A knowledge graph-based power service semantic recognition method and system

Abstract

The application provides a power service semantic recognition method and system based on a knowledge graph, the method comprising: collecting a plurality of groups of safety supervision field image data and text data, and performing cross-modal fusion; designing a graph node according to a multi-modal fusion vector, and obtaining a graph G=(E, R); according to the graph G=(E, R), combining a risk sensitivity coefficient, and reasoning out a potential risk factor set in a power safety supervision scene, and screening out a risk factor set in the graph G=(E, R); according to the risk factor set, generating a cause chain with an explicit cause-consequence relationship in the graph G=(E, R), and reasoning out a risk level of the whole operation through the cause chain; and based on the cause chain and the risk level, dynamically optimizing the structure and parameters of the graph G=(E, R). The application provides an innovative solution for realizing intelligent and interpretable safety risk identification and evaluation in a power operation scene.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of knowledge graph, and particularly relates to a power service semantic recognition method and system based on a knowledge graph. BACKGROUND

[0002] In the power industry, the safety risk control of the work site is always the focus and difficulty in production management. In order to cope with the complex and changeable work environment, the behavior of the work personnel and the use specification of the equipment in the power construction process, the existing technology generally uses computer vision (CV) technology to analyze the video images of the work site, and uses natural language processing (NLP) technology to analyze the text data such as work tickets and regulations. Some systems have tried to apply CV models and NLP models to the electric power safety supervision scene respectively, to identify the illegal behaviors of the site personnel, the appliances and the work environment through image recognition, and to understand the work plan and the safety regulations through text analysis, so as to realize the automatic risk identification.

[0003] However, the existing technology still has significant deficiencies: first, the CV and NLP models are usually operated separately in the application, and there is a lack of effective semantic fusion between the image model and the text model, it is difficult to establish a cross-modal causal relationship at the business level, and the image recognition result of the work site cannot be deeply matched and explained with the task content, risk control measures and other elements in the work ticket. Secondly, the existing system generally relies on preset rules or manually set risk scoring models, lacks automatic reasoning and explainable ability for risk causes, and is difficult to support safety supervision personnel to quickly and accurately identify potential systemic risks on site. For example, the system may only be able to find that “the personnel do not wear safety helmets”, but cannot combine the “high-altitude work” item in the work ticket to automatically reason out “high falling risk” and “risk cause chain”, and cannot judge the overall risk level in combination with multiple risk factors. In addition, the work environment, the behavior of the work personnel and the ticket information in the field of electric power safety supervision present the characteristics of diversification and dynamics, and the existing method is difficult to construct a complete knowledge system covering multiple scenes and multiple risk types, which limits the generalization ability and intelligent level of the model in the complex work environment. In summary, the existing technology has serious deficiencies in “multimodal data fusion, work scene business understanding, risk causal reasoning and explanation”, and it is difficult to meet the needs of the power work site for an intelligent safety supervision system with high precision, strong business semantic association and high explainability. SUMMARY

[0004] The purpose of the present application is to design a power service semantic recognition method and system based on a knowledge graph, to solve the problems of “CV and NLP model separation, insufficient semantic fusion, and missing risk cause chain” in the prior art.

[0005] In order to achieve the above object, the application provides a power service semantic recognition method based on a knowledge graph, which comprises the following steps:

[0006] Collecting a plurality of groups of safety supervision field image data and text data, introducing a risk sensitivity coefficient, and performing cross-modal fusion to output a multi-modal fusion vector; wherein the safety supervision field image data adopts a target detection model to extract specific image entities in a power safety supervision scene to obtain an image entity set E cv ={e i ,...,e n}, each e i is a safety supervision entity in an image; the text data adopts an NLP model to extract a work field in a bill to generate a text entity set E nlp ={e n+1 ,...,e m}; the risk sensitivity coefficient of each entity is obtained by training a power safety supervision historical accident library;

[0007] According to the multi-modal fusion vector, a graph node is designed, and five types of fixed relationship edges are designed to comprehensively cover cross-modal and rule relationships, form a complete safety supervision business link, and design a graph edge weight by using a fixed weight strategy to obtain a graph G=(E, R), wherein the nodes E and the edges R are all obtained from the entities extracted from the multi-modal fusion vector and the rule mapping of the knowledge base;

[0008] According to the graph G=(E, R), in combination with the risk sensitivity coefficient, a set of potential risk factors in the power safety supervision scene is inferred, and a risk factor set is screened out in the graph G=(E, R);

[0009] According to the risk factor set, a cause chain with an explicit cause-effect relationship is generated in the graph G=(E, R), and the risk level of the whole work is inferred through the cause chain;

[0010] Based on the cause chain and the risk level, the structure and parameters of the graph G=(E, R) are dynamically optimized to generate an adaptive evolution graph G'.

