Construction unit safety responsibility tracing method based on HSE atlas

Abstract

The invention discloses a construction unit safety responsibility tracing method based on an HSE map, and the method comprises the following steps: S1, collecting multi-source HSE data of a construction site, and constructing a structured data set; s2, constructing an HSE atlas containing multiple types of nodes and relationships based on the data set; s3, inputting the HSE atlas into the logic attention path network model, extracting a causal path and embedding nodes; s4, performing fuzzy responsibility scoring on the nodes in the extraction path to generate an attribution score; s5, adopting a cuckoo search algorithm to perform joint optimization on the model structure and the scoring parameters to improve the recognition effect; and S6, outputting a construction unit responsibility node and grade result based on the optimized model for responsibility tracing and decision making. According to the method, the recognition efficiency and the decision transparency are remarkably improved while the construction responsibility is accurately traced, and the method is suitable for various types of safety management scenes.

Description

Technical Field

[0001] The present invention relates to the technical field of safety management, and in particular to a construction unit safety responsibility tracing method based on an HSE graph. Background Art

[0002] In construction safety management, health, safety, and environment (HSE)—the core foundation of enterprise oversight and operational risk control—has long been the core of the HSE system. With the continuous expansion of construction projects, the increasing number of participating entities, the complex working environment, and the diverse collaborative processes, traditional HSE management models have gradually exposed problems such as data fragmentation, delayed response, and unclear responsibilities. Especially after a safety incident or accident, existing accountability methods mostly rely on manual analysis of construction records, on-site monitoring, and post-incident inquiries, lacking effective technical means to support the reconstruction of causal chains and the precise identification of responsible parties. This approach is not only inefficient, but the division of responsibilities often relies on empirical judgment and lacks quantitative standards, which can easily lead to disputes and management gaps.

[0003] Currently, some information systems attempt to incorporate HSE management into digital platforms, but most systems remain at the level of data recording and static display, lacking the ability to uniformly model the relationships between multi-source data such as work behavior, equipment status, approval processes, and environmental factors. The few studies that apply graph databases or knowledge graphs also focus mainly on static knowledge representation between entities, and have not yet achieved high-precision modeling of the dynamic evolution of construction behavior, accident propagation paths, and responsibility chains. At the same time, most existing graph models use general graph embedding or adjacent attention methods, which make it difficult to support the mining and identification of potential responsibility propagation across nodes and multi-hop paths. In addition, responsibility attribution generally lacks explainability and flexible mechanisms, and is often based on fixed rules or Boolean logic judgments, making it difficult to handle complex phenomena such as "multi-person collaboration," "cross-responsibility," and "non-linear behavior triggering" that are common in real construction scenarios.

[0004] When it comes to model optimization, mainstream methods often employ fixed structures or simple grid searches, making it difficult to achieve a globally optimal configuration in a nonlinear, high-dimensional parameter space. This is particularly true in responsibility identification models, which involve multi-dimensional, heterogeneous parameters such as the number of graph layers, the number of attention heads, and the weights of attribution functions. Traditional parameter tuning methods are ineffective and prone to falling into local optimality or poor generalization, resulting in insufficient adaptability across different project scenarios.

[0005] In summary, existing technologies for tracing construction unit safety responsibility have the following major flaws: First, they lack the ability to dynamically model the logical relationships between multi-source HSE data, making it difficult to accurately extract accident causal chains and responsibility paths; second, the responsibility scoring mechanism is rigid and lacks fuzzy expression, which cannot reflect the varying degrees of responsibility participation between nodes; third, the model structure and parameter configuration rely on manual settings and lack intelligent optimization methods, which affects the accuracy and adaptability of the identification results. Therefore, a responsibility tracing method that integrates graph intelligent modeling, fuzzy attribution mechanisms, and evolutionary optimization algorithms is urgently needed to achieve efficient, accurate, and explainable tracing of construction unit safety responsibility.

[0006] Therefore, how to provide a construction unit safety responsibility tracing method based on the HSE map is an urgent problem that technicians in this field need to solve. Summary of the Invention

[0007] One purpose of the present invention is to propose a construction unit safety responsibility tracing method based on HSE graph modeling and path reasoning. The present invention makes full use of heterogeneous graph modeling technology, logical attention path network and cuckoo search evolutionary algorithm, and describes in detail the complete process of realizing multi-source HSE data fusion mapping, causal path extraction, fuzzy responsibility scoring and structural optimization identification in construction scenarios. It has the advantages of high responsibility identification accuracy, strong attribution explanation ability and good model adaptability.

[0008] A method for tracing the safety responsibilities of construction units based on an HSE map according to an embodiment of the present invention includes the following steps:

[0009] S1. Collect health, safety and environment data from construction sites and build a structured HSE dataset.

