Electric power service semantic recognition method and system based on knowledge graph

Abstract

The invention provides an electric power service semantic recognition method and system based on a knowledge graph, and the method comprises the steps: collecting a plurality of groups of safety supervision scene image data and text data, and carrying out the cross-modal fusion; designing atlas nodes according to the multi-modal fusion vector, and obtaining an atlas G = (E, R); reasoning a potential risk factor set in the electric power safety supervision scene according to the map G = (E, R) in combination with a risk sensitivity coefficient, and screening a risk factor set in the map G = (E, R); according to the risk factor set, a cause chain with a definite cause-consequence relation is generated in the map G = (E, R), and the risk level of the whole operation is deduced through the cause chain; and based on the cause chain and the risk level, carrying out structure and parameter dynamic optimization on the graph G = (E, R). According to the invention, an innovative solution is provided for realizing intelligent and interpretable safety risk identification and evaluation in an electric power operation scene.

Description

Technical Field

[0001] The present invention belongs to the technical field of knowledge graphs, and particularly relates to a method and system for semantic recognition of power services based on knowledge graphs. Background Art

[0002] In the power industry, the safety risk control of the operation site has always been the focus and difficulty in production management. To cope with the complex and changeable operation environment, the behavior of operation personnel, and the equipment usage specifications during the power construction process, the existing technologies generally adopt computer vision (CV) technology to analyze video images of the operation site, and use natural language processing (NLP) technology to parse text data such as operation tickets and rules and regulations. Some systems have tried to apply the CV model and the NLP model to the power safety supervision scenario respectively, identify violations in the on-site personnel, equipment, and operation environment through image recognition, and at the same time use text analysis to assist in understanding the operation plan and safety regulations, to achieve partial automated risk recognition.

[0003] However, the existing technologies still have significant deficiencies: First, the CV and NLP models usually run separately in applications, and there is a lack of effective semantic fusion between the image model and the text model, making it difficult to establish cross-modal causal associations at the business level, resulting in the inability to deeply match and interpret the image recognition results at the operation site with elements such as the task content and risk control measures in the operation ticket. Second, the existing systems generally rely on preset rules or manually set risk scoring models, lacking the automated reasoning and interpretability of risk causes, and it is difficult to support safety supervision personnel to quickly and accurately identify potential systematic risks on site. For example, the system may only be able to detect that "the person is not wearing a safety helmet", but cannot combine the "high-altitude operation" item in the operation ticket to automatically infer that "the risk of falling is high" and the "risk cause chain", let alone make a global risk level judgment by combining multiple risk factors. In addition, the operation environment, the behavior of operation personnel, and the ticket information in the power safety supervision field present the characteristics of diversification and dynamism. The existing methods are difficult to construct a complete knowledge system covering multiple scenarios and multiple risk types, restricting the generalization ability and intelligent level of the model in complex operation environments. In summary, the existing technologies have serious deficiencies in "multi-modal data fusion, business understanding of operation scenarios, risk causal reasoning and interpretation", and it is difficult to meet the requirements of the intelligent safety supervision system for high-precision, strong business semantic association, and high interpretability in the power operation site. Summary of the Invention

[0004] The objective of the present invention is to design a method and system for semantic recognition of power services based on knowledge graphs to solve the problems in the existing technologies such as "separation of CV and NLP models, insufficient semantic fusion, and lack of risk cause chain".

[0005] To achieve the above object, in the first aspect of the present invention, a method for semantic recognition of power services based on a knowledge graph is provided. The method includes the following steps:

[0006] Collect several groups of safety supervision on-site image data and text data, introduce a risk sensitivity coefficient, and perform cross-modal fusion to output a multi-modal fusion vector. Among them, the safety supervision on-site image data uses an object detection model to extract specific image entities in the power safety supervision scenario to obtain an image entity set E cv ={e1, e2,.., e i ,.., e n}}, and each e i is a safety supervision entity in an image; the text data uses an NLP model to extract operation fields in the 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 with a power safety supervision historical accident database;

[0007] Design graph nodes according to the multi-modal fusion vector, and design five types of fixed relationship edges to comprehensively cover cross-modal and rule relationships, forming a complete safety supervision business link. Use a fixed weight strategy to design graph edge weights to obtain a graph G=(E, R), where the nodes E and the edges R all come from the mapping between the entities extracted from the multi-modal fusion vector and the knowledge base rules;

