Safety production risk identification method and system based on knowledge graph

Abstract

The invention discloses a safety production risk identification method and system based on a knowledge graph, and relates to the technical field of safety production risk identification. Entity nodes and relation edge data of the knowledge graph are obtained, feature vectors are extracted, and embedded representation is generated by adopting a graph neural network; calculating a node weight by using an attention mechanism to determine a risk mode, traversing an association path to generate a risk propagation sequence, fusing time sequence features to update an entity state and determine a dynamic propagation path, extracting a key node sub-graph to adjust an edge weight to optimize the risk mode, and finally integrating environment features through iterative query to identify a complete risk propagation chain. According to the invention, dynamic tracking of equipment, personnel and environment network risks and cross-dimension cascade risk identification are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of safety production risk identification, and in particular to a safety production risk identification method and system based on a knowledge graph. Background Art

[0002] Safety production is a critical component of industrial safety, ensuring personnel safety, equipment stability, and production continuity. It is directly linked to the stability of business operations and the sustainable development of the socio-economic system. With the increasing complexity of industrial systems, the sources of safety risks have become more diverse, encompassing multiple dimensions, including equipment operating status, human operator behavior, and environmental factors. Existing safety risk identification methods often rely on single-dimensional monitoring or rule-making, making it difficult to capture the dynamic correlations between different factors. For example, traditional methods may focus solely on equipment failure signals, ignoring the potential impact of human operational errors or environmental anomalies on the failure.

[0003] Furthermore, many solutions lack a comprehensive analysis of risk propagation pathways when dealing with complex systems, resulting in delayed discovery of potential hazards and insufficient early warning capabilities. Knowledge graphs have become a key technical tool for identifying production safety risks due to their ability to integrate multi-source heterogeneous data and reveal relationships between entities. However, after constructing a knowledge graph encompassing equipment status, human behavior, and environmental factors, the core challenge remains how to effectively analyze the characteristics of entity nodes and relationship edges to uncover hidden risk patterns.

[0004] The primary technical challenge lies in extracting deep correlation features from multi-dimensional feature vectors. For example, in chemical production scenarios, abnormal equipment temperature may be related to operators failing to adjust parameters according to regulations. However, this correlation is often hidden in the data and difficult to capture directly. Furthermore, the complexity of this feature extraction leads to another challenge: how to accurately identify the transmission path of risks between different entities based on the topological structure of the graph. For example, a minor equipment failure may spread through human error, eventually causing environmental anomalies and leading to major safety accidents. Existing technology makes it difficult to accurately predict this chain reaction.

[0005] Therefore, how to automatically extract hidden risk patterns and identify risk propagation paths by deeply analyzing the multi-dimensional features of entity nodes and relationship edges in the knowledge graph and combining them with the graph topology has become a key issue in the identification of production safety risks. Summary of the Invention

[0006] To this end, the technical problem to be solved by the present invention is to overcome the defects of the existing technology in dealing with complex systems, which lacks a comprehensive analysis of the risk propagation path, resulting in untimely discovery of potential hidden dangers and insufficient early warning capabilities. A knowledge graph-based production safety risk identification method is provided, which can realize dynamic tracking of risks along equipment, personnel, and environmental networks through a multi-level collaborative processing architecture, and automatically identify cascading risk patterns generated by cross-dimensional interactions.

[0007] To solve the above technical problems, the present invention provides a method for identifying production safety risks based on a knowledge graph, comprising: Obtain entity node and relationship edge data in the knowledge graph and extract multi-dimensional feature vectors; A graph neural network is used to perform convolution operations on entity nodes and relationship edges, aggregate neighborhood information, and generate embedded representations; Through embedding representation, the attention mechanism is used to calculate the weight distribution between entity nodes and determine the risk pattern; According to the risk pattern, traverse the associated paths in the knowledge graph to generate a risk propagation sequence; According to the risk propagation sequence, the temporal changes of multi-dimensional feature vectors are integrated to update the entity status and determine the dynamic propagation path; Extracting key node subgraphs from the dynamic propagation path, adjusting path edge weights, highlighting the interactive impact between device status and human behavior, and generating refined risk models; Based on the refined risk pattern representation, the knowledge graph is iteratively queried, additional features of environmental factors are integrated, the complete transmission chain is judged, and the final risk identification result is obtained.

[0008] In one embodiment of the present invention, extracting a multi-dimensional feature vector includes: By querying the preset knowledge graph database interface, an initial multi-dimensional feature vector is obtained from the entity nodes of device status, human behavior, and environmental factors, including temperature readings that record device status, operation logs that record human behavior, and humidity indicators that record environmental factors. The temperature readings, operation logs, and humidity indicators constitute a feature set; The k-nearest neighbor algorithm is used to perform cluster analysis on the feature set, and the temperature readings, humidity indicators, and operation logs are grouped according to their similarities to obtain the feature clustering results; If there are abnormal groups in the feature clustering results, the abnormal feature vectors are isolated through the anomaly detection algorithm, and the anomaly is judged based on the preset threshold to obtain the normal feature set; Based on the normal feature set, contextual relationships are extracted from the knowledge graph to generate an association rule set. Based on the association rule set, a decision tree algorithm is used to classify and predict temperature readings, humidity indicators, and operation logs to obtain device status prediction results. If the device status prediction result deviates from the normal range, the entity node attributes in the knowledge graph are updated through the database interface to obtain the updated feature vector. Based on the updated feature vector, cluster analysis is re-executed to verify consistency and obtain the optimized feature set.

[0009] In one embodiment of the present invention, a graph neural network is used to perform convolution operations on entity nodes and relationship edges, aggregate neighborhood information, and generate an embedded representation, including: Through the preset knowledge graph database interface, we obtain entity node and relationship edge data, use graph neural network to perform convolution operations, aggregate neighborhood information, and generate initial embedding vectors; According to the initial embedding vector, the attention mechanism is used to assign weights, highlight the local features of key entity nodes and relationship edges, and generate a weighted embedding vector; If the similarity between the weighted embedding vector and the preset topological structure template is lower than the preset threshold, the embedding vectors are grouped through cluster analysis to determine the abnormal embedding vectors and obtain a set of normal embedding vectors; According to the normal embedding vector set, context information is extracted from the knowledge graph to generate an association rule set. Based on the association rule set, the normal embedding vector is classified using the random forest algorithm to determine the state prediction result of the entity node; If the state prediction result deviates from the preset normal range, the entity node attributes in the knowledge graph are updated through the database interface to generate an updated embedding vector. Based on the updated embedding vector, the graph neural network convolution operation is re-executed to verify the consistency and obtain the optimized embedding vector set.