[0010] S2. constructing an HSE atlas based on the HSE dataset;

[0011] S3. Input the HSE graph into the logical attention path network model, perform embedding coding based on node features and edge relationships in the graph, and extract the causal path set associated with the specified security event node through a multi-hop path attention mechanism;

[0012] S4. Applying a fuzzy attribution enhancement mechanism to the causal path set, performing a fuzzy responsibility score on each operation behavior node and construction unit node based on the position weight, node centrality, and behavior risk level in the path, and generating an attribution credibility matrix;

[0013] S5. Use the cuckoo search evolutionary algorithm to perform multi-objective joint optimization on the depth of the graph convolution layer, the attention weight distribution ratio, and the fuzzy responsibility function parameters in the logical attention path network model, and output the responsibility recognition model after structural optimization.

[0014] S6. Based on the responsibility identification model after structural optimization, path aggregation analysis and node score calculation are performed on the designated safety event nodes, and the output includes a set of responsibility level results of the construction unit nodes and the corresponding quantitative responsibility index as the basis for responsibility tracing.

[0015] Optionally, the S1 constructs a structured HSE data set including: performing time alignment, field unification and semantic standardization on operator identity information, operation behavior logs, equipment status data, operation approval records, environmental monitoring data and safety event logs to form an HSE data set in a unified format.

[0016] Optionally, the S2 constructs an HSE graph based on the HSE data set specifically including: taking construction units, workers, work behaviors, equipment and safety events as graph nodes respectively, establishing graph edges based on personnel affiliation relationships, behavior triggering relationships, equipment participation relationships and event causal relationships, and generating a heterogeneous graph structure with multiple types of entity nodes and multiple types of edge relationships.

[0017] Optionally, the S3 specifically includes:

[0018] S31. The constructed HSE graph is represented as a graph structure G = (V, E), where V is a node set, including the construction unit node v u , operator node v p , job behavior node v a 、Device node v d and security event node v s , E is the edge set, including the person-attribution edge e p,u , behavior trigger edge a,s , device association edge e d,a and behavioral causal edges a,a′ ;

[0019] S32. Construct an initial feature vector h0(v) for each node v∈V, wherein the feature vector includes a node identity code, a category identifier, a timestamp information, and a centrality index, forming an initial feature set {h0(v)};

[0020] S33. Input the graph structure G and the initial feature set {h0(v)} into the logical attention path network model, perform the graph path convolution operation, and calculate the embedding vector of the l-th layer node using the following embedding update formula:

[0021]

[0022] in, represents the first-order neighbor set of node v, α v,u is the attention weight from node u to node v, W is a trainable linear transformation matrix, and σ is a nonlinear activation function;

[0023] S34, for the specified target security event node v s , perform multi-hop path extraction operation, construct a i to v s The set of causal paths between them P={p1,p2,…,p n}, each path p i is an ordered sequence of nodes:

[0024] p i = <v1,v2,…,v s >

[0025] S35. For each path p i ∈P calculates the path attention score β i , indicating the path to the target event node v s The causal influence weight of the path score is calculated as follows:

[0026]

[0027] Among them, |p i | represents the path p i The number of nodes in , w is the attention score vector, h l (v j ) is the embedding representation of the j-th node in the path at layer l;

[0028] S36, output causal path set P and corresponding attention score set {β i}, as the target security event node v s The collection of associated causal paths.

[0029] Optionally, the path attention mechanism of step S35 not only performs weighted aggregation on the node embedding information in each causal path, but also introduces the temporal position of the path node, the risk level of the behavior node and the structural position weight as combination factors to participate in the overall path scoring process. The global attention scoring vector and the path feature mapping function are dynamically learned through the training phase, so that the score of each path has differentiated semantic expression capabilities while maintaining structural consistency.

[0030] Optionally, the S4 specifically includes:

[0031] S41, for the causal path set P = {p1, p2, ..., p n} and the corresponding path attention score set {β1,β2,…,β n}, for each path p i Each node v in ∈P j Extract three-dimensional evaluation indicators: node centrality C(vj ), behavioral risk level R(v j ), path position factor δ(v j ,p i ), which respectively represent the structural importance of the node in the graph, the risk level of the operation behavior represented by the node, and the relative position of the node in the sequence of the path;

[0032] S42. Construct a fuzzy responsibility scoring function to calculate the attribution degree of the node in the path. j On path p i The fuzzy responsibility score μ(v j ,p i ) is defined as follows:

[0033] μ(v j ,p i )=γ1·C(v j )+γ2·R(v j )+γ3·δ(v j ,p i );

[0034] Among them, γ1, γ2, and γ3 are the structural importance weight, risk level weight, and path position weight, respectively, satisfying the constraint condition γ1+γ2+γ3=1. The weights are dynamically updated by the optimization algorithm during the model training process.