[0008] According to the graph G=(E, R), combined with the risk sensitivity coefficient, infer the set of potential risk factors in the power safety supervision scenario, and screen out the risk factor set in the graph G=(E, R);

[0009] According to the risk factor set, generate a cause chain with a clear cause-consequence relationship in the graph G=(E, R), and infer the risk level of the overall operation through the cause chain;

[0010] Based on the cause chain and the risk level, perform dynamic optimization of the structure and parameters of the graph G=(E, R) to generate an adaptive evolutionary graph G';

[0011] Preferably, the graph nodes include safety supervision entities in the image, business fields in the operation ticket, and domain knowledge nodes;

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

[0013] Image entity - text entity relationship;

[0014] Spatial and operation scenario co-occurrence relationship between image entities;

[0015] Logical relationship between operation ticket fields;

[0016] Bidirectional constraint relationship between image / text entities and rule nodes.

[0017] Preferably, for the graph G=(E, R), edges will be preferentially built for high-risk entity pairs to ensure the integrity of the high-risk path structure of G;

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

[0019] Preferably, according to the graph G=(E, R), combined with the risk sensitivity coefficient, a set of potential risk factors in the power safety supervision scenario is deduced, and the set of risk factors 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, use the risk discrimination function to screen the set of risk factors.

[0022] Preferably, the screening formula is:

[0023]

[0024] where E risk is the set of risk factors; θ is the risk coefficient screening threshold; ρ is the feature strength threshold;

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

[0026] Preferably, according to the set of risk factors, a cause chain with a clear cause-consequence relationship is generated in the graph G=(E, R), specifically:

[0027] Based on the risk types, risk sensitivity coefficients, and domain knowledge nodes of the nodes in the set of risk factors, a directed cause relationship edge set R c is formed to form a cause chain;

[0028] Use the preset threshold of the edge weight to screen all edges for strong associations to form a risk cause path.

[0029] Preferably, the cause chain also includes:

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

[0031] c ij =λ i ·w ij +δ ij

[0032] Among them, λ i is the risk sensitivity coefficient of node e i ; w ij is the causal rule weight defined according to the knowledge base in the field of work safety supervision; δ ij ∈ {0, 0.2}: If node e i and node e j are in the same business scenario, then δ ij = 0.2 to enhance the causal consistency within the scenario.

[0033] Among them, the risk level of the overall operation is inferred through the cause - effect chain, specifically:

[0034] According to the edge weights on the cause - effect path in the cause - effect chain, calculate the overall link risk aggregation value, and determine the risk level based on this: general risk, relatively large risk, and major risk.

[0035] Preferably, based on the cause - effect chain and the risk level, perform dynamic optimization of the structure and parameters of the graph G=(E, R) to generate an adaptive evolutionary graph G′, specifically:

[0036] For all edges in the graph associated with the nodes and risk factors in the cause - effect chain, analyze the cause - effect chain and the risk level through the indicator function, and dynamically adjust the edge weights based on the indicator function to determine whether the node is in the risk set;

[0037] For potential causal relationships in the cause - effect chain that are not shown in the graph, supplement cross - modal or cross - level implicit risk edges;

[0038] Among them, the dynamic adjustment of the edge weights based on the indicator function to determine whether the node is in the risk set includes:

[0039] If node e i or e j belongs to the risk factor set, amplify the edge weight to strengthen the role of risk - related connections;

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

[0041] Preferably, for potential causal relationships in the cause - effect chain that are not shown in the graph, supplement cross - modal or cross - level implicit risk edges, and the specific rules are:

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

[0043] r pq ′ = β·cpq +μ·η pq

[0044] Among them, β is the complementary chain factor of the cause-and-effect chain; c pq is the causal edge strength in the cause-and-effect chain; η pq is the cross-modal consistency factor. If e p and e q are different modalities respectively, then η pq = 0.3, otherwise it is 0; μ is the cross-modal complementary chain gain factor.