[0010] In one embodiment of the present invention, the weight distribution between entity nodes is calculated through an attention mechanism to determine the risk pattern, including: Extract node relationships from the embedded representation, use the attention mechanism to calculate the weight distribution between entity nodes, obtain the connection strength, and analyze risk patterns based on the connection strength; If the connection strength exceeds the preset threshold, it is judged as a high-risk association and a risk pattern set is obtained; For the risk pattern set, a clustering algorithm is used to mine potential patterns and obtain pattern classification results. Features are extracted from the pattern classification results, and data associations are analyzed to obtain potential hidden patterns. Through the potential hidden patterns, the association strength between patterns is calculated to obtain the final pattern distribution. Based on the final pattern distribution, a visual representation of the risk pattern is generated to obtain the analysis results.

[0011] In one embodiment of the present invention, according to the risk pattern, traversing the associated paths in the knowledge graph to generate a risk propagation sequence includes: Extract related entities from the knowledge graph, use breadth-first search to traverse the connection paths between entities, and obtain a preliminary risk propagation sequence. Obtain the attribute value of each entity in the path, determine the risk transmission intensity between entities through weighted calculation, and generate a risk transmission matrix. If there is a high-intensity transmission path in the risk transmission matrix, extract the main risk transmission direction through matrix decomposition algorithm to obtain the key transmission subsequence; Based on the key propagation subsequence, cluster analysis is used to group entities, obtain entity clusters with concentrated risks, and determine the risk concentration area. If the number of entities in the risk concentration area exceeds the preset threshold, the entity cluster is mapped to a low-dimensional space through a graph embedding algorithm to obtain a semantic relationship vector; According to the semantic relationship vector, anomaly detection algorithm is used to identify potential abnormal entities, obtain the abnormal entity set, and traverse the association path in the knowledge graph through the abnormal entity set to generate the final risk propagation sequence.

[0012] In one embodiment of the present invention, based on the risk propagation sequence, the temporal changes of the multi-dimensional feature vectors are integrated, the entity state is updated, and the dynamic propagation path is determined, including: Obtain entity data in the risk propagation sequence, extract feature vectors from multi-dimensional time series data, obtain a set of entity feature vectors, construct a network topology structure through a graph neural network, and update the topological connections based on the set of entity feature vectors to obtain the network topology structure; A recursive update mechanism is used to iteratively calculate entity state changes based on the network topology to obtain an entity state sequence. If any entity state in the entity state sequence exceeds a preset threshold, the propagation path weight is adjusted through the graph neural network to obtain a propagation path with adjusted weights. According to the propagation path after weight adjustment, the temporal change trend of the propagation path prediction is analyzed to obtain the propagation path prediction result. Through the propagation path prediction result, the state update rule is integrated to iteratively optimize the dynamic propagation path to obtain the optimized dynamic propagation path; The optimized dynamic propagation path is adopted and combined with time series data fusion to determine the final dynamic propagation path.

[0013] In one embodiment of the present invention, a key node subgraph is extracted from the dynamic propagation path, path edge weights are adjusted, and the interactive impact between device status and human behavior is highlighted to generate a refined risk model, including: Extract key node subgraphs from dynamic propagation paths, use graph segmentation algorithms to determine node connection strengths, and obtain key node subgraphs. For key node subgraphs, use attention mechanisms to calculate the interaction influence relationships between nodes, adjust path edge weights, and generate weighted path subgraphs. The interaction features between device status data and personnel behavior data are obtained from the weighted path subgraph. Principal component analysis is used to extract the main interaction patterns and determine the interaction influence relationships. If the intensity of the interaction influence relationship exceeds a preset threshold, the interaction features are grouped using a clustering algorithm to obtain the association pattern between the behavior triggering conditions and the device abnormality signals. According to the correlation pattern, the correspondence between the abnormal signals of the equipment and the behavior triggering conditions is extracted to generate the risk assessment index. Through the risk assessment index, the weighted summation method is used to fuse the abnormal signals of the equipment and the behavior triggering conditions to generate a refined risk model.

[0014] In one embodiment of the present invention, based on the refined risk pattern representation, the knowledge graph is iteratively queried, additional features of environmental factors are integrated, and the complete propagation chain is determined to obtain the final risk identification result, including: The propagation chain features are obtained from the refined risk model. The knowledge graph query method is used to extract the node attributes related to the environmental factor data to obtain the environmental correlation feature set. Based on the environmental correlation feature set, the graph convolutional network is used to analyze the propagation relationship between nodes, and the weights of the propagation chain features are updated to obtain the weighted propagation features. The correlation data between device status interaction and human behavior interaction is extracted from the weighted propagation features. The main interaction patterns are extracted using principal component analysis, and the interaction feature set is determined. If the correlation strength of the interaction feature set exceeds the preset threshold, the K-means clustering method is used to group the interaction features to obtain the correlation pattern between the behavior triggering conditions and the device abnormal signals. According to the association pattern, the correspondence between abnormal equipment signals and behavior triggering conditions is extracted, and the weighted summation method is used to integrate environmental factor data to generate a comprehensive risk index. The changing trend of the risk propagation path is extracted from the comprehensive risk index. The time series analysis method is used to predict the dynamic changes of the risk propagation path and obtain dynamic risk prediction results. According to the dynamic risk prediction results, the knowledge graph update method is used to adjust the node weights of the propagation chain characteristics and determine the final risk identification results.