[0011] Preferably, the graph node comprises a safety supervision entity in an image, a business field in a work bill, and a domain knowledge node;

[0012] The five types of fixed relationship edges comprise:

[0013] An image entity-text entity relationship;

[0014] A spatial and work scene co-occurrence relationship between image entities;

[0015] A logical relationship between work bill fields;

[0016] The bidirectional constraint relationship between the image / text entity and the rule node.

[0017] Preferably, the graph G=(E, R) will prioritize the high-risk entity pair to build edges, to ensure the high-risk path structure of G is complete;

[0018] The fixed weight strategy sets the edge weight as a static parameter.

[0019] Preferably, the potential risk factor set in the power safety monitoring scenario is inferred according to the graph G=(E, R) combined with the risk sensitivity coefficient, and the risk factor set is screened out in the graph G=(E, R), specifically:

[0020] Based on the risk sensitivity coefficient, risk-sensitive feature propagation is performed in the graph G=(E, R) to obtain node features;

[0021] After completing the L-layer feature propagation, the risk factor set is screened out by using the risk discrimination function.

[0022] Preferably, the screening formula is:

[0023]

[0024] Wherein, E risk is a risk factor set; θ is a risk coefficient screening threshold; ρ is a feature intensity threshold;

[0025] Wherein, the node e i that meets the dual requirements of risk sensitivity and feature intensity is identified as a risk factor.

[0026] Preferably, the cause chain of explicit cause-effect relationship is generated in the graph G=(E, R) according to the risk factor set, specifically:

[0027] Based on the risk type, risk sensitivity coefficient and domain knowledge node of the node in the risk factor set, a directed cause relationship edge set R c is constructed to form a cause chain;

[0028] All edges are screened and strongly associated using a preset threshold of edge weight, to form a risk cause path.

[0029] Preferably, the cause chain further includes:

[0030] The direction of the edge represents the cause-result relationship, and the edge weight c ij characterizes the cause intensity, which is calculated as

[0031] c ij =λ i ·w ij +δ ij

[0032] where λ i is the risk sensitivity coefficient of node e i ; w ij is the weight of the causal rule defined according to the safety supervision domain knowledge base; δ ij ∈{0,0.2}: if node e i is in the same business scenario as node e j , then δ ij =0.2, which enhances the consistency of the causal relationship within the scenario.

[0033] wherein the risk level of the overall operation is inferred through the cause chain, and specifically:

[0034] According to the edge weight on the cause path in the cause chain, the overall link risk aggregation value is calculated, and the risk level is determined: general risk, greater risk and major risk.

[0035] Preferably, based on the cause chain and the risk level, the structure and parameters of the graph G=(E,R) are dynamically optimized to generate an adaptive evolution graph G', and specifically:

[0036] For all edges in the graph associated with nodes and risk factors in the cause chain, the cause chain and the risk level are analyzed through the indicator function, and the indicator function is used to determine whether the node is in the risk set, and the edge weight is dynamically adjusted;

[0037] For potential cause relationships in the cause chain that are not shown in the graph, implicit risk edges across modalities or across levels are supplemented;

[0038] wherein the indicator function is used to determine whether the node is in the risk set, and the edge weight is dynamically adjusted, including:

[0039] If node e i or e j belongs to the risk factor set, then the edge weight is amplified to strengthen the effect of the risk-related connection;

[0040] If the condition is not met, the edge weight remains unchanged.

[0041] Preferably, for potential cause relationships in the cause chain that are not shown in the graph, implicit risk edges across modalities or across levels are supplemented, and the specific rules are:

[0042] If there is a node e p →e q in the cause chain and the graph G has no edge (e p , e q ) or edge weight r pq <∈, then a new edge is supplemented or the edge weight is reset:

[0043] r pq ′=β·cpq + μ·η pq

[0044] wherein β is a causal chain supplement factor; c pq is a causal edge strength in a causal chain; η pq is a cross-modal consistency factor, if e p and e q are different modalities, then η pq = 0.3, otherwise 0; μ is a cross-modal supplement gain factor.