[0045] In the second aspect of the present invention, a semantic recognition system for power services based on a knowledge graph is provided. The system includes:

[0046] A data acquisition unit, which is used to collect several groups of safety supervision on-site image data and text data, introduce a risk sensitivity coefficient, and perform cross-modal fusion to output a multi-modal fusion vector. Among them, the safety supervision on-site image data uses an object detection model to extract specific image entities in the power safety supervision scenario to obtain an image entity set E cv ={e1, e2,.., e i .,, e n}, and each e i is a safety supervision entity in an image; the text data uses an NLP model to extract operation fields in the bills to generate a text entity set E nlp ={e n+1 ,..., e m}; the risk sensitivity coefficient of each entity is obtained by training with a power safety supervision historical accident database;

[0047] A graph construction unit, which is used to design graph nodes according to the multi-modal fusion vector, and design five types of fixed relationship edges to comprehensively cover cross-modal and rule relationships, form a complete safety supervision business link, and design graph edge weights using a fixed weight strategy to obtain a graph G=(E, R), where the nodes E and the edges R all come from the mapping between the entities extracted from the multi-modal fusion vector and the knowledge base rules;

[0048] A risk analysis unit, which is used to infer a set of potential risk factors in the power safety supervision scenario according to the graph G=(E, R) and in combination with the risk sensitivity coefficient, and screen out the set of risk factors in the graph G=(E, R);

[0049] A risk reasoning unit, which is used to generate a cause-and-effect chain with clear cause-consequence relationships in the graph G=(E, R) according to the set of risk factors, and infer the risk level of the overall operation through the cause-and-effect chain;

[0050] The knowledge graph updating unit is used to dynamically optimize the structure and parameters of the graph G=(E,R) based on the causal chain and risk level to generate an adaptive evolutionary graph G′.

[0051] The beneficial technical effects of the present invention are at least as follows:

[0052] The present invention designs a set of cross-modal data fusion mechanisms, and for the first time integrates and models the image recognition results of the work site and the text information of the work ticket in a unified knowledge graph structure, opens up the semantic association link between images and texts in business scenarios, and constructs a multimodal knowledge system for power safety supervision, which improves the model's comprehensive understanding of complex working environments. On this basis, the system further combines the risk causal reasoning mechanism, and can realize automatic reasoning and interpretable output of risk causal relationships based on the associated multimodal entities, business scenario elements and safety regulations in the knowledge graph, and comprehensively improves the scientificity, rationality and automation level of risk assessment. Through the above mechanism, the present invention breaks through the technical bottlenecks of insufficient multimodal data fusion in the prior art, the inability to model business logic, and the lack of causal support for risk inference, and provides an innovative solution for realizing intelligent and interpretable safety risk identification and assessment in power operation scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] The present invention is further described using the accompanying drawings, but the embodiments in the accompanying drawings do not constitute any limitation to the present invention. A person skilled in the art can obtain other drawings based on the following drawings without creative work.

[0054] Figure 1 This is a flow chart of a method for semantic recognition of electric service based on knowledge graph according to an embodiment of the present invention.

[0055] Figure 2 This is a framework diagram of a power service semantic recognition system based on a knowledge graph according to an embodiment of the present invention. DETAILED DESCRIPTION

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

[0057] In one or more embodiments, Figure 1 As shown, a method for semantic recognition of electric service based on knowledge graph is disclosed in the present invention, and the method includes steps S1 to S5, including:

[0058] S1. Collect several groups of on-site safety supervision image data and text data, introduce a risk sensitivity coefficient, and perform cross-modal fusion to output a multi-modal fusion vector. Among them, the on-site safety supervision image data uses an object detection model to extract specific image entities in the power safety supervision scenario to obtain an image entity set E cv ={e1, e2,.., e i ·, e n}, and each e i is a safety supervision entity in an image; the text data uses an NLP model to extract the operation fields in the ticket to generate a text entity set E nlp ={e n+1 ,..., e m}; the risk sensitivity coefficient of each entity is obtained by training from the power safety supervision historical accident database.

[0059] Specifically, collect:

[0060] D cv : On-site safety supervision image data, including personnel, equipment, environment, safety supervision protection facilities, etc.

[0061] D nlp : Text data such as work tickets and safety regulations clauses, including fields such as work types, measures, and risk levels.

[0062] In view of the data characteristics of the power safety supervision scenario, a "Multi-modal Safety Supervision Feature Fusion Engine (MAFE)" is proposed to uniformly fuse the safety supervision risk entities in the image and text data and output a cross-modal feature vector V m .