[0035] S43, based on the above fuzzy responsibility scoring function, for all paths p i ∈P and the corresponding node v j ∈p i Perform attribution scoring and construct the attribution credibility matrix M, where M i,j =μ(v j ,p i ), n is the number of paths, and m is the number of nodes in a single path;

[0036] S44. Combine the path attention score and the fuzzy responsibility score to calculate the global attribution score of each node in all paths. Set the total responsibility score of node v to Γ(v). The calculation formula is as follows:

[0037]

[0038] Among them, β i Represents path p i The attention score, μ(v j ,p i ) is the node v j Fuzzy scoring in the path;

[0039] S45, output attribution score vector Γ={Γ(v1),Γ(v2),…,Γ(v k )}, each element in the score vector corresponds to a construction unit node or an operation behavior node in the HSE map.

[0040] Optionally, the fuzzy responsibility scoring function constructed in S42 is a weighted combination of node centrality, risk level and path position factor, and the weight parameters are trainable and meet normalization constraints to achieve attribution scoring under multi-dimensional features.

[0041] Optionally, the S5 specifically includes:

[0042] S51. Define the model optimization objective function J, which is used to measure the error between the responsibility score result generated by the model and the known responsibility label. The specific calculation formula is:

[0043]

[0044] Among them, U represents the set of nodes with labeled responsibility levels, w x represents the importance weight of node x, φ x Represents the responsibility scoring result output by the model, Indicates the true responsibility level of node x;

[0045] S52. Set the parameter vector to be optimized Θ = [L, H, α1, α2, α3], where L represents the number of graph convolution layers of the logical attention path network, H represents the number of attention sub-heads, and α1, α2, and α3 represent the node centrality weight, behavior risk level weight, and path position weight in the fuzzy responsibility scoring function, respectively, satisfying the normalization constraint:

[0046] α1+α2+α3=1;

[0047] S53, based on the cuckoo search evolution mechanism, construct the initial parameter population, and use the Levy flight strategy to generate each current solution vector Θ i A new candidate solution Θ i ′ , and perform integer projection on the layer number parameter L and sub-head number parameter H in the candidate solution, and perform normalization correction on the weight parameters α1, α2, and α3;

[0048] S54, the candidate solution vector Θ i ′ Applied to the responsibility identification model, the aforementioned graph embedding encoding, path attention calculation and fuzzy responsibility scoring processes are executed to obtain the model output responsibility scoring result φ x , and calculate the optimization objective function value J(Θ i ′ );

[0049] S55. Evaluate the fitness of the current population based on the objective function value, adopt a winner-replacement strategy to retain low-error candidate solutions, and introduce some random individuals to enhance population diversity to form the next generation solution set;

[0050] S56, repeat S53 to S55 until the maximum number of iterations or the objective function convergence condition is reached, and output the optimal parameter vector The responsibility identification model after final structural optimization is configured and applied to the calculation and output of node responsibility scores.

[0051] Optionally, the step S6 specifically includes: optimizing the parameter vector Applied to the logical attention path network model and fuzzy responsibility scoring function, combined with the aforementioned graph embedding representation, path attention score and node attribution score results, the global responsibility score φ of each construction unit node is calculated. x ; According to the responsibility score, interval mapping is performed within the set level threshold range, divided into multi-level responsibility level labels, forming a responsibility node set and level labeling results; the responsibility level results are associated with the original HSE map, and a responsibility traceability table containing node identification, responsibility value and level category is output for construction unit accountability decision-making and risk control management.

[0052] The beneficial effects of the present invention are:

[0053] First, this invention achieves unified modeling and causal structure representation of multi-source heterogeneous data at construction sites by constructing a multi-relational HSE graph that integrates construction units, work activities, equipment status, and incidents. This graph structure not only supports static entity relationship modeling but also possesses dynamic path reasoning capabilities, effectively restoring the chain of responsibility before and after an incident, significantly improving the logical integrity and data support for safety incident tracing.

[0054] Secondly, this paper introduces a logical attention path network and a fuzzy responsibility scoring mechanism to support flexible attribution analysis of responsible nodes across multi-hop causal paths. By integrating and weighting node centrality, behavioral risk level, and path position factors, a trainable fuzzy scoring function is constructed. This overcomes the inability of traditional hard-rule accountability methods to handle complex situations such as collaborative responsibility and nonlinear triggering, achieving a more interpretable and adjustable quantitative determination of responsibility.

