score, a risk graph perception score is generated. The keyword matching score, full-text search relevance score, vector search score, and risk graph perception score are fused according to a preset fusion weight to obtain the risk perception candidate score. When no risk qualifiers are parsed during the query, the risk-related main category score is used as the risk graph perception score.

8. The energy archive classification and retrieval method integrating risk early warning maps according to claim 7, characterized in that, Performing classification conflict correction and hierarchical sorting on the initial candidate archive set includes: Check the conflict markers between the static original tags and dynamic main classes of each candidate file, and determine the classification conflict relationship by combining the query-defined categories; When the dynamic main class is consistent with the query-defined category but inconsistent with the static original label, the corresponding candidate file is sorted upwards; when the static original label is consistent with the query-defined category but inconsistent with the dynamic main class, the corresponding candidate file is sorted downwards, thus obtaining the classification conflict correction coefficient. Based on the edge weight of the evidence edge between the candidate file anchor point and the query-related risk event anchor point, the maximum edge weight is extracted as the warning intensity. The final ranking score is generated based on the risk perception candidate score, classification conflict correction coefficient, and warning intensity. The initial candidate file set is then hierarchically sorted according to the final ranking score to obtain the classification retrieval result. When the query does not contain category qualifiers, the category conflict correction coefficient is set to a preset baseline value.

9. The energy archive classification and retrieval method integrating risk early warning maps according to claim 8, characterized in that, The classification and retrieval results are written to the results cache and a results page is generated. When the sorting results meet the early warning joint investigation conditions, the joint investigation results of related risk events and related files are expanded, including: The sorted candidate files are written to the results cache, and the results pages are organized according to file main class, device object, risk event, and time range; Each result is simultaneously output with the file number, static original label, dynamic main category, associated risk event, evidence fragment summary, and evidence image region number; Determine whether the warning intensity of the first file on the results page reaches the preset warning threshold, whether there are related risk event anchor points, and whether there are evidence edges; When the aforementioned early warning joint investigation conditions are met, the graph database interface is called to query the adjacent nodes of the first and second file anchor points, extract the direct risk events, upstream handling files, and downstream review files, generate the joint investigation results, and output them together with the results page.

10. An energy archive classification and retrieval system integrating risk early warning maps, characterized in that, include: The graph mapping module is used to read energy archives and historical risk events, construct a unified node relationship table corresponding to the risk warning graph, and calculate the graph node similarity between archive anchor points and risk event anchor points based on the unified node relationship table. The alignment representation module is used to perform bidirectional alignment of text evidence fragments and image evidence regions in energy archives based on archive anchor points, and generate dynamic classification representations based on the similarity of the map nodes and the bidirectional alignment results, and form a classification candidate set by combining the static original labels; The hierarchical recall module is used to receive query statements, perform keyword recall, full-text recall and vector recall based on the classification candidate set to obtain an initial candidate file set, and calculate the risk perception candidate score of the initial candidate file set by combining the risk event anchor point set corresponding to the query and the similarity of the graph nodes. The conflict correction and sorting module is used to perform classification conflict correction and hierarchical sorting on the initial candidate file set based on the conflict relationship between the static original labels and the dynamic classification representation, so as to obtain the classification retrieval results. The output query module is used to write the classification search results into the result cache and generate a result page. When the sorting results meet the early warning query conditions, the query results of related risk events and related files are expanded. Each module is used to implement the steps of the energy archive classification and retrieval method that integrates risk warning maps as described in any one of claims 1 to 9.

Images