[0045] In a second aspect of the present application, a power service semantic recognition system based on a knowledge graph is provided, and the system comprises:

[0046] a data acquisition unit, configured to acquire a plurality of groups of safety supervision field image data and text data, introduce a risk sensitivity coefficient, perform cross-modal fusion, and output a multi-modal fusion vector; wherein the safety supervision field image data adopts a target detection model to extract specific image entities in a power safety supervision scene, to obtain an image entity set E cv ={e1, e2,..,e i .,e n}, each e i is a safety supervision entity in an image; the text data adopts an NLP model to extract a work field in a bill, to generate a text entity set E nlp ={e n+1 ,...,e m}; the risk sensitivity coefficient of each entity is obtained by training a power safety supervision historical accident library;

[0047] a graph construction unit, configured to design graph nodes according to the multi-modal fusion vector, and design five types of fixed relationship edges, comprehensively cover cross-modal and rule relationships, form complete safety supervision business links, design graph edge weights by using a fixed weight strategy, and obtain a graph G=(E, R), wherein the nodes E and the edges R are all obtained from entities extracted by the multi-modal fusion vector and knowledge base rule mapping;

[0048] a risk analysis unit, configured to infer a potential risk factor set in a power safety supervision scene according to the graph G=(E, R) in combination with the risk sensitivity coefficient, and screen out the risk factor set in the graph G=(E, R);

[0049] a risk reasoning unit, configured to generate a causal chain with an explicit cause-effect relationship in the graph G=(E, R) according to the risk factor set, and infer a risk level of an overall work through the causal chain;

[0050] A knowledge graph updating unit is configured to perform dynamic optimization of the structure and parameters of the graph G=(E, R) based on the cause chain and the risk level, and generate an adaptive evolution graph G'.

[0051] The present application has at least the following advantages:

[0052] The present application designs a cross-modal data fusion mechanism, for the first time fuses the job site image recognition result and the job ticket text information in a unified knowledge graph structure, breaks through the semantic association link of image and text in the business scene, constructs a multi-modal knowledge system for power safety supervision, and improves the comprehensive understanding ability of the model to the complex job environment. On this basis, the system further combines the risk cause reasoning mechanism, can realize the automatic reasoning and interpretable output of the risk cause relationship based on the associated multi-modal entities, business scene elements and safety regulation rules in the knowledge graph, and comprehensively improves the scientificity, rationality and automation level of risk assessment. The present application breaks through the technical bottlenecks such as insufficient multi-modal data fusion, inability to model business logic, and lack of causal support for risk inference in the prior art, and provides an innovative solution for realizing intelligent and interpretable safety risk identification and evaluation in the power operation scene. BRIEF DESCRIPTION OF DRAWINGS

[0053] The present application will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present application. For those skilled in the art, other drawings can be obtained without creative labor on the premise that the following drawings are not limited.

[0054] Figure 1 A knowledge graph-based power service semantic identification method flowchart is provided for an embodiment of the present application.

[0055] Figure 2 A knowledge graph-based power service semantic identification system framework diagram is provided for an embodiment of the present application. DETAILED DESCRIPTION

[0056] The embodiments of the present application will be described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present application, and cannot be understood as a limitation on the present application.

[0057] In one or more embodiments, as Figure 1 shown, a knowledge graph-based power service semantic identification method is disclosed, which comprises steps S1-S5, comprising:

[0058] S1, collect a plurality of groups of safety supervision field image data and text data, and introduce a risk sensitivity coefficient, and perform cross-modal fusion to output a multi-modal fusion vector; wherein the safety supervision field image data extracts specific image entities in the power safety supervision scene by using a target detection model to obtain an image entity set E cv ={e1, e2,.., e i ·, e n} i , each e nlp is a safety supervision entity in an image; the text data extracts the work fields in the bill by using an NLP model to generate a text entity set E n+1 ={e m ,..., e cv} ; the risk sensitivity coefficient of each entity is obtained by training a power safety supervision historical accident library.

[0059] Specifically, the following is collected:

[0060] D nlp : safety supervision field image data, including personnel, equipment, environment, safety supervision protection facilities, etc.

[0061] D m : text data such as work tickets and safety regulations, including work types, measures, risk levels, etc.

[0062] According to the data characteristics of the power safety supervision scene, a “multi-modal safety supervision feature fusion device (MAFE)” is proposed, which is used to uniformly fuse safety supervision risk entities in image and text data, and output a cross-modal feature vector V cv .

