A safety inspection system and method based on hidden danger database model

Abstract

The present invention discloses a safety inspection system and method based on a hidden danger database model, comprising the following steps: S1, establishing a hidden danger table, an inspection record table, and a rectification record table, and constructing a hidden danger rule base; S2, collecting hidden danger data, and storing it in the hidden danger database model after standardization; S3, constructing a hidden danger knowledge graph and generating a feature matrix; S4, inputting the data and the feature matrix into a time series model to generate a risk scoring matrix; S5, optimizing the hidden danger classification and updating the rule base based on the risk score and knowledge graph; S6, inputting a multimodal hidden danger identification model to classify the hidden dangers; S7, optimizing task allocation, updating the hidden danger status and storing it in the database; S8, integrating data to generate dynamic feature labels, optimizing the model, and outputting a heat map and trend report. Through multimodal data fusion and dynamic analysis, the present invention realizes intelligent management of hidden danger detection, risk assessment, task allocation, and model optimization, significantly improving the efficiency and accuracy of safety inspections.

Description

Technical Field

[0001] The present invention relates to the field of database technology, and in particular to a safety inspection system and method based on a hidden danger database model. Background Art

[0002] In recent years, with the rapid development of industrial production and social development, safety hazards have received increasing attention. Hazard management, a key component of safety inspections, aims to promptly identify potential risks and implement appropriate measures to address them, thereby reducing the likelihood of accidents. However, existing hazard inspection and management methods are generally inefficient, lack intelligence, and lack dynamic adaptability, making them difficult to meet safety management needs in complex environments.

[0003] Traditional hidden danger management systems rely primarily on manual inspections and static record-keeping. This approach is not only labor-intensive and time-consuming, but also susceptible to human influence, resulting in a high rate of missed hidden danger detection. Furthermore, traditional systems are typically based on a single data type, such as text records or single sensor data. This approach struggles to fully reflect the multidimensional nature of hidden dangers and effectively capture their dynamic evolution. Especially when hidden dangers are highly contagious or complexly interconnected, traditional systems are unable to promptly identify their transmission chains and potential threats to other areas, severely limiting the comprehensiveness and accuracy of hidden danger management.

[0004] When it comes to hazard risk assessment, existing technologies typically employ static, rule-based assessment methods. For example, risks are simply classified and graded using predefined hazard levels and handling processes. While this approach has played a role in standardized management, its limitations are also significant. On the one hand, static assessment methods cannot adapt to the dynamic changes in hazards and struggle to capture the real-time evolution of risk levels over time and environmental changes. On the other hand, this approach relies too heavily on a fixed rule base, making it difficult to handle complex multimodal hazard information, such as the dynamic connections between hazard nodes, the execution status of remediation tasks, and the dynamic changes in environmental variables.

[0005] In addition, existing technologies also have shortcomings in hidden danger classification and task allocation. Traditional classification methods usually classify hidden dangers based on a single data dimension or static rules, and are unable to integrate multimodal data characteristics to improve the accuracy and reliability of classification. This leads to low accuracy of hidden danger classification results, which further affects the efficiency of rectification task allocation and the rationality of resource scheduling. At the same time, the task allocation process is mostly based on manual experience or simple priority rules, which makes it difficult to achieve dynamic optimization and efficient use of resources. Especially when hidden dangers are highly correlated, task allocation lacks in-depth analysis of the hidden danger propagation chain, which can easily lead to unbalanced resource allocation and delayed processing of some high-risk hidden dangers.

[0006] In terms of the analytical capabilities of hidden danger management systems, existing technologies lack the ability to deeply model and comprehensively analyze the dynamic characteristics of hidden danger nodes. For example, while some systems utilize hidden danger heat maps or risk distribution diagrams, these heat maps are mostly based on static data or simple correlation models, making it difficult to reflect the dynamic changes and propagation paths of hidden dangers. Furthermore, existing systems lack the ability to predict and analyze hidden danger trends, failing to provide targeted trend analysis reports. This makes it difficult for managers to accurately determine the potential risks of hidden danger development and the priority of remediation tasks, further increasing the complexity and uncertainty of hidden danger management.

[0007] Existing hidden danger management systems also have significant shortcomings in terms of model optimization and feedback mechanisms. Traditionally, model optimization is typically performed through offline training, lacking dynamic adaptability to real-time data. This makes it difficult for model performance to adapt to changes in hidden danger scenarios, thus impacting the accuracy of hidden danger identification and task assignment. Furthermore, an imperfect feedback mechanism is a major issue. After completing task assignments, existing systems often lack a comprehensive analysis of the results of corrective action execution and are unable to iteratively optimize model performance based on historical data. This can easily lead to inefficient handling of repeated issues when multiple hidden dangers occur.

[0008] Therefore, how to provide a safety inspection system and method based on a hidden danger database model is an urgent problem that those skilled in the art need to solve. Summary of the Invention

[0009] One purpose of the present invention is to propose a safety inspection system and method based on a hidden danger database model. The present invention makes full use of advanced technologies such as multimodal data fusion, dynamic risk assessment and task optimization allocation, and describes in detail the full-process method of hidden danger information collection, knowledge graph construction, dynamic characteristic analysis, risk assessment, task allocation and model optimization. It has the advantages of high intelligence, strong dynamic adaptability, high hidden danger handling efficiency and reasonable resource allocation.

[0010] A safety inspection method based on a hidden danger database model according to an embodiment of the present invention includes the following steps:

[0011] S1. Establish a hidden danger table, inspection record table, and rectification record table, and build a hidden danger rule library based on historical accident cases and industry standards;

[0012] S2. Collect multi-source hidden danger data through the data acquisition module, perform standardization processing, form a hidden danger data set, and store it in the hidden danger database model;

[0013] S3. Based on the hidden danger dataset and hidden danger rule base, build a hidden danger association knowledge graph, extract the potential features of hidden danger nodes and edges, and generate a knowledge graph feature matrix;

[0014] S4. Input the hidden danger dataset and knowledge graph feature matrix into the time series model to perform dynamic risk analysis and generate a hidden danger risk score matrix;

[0015] S5. Combine the hidden danger risk scoring matrix and the knowledge graph feature matrix, perform causal inference through the Bayesian network model, analyze the hidden danger propagation chain and impact range, optimize the hidden danger risk classification, and update the hidden danger rule library;

[0016] S6. Combine the hidden danger dataset, knowledge graph feature matrix, and time series features, input into the multimodal hidden danger identification model, and classify the hidden dangers;

[0017] S7. Combine the hidden danger risk classification results, classified hidden danger results, and knowledge graph feature matrix to optimize the rectification task allocation strategy, update the hidden danger status in real time, and store the rectification plan in the hidden danger database model;

[0018] S8. Integrate the hidden danger risk grading results, hidden danger classification results, knowledge graph feature matrix and rectification task execution data to generate hidden danger dynamic characteristic labels, input the time series model and multimodal hidden danger identification model, iterate the model performance, and generate hidden danger heat map and trend analysis report.