[0055] Furthermore, this paper employs a cuckoo search evolutionary algorithm to globally optimize the model structure and parameters, addressing the limitations of traditional graph model structures, which rely on empirically defined parameters and have limited optimization capabilities. By searching for optimal combinations across the parameter space of graph convolution depth, number of attention heads, and attribution weights, the model's generalization and responsibility identification accuracy across diverse construction project scenarios are enhanced, providing a technical foundation for greater adaptability and decision-making support in construction safety management systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] 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:

[0057] Figure 1 This is the overall flow chart of the construction unit safety responsibility tracing method based on HSE map proposed by the present invention;

[0058] Figure 2 This is a logical attention path network model architecture diagram of a construction unit safety responsibility tracing method based on HSE graph proposed in this invention;

[0059] Figure 3 This is a cuckoo search evolutionary optimization flowchart of the construction unit safety responsibility tracing method based on HSE graph proposed in this invention. DETAILED DESCRIPTION

[0060] 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.

[0061] refer to Figure 1-3 A construction unit safety responsibility tracing method based on HSE atlas includes the following steps:

[0062] S1. Collect health, safety and environment data from construction sites and build a structured HSE dataset.

[0063] S2. constructing an HSE atlas based on the HSE dataset;

[0064] S3. Input the HSE graph into the logical attention path network model, perform embedding coding based on node features and edge relationships in the graph, and extract the causal path set associated with the specified security event node through a multi-hop path attention mechanism;

[0065] S4. Applying a fuzzy attribution enhancement mechanism to the causal path set, performing a fuzzy responsibility score on each operation behavior node and construction unit node based on the position weight, node centrality, and behavior risk level in the path, and generating an attribution credibility matrix;

[0066] S5. Use the cuckoo search evolutionary algorithm to perform multi-objective joint optimization on the depth of the graph convolution layer, the attention weight distribution ratio, and the fuzzy responsibility function parameters in the logical attention path network model, and output the responsibility recognition model after structural optimization.

[0067] S6. Based on the responsibility identification model after structural optimization, path aggregation analysis and node score calculation are performed on the designated safety event nodes, and the output includes a set of responsibility level results of the construction unit nodes and the corresponding quantitative responsibility index as the basis for responsibility tracing.

[0068] This paper proposes a construction unit safety responsibility tracing method based on HSE graphs. It establishes a complete technical process from data collection, graph modeling, causal path extraction, fuzzy responsibility scoring, to structural optimization, forming a reusable and quantifiable responsibility tracing system. Compared to traditional responsibility determination methods that rely on manual judgment, this paper effectively automates and quantifies responsibility identification through a structured graph model and path reasoning mechanism, enhancing the intelligent level of construction safety management.

[0069] In this embodiment, the S1 construction of a structured HSE data set includes: time alignment, field unification and semantic standardization of operator identity information, work behavior logs, equipment status data, work approval records, environmental monitoring data and safety event logs to form an HSE data set in a unified format.

[0070] This invention constructs a structured HSE dataset by preprocessing and uniformly modeling construction site HSE data, resolving the fragmented and inconsistent multi-source data issues in existing systems. Compared to traditional methods of managing information in table or log form, this dataset features time synchronization, field standardization, and semantic consistency, providing a solid foundation for subsequent graph construction and model input, significantly improving data processing efficiency and model usability.

[0071] In this embodiment, the S2 constructs an HSE graph based on the HSE data set, specifically including: taking construction units, workers, work behaviors, equipment and safety events as graph nodes respectively, establishing graph edges based on personnel affiliation relationships, behavior triggering relationships, equipment participation relationships and event causal relationships, and generating a heterogeneous graph structure with multiple types of entity nodes and multiple types of edge relationships.

[0072] This method constructs a multi-relationship heterogeneous graph structure based on HSE data, achieving a unified representation of the complex relationships between people, units, equipment, behaviors, and events. Compared with existing static knowledge graph approaches, this method supports dynamic temporal relationships and causal path modeling, providing structural support for subsequent graph neural network reasoning and responsibility propagation path identification, and enhancing the depth and completeness of responsibility chain analysis.

[0073] In this embodiment, S3 specifically includes:

[0074] S31. The constructed HSE graph is represented as a graph structure G = (V, E), where V is a node set, including the construction unit node v u , operator node v p , job behavior node v a 、Device node v d and security event node v s , E is the edge set, including the person-attribution edge e p,u , behavior trigger edge a,s , device association edge e d,a and behavioral causal edges a,a′ ;

[0075] S32. Construct an initial feature vector h0(v) for each node v∈V, wherein the feature vector includes a node identity code, a category identifier, a timestamp information, and a centrality index, forming an initial feature set {h0(v)};

[0076] S33. Input the graph structure G and the initial feature set {h0(v)} into the logical attention path network model, perform the graph path convolution operation, and calculate the embedding vector of the l-th layer node using the following embedding update formula:

[0077]

[0078] in, represents the first-order neighbor set of node v, α v,u is the attention weight from node u to node v, W is a trainable linear transformation matrix, and σ is a nonlinear activation function;