[0063] Image data processing path: D cv input a safety supervision special target detection model f cv , extract specific image entities in the power safety supervision scene to obtain an image entity set E n ={e1, e2,.., e i}; each e i is a safety supervision entity in an image (such as e1 = “personnel without safety belt” and e2 = “no fence”). Each image entity e i is mapped to a d-dimensional vector

[0064] Text data processing path: D nlp input a work ticket special NLP model f nlp , extract the work fields in the bill to generate a text entity set E nlp ={e n+1 ,..., e m}, such as e n+1= "high-altitude operation", e n+2 = "ticket without warning measures". Each text entity e j j is mapped to a d-dimensional vector

[0065] Further, the "risk sensitivity coefficient" λ is introduced in the power safety scene. The risk sensitivity coefficient λ i of each entity is j trained by the power safety historical accident library, specifically:

[0066] The high-risk target (such as "unfasten safety belt") is assigned λ i = 0.9;

[0067] The low-risk target (such as "not wearing reflective clothing") is assigned λ i = 0.5;

[0068] Similarly, the λ j of the ticket entity e j is assigned according to the "operation type risk level", such as "high-altitude operation" λ j = 0.8.

[0069] Finally, the "risk sensitive weighted fusion mechanism" is proposed to output the cross-modal embedding V m , the formula is as follows:

[0070]

[0071] Where n is the number of image entities, m-n is the number of text entities; φ(e i ) is the image feature embedding of e i ; ψ(e j ) is the text feature embedding of e j ; λ i is the risk sensitivity coefficient of the image entity e i ; λ j is the risk sensitivity coefficient of the text entity e j .

[0072] This mechanism ensures that high-risk targets in V m have greater influence in the embedding space, improving the attention to "high-risk entity nodes" in the subsequent graph modeling stage. Output: V m , as the only input for the subsequent graph construction step.

[0073] ​S2, design the graph node according to the multi-modal fusion vector, and design five types of fixed relationship edges, comprehensively cover cross-modal and rule relationship, form a complete safety supervision business link, adopt a fixed weight strategy to design the graph edge weight, obtain the graph G=(E, R), wherein the node E and the edge R are all from the entity and knowledge base rule mapping extracted by the multi-modal fusion vector.

[0074] Specifically, the input is the multi-modal fusion vector V output by step S1 m , wherein The unified representation of image and text information in the power safety supervision scene has been completed. The goal of this step is to construct a cross-modal safety supervision knowledge graph G based on V m , which provides a structured graph input for the subsequent reasoning module (step 3). The scheme is strictly limited to the "graph building" task and does not involve subsequent functions such as reasoning and regularization optimization.

[0075] The present application proposes a "power safety supervision multi-modal relationship graph modeling mechanism (EARM-Graph)", which is suitable for the typical "image-job ticket-safety regulation knowledge" scene of the Southern Power Grid, and completes the structured construction of nodes and relationships.

[0076] Graph node design: node set E=E cv ∪E nlp ∪E rule , including three types:

[0077] E cv : safety supervision entities in images, such as "no safety helmet" and "no fence in construction area";

[0078] E nlp : business fields in job tickets, such as "high-altitude work" and "no warning measures";

[0079] E rule : domain knowledge nodes, such as "Article 45 of the Safety Regulations" and "Article 12 of the Safety Regulations", all of which are from the Southern Power Grid safety supervision system knowledge base.

[0080] Graph relationship design: five types of fixed relationship edges are designed to comprehensively cover cross-modal and rule relationships and form a complete safety supervision business link:

[0081] R cv-nlp : image entity-text entity relationship (such as "no safety belt" <-> "high-altitude work");

[0082] R cv-cv : spatial and work scene co-occurrence relationship between image entities (such as "personnel without safety belt" <-> "no fence");

[0083] R nlp-nlp : logical relationship between job ticket fields (such as "high-altitude work" <-> "need safety protection");

[0084] R cv-role With R nlp-role : The two-way constraint relationship between image / text entities and rule nodes, such as "not wearing a seat belt" → "Safety Regulation 45".

[0085] Furthermore, the risk sensitivity coefficient λ from step S1 is retained. i and λ j This serves as auxiliary information for graph construction. It does not directly affect the edge weights of the graph, but it is used to control the priority order of edge construction.

[0086] High-risk entities pair (e p ,e q (e.g., λ) p +λ q >1.5) Edges will be built first to ensure the integrity of G's "high-risk path" structure.

[0087] Graph edge weight design: This stage only uses a fixed weight strategy and does not perform adaptive learning.

[0088] r cv-nlp =1, r cv-cv =1, r nlp-nlp =1;

[0089] r cv-rule =r nlp-rule =1.2, used to highlight the relative importance of "safety regulations".

[0090] The edge weights are set as static parameters, and the GNN inference in step 3 is responsible for "information propagation and weight optimization".