[0019] Optionally, the S3 specifically includes:

[0020] S31. Extract hidden danger features from the hidden danger dataset, including the hidden danger description information, risk level, occurrence frequency, and associated inspection and rectification records;

[0021] S32. Classify the hidden danger data and mark the type and attribute of the hidden danger according to the rules of the hidden danger rule library;

[0022] S33. Establishing a knowledge graph of hidden dangers G= ( V,E ) , where V represents the set of hidden danger nodes, including the type, occurrence time, and location of the hidden danger, and E represents the relationship between hidden danger nodes, including causal relationship, temporal association, and spatial association;

[0023] S34. In the constructed hidden danger association knowledge graph G, an embedding optimization algorithm combining the dynamic characteristics of time series and the hidden danger association weight is used to embed the hidden danger node v i The eigenvector of ∈V is expressed as:

[0024]

[0025] Among them, f ( v i) For node v i The final embedding vector representation of W1, W2 and W3 are the trainable weight matrices of node features, time series features and neighborhood features respectively, T is the total number of time steps, αt is the attention weight at time step t, x i,t For node v i Time series features at time step t, β ij For node v i and neighboring node v j The association weight between ( v i ) For node v i The set of neighboring nodes, x j,t For node v j Time series features at time step t, h j is the neighborhood node v j The original eigenvector, h i is the neighborhood node v i The original eigenvector of g ( h j ,x j,t ) is a nonlinear combination function;

[0026] S35. Extract the hidden danger node features and edge features after graph embedding to generate the knowledge graph feature matrix F G , where F G =[f(v1),f(v2),...,f(v n )], each f(v i ) is the node v i The embedded feature vector of .

[0027] Optionally, the S4 specifically includes:

[0028] S41. Define the hidden danger dataset and knowledge graph feature matrix as X={x1,x2,…,x n} and F G ={f1,f2,…,f n}, where n represents the number of nodes with hidden dangers, x i is the original feature vector of the potential danger node i, f i is the knowledge graph feature vector of hidden danger node i;

[0029] S42, integrate the hidden danger dataset features and knowledge graph features to form a comprehensive feature vector z i , which is expressed as:

[0030] z i =σ(W1x i +W2f i );

[0031] Among them, W1 and W2 are trainable weight matrices, and σ is the activation function;

[0032] S43, the comprehensive feature vector z i Input time series model, define the input sequence Z of the time series model t ={z 1,t ,z 2,t ,…,z n,t}, where t is the time step, z i,t is the comprehensive characteristics of the potential danger node i at time step t;

[0033] S44. Calculate the dynamic risk score of hidden danger nodes through the time series model:

[0034]

[0035] Among them, r i,t is the risk score of potential danger node i at time step t, W3 and W4 are trainable weight matrices, Softmax is the normalization function, z i,k is the comprehensive feature of hidden danger node i at time step k, T is the total number of time steps in the time series, α k is the attention weight at time step k:

[0036]

[0037] in, is the attention score at time step k, is the attention score at time step j, exp is the exponential function;

[0038] S45, normalize the dynamic risk scores of all hidden danger nodes to generate a hidden danger risk score matrix R = {r i,t}.

[0039] Optionally, the S5 specifically includes:

[0040] S51, the hidden danger risk score matrix R and the knowledge graph feature matrix F G Input the Bayesian network model to construct a causal relationship diagram of hidden danger propagation;

[0041] S52. Define the Bayesian network as BN = (N, L, P), where the set of potential danger nodes N = {n1, n2, ..., n n}, the directed edge L between the potential hazard nodes = {l ij}, P is the conditional probability distribution, defined as:

[0042]

[0043] Among them, Pa(n i ) is the potential danger node n i The parent node set of θ kis the conditional probability of the k-th parent node, x k is a state variable;

[0044] S53, according to the knowledge graph feature matrix F G And the hidden danger propagation causal relationship diagram, calculate the hidden danger node n i The propagation chain weight w i :

[0045]

[0046] Among them, β ij For node n i and parent node n j The correlation strength, r j is the parent node n j Dynamic risk score, f j is the parent node n j The knowledge graph feature vector, j is the index, g(r j ,f j ) is the parent node n j The nonlinear combination function of the risk score and feature matrix:

[0047] g(r j ,f j )=ReLU(W q r j +W e f j );

[0048] Among them, ReLU is the activation function, W q and W e is a trainable weight matrix;

[0049] S54, normalize the propagation chain weights to generate the hidden danger propagation chain matrix W = {w i}, where hidden danger node n i The propagation chain weight w i Expressed as:

[0050]

[0051] Among them, δ is the adjustment coefficient, h l For node n i The static propagation weight, l, k and m are indexes, n k is the node, h m For node n m The static propagation weight, N is the set of all nodes, β kj is the propagation weight, Pa(n k ) is node n k The parent node set of

[0052] S55, transform the hidden danger propagation chain matrix W={w i Hidden danger risk scoring matrix R and knowledge graph feature matrix F G Combined to generate optimized hidden danger risk classification results i :

[0053]

[0054] Among them, γ1, γ2, γ3 are the adjustment coefficients of the transmission chain weight, dynamic risk score and neighborhood comprehensive characteristics respectively, Softmax is the normalization function, N(n i ) is the neighborhood node set;

[0055] S56. Update the optimized hidden danger risk classification results to the hidden danger rule library.

[0056] Optionally, the S6 specifically includes:

[0057] S61, extracting a multimodal feature set of hidden danger nodes, including hidden danger dataset feature X={x1,x2,…,x n}、Knowledge graph feature matrix F G ={f1,f2,…,f n} and time series feature matrix T={t1,t2,…,t n}, for each node n i Construct feature fusion input:

[0058]

[0059] in, For node n i The initial multimodal feature vector of , ReLU is the activation function, R1, R2 and R3 are trainable weight matrices;

[0060] S62. Update the feature vector of the potential danger node through the multi-layer implicit feature propagation module:

[0061]

[0062] in, For node n i The feature vector after propagation in the l+1 layer, N(n i ) is the set of neighborhood nodes, is the attention weight, For node n j The feature vector after propagation in layer l, For node n iThe feature vector after propagation in layer l, R4 and R5 are trainable weight matrices, and j is the index;

[0063] S63. Combined with the dynamic feature adjustment mechanism, according to the time series feature T, the attention weight is updated through the dynamic adjustment mechanism to generate the node feature propagation matrix:

[0064]

[0065] in, is the final attention weight, e ij and e ik is the basic attention score, is the time dynamic characteristic weight, ψ j and ψ k is the global importance weight, k1 and k2 are adjustment coefficients, exp is the exponential function, and k is the index;

[0066] S64 takes the final feature vector of the hidden danger node Input the multimodal hidden danger identification model and generate classification results based on the multi-objective optimization mechanism and dynamic characteristic enhancement mechanism:

[0067]

[0068] Among them, y i is the classification result, Softmax is the normalization function, g(z j ,r j ) is the neighborhood feature aggregation function, is the prediction auxiliary module function, R6 is the trainable weight matrix, λ1 and λ2 are the adjustment coefficients, α ij is the attention weight, z j is the eigenvector, r j is the risk score, p i To predict auxiliary characteristics;

[0069] S65. Output the hidden danger classification results, including hidden danger category, hidden danger risk level and key neighborhood influencing factors, and update the classification fields of the hidden danger database model.

[0070] Optionally, the S7 specifically includes:

[0071] S71. Extract hidden danger risk classification result S = {s1,s2,…,s n}、Hazard classification result Y={y1,y2,…,y n} and the knowledge graph feature matrix F G ={f1,f2,…,f n};

[0072] S72. Build an optimized task allocation model to determine the task priority of hidden danger nodes based on risk classification and hidden danger category:

[0073]

[0074] Among them, p i is the task priority, Softmax is the normalization function, B1, B2 and B3 are adjustment coefficients, g(f j ) is the characteristic mapping function, β ij is the attention weight, s i is the risk grading score, y i is the classification result, f j is the knowledge graph feature vector, h(y i B is the hidden danger classification score function, N(n i ) is the neighborhood node set, j is the index;

[0075] S73, based on task priority p i , generate an optimized task allocation plan:

[0076]

[0077] Among them, Y is the task allocation matrix, c i is the rectification cost, d ij is the correlation distance, ξ ij is the task allocation constraint variable, N is the total number of nodes, argmin is the matrix Y for adjusting the task allocation scheme to achieve the optimal balance between the total task cost and the coordination of task allocation, and i and j are indexes;

[0078] S74. Optimize the task allocation strategy through reinforcement learning and use the policy gradient method to update the task allocation model parameters. The loss function is defined as:

[0079]

[0080] in, The objective loss for task allocation optimization, π(a i |s i ) is in state s i Next select action a i The probability of ln is the logarithmic function, r i is the reward function, E is the expected symbol;

[0081] S75. Output the optimized task allocation result and update the task execution field of the hidden danger database model.

[0082] Optionally, the S8 specifically includes:

[0083] S81. Integrate the hidden danger risk classification results S = {s1, s2, ..., s n}、Hazard classification result Y={y1,y2,…,y n}、Rectification task execution data T={t1,t2,…,t n} and the knowledge graph feature matrix F G ={f1,f2,…,f n}, construct the hidden danger dynamic characteristic label L d ={l1,l2,…,l n}:

[0084]

[0085] Among them, Q1, Q2, Q3 and Q4 are trainable weight matrices, ReLU is the activation function, s i is the risk grading score, y i is the hidden danger classification result, t i The data vector for the rectification task, f i is the knowledge graph feature vector, is the final feature vector, l i is a dynamic feature label, h(y i ) is the hidden danger classification score function;

[0086] S82, label the hidden danger dynamic characteristic L d Input the hidden danger feature propagation module, combine the time series model and the multimodal hidden danger identification model, and dynamically update the hidden danger node features:

[0087]

[0088] in, is the potential danger node n i The updated feature vector of is the final feature vector of the hidden danger node, Q5 and Q6 are trainable weight matrices, α ij is the attention weight, g(z j ,f j ) is the neighborhood feature combination function, ReLU is the nonlinear activation function, z j is the characteristic vector of the neighborhood node, f j is the knowledge graph feature, N(n i ) is the neighborhood node set, j is the index;

[0089] S83. Based on the updated hidden danger node characteristics, calculate the dynamic correlation heat between the hidden danger nodes and generate a hidden danger heat map H = {H ij}:

[0090]

[0091] in, is the potential danger node n j The updated feature vector, cosine_similarity is the cosine similarity function, H ij is the potential danger node n i and n j Dynamic correlation heat;

[0092] S84. Using the hidden danger heat map H and the hidden danger dynamic characteristic label L d , analyze the dynamic evolution trend of hidden danger nodes and generate trend analysis reports:

[0093]

[0094] Among them, Trend_Score i is the potential danger node n i Dynamic trend score, is the dynamic risk change rate, λ q ,λ w and λ e is the weight parameter, cosine_similarity(l i ,l j ) is the dynamic feature label similarity, t i The time for completing the rectification task;

[0095] S85. Combine the hidden danger heat map and trend analysis report to update the time series model and multimodal hidden danger identification model and iterate the model performance:

[0096]

[0097] in, is the loss function for iterative optimization, E is the expected symbol, is the true eigenvector, is the weight parameter, is the predicted distribution, is the actual distribution, N is the total number of nodes, is the KL divergence between the predicted distribution and the true distribution, and i is the index.

[0098] The safety inspection system based on the hidden danger database model according to an embodiment of the present invention includes the following modules:

[0099] Data collection and storage module, used to collect multi-source hidden danger data;

[0100] The knowledge graph construction module is used to construct a hidden danger related knowledge graph and extract node features and association relationships;

[0101] Risk assessment and hidden danger classification module, used to generate hidden danger risk scores and classify hidden danger nodes;

[0102] Dynamic characteristic analysis and heat map generation module, used to analyze the dynamic characteristic labels of hidden danger nodes and generate hidden danger heat maps;

[0103] Risk assessment and hidden danger classification module, used to perform dynamic risk analysis and classification of hidden danger nodes;

[0104] Task allocation and optimization module, used to optimize task allocation strategies and dynamically adjust resource allocation and priorities;

[0105] Trend analysis and report generation module, used to generate hidden danger trend analysis reports and decision support information;

[0106] Model training and optimization module, used to optimize the performance of time series models and multimodal hidden danger identification models.

[0107] The beneficial effects of the present invention are:

[0108] The present invention significantly improves the intelligence level and dynamic adaptability of hidden danger management by constructing a safety inspection system and method based on a hidden danger database model. Compared with the existing technology, the present invention integrates multiple technical means such as multimodal data fusion, knowledge graph construction, dynamic characteristic analysis, time series risk assessment and task optimization allocation, and realizes full-process intelligent management from hidden danger information collection to analysis, classification, allocation and feedback optimization. Through the combination of dynamic characteristic labels and hidden danger heat maps, the system can accurately capture the dynamic changes of hidden danger nodes and conduct in-depth analysis of hidden danger propagation paths and potential impact ranges, solving the problem that traditional hidden danger management systems are difficult to dynamically adapt to risk evolution.