[0015] To solve the above technical problems, the present invention also provides a knowledge graph-based safety production risk identification system, comprising: The data acquisition module is used to extract entity nodes and relationship edge data from the knowledge graph and generate multi-dimensional feature vectors; A graph neural network processing module, which performs convolution operations on entity nodes and relationship edges, aggregates neighborhood information, and generates embedded representations; A risk pattern calculation module, based on the embedded representation, uses an attention mechanism to calculate the weight distribution between entity nodes and determine the risk pattern; A risk propagation sequence generation module is used to traverse the associated paths in the knowledge graph according to the risk pattern to generate a risk propagation sequence; Dynamic path update module, used to integrate the temporal changes of multi-dimensional feature vectors, update entity states and determine dynamic propagation paths; A key subgraph extraction module is used to extract key node subgraphs from the dynamic propagation path, adjust the path edge weights to highlight the interactive impact between device status and human behavior, and generate a refined risk model; The risk identification module is used to perform iterative queries based on refined risk patterns, integrate additional features of environmental factors, determine the complete transmission chain and output the final risk identification results.

[0016] The above technical method of the present invention has the following advantages over the prior art: The present invention discloses a method for identifying production safety risks based on knowledge graphs. It aims to solve the problem of business scenarios where hidden risk patterns are difficult to capture under the interactive influence of equipment status, personnel behavior and environmental factors. By integrating multi-dimensional feature vector extraction, graph neural network convolution and attention mechanism, it solves the difficult problem of dynamic tracking and accurate identification of risk propagation paths.

[0017] The present invention first extracts multidimensional feature vectors from the knowledge graph, employs a graph neural network to aggregate neighborhood information, and generates a deep, associative embedded representation. An attention mechanism is used to calculate inter-entity weights, highlighting risk-related connections. If abnormal weights are found, a breadth-first search is used to track the propagation path, recursively updating entity states based on temporal changes and refining key node subgraphs. Finally, an iterative query integrates environmental factors to construct a complete propagation chain. By dynamically capturing the interactive risk patterns of equipment and human behavior, the present invention achieves high-precision risk identification, significantly improving operational safety and management efficiency in complex industrial scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to make the content of the present invention more clearly understood, the present invention is further described in detail below based on specific embodiments of the present invention in conjunction with the accompanying drawings, wherein: Figure 1 It is a flowchart of the steps of the method for identifying production safety risks based on knowledge graph of the present invention; Figure 2 It is a flowchart of the steps of extracting multi-dimensional feature vectors according to the present invention; Figure 3 is a flow chart of the steps of generating an embedded representation according to the present invention; Figure 4 is a flow chart of the steps for determining a risk model according to the present invention; Figure 5 is a flow chart of the steps of generating a risk propagation sequence according to the present invention; Figure 6 is a flow chart of the steps of determining a dynamic propagation path according to the present invention; Figure 7 is a flow chart of the steps for generating a refined risk model according to the present invention; Figure 8 This is a flowchart of the steps for obtaining the final risk identification result of the present invention; Figure 9 This is a structural framework diagram of the knowledge graph-based production safety risk identification system of the present invention. DETAILED DESCRIPTION

[0019] The present invention will be further described below with reference to the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it. However, the embodiments are not intended to limit the present invention.

[0020] Reference Figure 1 As shown, the present invention discloses a method for identifying production safety risks based on a knowledge graph, comprising the following steps: Obtain entity nodes and relationship edge data in the knowledge graph and extract multi-dimensional feature vectors. This step is the basis for risk identification, because safe production involves multi-dimensional data such as equipment status, personnel operations, and environmental parameters. Traditional methods often process each type of data separately, resulting in the loss of related information. In this embodiment, by uniformly extracting the feature vectors of entities and relationships, subsequent analysis can retain the complex correlations in the original data. For example, the time series data of the temperature sensor and the event data of the maintenance record are uniformly represented, laying a data foundation for building a complete risk propagation model.

[0021] A graph neural network is used to perform convolution operations on entity nodes and relationship edges, aggregate neighborhood information, and generate an embedded representation. Traditional risk identification methods only consider the attributes of the node itself, while ignoring its contextual relationship in the knowledge graph. This embodiment uses the neighborhood convolution of the graph neural network to capture the potential association between equipment failures and adjacent nodes, forming an embedded representation that includes topological structure. This aggregation mechanism enables the model to identify local risk features that cannot be discovered by traditional statistical analysis.

[0022] Through embedded representation, the attention mechanism is used to calculate the weight distribution between entity nodes and determine the risk pattern. Static knowledge graph analysis often treats all connection relationships equally, while actual risk propagation is directional and selective. This embodiment uses the attention mechanism to dynamically allocate weights between nodes, which can identify truly high-risk relationship paths rather than irrelevant ordinary inspection records. This adaptive weight distribution enables the system to focus on the most threatening risk transmission links, reducing false positives and missed reports.

[0023] Based on the risk pattern, the associated paths in the knowledge graph are traversed to generate a risk propagation sequence. Simply identifying local risk patterns is not enough to predict systemic accidents, and the complete propagation chain must be restored. This embodiment uses graph traversal to connect scattered risk signals in series. For example, isolated events such as "sensor drift → control parameter distortion → automatic adjustment failure → manual intervention delay" can be integrated into an ordered sequence. This global perspective enables security personnel to predict how risks will evolve from local failures to global accidents, rather than just passively responding to alarms that have already occurred.

[0024] Based on the risk propagation sequence, the temporal changes of multi-dimensional feature vectors are integrated to update entity states and determine the dynamic propagation path. Risk transmission is time-dependent, and traditional static graph analysis cannot capture this dynamic characteristic. This embodiment, by integrating temporal features, can identify key state change nodes. This dynamic update mechanism enables the model to adapt to real-time changing production environments and identify gradual risks such as "slow corrosion → sudden rupture" that are difficult to capture with traditional methods.

[0025] Key node subgraphs are extracted from the dynamic propagation paths, and path edge weights are adjusted to highlight the interactive impact between equipment status and personnel behavior, generating a refined risk model. Accident investigations have shown that most major accidents stem from failures in "equipment-personnel" interactions. This embodiment, by constructing key subgraphs and adjusting weights, quantifies this cross-dimensional impact. Local networks strongly correlated with the current risk are extracted from the vast knowledge graph, and the connection strength between these elements is dynamically adjusted through an algorithm. For example, in the path "night shift fatigue → inspection negligence → worsening cable wear," the system strengthens the edge weights between "fatigue" and "negligence" while weakening the influence of irrelevant factors. This refinement removes noise from the graph, allowing the risk model to present a clearer causal relationship and support precise intervention decisions.