[0091] Mapping formula: Define the final map G = (E, R) as:

[0092] G = {E, R = R} cv-nlp ∪R cv-cv ∪R nlp-nlp ∪R cv-rule ∪R nlp-rule}(2)

[0093] Both E and R come from V. m The extracted entities are mapped to knowledge base rules.

[0094] The completed cross-modal safety supervision knowledge graph G will be used as input for the next step, GNN, to infer cross-modal risk factors.

[0095] S3. Based on the graph G=(E,R) and combined with the risk sensitivity coefficient, deduce the set of potential risk factors in the power safety supervision scenario, and select the set of risk factors from the graph G=(E,R).

[0096] Specifically, the input is the cross-modal knowledge graph G = (E, R) output by step 2, the graph is composed of image entities E cv , text entities E nlp and rule entities E rule , the nodes carry risk sensitivity coefficients λ i from step 1, and the graph relationship R contains structured edge information between multi-modal entities and rules. The goal of this step is to infer the potential risk factor set E risk in the power safety monitoring scene on G, and complete the single-point identification from the graph G to the risk factor as the input of the subsequent cause chain generation. For this purpose, the present application proposes an "electric power safety monitoring cross-modal graph factor inference network (E-RiskNet)", which designs a "risk-sensitive feature propagation mechanism" to highlight the feature enhancement of high-risk nodes in the power safety monitoring graph, helping to infer potential high-risk factors.

[0097] Further, the risk-sensitive feature propagation mechanism: for each node e i in G, combined with its initial risk sensitivity coefficient λ i , perform risk-enhanced feature propagation within the graph structure, the specific formula is as follows:

[0098]

[0099] Where, is the feature representation of node e i at layer l; is the neighbor node set of e i ; r ij is the static edge weight in step 2; λ i is the risk sensitivity factor of e i , which comes from step 1; γ is the risk enhancement factor (for example, γ = 1.0); σ(·) is the activation function (such as ReLU); W (l) is a learnable parameter matrix.

[0100] Unlike ordinary GNNs, this mechanism allows high-risk nodes with high λ i to have stronger propagation capabilities during feature propagation; this can improve the information aggregation of high-risk nodes such as "not wearing safety belts + high-altitude operations" in the graph, and solve the problem of dilution of high-risk paths in the power safety monitoring graph.

[0101] Risk factor identification mechanism: after completing the L-layer feature propagation, directly calculate the risk intensity of the final node feature i of all e ; through the risk discrimination function, filter the risk factor set E risk , the filtering formula is as follows:

[0102]

[0103] Where θ is the risk coefficient screening threshold (e.g., θ = 0.7); ρ is the feature strength threshold (e.g., ρ = 0.8); and the node e that satisfies both risk sensitivity and feature strength requirements... i It was identified as a risk factor.

[0104] Traditional graph inference typically selects nodes solely based on feature propagation strength; this scheme additionally incorporates a "risk-sensitive weight" λ. i This establishes a risk-priority factor reasoning mechanism, which is more in line with the actual business characteristics of power safety supervision scenarios; it ensures the priority of "high-risk nodes" in information dissemination and factor discrimination, and improves the pertinence of cross-modal graph reasoning.

[0105] S4. Based on the set of risk factors, generate a causal chain with a clear cause-effect relationship in the graph G=(E,R), and infer the risk level of the overall operation through the causal chain.

[0106] Specifically, the input is the set of high-risk factors E output from step S3. risk Each node e in this set i All data originate from graph G and have a known risk sensitivity coefficient λ. i The types include image entities (such as "not wearing a safety belt"), work order fields (such as "lacking height protection"), and safety regulation entities (such as "Safety Regulation Article 45"). No causal relationships have yet been established between risk factors in this set. The goal of this step is: based on E... risk This method generates a risk causal chain C with a clear cause-effect structure, and infers the overall operational risk level L from the causal chain. This causal chain must reflect the risk evolution logic in multimodal scenarios of power safety supervision, such as the causal path between image-based behavioral violations, work permit defects, and safety regulations. To achieve this goal, the "Southern Power Grid Safety Supervision Multimodal Causal Chain Inferencer (NAM-Causal)" is proposed, integrating knowledge from the power safety supervision field to construct a cross-modal risk causal chain. This mechanism includes two core steps: causal chain construction and risk level quantification.