[0063] Image data processing path: D cv Input the safety supervision dedicated object detection model f cv , extract specific image entities in the power safety supervision scenario to obtain an image entity set E cv ={e1, e2,..., e n}; each e i is a safety supervision entity in an image (such as e1 = "person not wearing a safety belt", e2 = "no fence"). Each image entity e i passes through the image dedicated feature encoder φ(e i ), and is mapped to a d-dimensional vector

[0064] Text data processing path: D nlp Input the work ticket dedicated NLP model f nlp , extract the operation fields in the ticket to generate a text entity set E nlp ={e n+1 ,..., e m}, such as e n+1= "Operation type = aerial work", e n+2 = "There is no warning measure for the bill". Each text entity e j Passes through the text feature encoder ψ(e j ), and is mapped to a d-dimensional vector

[0065] Furthermore, introduce the "risk sensitivity coefficient" λ in the power safety supervision scenario. The risk sensitivity coefficient λ of each entity i or λ j Is obtained by training from the power safety supervision historical accident database. Specifically:

[0066] High-risk targets (such as "not wearing a seat belt") are assigned λ i = 0.9;

[0067] Low-risk targets (such as "not wearing a reflective vest") are assigned λ i = 0.5;

[0068] Similarly, for the bill entity e j 's λ j Is assigned according to the "risk level of operation type". For example, for "aerial work", λ j = 0.8.

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

[0070]

[0071] Among them, n is the number of image entities, and 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 a greater influence in the embedding space, and improve 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 graph nodes based on the multi-modal fusion vector, and design five types of fixed relationship edges to comprehensively cover cross-modal and rule relationships, forming a complete work safety supervision business link. Use the fixed weight strategy to design the edge weights of the graph to obtain the graph G=(E, R), where all nodes E and edges R come from the entities extracted from the multi-modal fusion vector and the mapping of knowledge base rules.

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

[0075] The present invention proposes the "Electric Power Work Safety Supervision Multi-modal Relationship Graph Modeling Mechanism (EARM-Graph)", which adapts to the typical "image - work ticket - safety regulations knowledge" scenario of the Southern Power Grid and completes the structured construction of nodes and relationships.

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

[0077] E cv : Safety supervision entities in the image, such as "not wearing a safety helmet" and "no fence in the construction area";

[0078] E nlp : Business fields in the work ticket, such as "operation type = high altitude" 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", etc., all come from the Southern Power Grid work safety supervision system knowledge base.

[0080] Graph relationship design: Design five types of fixed relationship edges to comprehensively cover cross-modal and rule relationships, forming a complete work safety supervision business link:

[0081] R cv-nlp : Image entity - text entity relationship (such as "not wearing a seat belt" ←→ "high altitude operation");

[0082] R cv-cv : Spatial and operation scenario co-occurrence relationship between image entities (such as "person not wearing a seat belt" ←→ "no fence");

[0083] R nlp-nlp : Logical relationship between work ticket fields (such as "high altitude operation" ←→ "require safety protection");

[0084] R cv-role With R nlp-role : The bidirectional constraint relationship between the image / text entity and the rule node, such as "not wearing a seat belt" → "Safety Regulation Article 45".

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

[0086] High-risk entity pairs (e p , e q )(such as λ p +λ q >1.5) will have edges constructed first to ensure the integrity of the "high-risk path" structure of G.

[0087] Graph edge weight design: Only the fixed weight strategy is adopted in this stage, and no adaptive learning is performed:

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

[0089] r cv-rule =r nlp-rule =1.2, which is used to highlight the relative importance of the "safety regulation constraint";

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

[0091] Graph construction formula: Define the final graph 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] where both E and R come from the entity extraction from V m and the mapping of the knowledge base rules.

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

[0095] S3. According to the graph G=(E,R), combined with the risk sensitivity coefficient, infer the set of potential risk factors in the power safety supervision scenario, and screen out the set of risk factors in the graph G=(E,R).