[0026] Based on the refined risk model representation, the knowledge graph is iteratively queried, integrating additional features of environmental factors to determine the complete transmission chain and ultimately identify the risk. A single analysis cycle may miss long-term environmental factors (such as seasonal moisture-accelerated corrosion). This embodiment uses iterative queries to supplement environmental data, allowing the system to refine the risk chain. For example, this system can incorporate the "rainy weather → cable insulation degradation → short-circuit spark" pattern into the existing analysis framework. This closed-loop optimization mechanism ensures the completeness of risk identification. The final output is not an isolated alert, but a systematic risk assessment that includes the root cause, transmission path, and evolutionary trend, providing a comprehensive decision-making basis for proactive prevention and control.

[0027] Reference Figure 2As shown, this embodiment constructs a complete feature engineering process for the step of "obtaining entity node and relationship edge data from the knowledge graph and extracting multi-dimensional feature vectors" proposed in the above embodiment. In its specific implementation, multi-source data is first collected through a preset knowledge graph database interface. Time-series monitoring data such as temperature readings are obtained from the device status entity node, behavioral records such as operation logs are extracted from the personnel behavior node, and operating parameters such as humidity indicators are read from the environmental factor node. This structured data acquisition method ensures the integrity and standardization of feature sources. After obtaining the initial feature set, a k-nearest neighbor algorithm is used for cluster analysis. This step, based on the synergistic effect of device temperature, ambient humidity, and personnel operation behavior in risk formation, automatically groups data with similar risk characteristics using a distance metric in a multi-dimensional feature space. This unsupervised learning method can discover potential risk patterns without relying on historical labeled data. When an anomalous group significantly deviates from the main group in the clustering results, an anomaly detection algorithm based on statistical distribution is activated to identify and isolate the outlier by calculating metrics such as the Mahalanobis distance. This mechanism effectively eliminates interference from sensor failure data or human recording errors in subsequent analysis. After data cleaning, contextual relationships between entities are extracted from the knowledge graph to generate a set of association rules with support and confidence. This step leverages the unique semantic associations of the knowledge graph, identifying, for example, risk-indicating combinations such as "high temperature environment + new employee operation." A decision tree algorithm is then used to classify and predict the cleaned features. Its technical advantage lies in its ability to automatically prioritize the importance of different feature dimensions. For example, it may be found that a sudden change in temperature readings within a specific humidity range is more indicative of risk than the absolute temperature value. When a prediction exceeds a preset safety threshold, real-time feedback is fed back to the knowledge graph to update the attribute status of the corresponding entity node. This dynamic update mechanism overcomes the limitations of traditional static knowledge graphs and can capture the evolving trends of risk factors. Finally, iterative clustering verification ensures the consistency of the feature set. This closed-loop process maintains the stability of feature engineering while retaining flexibility to respond to unexpected situations.

[0028] From a technical perspective, this approach improves the reliability of risk identification through four key innovations: First, the comprehensive collection of multi-source features constructs a panoramic risk observation dimension; second, clustering-based anomaly detection ensures the quality of input data; third, association rule mining reveals potential connections between cross-domain features; and finally, a dynamic update mechanism ensures that the knowledge graph remains up-to-date. Through a progressive process of "data collection-cleaning-association-prediction-update," this approach transforms raw industrial data into feature representations with clear risk indicative significance. This provides a high-quality input foundation for subsequent graph neural network analysis, enabling subsequent risk propagation analysis to be based on more accurate and reliable feature representations.

[0029] Reference Figure 3 As shown, this embodiment constructs a closed-loop "generate-verify-optimize" process for the "using graph neural networks to generate embedded representations" step proposed in the above embodiment. First, when acquiring raw graph data through a pre-set interface, the graph neural network performs more than simple neighborhood aggregation. Instead, it establishes a basic feature space by generating initial embedding vectors. The introduction of the attention mechanism plays a crucial role here—it identifies key entities and amplifies their local features through weighted embedding vectors, preventing dilution of the overall representation by common node features. When the weighted vectors lack similarity to a pre-set template (e.g., a standard topological pattern for device health status), rather than simply discarding the data, anomaly detection is performed through clustering analysis. This grouping mechanism distinguishes true device failure characteristics from temporary noise, for example, classifying short-term temperature fluctuations and persistent overtemperatures into different clusters. After obtaining a set of normal embedding vectors, the underlying association rules are further mined. These rules are converted into classification boundaries using a random forest algorithm, enabling state prediction to not only output discrete labels but also quantify the degree of risk. When the prediction results are abnormal, a simple alarm is not issued but a real-time update mechanism of the knowledge graph is triggered. This dynamic feedback ensures that the newly generated embedding vectors reflect the latest system status. The final set of optimized vectors is verified for consistency through secondary convolution to ensure that the newly added information does not destroy the original topological structure.

[0030] The technical benefits of this closed-loop processing are reflected in three aspects: first, the weighted embedding mechanism enables the model to focus on truly high-risk nodes, avoiding the waste of resources caused by uniform calculations on the entire graph; second, the combination of anomaly detection and classification algorithms can identify gradual failure modes (such as slow aging of equipment) that are difficult to capture with traditional methods; finally, the dynamic update mechanism enables the system to have the ability to continuously evolve. When the factory adds new sensors or revises operating procedures, the embedded representation can automatically adapt to the new data distribution.

[0031] Reference Figure 4As shown, for the "identifying risk patterns through an attention mechanism" step proposed in the above embodiment, this embodiment constructs a progressive process of "strength analysis - pattern mining - visual presentation." First, the node relationships extracted from the embedded representation are not simply calculated by calculating the connection weights between each pair. Instead, the attention mechanism comprehensively analyzes the overall impact of upstream and downstream nodes. For example, in a power system, not only is the direct connection strength of "transformer-circuit breaker" calculated, but the composite weight of the "transformer-relay protection-circuit breaker" path is also evaluated. When certain connection strengths exceed a preset threshold, they are not mechanically marked as high risk, but are instead incorporated into a risk pattern set for deeper analysis. For example, the multi-node association of "excessive oil temperature - cooling system failure - sudden load increase" at a substation is identified as a high-risk combination. Within this initially identified risk pattern set, a clustering algorithm is further employed to mine potential patterns. The innovation of this step lies in its ability to discover seemingly unrelated but inherently interconnected fault modes. For example, in power grid analysis, it may be discovered that seemingly independent events, such as "lightning tripping" and "insulator flashover," are actually related to the hidden pattern of "humid weather + equipment aging." By calculating the strength of correlations between patterns, a network structure of risk patterns can be constructed, rather than a list of isolated patterns. This in-depth analysis allows the resulting visualization to not only showcase surface risks but also reveal the underlying mechanisms of risk transmission. For example, by overlaying a heat map with a network diagram, it can be intuitively shown which areas of the power grid have highly correlated risk patterns and which are isolated localized failures.