[0107] Cause chain construction mechanism: based on E risk Risk type (image / text / rule) and risk level λ of the middle node i And domain knowledge rules, construct a directed causal relation edge set R c , forming the causal chain C = (E risk ,R c The direction of an edge represents the cause-effect relationship, and the edge weight c ij Characterizing causal strength. The edge weight of causal strength is defined as:

[0108] c ij =λi ·w ij +δ ij (5)

[0109] where λ i : risk sensitivity coefficient of node e i ; w ij : weight of causal rule defined according to safety domain knowledge base, such as w ij = 0.9 of "no protection → falling risk"; δ ij ∈{0,0.2}: if e i and e j are in the same business scenario (such as belonging to "high-altitude operation"), then δ ij = 0.2, to enhance the consistency of intra-scenario causality. This mechanism ensures that high-risk nodes are prioritized to form cause chains, and intra-scenario factors are easy to form strong cause paths, meeting the aggregation characteristics of power safety risks.

[0110] Cause chain generation rule: based on all c ij , strong correlation edges e i → e j are screened using threshold τ = 1.0 to form risk cause paths. For example: "operation ticket lacks protective measures (λ = 0.8)" → "personnel does not wear safety belt (λ = 0.9)" → "falling risk", forming a complete cause chain.

[0111] Risk level quantification mechanism: according to the edge weights on the cause path in C, the overall link risk aggregation value S C is calculated, and the risk level L is determined accordingly:

[0112]

[0113] The risk level segmentation standard is set according to the South Grid safety regulations, reflecting the relationship between the aggregation strength of the cause chain and the severity of the operation risk.

[0114] Further, the output of this step is as follows:

[0115] Cause chain C: a directed graph composed of E risk and R c , describing the cause-result path of risk factors;

[0116] Risk level L: used for on-site operation risk classification and subsequent optimization of the graph.

[0117] S5, based on the cause chain and the risk level, dynamically optimizing the structure and parameters of the graph G = (E, R) to generate an adaptive evolutionary graph G'.

[0118] Specifically, the input is the risk cause chain C = (E risk , Rc ) and the corresponding risk level L, and the initial cross-modal knowledge graph G = (E, R) generated in step 2. Where C has described E risk the cause-effect relationship between high-risk factors, L is the risk level of the work site (such as general, greater, major risk), and G is still the static modeled graph structure. The goal of this step is to: based on C and L, dynamically optimize the structure and parameters of G, generate an adaptive evolving graph G', to improve the risk reasoning ability and business sensitivity of G' in the next work scenario, and form a real "business-graph-risk reasoning-graph evolution" closed loop. For this purpose, the "Nanwang safety monitoring causal feedback graph evolution mechanism (NAF-GE)" is proposed, which combines risk factors in the cause chain, cause edge strength and risk level to complete graph structure optimization and parameter adaptive update.

[0119] Mechanism one: risk level guided global edge weight redistribution:

[0120] Through the risk level L to drive the structure adjustment of the whole graph G, the problem that the safety monitoring graph is not sensitive to the risk level and leads to the reasoning not focusing on the high-risk scene is solved.

[0121] First, for all edges associated with the nodes E risk in C, perform "global risk enhancement" uniformly, defined as follows:

[0122]

[0123] Where α is the enhancement base factor of the risk level (such as α = 0.5); γ L is the risk level factor driven by L (such as γ L = 0.5, if L = "greater risk", γ L = 1.0, if L = "major risk"); is the indicator function of the edges connected to the high-risk factor E risk in the graph G.

[0124] We not only enhance the weight of the cause edge in C, but also uniformly enhance the multi-modal relationship edge related to E risk in the whole graph according to L, focusing on the whole graph feature propagation ability of the high-risk scene; combined with the risk level L, realize the dynamic update of the graph "level sensitive" structure, and strengthen the cause chain path under the high-risk level.

[0125] Mechanism two: implicit structure supplement driven by cause chain:

[0126] For the potential cause relationship in C that does not exist explicitly in G, the "cross-modal chain supplement mechanism" is innovatively designed to supplement the implicit risk edges across modalities or across levels. The specific rules are as follows:

[0127] If e exists in C p →e q And G has no (e) p ,e q ) or r pq If the edge weight is less than or equal to ∈ (e.g., ∈ = 0.5), then add a new edge or reset the edge weight:

[0128] r pq ′=β·c pq +μ·η pq (8)

[0129] Where β is the causal chain complement factor (e.g., β = 0.4); c pq η represents the causal edge strength in the causal chain. pq As the cross-modal consistency factor, if e p and e q For different modalities (e.g., image + ticket, ticket + rule), η pq =0.3, otherwise 0; μ is the cross-modal complement gain factor (e.g., μ = 0.2).