[0109] Furthermore, the present invention utilizes a time series model and a multimodal hidden danger identification model to dynamically assess and accurately classify hidden danger risks, improving the accuracy and real-time nature of risk grading. Simultaneously, combined with the task allocation and optimization module, the system can dynamically adjust the priority of rectification tasks and resource allocation strategies based on risk grading and classification results, achieving efficient resource utilization and prioritizing high-risk hidden dangers. Through an iterative optimization mechanism, the present invention can continuously improve model performance based on historical data, giving the system greater robustness and predictive capabilities when dealing with complex hidden danger scenarios.

[0110] This invention effectively overcomes the shortcomings of existing technologies, including insufficient hidden danger data analysis capabilities, static risk assessment methods, inefficient task allocation, and imperfect model optimization and feedback mechanisms. It significantly improves the overall efficiency and reliability of hidden danger management. The hidden danger trend analysis reports generated by the system provide decision-making support for managers, helping to identify and prevent further expansion of safety hazards earlier, ultimately achieving intelligent, precise, and efficient hidden danger management. BRIEF DESCRIPTION OF THE DRAWINGS

[0111] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0112] Figure 1 This is a flow chart of a safety inspection method based on a hidden danger database model proposed by the present invention;

[0113] Figure 2 This is a structural diagram of a safety inspection system based on a hidden danger database model proposed by the present invention. DETAILED DESCRIPTION

[0114] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0115] refer to Figure 1 , a safety inspection method based on a hidden danger database model, comprising the following steps:

[0116] S1. Establish a hidden danger table, inspection record table, and rectification record table, and build a hidden danger rule library based on historical accident cases and industry standards;

[0117] S2. Collect multi-source hidden danger data through the data acquisition module, perform standardization processing, form a hidden danger data set, and store it in the hidden danger database model;

[0118] S3. Based on the hidden danger dataset and hidden danger rule base, build a hidden danger association knowledge graph, extract the potential features of hidden danger nodes and edges, and generate a knowledge graph feature matrix;

[0119] S4. Input the hidden danger dataset and knowledge graph feature matrix into the time series model to perform dynamic risk analysis and generate a hidden danger risk score matrix;

[0120] S5. Combine the hidden danger risk scoring matrix and the knowledge graph feature matrix, perform causal inference through the Bayesian network model, analyze the hidden danger propagation chain and impact range, optimize the hidden danger risk classification, and update the hidden danger rule library;

[0121] S6. Combine the hidden danger dataset, knowledge graph feature matrix, and time series features, input into the multimodal hidden danger identification model, and classify the hidden dangers;

[0122] S7. Combine the hidden danger risk classification results, classified hidden danger results, and knowledge graph feature matrix to optimize the rectification task allocation strategy, update the hidden danger status in real time, and store the rectification plan in the hidden danger database model;

[0123] S8. Integrate the hidden danger risk grading results, hidden danger classification results, knowledge graph feature matrix and rectification task execution data to generate hidden danger dynamic characteristic labels, input the time series model and multimodal hidden danger identification model, iterate the model performance, and generate hidden danger heat map and trend analysis report.

[0124] In this embodiment, S3 specifically includes:

[0125] S31. Extract hidden danger features from the hidden danger dataset, including the hidden danger description information, risk level, occurrence frequency, and associated inspection and rectification records;

[0126] S32. Classify the hidden danger data and mark the type and attribute of the hidden danger according to the rules of the hidden danger rule library;

[0127] S33. Establishing a knowledge graph of hidden dangers G= ( V,E ) , where V represents the set of hidden danger nodes, including the type, occurrence time, and location of the hidden danger, and E represents the relationship between hidden danger nodes, including causal relationship, temporal association, and spatial association;

[0128] S34. In the constructed hidden danger association knowledge graph G, an embedding optimization algorithm combining the dynamic characteristics of time series and the hidden danger association weight is used to embed the hidden danger node v i The eigenvector of ∈V is expressed as:

[0129]

[0130] Among them, f(v i ) is the node v i The final embedding vector representation of W1, W2 and W3 are the trainable weight matrices of node features, time series features and neighborhood features respectively, T is the total number of time steps, α t is the attention weight at time step t, x i,t For node v i Time series features at time step t, β ij For node v i and neighboring node v j The association weight between i ) is the node v i The set of neighboring nodes, x j,tFor node v j Time series features at time step t, h j is the neighborhood node v j The original eigenvector, h i is the neighborhood node v i The original eigenvector of g(h j ,x j,t ) is a nonlinear combination function;

[0131] S35. Extract the hidden danger node features and edge features after graph embedding to generate the knowledge graph feature matrix F G , where F G =[f(v1),f(v2),...,f(v n )], each f(v i ) is the node v i The embedded feature vector of .

[0132] In this embodiment, the S4 specifically includes:

[0133] S41. Define the hidden danger dataset and knowledge graph feature matrix as X={x1,x2,…,x n} and F G ={f1,f2,…,f n}, where n represents the number of nodes with hidden dangers, x i is the original feature vector of the potential danger node i, f i is the knowledge graph feature vector of hidden danger node i;

[0134] S42, integrate the hidden danger dataset features and knowledge graph features to form a comprehensive feature vector z i , which is expressed as:

[0135] z i =σ(W1x i +W2f i );

[0136] Among them, W1 and W2 are trainable weight matrices, and σ is the activation function;

[0137] S43, the comprehensive feature vector z i Input time series model, define the input sequence Z of the time series model t ={z 1,t ,z 2,t ,…,z n,t}, where t is the time step, z i,t is the comprehensive characteristics of the potential danger node i at time step t;

[0138] S44. Calculate the dynamic risk score of hidden danger nodes through the time series model:

[0139]

[0140] Among them, r i,t is the risk score of potential danger node i at time step t, W3 and W4 are trainable weight matrices, Softmax is the normalization function, z i,k is the comprehensive feature of the potential danger node i at time step k, T is the total number of time steps in the time series, α k is the attention weight at time step k:

[0141]

[0142] in, is the attention score at time step k, is the attention score at time step j, exp is the exponential function;

[0143] S45. Normalize the dynamic risk scores of all hidden danger nodes to generate a hidden danger risk score matrix R= { r i,t } .