[0032] The technical benefits of this design are primarily reflected in three aspects: First, multi-level pattern analysis can identify systemic risks that are difficult to detect with traditional methods, such as cross-regional cascading failure patterns in power systems; second, clustering and correlation strength calculations enable the system to automatically categorize risks, reducing reliance on manual empirical rules; and finally, innovative visualization methods enable operations and maintenance personnel to quickly understand complex risk networks. For example, color coding and line thickness can be used to intuitively determine the similarity and degree of correlation between risk patterns at different substations. Particularly in systems like power grids, which have numerous devices and complex connections, the technical methods provided in this embodiment can effectively address the industry pain point of "the difficulty of correlating and analyzing massive amounts of alarms." Through systematic pattern mining and intuitive visualization, dispersed risk signals are integrated into an actionable risk awareness framework, providing decision support for preventing major incidents such as large-scale power outages.

[0033] Reference Figure 5As shown, for the "generating a risk propagation sequence" step proposed in the above embodiment, this embodiment constructs a progressive analysis framework of "path search - strength analysis - regional location - anomaly detection." First, using breadth-first search to extract the connection paths of related entities from the knowledge graph, not only does it capture topological connectivity but, more importantly, it also performs weighted calculations based on the dynamic attribute values ​​of each entity. For example, in a power system, when analyzing the risk of "transformer overheating," not only are directly connected devices such as circuit breakers and busbars identified, but the probability intensity of risk propagation along different paths is also calculated based on attribute values ​​such as real-time load factor and insulation status, forming a risk transmission matrix. This quantitative analysis effectively distinguishes the actual risk transmission differences between the "transformer-circuit breaker-feeder" path and the "transformer-cooling system-control cabinet" path, avoiding the drawback of traditional methods that treat all paths equally.

[0034] After identifying high-intensity transmission paths, a matrix factorization algorithm is used to extract the primary propagation direction. The innovation of this step lies in its ability to focus on the most critical risk transmission chain from a large number of possible paths. For example, in power grid analysis, the subsequence of "generator excitation failure → voltage fluctuation → protection malfunction → large-scale power outage" may be identified as the core transmission path of concern, rather than other secondary impact paths. Cluster analysis is then used to identify risk concentration areas. This not only considers the number of entities but also analyzes their topological characteristics and attribute similarity. For example, multiple overloaded devices within the same substation are automatically grouped into a high-risk cluster, rather than simply grouping them by physical distance. This intelligent grouping enables more precise location of risk areas. In particular, by mapping entity clusters into a low-dimensional space using a graph embedding algorithm, semantic relationships previously difficult to detect in a high-dimensional knowledge graph become clearly discernible. For example, it may be possible to discover that "improper relay protection settings" and "frequent malfunctions" are highly correlated in the semantic space.

[0035] Ultimately, when anomaly detection algorithms identify potentially abnormal entities, they can uncover marginal risk points that are easily overlooked using traditional methods. For example, in a power grid, a circuit breaker located far from a fault point and with a low load factor may exhibit abnormal characteristics in the semantic relationship vector due to its unique protection setting, thus being identified as a potential risk amplifier. This in-depth analysis allows the resulting risk propagation sequence to not only include explicit primary propagation paths but also capture implicit risk-enhancing factors, providing more comprehensive risk warnings for the safe operation of the power system.

[0036] The technical benefits of this method are primarily reflected in the following: First, a weighted risk transfer matrix enables the system to quantitatively assess the actual threat level of different paths; second, the combined use of matrix decomposition and cluster analysis can intelligently extract key risk propagation frameworks from complex power grid topologies; and finally, graph embedding-based anomaly detection can identify hidden risk nodes that are difficult to identify with traditional monitoring methods. These innovations enable power system operators to not only see how risks propagate, but also understand why certain paths are more likely to become risk transmission channels, providing a scientific basis for formulating precise prevention and control strategies.

[0037] Reference Figure 6 As shown, in response to the step of "determining the dynamic propagation path" proposed in the above embodiment, this embodiment constructs a closed-loop optimization system of "dynamic perception-real-time adjustment-trend prediction". First, when extracting entity data from the risk propagation sequence, special emphasis is placed on the fusion processing of multi-dimensional time series features. For example, in the power system scenario, not only conventional monitoring data such as transformer oil temperature and load rate are collected, but also time series features such as meteorological environment changes and maintenance record updates are integrated. These dynamically changing feature vectors are mapped to the network topology structure in real time through the graph neural network. This processing method allows the originally static device connection relationship to dynamically adjust the weight according to the operating status. For example, when the insulator contamination level of a transmission line is monitored to continue to rise, the connection strength of the line in the risk propagation network will be automatically enhanced, while traditional methods can often only maintain fixed topological relationships.

[0038] A recursive update mechanism, which calculates entity state changes, enables truly dynamic risk assessment. For example, for a substation circuit breaker, this system not only determines whether the current number of trips exceeds the specified limit but also predicts the remaining lifespan based on historical operating data. When the predicted state exceeds a threshold, it immediately triggers an adjustment to the propagation path weight. This mechanism can capture gradual changes in power equipment performance. For example, the aging of capacitor banks can escalate the previously minor "harmonic amplification-protection misoperation" path into a major risk channel. When analyzing propagation path predictions, a state update rule base is innovatively introduced, incorporating structured knowledge such as industry standards and operational experience into the optimization process. For example, if the predicted risk probability for the "typhoon weather → conductor galloping → phase-to-phase short circuit" path continues to increase, the monitoring level of the relevant line will be automatically increased in accordance with the "Guidelines for Power System Safety and Stability," and the protection settings of surrounding equipment will be adjusted accordingly.