[0096] Specifically, the input is the cross-modal knowledge graph G=(E, R) output from Step 2. The graph is composed of image entities E cv , text entities E nlp , and rule entities E rule . The nodes already carry the risk sensitivity coefficient λ i from Step 1. The graph relationship R contains the structured edge information between multi-modal entities and rules. The goal of this step is to infer the set of potential risk factors E risk in the power safety supervision scenario on G, completing the single-point identification from the graph G to the risk factors, which serves as the input for the subsequent cause chain generation. For this purpose, the present invention proposes the "Power Safety Supervision 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 supervision graph and help infer potential high-risk factors.

[0097] Furthermore, 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] Among them, is the feature representation of node e i at layer l; is the set of neighbor nodes 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 the learnable parameter matrix.

[0100] Different from the equal-weight propagation of ordinary GNNs, this mechanism enables high-risk nodes with high λ i to obtain stronger propagation ability during the feature propagation process; this can improve the information aggregation degree of high-risk nodes such as "not wearing a seat belt + working at height" in the graph and solve the problem of dilution of high-risk paths in the power safety supervision graph.

[0101] Risk factor identification mechanism: After completing the feature propagation of L layers, directly calculate the risk intensity of the final node features i of all e ; through the risk discrimination function, screen the risk factor set E risk , and the screening formula is as follows:

[0102]

[0103] Among them, θ is the risk coefficient screening threshold (e.g., θ = 0.7); ρ is the feature strength threshold (e.g., ρ = 0.8); the node e that meets the dual requirements of risk sensitivity and feature strength i is identified as a risk factor.

[0104] Traditional graph reasoning usually only screens nodes based on the feature propagation strength. This solution additionally integrates the "risk-sensitive weight" λ i , forming a factor reasoning mechanism with risk priority, which is more in line with the actual business characteristics in the power safety supervision scenario; 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. According to the set of risk factors, generate a cause chain with clear cause-consequence relationships in the graph G=(E, R), and infer the risk level of the overall operation through the cause chain.

[0106] Specifically, the input is the set of high-risk factors E output by step S3 risk , and each node e in this set i comes from the graph G and has a determined risk sensitivity coefficient λ i , and the types include image entities (such as "not wearing a safety belt"), work ticket fields (such as "lack of high-altitude protection"), and safety regulation rule entities (such as "Article 45 of the safety regulations"). There is no cause-consequence relationship formed among the risk factors in this set. The goal of this step is: based on E risk , generate a risk cause chain C with a clear cause-consequence structure, and infer the overall operation risk level L through the cause chain. This cause chain needs to reflect the risk evolution logic in the multi-modal scenario of power safety supervision, such as the causal path between image behavior violations, work ticket defects, and safety regulation constraints. To achieve the above goal, a "Southern Power Grid Safety Supervision Multi-modal Cause Chain Inference Machine (NAM-Causal)" is proposed, which integrates the knowledge in the field of power safety supervision to construct a cross-modal risk cause chain. This mechanism includes two core steps: cause chain construction and risk level quantification.

[0107] Cause chain construction mechanism: According to the risk types (image / text / rule), risk levels λ risk of the nodes in E i and the domain knowledge rules, construct a directed cause relationship edge set R c , forming a cause chain C=(E risk , R c ). The direction of the edge represents the cause-result relationship, and the edge weight c ij represents the cause strength. The cause strength edge weight is defined as:

[0108] c ij =λi ·w ij +δ ij (5)

[0109] Among them, λ i : the risk sensitivity coefficient of node e i ; w ij : the causal rule weight defined according to the knowledge base in the field of work safety supervision, such as the w of "no protection → fall risk" ij = 0.9; δ ij ∈ {0, 0.2}: If e i and e j are in the same business scenario (such as both belonging to "working at height"), then δ ij = 0.2 to enhance the causal consistency within the scenario. This mechanism ensures that high-risk nodes preferentially form cause-effect chains, and factors within the same scenario are likely to form strong cause-effect paths, meeting the characteristics of power safety supervision risk aggregation.

[0110] Cause-effect chain generation rule: Based on all c ij , use the threshold τ = 1.0 to screen the strongly associated edges e i → e j , and form a risk cause-effect path. For example: "The work permit lacks protective measures (λ = 0.8)" → "The personnel did not fasten the safety belt (λ = 0.9)" → "Fall risk", forming a complete cause-effect chain.