[0144] In this embodiment, the S5 specifically includes:

[0145] S51, the hidden danger risk score matrix R and the knowledge graph feature matrix F G Input the Bayesian network model to construct a causal relationship diagram of hidden danger propagation;

[0146] S52. Define Bayesian network as BN= ( N,L,P ) , where the hidden danger node set N={n1,n2,…,n n } , the directed edge L between the potential hazard nodes = { l ij } , P is the conditional probability distribution, defined as:

[0147]

[0148] Among them, Pa ( n i) is the potential danger node n i The parent node set of θ k is the conditional probability of the k-th parent node, x k is a state variable;

[0149] S53, according to the knowledge graph feature matrix F G And the hidden danger propagation causal relationship diagram, calculate the hidden danger node n i The propagation chain weight wi :

[0150]

[0151] Among them, β ij For node n i and parent node n j The correlation strength, r j is the parent node n j Dynamic risk score, f j is the parent node n j The knowledge graph feature vector, j is the index, g ( r j ,f j ) is the parent node n j The nonlinear combination function of the risk score and feature matrix:

[0152] g ( r j ,f j ) =ReLU ( W q r j +W e f j ) ;

[0153] Among them, ReLU is the activation function, W q and W e is a trainable weight matrix;

[0154] S54. Normalize the propagation chain weights to generate the hidden danger propagation chain matrix W = { w i} , where the hidden danger node n i The propagation chain weight w i Expressed as:

[0155]

[0156] Among them, δ is the adjustment coefficient, h l For node n i The static propagation weight, l, k and m are indexes, n k is the node, h m For node n m The static propagation weight, N is the set of all nodes, β kj is the propagation weight, Pa ( n k ) is node n k The parent node set of

[0157] S55. The hidden danger transmission chain matrix W ={ w i} , hidden danger risk scoring matrix R and knowledge graph feature matrix F G Combined to generate optimized hidden danger risk classification results i :

[0158]

[0159] Among them, γ1, γ2, γ3 are the adjustment coefficients of the transmission chain weight, dynamic risk score and neighborhood comprehensive characteristics respectively, Softmax is the normalization function, N ( n i) is the set of neighborhood nodes;

[0160] S56. Update the optimized hidden danger risk classification results to the hidden danger rule library.

[0161] In this embodiment, S6 specifically includes:

[0162] S61, extracting a multimodal feature set of hidden danger nodes, including hidden danger dataset feature X={x1,x2,…,x n} , knowledge graph feature matrix F G = { f1,f2,…,f n} And the time series feature matrix T={t1,t2,…,t n} , for each node n i Construct feature fusion input:

[0163]

[0164] in, For node n i The initial multimodal feature vector of , ReLU is the activation function, R1, R2 and R3 are trainable weight matrices;

[0165] S62. Update the feature vector of the potential danger node through the multi-layer implicit feature propagation module:

[0166]

[0167] in, For node n i The feature vector after propagation in the l+1th layer, N ( n i) is the set of neighborhood nodes, is the attention weight, For node n j The feature vector after propagation in layer l, For node n iThe feature vector after propagation in layer l, R4 and R5 are trainable weight matrices, and j is the index;

[0168] S63. Combined with the dynamic feature adjustment mechanism, according to the time series feature T, the attention weight is updated through the dynamic adjustment mechanism to generate the node feature propagation matrix:

[0169]

[0170] in, is the final attention weight, e ij and e ik is the basic attention score, is the time dynamic characteristic weight, ψ j and ψ k is the global importance weight, k1 and k2 are adjustment coefficients, exp is the exponential function, and k is the index;

[0171] S64 takes the final feature vector of the hidden danger node Input the multimodal hidden danger identification model and generate classification results based on the multi-objective optimization mechanism and dynamic characteristic enhancement mechanism:

[0172]

[0173] Among them, y i is the classification result, Softmax is the normalization function, g ( z j ,r j ) is the neighborhood feature aggregation function, is the prediction auxiliary module function, R6 is the trainable weight matrix, λ1 and λ2 are the adjustment coefficients, α ij is the attention weight, z j is the eigenvector, r j is the risk score, p i To predict auxiliary characteristics;

[0174] S65. Output the hidden danger classification results, including hidden danger category, hidden danger risk level and key neighborhood influencing factors, and update the classification fields of the hidden danger database model.

[0175] In this embodiment, the S7 specifically includes:

[0176] S71. Extract hidden danger risk classification results S= { s1,s2,…,s n} 、Hazard classification result Y={y1,y2,…,y n} and the knowledge graph feature matrix F G = { f1,f2,…,f n} ;

[0177] S72. Build an optimized task allocation model to determine the task priority of hidden danger nodes based on risk classification and hidden danger category:

[0178]

[0179] Among them, p i is the task priority, Softmax is the normalization function, B1, B2 and B3 are adjustment coefficients, g ( f j ) is the characteristic mapping function, β ij is the attention weight, s i is the risk grading score, y i is the classification result, f j is the knowledge graph feature vector, h ( y i) is the hidden danger classification score function, N ( n i) is the set of neighborhood nodes, j is the index;

[0180] S73, based on task priority p i , generate an optimized task allocation plan:

[0181]

[0182] Among them, Y is the task allocation matrix, c i is the rectification cost, d ij is the correlation distance, ξ ij is the task allocation constraint variable, N is the total number of nodes, argmin is the matrix Y for adjusting the task allocation scheme to achieve the optimal balance between the total task cost and the coordination of task allocation, and i and j are indexes;

[0183] S74. Optimize the task allocation strategy through reinforcement learning and use the policy gradient method to update the task allocation model parameters. The loss function is defined as:

[0184]

[0185] in, The objective loss optimized for task allocation, π ( a i |s i) For state s i Next select action a i The probability of ln is the logarithmic function, r i is the reward function, E is the expected symbol;

[0186] S75. Output the optimized task allocation result and update the task execution field of the hidden danger database model.