[0130] This mechanism is specifically designed for power safety supervision scenarios. It takes into account the problem of missing links in the triple relationship between "image-text-safety regulations" and automatically supplements high-risk cross-modal links. It avoids path breaks caused by the "missing of key cross-modal risk links" in the subsequent reasoning stage and enhances the cross-modal connectivity of the graph.

[0131] In one or more embodiments, such as Figure 2 As shown, this invention discloses a knowledge graph-based semantic recognition system for power services, the system comprising:

[0132] The data acquisition unit 101 is used to acquire several sets of on-site safety supervision image data and text data, introduce risk sensitivity coefficients, perform cross-modal fusion, and output a multimodal fusion vector; among them, the on-site safety supervision image data uses a target detection model to extract specific image entities in the power safety supervision scene, obtaining an image entity set E. cv ={e1,e2,..,e i .,e n}, each e i The text data is represented by a safety supervision entity in an image; the text data is extracted using an NLP model to extract the work field from the ticket, generating a text entity set E. nlp ={e n+1 ,...,e m The risk sensitivity coefficient of each entity is obtained by training the historical accident database of power safety monitoring.

[0133] The graph construction unit 102 is used for designing a graph node according to a multi-modal fusion vector, and designing five types of fixed relationship edges, comprehensively covering cross-modal and rule relationships, forming a complete safety supervision business link, and designing graph edge weights by using a fixed weight strategy, to obtain a graph G=(E, R), wherein the nodes E and the edges R are all obtained from entity extraction and knowledge base rule mapping of the multi-modal fusion vector;

[0134] The risk analysis unit 103 is used for reasoning out a potential risk factor set in the power safety supervision scene according to the graph G=(E, R) and in combination with a risk sensitivity coefficient, and screening out the risk factor set from the graph G=(E, R);

[0135] The risk reasoning unit 104 is used for generating a cause chain with an explicit cause-consequence relationship in the graph G=(E, R) according to the risk factor set, and reasoning out a risk level of the whole operation through the cause chain;

[0136] The knowledge graph updating unit 105 is used for dynamically optimizing the structure and parameters of the graph G=(E, R) based on the cause chain and the risk level, and generating an adaptive evolution graph G'.

[0137] The above-mentioned embodiments are some preferred embodiments of the present application, and of course cannot limit the scope of the present application, and those skilled in the art can understand that all or part of the processes of the above-mentioned embodiments can be implemented, and equivalent changes made according to the claims of the present application still belong to the scope covered by the present application.

Claims

1. A knowledge graph-based semantic recognition method for electricity services, characterized in that, The method includes the following steps: Several sets of on-site safety supervision image data and text data were collected, and a risk sensitivity coefficient was introduced. Cross-modal fusion was then performed to output a multimodal fusion vector. Specifically, the on-site safety supervision image data was used to extract specific image entities in the power safety supervision scene using a target detection model, resulting in an image entity set. each The text data is represented by a safety supervision entity in an image; the text data is extracted using an NLP model to extract the work field from the ticket, generating a set of text entities. The risk sensitivity coefficient of each entity is obtained by training the historical accident database of power safety supervision. Graph nodes are designed based on multimodal fusion vectors, and five types of fixed relationship edges are designed to fully cover cross-modal and rule-based relationships, forming a complete safety supervision business link. A fixed weight strategy is used to design graph edge weights to obtain the graph. , among which, nodes With edge All entities are derived from multimodal fusion vector extraction and their mapping to knowledge base rules; According to the map Based on the risk sensitivity coefficient, a set of potential risk factors in the power safety supervision scenario is deduced, and this is presented in the graph. Select the set of risk factors; Based on the set of risk factors, in the graph A causal chain with a clear cause-effect relationship is generated, and the risk level of the overall operation is inferred from the causal chain; Based on the aforementioned causal chain and risk level, the map... Dynamic optimization of structure and parameters is performed to generate an adaptive evolutionary map. ; Among them, the graph is based on the causal chain and risk level. Dynamic optimization of structure and parameters is performed to generate an adaptive evolutionary map. Specifically, these include: For all edges in the graph that are associated with nodes and risk factors in the causal chain, the causal chain and risk level are analyzed by indicator functions. Based on the indicator functions, it is determined whether a node is in a risk set, and the edge weights are dynamically adjusted. For potential causal relationships in the causal chain that are not explicitly present in the graph, implicit risk edges across modalities or levels are added. The step of determining whether a node is in the risk set based on an indicator function and dynamically adjusting the edge weights includes: If node or If it belongs to a set of risk factors, then the side weights are amplified to strengthen the role of risk-related connections; If the condition is not met, the edge weight remains unchanged; Specifically, for potential causal relationships not explicitly shown in the graph within the causal chain, implicit risk edges across modalities or levels are added, with the following specific rules: If there are nodes in the causal chain And the map Boundless or side rights < Then add a new side. Or reset edge weights: ； in, It is the complementary factor of the causal chain; The strength of the causal edge in the causal chain; As the cross-modal consistency factor, if and If they are different modes, then Otherwise, it is 0; This is the cross-modal complement gain factor.