[0111] Risk level quantification mechanism: According to the edge weights on the cause-effect path in C, calculate the overall link risk aggregation value S C , and determine the risk level L accordingly:

[0112]

[0113] The risk level segmentation standard is set according to the safety supervision regulations of China Southern Power Grid, reflecting the relationship between the aggregation intensity of the cause-effect chain and the severity of the operation risk.

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

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

[0116] Risk level L: Used for on-site operation risk grading determination and as a basis for subsequent optimization of the map.

[0117] S5. Based on the cause-effect chain and risk level, perform dynamic optimization on the structure and parameters of the map G = (E, R) to generate an adaptive evolutionary map G'.

[0118] Specifically, the input is the risk cause-effect chain C = (E risk , Rc ) and the corresponding risk level L, as well as the initial cross-modal knowledge graph G = (E, R) generated in step 2. Among them, C has described E risk The causal relationship between high-risk factors within, L is the risk level of the operation site (such as general, relatively large, major risk), and G is still the graph structure of static modeling. The goal of this step is: based on C and L, dynamically optimize the structure and parameters of G to generate an adaptive evolutionary graph G′, so as to improve the risk reasoning ability and business sensitivity of G′ in the next operation scenario, and form a real "business-graph-risk reasoning-graph evolution" closed loop. For this reason, the "Southern Power Grid Safety Supervision Causal Feedback Graph Evolution Mechanism (NAF-GE)" is proposed, which combines the risk factors, causal edge strengths and risk levels in the causal chain to complete the graph structure optimization and parameter adaptive update.

[0119] Mechanism 1: Global edge weight reallocation guided by risk level:

[0120] Drive the structural adjustment of the entire graph G through the risk level L to solve the problem that the safety supervision graph's reasoning does not focus on high-risk scenarios due to "risk level insensitivity".

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

[0122]

[0123] Among them, α is the enhancement basic 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 = "relatively large risk", γ L = 1.0 if L = "major risk"); is the indicator function of the edges connected to the high-risk factor E in the graph G risk .

[0124] We not only enhance the weights of the causal edges in C alone, but uniformly enhance the multi-modal relationship edges related to E in the entire graph according to L risk to focus on the full-graph feature propagation ability of high-risk scenarios; combined with the risk level L, realize the dynamic update of the "level-sensitive" structure of the graph, and strengthen the causal chain path under high-risk levels.

[0125] Mechanism 2: Implicit structure supplementation driven by the causal chain:

[0126] For the potential causal relationships that do not explicitly exist in G in C, innovatively design a "cross-modal chain supplementation mechanism" to supplement cross-modal or cross-level implicit risk edges. The specific rules are as follows:

[0127] If e exists in C p →e q and there is no (e p ,e q ) or r pq in G and the edge weight <∈ (e.g., ∈ = 0.5), then a new edge is added or the edge weight is reset:

[0128] r pq ′ = β · c pq + μ · η pq (8)

[0129] where β is the causal chain complementary chain factor (e.g., β = 0.4); c pq is the causal edge strength in the causal chain; η pq is the cross-modal consistency factor. If e p and e q are in different modalities respectively (such as image + bill, bill + rule), η pq = 0.3, otherwise it is 0; μ is the cross-modal complementary chain gain factor (e.g., μ = 0.2).

[0130] This mechanism is designed specifically for the power safety supervision scenario, considering the problem of the lack of the "image-text-safety regulations" triple relationship link, automatically supplementing the high-risk cross-modal chain; avoiding the path breakage caused by the "lack of key cross-modal risk links" in the subsequent reasoning stage, and enhancing the cross-modal connectivity of the knowledge graph.

[0131] In one or more embodiments, as Figure 2 shown, the present invention discloses a power service semantic recognition system based on a knowledge graph, and the system includes:

[0132] A data acquisition unit 101, configured to collect several groups of safety supervision on-site image data and text data, introduce a risk sensitivity coefficient, and perform cross-modal fusion to output a multi-modal fusion vector; wherein, the safety supervision on-site image data uses an object detection model to extract specific image entities in the power safety supervision scenario to obtain an image entity set E cv ={e1, e2,.., e i .,, e n}, and each e i is a safety supervision entity in an image; the text data uses an NLP model to extract operation fields in the 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 from a power safety supervision historical accident database;