[0187] In this embodiment, the S8 specifically includes:

[0188] S81. Integrate the hidden danger risk classification results S= { s1,s2,…,s n} 、Hazard classification result Y={y1,y2,…,y n} , rectification task execution data T= { t1,t2,…,t n} and the knowledge graph feature matrix F G ={f1,f2,…,f n} , construct the hidden danger dynamic characteristic label L d = { l1,l2,…,l n} :

[0189]

[0190] Among them, Q1, Q2, Q3 and Q4 are trainable weight matrices, ReLU is the activation function, s i is the risk grading score, y i is the hidden danger classification result, t i The data vector for the rectification task, f i is the knowledge graph feature vector, is the final feature vector, l i is the dynamic feature label, h ( y i) is the hidden danger classification score function;

[0191] S82, label the hidden danger dynamic characteristic L d Input the hidden danger feature propagation module, combine the time series model and the multimodal hidden danger identification model, and dynamically update the hidden danger node features:

[0192]

[0193] in, is the potential danger node n i The updated feature vector of is the final feature vector of the hidden danger node, Q5 and Q6 are trainable weight matrices, α ij is the attention weight, g ( z j ,f j ) is the neighborhood feature combination function, ReLU is the nonlinear activation function, z j is the characteristic vector of the neighborhood node, f jis the knowledge graph feature, N ( n i ) is the set of neighborhood nodes, j is the index;

[0194] S83. Based on the updated hidden danger node characteristics, calculate the dynamic correlation heat between the hidden danger nodes and generate a hidden danger heat map H= { H ij } :

[0195]

[0196] in, is the potential danger node n j The updated feature vector, cosine_similarity is the cosine similarity function, H ij is the potential danger node n i and n j Dynamic correlation heat;

[0197] S84. Using the hidden danger heat map H and the hidden danger dynamic characteristic label L d , analyze the dynamic evolution trend of hidden danger nodes and generate trend analysis reports:

[0198]

[0199] Among them, Trend_Score i is the potential danger node n i Dynamic trend score, is the dynamic risk change rate, λ q ,λ w and λ e is the weight parameter, cosine_similarity(l i ,l j ) is the dynamic feature label similarity, t i The time for completing the rectification task;

[0200] S85. Combine the hidden danger heat map and trend analysis report to update the time series model and multimodal hidden danger identification model and iterate the model performance:

[0201]

[0202] in, is the loss function for iterative optimization, E is the expected symbol, is the true eigenvector, is the weight parameter, is the predicted distribution, is the actual distribution, N is the total number of nodes, is the KL divergence between the predicted distribution and the true distribution, and i is the index.

[0203] refer to Figure 2 , a safety inspection system based on a hidden danger database model, including the following modules:

[0204] Data collection and storage module, used to collect multi-source hidden danger data;

[0205] The knowledge graph construction module is used to construct a hidden danger related knowledge graph and extract node features and association relationships;

[0206] Risk assessment and hidden danger classification module, used to generate hidden danger risk scores and classify hidden danger nodes;

[0207] Dynamic characteristic analysis and heat map generation module, used to analyze the dynamic characteristic labels of hidden danger nodes and generate hidden danger heat maps;

[0208] Risk assessment and hidden danger classification module, used to perform dynamic risk analysis and classification of hidden danger nodes;

[0209] Task allocation and optimization module, used to optimize task allocation strategies and dynamically adjust resource allocation and priorities;

[0210] Trend analysis and report generation module, used to generate hidden danger trend analysis reports and decision support information;

[0211] Model training and optimization module, used to optimize the performance of time series models and multimodal hidden danger identification models.

[0212] Example 1:

[0213] To verify the feasibility of this invention, we applied it to a large industrial park covering approximately 5 square kilometers, encompassing multiple production workshops, storage areas, and logistics sites. This park presents numerous potential safety hazards, including aging equipment, accumulation of flammable materials, and pipeline leaks. Traditional safety inspections primarily rely on manual inspections, with hazard records stored in paper files or a single database. This approach struggles to adapt to the dynamic nature of these hazards and the management demands of complex environments.

[0214] To overcome this problem, the system of the present invention first deploys multimodal data acquisition equipment, including sensors, cameras, and infrared detectors, at key locations within the industrial park to collect real-time information related to hidden dangers. This data capture covers equipment operating status, temperature and humidity, smoke concentration, and the storage of hazardous materials. After standardization, this data is stored in a hidden danger database model. Subsequently, the system constructs a hidden danger rule library based on historical accident cases and the park's safety management regulations, and uses a knowledge graph construction module to generate a hidden danger association network.

[0215] In a real-world application, multimodal equipment in a production workshop detected an abnormally high temperature, and smoke sensors sounded an alarm. The system analyzed dynamic feature labels, marking this potential risk event as a high-risk node and analyzing its potential propagation path using a potential risk heat map. The results showed that high-temperature areas could affect nearby stacked items, further posing a fire risk. Simultaneously, the system used a time series risk assessment model to predict the changing risk level of the potential risk, assigning it a risk score of 0.85 (on a scale of 0 to 1, with 1 being extremely high risk).

[0216] To ensure a rapid response, the system automatically assigned corrective actions based on hidden danger risk scores and knowledge graph features. The system prioritized the dispatch of two professional maintenance personnel and, using the task allocation and path planning scheme calculated by the optimization module, ensured they arrived at the scene within 10 minutes. Simultaneously, the system generated a dynamic hidden danger report and corrective action plan, uploading a hidden danger heat map, corrective action steps, and resource requirements to the management platform. Based on the system's provided plan, management mobilized the fire brigade for a comprehensive on-site inspection, completing the transfer of hazardous materials and equipment cooling within 30 minutes.

[0217] After the rectification was completed, the system dynamically integrated the rectification task execution data with the updated characteristics of the potential risk node and recorded the potential risk's evolution through a trend analysis report. The analysis showed that the risk score of the potential risk node decreased from the initial 0.85 to 0.15, demonstrating that the system effectively mitigated the security threat. Furthermore, the system updated the potential risk rule base in the knowledge graph based on task execution efficiency, providing data support for the subsequent management of similar potential risks.

[0218] Table 1 Comparison data of hidden danger management systems

[0219]

[0220]

[0221] The table above demonstrates the significant advantages of the system presented in this paper over traditional hazard management methods. In terms of hazard detection time, the system of this paper, through multimodal data fusion, completes detection in just 10 minutes, 75% faster than the 40 minutes of traditional methods. The accuracy of hazard identification has increased from 70% with traditional methods to 92%, significantly reducing false positives and missed detections. Furthermore, traditional methods are unable to analyze the propagation paths of hazards, while the system of this paper achieves this through hazard heat maps and knowledge graphs.

[0222] In terms of dynamic risk assessment, traditional static scoring methods fail to reflect real-time changes in hidden dangers. However, the proposed system utilizes time series models for dynamic risk analysis, enabling real-time assessment. Regarding task allocation efficiency, the proposed system reduces task allocation time from 20 minutes to 5 minutes through an optimization algorithm. It also supports real-time feedback and task tracking, addressing the lack of corrective action tracking in traditional methods.

[0223] Furthermore, the system, through its hazard heatmap generation and trend analysis capabilities, provides more accurate hazard propagation and trend predictions, assisting managers in making informed decisions. Compared to traditional static models, the system possesses dynamic optimization capabilities, enabling continuous improvement of model performance through online learning and adaptability to complex scenarios. These improvements fully demonstrate the system's superiority in efficiency, accuracy, and intelligence.