[0079] S34, for the specified target security event node v s , perform multi-hop path extraction operation, construct a i to v s The set of causal paths between them P={p1,p2,…,p n}, each path p i is an ordered sequence of nodes:

[0080] p i = <v1,v2,…,v s >

[0081] S35. For each path p i ∈P calculates the path attention score β i , indicating the path to the target event node v s The causal influence weight of the path score is calculated as follows:

[0082]

[0083] Among them, |p i | represents the path p i The number of nodes in , w is the attention score vector, h l (v j ) is the embedding representation of the j-th node in the path at layer l;

[0084] S36, output causal path set P and corresponding attention score set {β i}, as the target security event node v s The collection of associated causal paths.

[0085] This paper introduces a logical attention path network to perform node embedding and multi-hop path reasoning on HSE graphs. Compared to conventional graph convolution methods, this method offers path-level reasoning capabilities and a multi-layer attention fusion mechanism, effectively capturing potential responsibility links within the graph. This mechanism significantly improves the ability to model complex accident causal relationships, providing a more accurate graph representation foundation for subsequent attribution analysis.

[0086] In this embodiment, the path attention mechanism of step S35 not only performs weighted aggregation on the node embedding information in each causal path, but also introduces the temporal position of the path node, the risk level of the behavior node and the structural position weight as combination factors to participate in the overall path scoring process. The global attention scoring vector and the path feature mapping function are dynamically learned through the training phase, so that the score of each path has differentiated semantic expression capabilities while maintaining structural consistency.

[0087] This paper introduces an innovative multi-factor weighting design into the path attention mechanism, combining node embedding representations with overall path characteristics. This overcomes the limitations of traditional path scoring methods, which rely solely on local adjacency information. This mechanism accurately characterizes causal strength at the path granularity, improving the accuracy of identifying responsibility propagation paths within the graph and supporting analytical decision-making in complex collaborative responsibility environments.

[0088] In this embodiment, the S4 specifically includes:

[0089] S41, for the causal path set P = {p1, p2, ..., p n} and the corresponding path attention score set {β1,β2,…,β n}, for each path p i Each node v in ∈P j Extract three-dimensional evaluation indicators: node centrality C(v j ), behavioral risk level R(v j ), path position factor δ(v j ,pi ), which respectively represent the structural importance of the node in the graph, the risk level of the operation behavior represented by the node, and the relative position of the node in the sequence of the path;

[0090] S42. Construct a fuzzy responsibility scoring function to calculate the attribution degree of the node in the path. j On path p i The fuzzy responsibility score μ(v j ,p i ) is defined as follows:

[0091] μ(v j ,p i )=γ1·C(v j )+γ2·R(v j )+γ3·δ(v j ,p i );

[0092] Among them, γ1, γ2, and γ3 are the structural importance weight, risk level weight, and path position weight, respectively, satisfying the constraint condition γ1+γ2+γ3=1. The weights are dynamically updated by the optimization algorithm during the model training process.

[0093] S43, based on the above fuzzy responsibility scoring function, for all paths p i ∈P and the corresponding node v j ∈p i Perform attribution scoring and construct the attribution credibility matrix M, where M i,j =μ(v j ,p i ), n is the number of paths, and m is the number of nodes in a single path;

[0094] S44. Combine the path attention score and the fuzzy responsibility score to calculate the global attribution score of each node in all paths. Set the total responsibility score of node v to Γ(v). The calculation formula is as follows:

[0095]

[0096] Among them, β i Represents path p i The attention score, μ(v j ,p i ) is the node v j Fuzzy scoring in the path;

[0097] S45, output attribution score vector Γ={Γ(v1),Γ(v2),…,Γ(v k )}, each element in the score vector corresponds to a construction unit node or an operation behavior node in the HSE map.

[0098] The proposed fuzzy responsibility scoring mechanism integrates node centrality, behavioral risk level, and path position factors to construct a trainable weighted scoring function. This effectively addresses the rigid attribution and inability to express differences in responsibility between nodes in traditional responsibility assessments. This fuzzy scoring mechanism enables flexible modeling of responsibility participation, enhancing the interpretability and robustness of responsibility results.

[0099] In this embodiment, the fuzzy responsibility scoring function constructed in S42 is a weighted combination of node centrality, risk level and path position factor. The weight parameters are trainable and meet normalization constraints to achieve attribution scoring under multi-dimensional features.