[0133] The graph construction unit 102 is used to design graph nodes according to the multi-modal fusion vector, and design five types of fixed relationship edges, comprehensively covering cross-modal and regular relationships, forming a complete work safety supervision business link, and designing graph edge weights using a fixed weight strategy to obtain the graph G=(E, R), where the nodes E and the edges R all come from the entities extracted from the multi-modal fusion vector and the mapping of knowledge base rules;

[0134] The risk analysis unit 103 is used to infer the set of potential risk factors in the power work safety supervision scenario according to the graph G=(E, R) in combination with the risk sensitivity coefficient, and screen out the set of risk factors in the graph G=(E, R);

[0135] The risk reasoning unit 104 is used to generate a cause chain with a clear cause-consequence relationship in the graph G=(E, R) according to the set of risk factors, and infer the risk level of the overall operation through the cause chain;

[0136] The knowledge graph update unit 105 is used to dynamically optimize the structure and parameters of the graph G=(E, R) based on the cause chain and the risk level, and generate an adaptive evolutionary graph G'.

[0137] These are only some preferred embodiments of the present invention. Of course, the scope of the rights of the present invention cannot be limited thereby. Those of ordinary skill in the art can understand all or part of the processes of implementing the above embodiments, and the equivalent changes made according to the claims of the present invention still fall within the scope covered by the invention.

Claims

1. A semantic recognition method for power services based on a knowledge graph, characterized in that, The method includes the following steps: Collect several groups of safety supervision on-site image data and text data, introduce a risk sensitivity coefficient, and perform cross-modal fusion to output a multi-modal fusion vector; among them, the safety supervision on-site image data uses an object detection model to extract specific image entities in the power safety supervision scenario 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 uses an NLP model to extract the operation fields in the 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 with a power safety supervision historical accident database; Design graph nodes based on the multi-modal fusion vector, and design five types of fixed relationship edges to comprehensively cover cross-modal and rule relationships, forming a complete safety supervision business link. Use the fixed weight strategy to design the edge weights of the graph to obtain the graph G=(E, R), where the nodes E and the edges R all come from the entities extracted from the multi-modal fusion vector and the mapping of knowledge base rules; Based on the graph G=(E, R), combined with the risk sensitivity coefficient, infer the set of potential risk factors in the power safety supervision scenario, and screen out the set of risk factors in the graph G=(E, R); According to the set of risk factors, generate a cause chain with clear cause-consequence relationships in the graph G=(E, R), and infer the risk level of the overall operation through the cause chain; Based on the cause chain and the risk level, perform dynamic optimization of the structure and parameters of the graph G=(E, R) to generate an adaptive evolutionary graph G'; 2. The semantic recognition method of power service based on knowledge graph according to claim 1, characterized in that, The graph nodes include safety supervision entities in the image, business fields in the work ticket, and domain knowledge nodes; The five types of fixed relationship edges include: Image entity - text entity relationship; Spatial and operation scenario co-occurrence relationships between image entities; Logical relationships between work ticket fields; Bidirectional constraint relationships between image / text entities and rule nodes.

3. The semantic recognition method of power service based on knowledge graph according to claim 2, characterized in that, The graph G=(E, R) will preferentially establish edges for high-risk entity pairs to ensure the integrity of the high-risk path structure of the graph G; The fixed weight strategy uses the edge weight set as a static parameter.

4. A semantic recognition method for power services based on a knowledge graph according to claim 1, characterized in that, Based on the graph G=(E, R), combined with the risk sensitivity coefficient, infer the set of potential risk factors in the power safety supervision scenario, and screen out the set of risk factors in the graph G=(E, R). Specifically: Perform risk-sensitive feature propagation in the graph G=(E, R) based on the risk sensitivity coefficient to obtain node features; After completing the L-layer feature propagation, use the risk discrimination function to screen out the set of risk factors.

5. The semantic recognition method of power service based on knowledge graph according to claim 4, characterized in that, The screening formula is: Among them, E risk is the set of risk factors; θ is the screening threshold of risk coefficients; ρ is the feature intensity threshold; Among them, the node e that meets the dual requirements of risk sensitivity and feature strength i is identified as a risk factor.