[0039] Ultimately, the technology demonstrates unique adaptive capabilities when determining dynamic propagation paths through the fusion of time-series data. In actual grid operation, this technology can identify risk transmission patterns that are invisible to traditional static analysis, such as how the shift of load centers changes the direction of fault propagation or how the intermittent output of renewable energy power plants affects the distribution of system vulnerabilities.

[0040] The technical benefits of this method are primarily reflected in three aspects: First, the fusion of time-series features enables risk path assessment to closely track changes in system operating status, resolving the issue of outdated assessment results associated with traditional methods. Second, a recursive update mechanism empowers the system to identify evolving risks early, providing early warning of potential propagation paths before complete equipment failure. Finally, the integration of industry rules into the optimization process ensures that the technical solution is both data-driven and forward-looking while also meeting the practical constraints of power safety production. This dynamic risk assessment mechanism is particularly well-suited to the complex and highly uncertain environment of modern power systems. It provides dispatchers with a real-time, updated risk landscape, rather than a static snapshot, significantly enhancing the ability to prevent cascading failures.

[0041] Reference Figure 7 As shown, in response to the "generating a refined risk model" step proposed in the above embodiment, this embodiment constructs a quantitative analysis framework for "equipment-personnel" cross-dimensional risk interactions. First, when extracting key node subgraphs from dynamic propagation paths, a graph segmentation algorithm is used that not only considers the strength of topological connections but also innovatively introduces cross-dimensional correlation analysis. For example, in power system operation and maintenance scenarios, transformer oil temperature anomaly nodes and relevant operator inspection record nodes are included in the same subgraph for analysis. When adjusting path edge weights through the attention mechanism, it is possible to automatically identify which human behaviors have a substantial impact on equipment status. For example, in substation operations, it may be found that the behavior of "skipping the infrared temperature measurement step" is strongly correlated with the "bushing joint overheating" fault, while conventional inspection frequency adjustments have a weaker impact. This fine-grained weight distribution makes risk analysis more accurate.

[0042] Using principal component analysis to extract key interaction patterns enables the extraction of core risk patterns from massive amounts of monitoring data. For example, in relay protection devices, the correlation between "improper protection setting modification" and "protection malfunction" is not a simple linear relationship, but rather a specific combination of conditions. For example, the impact is particularly significant when the system load factor exceeds 85%. When the intensity of the interaction exceeds a threshold, a clustering algorithm can intelligently group these complex relationships, identifying typical risk patterns in different scenarios, such as "nighttime operation," "new employee operation," and "shift handover." This analytical depth goes far beyond the simple cause-and-effect statistics of traditional methods, revealing the underlying mechanisms by which human behavior influences equipment status. For example, it may be discovered that operators are more likely to overlook certain inspection steps in high-temperature environments, and this behavioral pattern, in turn, increases the missed detection rate of specific equipment defects.

[0043] When generating refined risk models, a creative risk assessment indicator is constructed that integrates device anomaly signals and behavioral trigger conditions. In a power dispatch scenario, this indicator might be expressed as a weighted combination of "dispatch instruction response delay" and "renewable energy output fluctuation range," rather than evaluating either aspect in isolation. This cross-dimensional integration makes the resulting risk model more interpretable and actionable. For example, it can clearly indicate that "the risk of wind farm collector line overload increases significantly when wind speed changes suddenly by more than 30% and the operator has less than three years of experience."

[0044] The technical benefits of this method are primarily reflected in the following: First, the combined use of graph segmentation and attention mechanisms enables the system to accurately locate key interaction nodes within complex production networks; second, the combined application of principal component analysis and clustering algorithms extracts meaningful "equipment-personnel" interaction patterns; and finally, the risk assessment indicators generated through weighted fusion provide a quantitative basis for developing targeted prevention and control measures. This technology is particularly well-suited for the complex environment of modern power systems, which rely heavily on human-machine collaboration. It helps safety managers not only identify equipment anomalies but also understand the human factors behind them, thereby shifting from passive response to proactive prevention and effectively reducing systemic risks caused by human factors.

[0045] Reference Figure 7 As shown, for the "determining the final risk identification result" step proposed in the above embodiment, this embodiment proposes a three-dimensional risk identification system combining "environmental factor integration - dynamic prediction - knowledge evolution." First, when extracting propagation chain features from refined risk models, this innovatively incorporates deep integration of environmental factor data. In the power system scenario, this includes not only conventional meteorological data such as temperature and humidity, but also spatial environmental information such as construction activities and vegetation growth around the power grid. By analyzing inter-node propagation relationships through a graph convolutional network, it is possible to dynamically assess the modulation of environmental factors on risk transmission paths. For example, it discovered a hidden correlation between the "risk of conductor galloping during typhoon season" and the "daily tree barrier clearing cycle," a correlation often overlooked in traditional static analysis. When vegetation growth in a region is detected to exceed a threshold, the connection weight between the device nodes in that region and the tree barrier risk features is automatically increased, enabling proactive awareness of environmental risks.

[0046] When extracting interactive features between equipment status and human behavior, principal component analysis is used to remove redundant information and focus on core risk drivers. For example, during a substation inspection, it may be found that the environmental condition of "inadequate nighttime lighting" exacerbates the impact of the behavior of "streamlining inspection procedures" on equipment defect detection rates. This interactive pattern is difficult to detect using conventional data analysis. When the correlation strength of the interactive features exceeds a threshold, the K-means clustering algorithm can intelligently identify typical risk combinations under different environmental scenarios. For example, it can distinguish between behavioral deviations and corresponding equipment vulnerabilities under different meteorological conditions, such as "hot and dry," "humid and rainy," and "low and freezing temperatures." This contextual awareness capability enables risk analysis to move from a two-dimensional perspective to a three-dimensional one. For example, in power grid dispatching, it can identify the risk signature of transformer overload caused by the combination of "high summer temperatures" and "surges in air conditioning load."

[0047] The generation of comprehensive risk indicators creatively integrates the three dimensions of equipment status, human behavior, and environmental factors. This process is not a simple data overlay; rather, it dynamically adjusts the contribution of each dimension through weighted summation. For example, when assessing the lightning strike risk of a transmission line, the system considers the line's lightning protection configuration (equipment status), the quality of inspection and maintenance (human behavior), and thunderstorm activity forecasts (environmental factors), ultimately outputting a risk indicator with temporal and spatial characteristics. Time series analysis based on this indicator can predict the evolution of risk transmission paths, for example, predicting the speed and scope of the chain reaction: "persistent drought → vegetation desiccation → wildfire risk → line tripping." Finally, node weight adjustments, implemented through a knowledge graph update mechanism, possess continuous evolutionary capabilities, enabling rapid adjustments to the risk identification model when new environmental risk factors emerge, such as abnormal insulator contamination caused by new pollutants.