2. The knowledge graph-based semantic recognition method for power services according to claim 1, characterized in that, The graph nodes include safety supervision entities in the image, business fields in the work order, and domain knowledge nodes; The five types of fixed relationship edges include: Image entity-text entity relationship; Spatial and operational scene co-occurrence relationships between image entities; Logical relationships between fields on the work order; A two-way constraint relationship between image or text entities and rule nodes.

3. The knowledge graph-based semantic recognition method for power services according to claim 2, characterized in that, The map Edges will be prioritized for high-risk entities to ensure the security of the graph. The high-risk path structure is complete; The fixed weight strategy uses edge weights as static parameters.

4. The knowledge graph-based semantic recognition method for power services according to claim 1, characterized in that, According to the map Based on the risk sensitivity coefficient, a set of potential risk factors in the power safety supervision scenario is deduced, and this is presented in the graph. The set of risk factors was selected, specifically including: Based on the risk sensitivity coefficient in the map Propagate risk-sensitive features to obtain node characteristics; After completion After layer feature propagation, the risk factor set is screened using a risk discrimination function.

5. The knowledge graph-based semantic recognition method for power services according to claim 4, characterized in that, The screening formula is: ； in, A set of risk factors; The threshold for risk coefficient screening; The feature intensity threshold; For nodes Risk sensitivity factors For nodes Feature representation at layer l; Among them, nodes that meet both risk sensitivity and feature strength requirements It was identified as a risk factor.

6. The knowledge graph-based semantic recognition method for power services according to claim 2, characterized in that, Based on the set of risk factors, in the graph The formation of causal chains with clear cause-effect relationships includes: Based on the risk type, risk sensitivity coefficient, and domain knowledge nodes of the nodes in the risk factor set, a directed causal relationship edge set is constructed. This forms a causal chain; By using a preset threshold for edge weights, all edges are filtered to identify strong correlations, thus forming risk causal paths.

7. The knowledge graph-based semantic recognition method for power services according to claim 6, characterized in that, The causal chain also includes: The direction of an edge represents the cause-effect relationship, and the edge weight... Characterizing the intensity of the causal process, calculated as ； in, For nodes Risk sensitivity coefficient; The weights of causal rules are defined based on the knowledge base in the field of safety supervision; If node With nodes If they are in the same business scenario, then Enhance causal consistency within the scenario; Specifically, the method of inferring the overall risk level of the operation through the causal chain includes: Based on the edge weights along the causal paths in the causal chain, calculate the overall link risk aggregation value, and determine the risk level accordingly: , and .

8. A system for implementing the knowledge graph-based semantic recognition method for power services as described in claim 1, characterized in that, The system includes: The data acquisition unit is used to collect several sets of on-site safety supervision image data and text data, introduce risk sensitivity coefficients, perform cross-modal fusion, and output a multimodal fusion vector. Specifically, the on-site safety supervision image data uses a target detection model to extract specific image entities in the power safety supervision scene, resulting in an image entity set. each The text data is represented by a safety supervision entity in an image; the text data is extracted using an NLP model to extract the work field from the ticket, generating a set of text entities. The risk sensitivity coefficient of each entity is obtained by training the historical accident database of power safety supervision. The graph construction unit is used to design graph nodes based on multimodal fusion vectors and design five types of fixed relationship edges to fully cover cross-modal and rule-based relationships, forming a complete safety supervision business link. A fixed weight strategy is used to design graph edge weights to obtain the graph. , among which, nodes With edge All entities are derived from multimodal fusion vector extraction and their mapping to knowledge base rules; Risk analysis unit, used to analyze the map Based on the risk sensitivity coefficient, a set of potential risk factors in the power safety supervision scenario is deduced, and this is presented in the graph. Select the set of risk factors; Risk reasoning unit, used to, based on the set of risk factors, in the graph A causal chain with a clear cause-effect relationship is generated, and the risk level of the overall operation is inferred from the causal chain; The knowledge graph update unit is used to update the knowledge graph based on the causal chain and risk level. Dynamic optimization of structure and parameters is performed to generate an adaptive evolutionary map. .

Images