6. The semantic recognition method of power service based on knowledge graph according to claim 2, characterized in that According to the set of risk factors, generate a cause chain with clear cause-consequence relationships in the graph G=(E, R). Specifically: Construct a directed causal relationship edge set R based on the risk types, risk sensitivity coefficients of the nodes in the risk factor set, and domain knowledge nodes c , forming a causal chain Use the preset threshold of the edge weight to screen out strongly associated edges for all edges to form a risk cause path.

7. A semantic recognition method for power services based on a knowledge graph according to claim 6, characterized in that The cause chain also includes: The direction of the edge represents the cause-effect relationship, and the edge weight c ij represents the intensity of the cause and is calculated as c ij = λ i · w ij + δ ij Among them, λ i is the risk sensitivity coefficient of node e i ; w ij is the causal rule weight defined according to the safety supervision domain knowledge base; δ ij ∈{0, 0.2}: If node e i and node e j are in the same business scenario, then δ ij = 0.2 to enhance the causal consistency within the scenario; Among them, inferring the risk level of the overall operation through the cause chain. Specifically: Calculate the overall link risk aggregation value according to the edge weights on the cause path in the cause chain, and determine the risk level accordingly: general risk, relatively large risk, and major risk.

8. A semantic recognition method for power services based on a knowledge graph according to claim 1, characterized in that Based on the cause chain and the risk level, perform dynamic optimization of the structure and parameters of the graph G=(E, R) to generate an adaptive evolutionary graph G'. Specifically: For all edges in the graph associated with the nodes and risk factors in the cause chain, analyze the cause chain and the risk level through the indicator function, and dynamically adjust the edge weights based on the indicator function to determine whether the node is in the risk set; For potential cause relationships that are not explicitly present in the graph in the cause chain, supplement cross-modal or cross-level implicit risk edges; Among them, dynamically adjusting the edge weights based on the indicator function to determine whether the node is in the risk set includes: If node e i or e j belongs to the set of risk factors, then amplify the edge weight to strengthen the role of risk-related connections; If the condition is not met, the edge weight remains unchanged.

9. A semantic recognition method for power services based on a knowledge graph according to claim 8, characterized in that For the potential causal relationships in the causal chain that are not shown in the graph, implicit risk edges across modalities or hierarchical levels are supplemented. The specific rules are as follows: If there is a node e in the causal chain p →e q and the graph G has no edge (e p ,e q ) or the edge weight r pq <∈, then add a new edge or reset the edge weight: r pq ′ = β·c pq + μ·η pq Among them, β is the complementary chain factor of the causal chain; c pq is the causal edge strength in the causal chain; η pq is the cross-modal consistency factor. If e p and e q are different modalities respectively, then η pq = 0.3; otherwise it is 0; μ is the cross-modal complementary chain gain factor.

10. A semantic recognition system for power services based on a knowledge graph, characterized in that, The system includes: A data acquisition unit is used to collect several groups of safety supervision on-site image data and text data, introduce a risk sensitivity coefficient, and perform cross-modal fusion to output a multi-modal fusion vector. Among them, the safety supervision on-site image data uses an object detection model to extract specific image entities in the power safety supervision scenario to obtain an image entity set E cv ={e1, e2,.., e i ,.., e n}, and each e i is a safety supervision entity in an image; the text data uses an NLP model to extract operation fields in the 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 with a power safety supervision historical accident database; A graph construction unit, which is used to design graph nodes according to the multi-modal fusion vector, and design five types of fixed relationship edges to comprehensively cover cross-modal and rule relationships, form a complete safety supervision business link, design the edge weights of the graph using a fixed weight strategy, and obtain the graph G=(E, R), where the nodes E and the edges R all come from the entities extracted from the multi-modal fusion vector and the mapping of knowledge base rules; A risk analysis unit, which is used to infer the set of potential risk factors in the power safety supervision scenario according to the graph G=(E, R) and in combination with the risk sensitivity coefficient, and screen out the set of risk factors in the graph G=(E, R); A risk inference unit, which is used to generate a causal chain with clear cause-consequence relationships in the graph G=(E, R) according to the set of risk factors, and infer the risk level of the overall operation through the causal chain; A knowledge graph update unit, which is used to dynamically optimize the structure and parameters of the graph G=(E, R) based on the causal chain and the risk level, and generate an adaptive evolutionary graph G'.

Images