[0048] The technical benefits of this approach are primarily reflected in the following: First, the deep integration of environmental factors makes risk identification results more spatially and temporally specific; second, dynamic prediction capabilities enable the system to shift from passive response to active defense; and finally, the self-updating mechanism of the knowledge graph ensures the continued adaptability of the technical solution. This three-dimensional, integrated risk identification framework is particularly well-suited to the complex operating environments faced by modern power systems. It helps operations and maintenance personnel understand the evolution of risks from a more comprehensive perspective, provides a scientific basis for developing precise prevention and control strategies, and effectively enhances early warning capabilities for complex risks.

[0049] Reference Figure 9 As shown, in order to implement the above method, the present invention also provides a safety production risk identification system based on knowledge graph, including: The data acquisition module is used to extract entity nodes and relationship edge data from the knowledge graph and generate multi-dimensional feature vectors; A graph neural network processing module, which performs convolution operations on entity nodes and relationship edges, aggregates neighborhood information, and generates embedded representations; A risk pattern calculation module, based on the embedded representation, uses an attention mechanism to calculate the weight distribution between entity nodes and determine the risk pattern; A risk propagation sequence generation module is used to traverse the associated paths in the knowledge graph according to the risk pattern to generate a risk propagation sequence; Dynamic path update module, used to integrate the temporal changes of multi-dimensional feature vectors, update entity states and determine dynamic propagation paths; A key subgraph extraction module is used to extract key node subgraphs from the dynamic propagation path, adjust the path edge weights to highlight the interactive impact between device status and human behavior, and generate a refined risk model; The risk identification module is used to perform iterative queries based on refined risk patterns, integrate additional features of environmental factors, determine the complete transmission chain and output the final risk identification results.

[0050] Obviously, the above embodiments are merely examples for clarity of explanation and are not intended to limit the implementation methods. Those skilled in the art will appreciate that other variations or modifications can be made based on the above description. It is not necessary and impossible to enumerate all implementation methods here. Obvious variations or modifications arising therefrom remain within the scope of protection of the present invention.

Claims

1. A method for identifying production safety risks based on knowledge graph, characterized in that: include: Obtain entity node and relationship edge data in the knowledge graph and extract multi-dimensional feature vectors; A graph neural network is used to perform convolution operations on entity nodes and relationship edges, aggregate neighborhood information, and generate embedded representations; Through embedding representation, the attention mechanism is used to calculate the weight distribution between entity nodes and determine the risk pattern; According to the risk pattern, traverse the associated paths in the knowledge graph to generate a risk propagation sequence; According to the risk propagation sequence, the temporal changes of multi-dimensional feature vectors are integrated to update the entity status and determine the dynamic propagation path; Extracting key node subgraphs from the dynamic propagation path, adjusting path edge weights, highlighting the interactive impact between device status and human behavior, and generating refined risk models; Based on the refined risk pattern representation, the knowledge graph is iteratively queried, additional features of environmental factors are integrated, the complete transmission chain is judged, and the final risk identification result is obtained.

2. The method for identifying production safety risks based on knowledge graph according to claim 1, characterized in that: Extract multi-dimensional feature vectors, including: By querying the preset knowledge graph database interface, an initial multi-dimensional feature vector is obtained from the entity nodes of device status, human behavior, and environmental factors, including temperature readings that record device status, operation logs that record human behavior, and humidity indicators that record environmental factors. The temperature readings, operation logs, and humidity indicators constitute a feature set; The k-nearest neighbor algorithm is used to perform cluster analysis on the feature set, and the temperature readings, humidity indicators, and operation logs are grouped according to their similarities to obtain the feature clustering results; If there are abnormal groups in the feature clustering results, the abnormal feature vectors are isolated through the anomaly detection algorithm, and the anomaly is judged based on the preset threshold to obtain the normal feature set; Based on the normal feature set, contextual relationships are extracted from the knowledge graph to generate an association rule set. Based on the association rule set, a decision tree algorithm is used to classify and predict temperature readings, humidity indicators, and operation logs to obtain device status prediction results. If the device status prediction result deviates from the normal range, the entity node attributes in the knowledge graph are updated through the database interface to obtain the updated feature vector. Based on the updated feature vector, cluster analysis is re-executed to verify consistency and obtain the optimized feature set.

3. The method for identifying production safety risks based on knowledge graph according to claim 1, characterized in that: A graph neural network is used to perform convolution operations on entity nodes and relationship edges, aggregate neighborhood information, and generate embedded representations, including: Through the preset knowledge graph database interface, we obtain entity node and relationship edge data, use graph neural network to perform convolution operations, aggregate neighborhood information, and generate initial embedding vectors; According to the initial embedding vector, the attention mechanism is used to assign weights, highlight the local features of key entity nodes and relationship edges, and generate a weighted embedding vector; If the similarity between the weighted embedding vector and the preset topological structure template is lower than the preset threshold, the embedding vectors are grouped through cluster analysis to determine the abnormal embedding vectors and obtain a set of normal embedding vectors; According to the normal embedding vector set, context information is extracted from the knowledge graph to generate an association rule set. Based on the association rule set, the normal embedding vector is classified using the random forest algorithm to determine the state prediction result of the entity node; If the state prediction result deviates from the preset normal range, the entity node attributes in the knowledge graph are updated through the database interface to generate an updated embedding vector. Based on the updated embedding vector, the graph neural network convolution operation is re-executed to verify the consistency and obtain the optimized embedding vector set.