[0224] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A safety inspection method based on a hidden danger database model, characterized in that: The steps include: S1. Establish a hidden danger table, inspection record table, and rectification record table, and build a hidden danger rule library based on historical accident cases and industry standards; S2. Collect multi-source hidden danger data through the data acquisition module, perform standardization processing, form a hidden danger data set, and store it in the hidden danger database model; S3. Based on the hidden danger dataset and hidden danger rule base, build a hidden danger association knowledge graph, extract the potential features of hidden danger nodes and edges, and generate a knowledge graph feature matrix; S4. Input the hidden danger dataset and knowledge graph feature matrix into the time series model to perform dynamic risk analysis and generate a hidden danger risk score matrix; S5. Combine the hidden danger risk scoring matrix and the knowledge graph feature matrix, perform causal inference through the Bayesian network model, analyze the hidden danger propagation chain and impact range, optimize the hidden danger risk classification, and update the hidden danger rule library; S6. Combine the hidden danger dataset, knowledge graph feature matrix, and time series features, input into the multimodal hidden danger identification model, and classify the hidden dangers; S7. Combine the hidden danger risk classification results, classified hidden danger results, and knowledge graph feature matrix to optimize the rectification task allocation strategy, update the hidden danger status in real time, and store the rectification plan in the hidden danger database model; S8. Integrate the hidden danger risk grading results, hidden danger classification results, knowledge graph feature matrix and rectification task execution data to generate hidden danger dynamic characteristic labels, input the time series model and multimodal hidden danger identification model, iterate the model performance, and generate hidden danger heat map and trend analysis report.

2. A safety inspection method based on a hidden danger database model according to claim 1, characterized in that: The S3 specifically includes: S31. Extract hidden danger features from the hidden danger dataset, including the hidden danger description information, risk level, occurrence frequency, and associated inspection and rectification records; S32. Classify the hidden danger data and mark the type and attribute of the hidden danger according to the rules of the hidden danger rule library; S33. Establish a hidden danger association knowledge graph G = (V, E), where V represents a set of hidden danger nodes, including the type, occurrence time, and location of the hidden danger, and E represents the association between the hidden danger nodes, including causal relationships, temporal associations, and spatial associations; S34. In the constructed hidden danger association knowledge graph G, an embedding optimization algorithm combining the dynamic characteristics of time series and the hidden danger association weight is used to embed the hidden danger node v i The eigenvector of ∈V is expressed as: Among them, f(v i ) is the node v i The final embedding vector representation of W1, W2 and W3 are the trainable weight matrices of node features, time series features and neighborhood features respectively, T is the total number of time steps, α t is the attention weight at time step t, x i,t For node v i Time series features at time step t, β ij For node v i and neighboring node v j The association weight between i ) is the node v i The set of neighboring nodes, x j,t For node v j Time series features at time step t, h j is the neighborhood node v j The original eigenvector, h i is the neighborhood node v i The original eigenvector of g(h j ,x j,t ) is a nonlinear combination function; S35. Extract the hidden danger node features and edge features after graph embedding to generate the knowledge graph feature matrix F G , where F G =[f(v1),f(v2),...,f(v n )], each f(v i ) is the node v i The embedded feature vector of .

3. The safety inspection method based on the hidden danger database model according to claim 1 is characterized in that: The S4 specifically includes: S41. Define the hidden danger dataset and knowledge graph feature matrix as X={x1,x2,…,x n } and F G ={f1,f2,…,f n }, where n represents the number of nodes with hidden dangers, x i is the original feature vector of the potential danger node i, f i is the knowledge graph feature vector of hidden danger node i; S42, integrate the hidden danger dataset features and knowledge graph features to form a comprehensive feature vector z i , which is expressed as: z i =σ(W1x i +W2f i ); Among them, W1 and W2 are trainable weight matrices, and σ is the activation function; S43, the comprehensive feature vector z i Input time series model, define the input sequence Z of the time series model t ={z 1,t ,z 2,t ,…,z n,t }, where t is the time step, z i,t is the comprehensive characteristics of the potential danger node i at time step t; S44. Calculate the dynamic risk score of hidden danger nodes through the time series model: Among them, r i,t is the risk score of potential danger node i at time step t, W3 and W4 are trainable weight matrices, Softmax is the normalization function, z i,k is the comprehensive feature of the potential danger node i at time step k, T is the total number of time steps in the time series, α k is the attention weight at time step k: in, is the attention score at time step k, is the attention score at time step j, exp is the exponential function; S45, normalize the dynamic risk scores of all hidden danger nodes to generate a hidden danger risk score matrix R = {r i,t }.

4. The safety inspection method based on the hidden danger database model according to claim 1 is characterized in that: The S5 specifically includes: S51, the hidden danger risk score matrix R and the knowledge graph feature matrix F G Input the Bayesian network model to construct a causal relationship diagram of hidden danger propagation; S52. Define the Bayesian network as BN = (N, L, P), where the set of potential danger nodes N = {n1, n2, ..., n n }, the directed edge L between the potential hazard nodes = {l ij }, P is the conditional probability distribution, defined as: Among them, Pa(n i ) is the potential danger node n i The parent node set of θ k is the conditional probability of the k-th parent node, x k is a state variable; S53, according to the knowledge graph feature matrix F G And the hidden danger propagation causal relationship diagram, calculate the hidden danger node n i The propagation chain weight w i : Among them, β ij For node n i and parent node n j The correlation strength, r j is the parent node n j Dynamic risk score, f j is the parent node n j The knowledge graph feature vector, j is the index, g(r j ,f j ) is the parent node n j The nonlinear combination function of the risk score and feature matrix: g(r j ,f j )=ReLU(W q r j +W e f j ); Among them, ReLU is the activation function, W q and W e is a trainable weight matrix; S54, normalize the propagation chain weights to generate the hidden danger propagation chain matrix W = {w i }, where hidden danger node n i The propagation chain weight w i Expressed as: Among them, δ is the adjustment coefficient, h l For node n i The static propagation weight, l, k and m are indexes, n k is the node, h m For node n m The static propagation weight, N is the set of all nodes, β kj is the propagation weight, Pa(n k ) is node n k The parent node set of S55, transform the hidden danger propagation chain matrix W={w i }, hidden danger risk score matrix R and knowledge graph feature matrix F G Combined to generate optimized hidden danger risk classification results i : Among them, γ1, γ2, γ3 are the adjustment coefficients of the transmission chain weight, dynamic risk score and neighborhood comprehensive characteristics respectively, Softmax is the normalization function, N(n i ) is the neighborhood node set; S56. Update the optimized hidden danger risk classification results to the hidden danger rule library.