[0100] In this embodiment, the S5 specifically includes:

[0101] S51. Define the model optimization objective function J, which is used to measure the error between the responsibility score result generated by the model and the known responsibility label. The specific calculation formula is:

[0102]

[0103] Among them, U represents the set of nodes with labeled responsibility levels, w x represents the importance weight of node x, φ x Represents the responsibility scoring result output by the model, Indicates the true responsibility level of node x;

[0104] S52. Set the parameter vector to be optimized Θ = [L, H, α1, α2, α3], where L represents the number of graph convolution layers of the logical attention path network, H represents the number of attention sub-heads, and α1, α2, and α3 represent the node centrality weight, behavior risk level weight, and path position weight in the fuzzy responsibility scoring function, respectively, satisfying the normalization constraint:

[0105] α1+α2+α3=1;

[0106] S53, based on the cuckoo search evolution mechanism, construct the initial parameter population, and use the Levy flight strategy to generate each current solution vector Θ i A new candidate solution Θ i ′ , and perform integer projection on the layer number parameter L and sub-head number parameter H in the candidate solution, and perform normalization correction on the weight parameters α1, α2, and α3;

[0107] S54, the candidate solution vector Θ i ′ Applied to the responsibility identification model, the aforementioned graph embedding encoding, path attention calculation and fuzzy responsibility scoring processes are executed to obtain the model output responsibility scoring result φx , and calculate the optimization objective function value J(Θ i ′ );

[0108] S55. Evaluate the fitness of the current population based on the objective function value, adopt a winner-replacement strategy to retain low-error candidate solutions, and introduce some random individuals to enhance population diversity to form the next generation solution set;

[0109] S56, repeat S53 to S55 until the maximum number of iterations or the objective function convergence condition is reached, and output the optimal parameter vector The responsibility identification model after final structural optimization is configured and applied to the calculation and output of node responsibility scores.

[0110] This paper introduces a cuckoo search evolutionary algorithm to jointly optimize the graphical model structure and parameters, overcoming the bottleneck of traditional parameter tuning, which relies on empirical settings and has a limited search space. This optimization method possesses global optimization capabilities and adaptability to non-convex spaces, effectively improving the performance and generalization of the responsibility identification model and enhancing its deployment effectiveness in multiple construction scenarios.

[0111] In this embodiment, the step S6 specifically includes: Applied to the logical attention path network model and fuzzy responsibility scoring function, combined with the aforementioned graph embedding representation, path attention score and node attribution score results, the global responsibility score φ of each construction unit node is calculated. x ; According to the responsibility score, interval mapping is performed within the set level threshold range, divided into multi-level responsibility level labels, forming a responsibility node set and level labeling results; the responsibility level results are associated with the original HSE map, and a responsibility traceability table containing node identification, responsibility value and level category is output for construction unit accountability decision-making and risk control management.

[0112] This invention uses a structurally optimized responsibility identification model to categorize construction unit node responsibility scores and establish a quantitative accountability output system. Unlike traditional approaches that rely on textual descriptions or fuzzy classifications, this invention's output includes clear node identifiers, numerical scores, and grade labels, facilitating system integration and management, and improving the standardization and operability of safety accountability work.

[0113] Example 1:

[0114] To verify the feasibility of the present invention in practice, it was applied to a large-scale municipal infrastructure construction project. This project involved multiple construction units working in parallel, with complex operations, intensive equipment, frequent cross-work, and extremely difficult construction safety management. In this project, the project management has long faced problems such as unclear division of accident responsibility, long investigation cycles due to missing data, and difficulty in clarifying collaborative responsibilities. In particular, after a safety incident, conventional practices rely on monitoring playback and paper approval documents, which are inefficient and highly subjective, and cannot meet the needs of refined and real-time responsibility tracing.

[0115] In this project, the present invention uses daily collected construction site HSE data as the basis, including operator attendance records, dangerous operation approval data, special equipment operating status, on-site environmental monitoring parameters and historical hidden danger investigation logs, etc., covering a total of 23 key operation surfaces, involving 9 construction units and 218 operators, to construct a structured HSE data set with an average daily data volume of approximately 87.3MB. Relying on the method of the present invention, the data set is constructed into a dynamic heterogeneous HSE graph, with a total of 11,762 nodes and 42,381 edge relationships established, including 5,129 personnel behavior nodes, 2,870 equipment nodes, and 413 safety event nodes. The high-confidence path of the target event node is extracted through the logical attention path network model, with an average path hop number of 3.7 and a maximum hop number of 6, which can restore the multi-stage causal chain from behavior to event in complex scenarios.

[0116] In a specific application, a fall from height occurred in a certain work area. Traditional investigation methods are expected to take 9 days to complete the initial division of responsibilities. However, after accessing the relevant data, the system of the present invention completed full-map backtracking, path extraction, and fuzzy score calculation in just 4.6 hours, quickly locating 17 responsibility-related nodes and identifying 3 construction unit nodes with responsibility weights greater than 0.65. Combined with the review and feedback from the management, the consistency rate between the system output results and the final accountability reached 94.1%. The present invention completed 17 automatic responsibility analyses for similar accidents, with an average response time of 5.3 hours, significantly better than the average manual processing cycle of 7.5 days.