4. The method for identifying production safety risks based on knowledge graph according to claim 1, characterized in that: The attention mechanism is used to calculate the weight distribution between entity nodes and determine the risk pattern, including: Extract node relationships from the embedded representation, use the attention mechanism to calculate the weight distribution between entity nodes, obtain the connection strength, and analyze risk patterns based on the connection strength; If the connection strength exceeds the preset threshold, it is judged as a high-risk association and a risk pattern set is obtained; For the risk pattern set, a clustering algorithm is used to mine potential patterns and obtain pattern classification results. Features are extracted from the pattern classification results, and data associations are analyzed to obtain potential hidden patterns. Through the potential hidden patterns, the association strength between patterns is calculated to obtain the final pattern distribution. Based on the final pattern distribution, a visual representation of the risk pattern is generated to obtain the analysis results.

5. The method for identifying production safety risks based on knowledge graph according to claim 1, characterized in that: According to the risk pattern, the associated paths in the knowledge graph are traversed to generate a risk propagation sequence, including: Extract related entities from the knowledge graph, use breadth-first search to traverse the connection paths between entities, and obtain a preliminary risk propagation sequence. Obtain the attribute value of each entity in the path, determine the risk transmission intensity between entities through weighted calculation, and generate a risk transmission matrix. If there is a high-intensity transmission path in the risk transmission matrix, extract the main risk transmission direction through matrix decomposition algorithm to obtain the key transmission subsequence; Based on the key propagation subsequence, cluster analysis is used to group entities, obtain entity clusters with concentrated risks, and determine the risk concentration area. If the number of entities in the risk concentration area exceeds the preset threshold, the entity cluster is mapped to a low-dimensional space through a graph embedding algorithm to obtain a semantic relationship vector; According to the semantic relationship vector, anomaly detection algorithm is used to identify potential abnormal entities, obtain the abnormal entity set, and traverse the association path in the knowledge graph through the abnormal entity set to generate the final risk propagation sequence.

6. The method for identifying production safety risks based on knowledge graph according to claim 1, characterized in that: According to the risk propagation sequence, the temporal changes of multi-dimensional feature vectors are integrated to update the entity status and determine the dynamic propagation path, including: Obtain entity data in the risk propagation sequence, extract feature vectors from multi-dimensional time series data, obtain a set of entity feature vectors, construct a network topology structure through a graph neural network, and update the topological connections based on the set of entity feature vectors to obtain the network topology structure; A recursive update mechanism is used to iteratively calculate entity state changes based on the network topology to obtain an entity state sequence. If any entity state in the entity state sequence exceeds a preset threshold, the propagation path weight is adjusted through the graph neural network to obtain a propagation path with adjusted weights. According to the propagation path after weight adjustment, the temporal change trend of the propagation path prediction is analyzed to obtain the propagation path prediction result. Through the propagation path prediction result, the state update rule is integrated to iteratively optimize the dynamic propagation path to obtain the optimized dynamic propagation path; The optimized dynamic propagation path is adopted and combined with time series data fusion to determine the final dynamic propagation path.

7. The method for identifying production safety risks based on knowledge graph according to claim 1, characterized in that: Extract key node subgraphs from the dynamic propagation path, adjust path edge weights, highlight the interactive impact between device status and human behavior, and generate refined risk models, including: Extract key node subgraphs from dynamic propagation paths, use graph segmentation algorithms to determine node connection strengths, and obtain key node subgraphs. For key node subgraphs, use attention mechanisms to calculate the interaction influence relationships between nodes, adjust path edge weights, and generate weighted path subgraphs. The interaction features between device status data and personnel behavior data are obtained from the weighted path subgraph. Principal component analysis is used to extract the main interaction patterns and determine the interaction influence relationships. If the intensity of the interaction influence relationship exceeds a preset threshold, the interaction features are grouped using a clustering algorithm to obtain the association pattern between the behavior triggering conditions and the device abnormality signals. According to the correlation pattern, the correspondence between the abnormal signals of the equipment and the behavior triggering conditions is extracted to generate the risk assessment index. Through the risk assessment index, the weighted summation method is used to fuse the abnormal signals of the equipment and the behavior triggering conditions to generate a refined risk model.

8. The method for identifying production safety risks based on knowledge graph according to claim 1, characterized in that: Based on the refined risk pattern representation, the knowledge graph is iteratively queried, additional features of environmental factors are integrated, and the complete propagation chain is determined to obtain the final risk identification results, including: The propagation chain features are obtained from the refined risk model. The knowledge graph query method is used to extract the node attributes related to the environmental factor data to obtain the environmental correlation feature set. Based on the environmental correlation feature set, the graph convolutional network is used to analyze the propagation relationship between nodes, and the weights of the propagation chain features are updated to obtain the weighted propagation features. The correlation data between device status interaction and human behavior interaction is extracted from the weighted propagation features. The main interaction patterns are extracted using principal component analysis, and the interaction feature set is determined. If the correlation strength of the interaction feature set exceeds the preset threshold, the K-means clustering method is used to group the interaction features to obtain the correlation pattern between the behavior triggering conditions and the device abnormal signals. According to the association pattern, the correspondence between abnormal equipment signals and behavior triggering conditions is extracted, and the weighted summation method is used to integrate environmental factor data to generate a comprehensive risk index. The changing trend of the risk propagation path is extracted from the comprehensive risk index. The time series analysis method is used to predict the dynamic changes of the risk propagation path and obtain dynamic risk prediction results. According to the dynamic risk prediction results, the knowledge graph update method is used to adjust the node weights of the propagation chain characteristics and determine the final risk identification results.

9. A knowledge graph-based safety production risk identification system, characterized by: include: The data acquisition module is used to extract entity nodes and relationship edge data from the knowledge graph and generate multi-dimensional feature vectors; A graph neural network processing module, which performs convolution operations on entity nodes and relationship edges, aggregates neighborhood information, and generates embedded representations; A risk pattern calculation module, based on the embedded representation, uses an attention mechanism to calculate the weight distribution between entity nodes and determine the risk pattern; A risk propagation sequence generation module is used to traverse the associated paths in the knowledge graph according to the risk pattern to generate a risk propagation sequence; Dynamic path update module, used to integrate the temporal changes of multi-dimensional feature vectors, update entity states and determine dynamic propagation paths; A key subgraph extraction module is used to extract key node subgraphs from the dynamic propagation path, adjust the path edge weights to highlight the interactive impact between device status and human behavior, and generate a refined risk model; The risk identification module is used to perform iterative queries based on refined risk patterns, integrate additional features of environmental factors, determine the complete transmission chain and output the final risk identification results.

Images