5. The safety inspection method based on the hidden danger database model according to claim 1 is characterized in that: The S6 specifically includes: S61, extract the multimodal feature set of the hidden danger node, including the hidden danger data set feature X = x1, x2, ..., x n }、Knowledge graph feature matrix F G ={f1,f2,…,f n } and time series feature matrix T = t1, t2, ..., t n }, for each node n i Construct feature fusion input: in, For node n i The initial multimodal feature vector of , ReLU is the activation function, R1, R2 and R3 are trainable weight matrices; S62. Update the feature vector of the potential danger node through the multi-layer implicit feature propagation module: in, For node n i The feature vector after propagation in the l+1 layer, N(n i ) is the set of neighborhood nodes, is the attention weight, For node n j The feature vector after propagation in layer l, For node n i The feature vector after propagation in layer l, R4 and R5 are trainable weight matrices, and j is the index; S63. Combined with the dynamic feature adjustment mechanism, according to the time series feature T, the attention weight is updated through the dynamic adjustment mechanism to generate the node feature propagation matrix: in, is the final attention weight, e ij and e ik is the basic attention score, is the time dynamic characteristic weight, ψ j and ψ k is the global importance weight, k1 and k2 are adjustment coefficients, exp is the exponential function, and k is the index; S64 takes the final feature vector of the hidden danger node Input the multimodal hidden danger identification model and generate classification results based on the multi-objective optimization mechanism and dynamic characteristic enhancement mechanism: Among them, y i is the classification result, Softmax is the normalization function, g(z j ,r j ) is the neighborhood feature aggregation function, is the prediction auxiliary module function, R6 is the trainable weight matrix, λ1 and λ2 are the adjustment coefficients, α ij is the attention weight, z j is the eigenvector, r j is the risk score, p i To predict auxiliary characteristics; S65. Output the hidden danger classification results, including hidden danger category, hidden danger risk level and key neighborhood influencing factors, and update the classification fields of the hidden danger database model.

6. The safety inspection method based on the hidden danger database model according to claim 1 is characterized in that: The S7 specifically includes: S71. Extract hidden danger risk classification result S = {s1,s2,…,s n }、Hazard classification result Y=y1,y2,…,y n } and the knowledge graph feature matrix F G ={f1,f2,…,f n }; S72. Build an optimized task allocation model to determine the task priority of hidden danger nodes based on risk classification and hidden danger category: Among them, p i is the task priority, Softmax is the normalization function, B1, B2 and B3 are adjustment coefficients, g(f j ) is the characteristic mapping function, β ij is the attention weight, s i is the risk grading score, y i is the classification result, f j is the knowledge graph feature vector, h(y i ) is the hidden danger classification score function, N(n i ) is the neighborhood node set, j is the index; S73, based on task priority p i , generate an optimized task allocation plan: Among them, Y is the task allocation matrix, c i is the rectification cost, d ij is the correlation distance, ξ ij is the task allocation constraint variable, N is the total number of nodes, argmin is the matrix Y for adjusting the task allocation scheme to achieve the optimal balance between the total task cost and the coordination of task allocation, and i and j are indexes; S74. Optimize the task allocation strategy through reinforcement learning and use the policy gradient method to update the task allocation model parameters. The loss function is defined as: in, The objective loss for task allocation optimization, π(a i |s i ) is in state s i Next select action a i The probability of ln is the logarithmic function, r i is the reward function, E is the expected symbol; S75. Output the optimized task allocation result and update the task execution field of the hidden danger database model.

7. The safety inspection method based on the hidden danger database model according to claim 1 is characterized in that: The S8 specifically includes: S81. Integrate the hidden danger risk classification results S = {s1, s2, ..., s n }、Hazard classification result Y={y1,y2,…,y n }、Rectification task execution data T={t1,t2,…,t n } and the knowledge graph feature matrix F G ={f1,f2,…,f n }, construct the hidden danger dynamic characteristic label L d ={l1,l2,…,l n }: Among them, Q1, Q2, Q3 and Q4 are trainable weight matrices, ReLU is the activation function, s i is the risk grading score, y i is the hidden danger classification result, t i The data vector for the rectification task, f i is the knowledge graph feature vector, is the final feature vector, l i is a dynamic feature label, h(y i ) is the hidden danger classification score function; S82, label the hidden danger dynamic characteristic L d Input the hidden danger feature propagation module, combine the time series model and the multimodal hidden danger identification model, and dynamically update the hidden danger node features: in, is the potential danger node n i The updated feature vector of is the final feature vector of the hidden danger node, Q5 and Q6 are trainable weight matrices, α ij is the attention weight, g(z j ,f j ) is the neighborhood feature combination function, ReLU is the nonlinear activation function, z j is the characteristic vector of the neighborhood node, f j is the knowledge graph feature, N(n i ) is the neighborhood node set, j is the index; S83. Based on the updated hidden danger node characteristics, calculate the dynamic correlation heat between the hidden danger nodes and generate a hidden danger heat map H = {H ij }: in, is the potential danger node n j The updated feature vector, cosine_similarity is the cosine similarity function, H ij is the potential danger node n i and n j Dynamic correlation heat; S84. Using the hidden danger heat map H and the hidden danger dynamic characteristic label L d , analyze the dynamic evolution trend of hidden danger nodes and generate trend analysis reports: Among them, Trend_Score i is the potential danger node n i Dynamic trend score, is the dynamic risk change rate, λ q ,λ w and λ e is the weight parameter, cosine_similarity(l i ,l j ) is the dynamic feature label similarity, t i The time for completing the rectification task; S85. Combine the hidden danger heat map and trend analysis report to update the time series model and multimodal hidden danger identification model, and iterate the model performance: in, is the loss function for iterative optimization, E is the expected symbol, is the true eigenvector, is the weight parameter, is the predicted distribution, is the actual distribution, N is the total number of nodes, is the KL divergence between the predicted distribution and the true distribution, and i is the index.

8. A safety inspection system based on a hidden danger database model, a safety inspection method based on a hidden danger database model according to any one of claims 1 to 7, characterized in that: Includes the following modules: Data collection and storage module, used to collect multi-source hidden danger data; The knowledge graph construction module is used to construct a hidden danger related knowledge graph and extract node features and association relationships; Risk assessment and hidden danger classification module, used to generate hidden danger risk scores and classify hidden danger nodes; Dynamic characteristic analysis and heat map generation module, used to analyze the dynamic characteristic labels of hidden danger nodes and generate hidden danger heat maps; Risk assessment and hidden danger classification module, used to perform dynamic risk analysis and classification of hidden danger nodes; Task allocation and optimization module, used to optimize task allocation strategies and dynamically adjust resource allocation and priorities; Trend analysis and report generation module, used to generate hidden danger trend analysis reports and decision support information; Model training and optimization module, used to optimize the performance of time series models and multimodal hidden danger identification models.

Images