[0117] In addition, the graph network structure parameters and the weight factors of the fuzzy scoring function were jointly optimized through the cuckoo search algorithm, and converged to the optimal parameter combination after 10 rounds of iteration. The responsibility identification accuracy was improved from the initial 82.6% to 91.3%, and the path coverage was improved from 74.8% to 88.4%. The model still maintained a high degree of stability after cross-regional scheduling, verifying the adaptability and promotion value of the structural optimization scheme in actual scenarios.

[0118] During real-world project evaluations, the system assisted the safety management team in automatically generating maps 32 times, identifying responsible nodes 82 times, and conducting 12 hazard reviews and backtracking. It was successfully integrated into the project's HSE management platform for real-time decision-making. The management unit feedback system improved accident investigation efficiency by over 60%, clarified accountability, and effectively reduced coordination conflicts and accountability costs, demonstrating strong project feasibility and potential for expansion.

[0119] Table 1: Comparative evaluation of the effectiveness of HSE diagram-assisted responsibility identification

[0120] Indicator name Traditional processing methods Method of the present invention Improvement Average accident responsibility determination cycle (days) 7.5 0.22 (approximately 5.3 hours) ↓About 97.1% Responsible node identification accuracy (%) 82.6 91.3 ↑8.7％ Path backtracking coverage (%) 74.8 88.4 ↑13.6％ High-risk node identification time (hours) ≥48 4.6 ↓About 90.4% Consistency rate between system and manual accountability (%) — 94.1 — Number of rounds of structural parameter optimization — 10 — Number of data access operation surfaces — 23 — Total number of responsibility identification tasks (times) — 82 —

[0121] The above-mentioned "HSE Atlas Assisted Responsibility Identification Effect Comparative Evaluation Table" compares and demonstrates the performance differences between the method of the present invention and the traditional accident responsibility handling method in multiple key indicators, which can fully reflect the superiority and improvement effect of the present invention in practical applications.

[0122] First, in terms of incident handling efficiency, this invention significantly shortens the accident responsibility determination cycle. Traditional methods typically require 7.5 days to complete responsibility determination, while this invention averages only about 5.3 hours (or 0.22 days), achieving a reduction in processing time of over 97%. This significantly improves the real-time nature of security incident response and reduces the chain reaction risks that may arise from delayed handling.

[0123] In terms of recognition accuracy, the proposed method achieved a 91.3% accuracy rate for identifying responsible nodes, an 8.7 percentage point improvement over the 82.6% achieved by traditional methods. This demonstrates that, in complex collaborative environments, the proposed method, leveraging a logical path network and fuzzy attribution mechanism, can more accurately identify the true responsible nodes, reducing the risk of misjudgment or omission. Furthermore, the path backtracking coverage rate increased from 74.8% to 88.4%, demonstrating that the model can restore a more complete accident causal chain and enhance its ability to capture multi-level triggering behaviors.

[0124] Traditional methods typically require over two days to identify high-risk nodes. However, this new approach can complete retrospective analysis of responsible nodes in as little as 4.6 hours, reducing response time by over 90%. This saves valuable time for rapid resolution and on-site containment of high-risk scenarios. Furthermore, the system's output results achieve a 94.1% consistency rate with the final manual accountability results, further validating the model's reliability and practical decision-making value in real-world management scenarios.

[0125] From the perspective of model training and system deployment, the parameter optimization process of this invention achieved convergence within 10 rounds of evolution, demonstrating excellent structural adaptability. Furthermore, during the testing period, the system accessed 23 work surfaces, completed 82 responsibility identification tasks, and successfully supported multiple on-site risk review tasks, demonstrating strong processing capabilities and engineering scalability.

[0126] In summary, the tabular data fully verifies the application effectiveness of this invention in construction safety responsibility management. In particular, it is superior to traditional methods in terms of accident response efficiency, responsibility identification accuracy, causal path restoration capability, and system integration. It has clear technical advantages and engineering promotion value.

[0127] 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 construction unit safety responsibility tracing method based on HSE graph, characterized by: The steps include: S1. Collect health, safety and environment data from construction sites and build a structured HSE dataset. S2. constructing an HSE atlas based on the HSE dataset; S3. Input the HSE graph into the logical attention path network model, perform embedding coding based on node features and edge relationships in the graph, and extract the causal path set associated with the specified security event node through a multi-hop path attention mechanism; S4. Applying a fuzzy attribution enhancement mechanism to the causal path set, performing a fuzzy responsibility score on each operation behavior node and construction unit node based on the position weight, node centrality, and behavior risk level in the path, and generating an attribution credibility matrix; S5. Use the cuckoo search evolutionary algorithm to perform multi-objective joint optimization on the depth of the graph convolution layer, the attention weight distribution ratio, and the fuzzy responsibility function parameters in the logical attention path network model, and output the responsibility recognition model after structural optimization. S6. Based on the responsibility identification model after structural optimization, path aggregation analysis and node score calculation are performed on the designated safety event nodes, and the output includes a set of responsibility level results of the construction unit nodes and the corresponding quantitative responsibility index as the basis for responsibility tracing.

2. The method for tracing the safety responsibility of construction units based on HSE graph according to claim 1 is characterized in that: The S1 construction of a structured HSE dataset includes: time alignment, field unification and semantic standardization of operator identity information, operation behavior logs, equipment status data, operation approval records, environmental monitoring data and safety event logs to form an HSE dataset in a unified format.

3. The method for tracing the safety responsibility of construction units based on HSE graph according to claim 1 is characterized in that: The S2 constructs an HSE graph based on the HSE data set, specifically including: taking construction units, workers, work behaviors, equipment and safety events as graph nodes respectively, establishing graph edges according to personnel affiliation relationships, behavior triggering relationships, equipment participation relationships and event causal relationships, and generating a heterogeneous graph structure with multiple types of entity nodes and multiple types of edge relationships.

4. The method for tracing the safety responsibility of construction units based on HSE atlas according to claim 1 is characterized in that: The S3 specifically includes: S31. The constructed HSE graph is represented as a graph structure G = (V, E), where V is a node set, including the construction unit node v u , operator node v p , job behavior node v a 、Device node v d and security event node v s , E is the edge set, including the person-attribution edge e p,u , behavior trigger edge a,s , device association edge e d,a and behavioral causal edges a,a′ ; S32. Construct an initial feature vector h0(v) for each node v∈V, wherein the feature vector includes a node identity code, a category identifier, a timestamp information, and a centrality index, forming an initial feature set {h0(v)}; S33. Input the graph structure G and the initial feature set {h0(v)} into the logical attention path network model, perform the graph path convolution operation, and calculate the embedding vector of the l-th layer node using the following embedding update formula: in, represents the first-order neighbor set of node v, α v,u is the attention weight from node u to node v, W is a trainable linear transformation matrix, and σ is a nonlinear activation function; S34, for the specified target security event node v s , perform multi-hop path extraction operation, construct a i to v s The set of causal paths between them P={p1,p2,…,p n }, each path p i is an ordered sequence of nodes: p i =<v1,v2,…,v s > S35. For each path p i ∈P calculates the path attention score β i , indicating the path to the target event node v s The causal influence weight of the path score is calculated as follows: Among them, |p i | represents the path p i The number of nodes in , w is the attention score vector, h l (v j ) is the embedding representation of the j-th node in the path at layer l; S36, output causal path set P and corresponding attention score set {β i }, as the target security event node v s The collection of associated causal paths.

5. The method for tracing the safety responsibility of construction units based on HSE graph according to claim 1 is characterized in that: The path attention mechanism of step S35 not only performs weighted aggregation on the node embedding information in each causal path, but also introduces the temporal position of the path node, the risk level of the behavior node, and the structural position weight as combination factors to participate in the overall path scoring process. The global attention score vector and the path feature mapping function are dynamically learned through the training phase, so that the score of each path has differentiated semantic expression capabilities while maintaining structural consistency.

6. The method for tracing the safety responsibility of construction units based on HSE graph according to claim 1 is characterized in that: The S4 specifically includes: S41, for the causal path set P = {p1, p2, ..., p n } and the corresponding path attention score set {β1,β2,…,β n }, for each path p i Each node v in ∈P j Extract three-dimensional evaluation indicators: node centrality C(v j ), behavioral risk level R(v j ), path position factor δ(v j ,p i ), which respectively represent the structural importance of the node in the graph, the risk level of the operation behavior represented by the node, and the relative position of the node in the sequence of the path; S42. Construct a fuzzy responsibility scoring function to calculate the attribution degree of the node in the path. j On path p i The fuzzy responsibility score μ(v j ,p i ) is defined as follows: μ(v j ,p i )=γ1·C(v j )+γ2·R(v j )+γ3·δ(v j ,p i ); Among them, γ1, γ2, and γ3 are the structural importance weight, risk level weight, and path position weight, respectively, satisfying the constraint condition γ1+γ2+γ3=1. The weights are dynamically updated by the optimization algorithm during the model training process. S43, based on the above fuzzy responsibility scoring function, for all paths p i ∈P and the corresponding node v j ∈p i Perform attribution scoring and construct the attribution credibility matrix M, where M i,j =μ(v j ,p i ), n is the number of paths, and m is the number of nodes in a single path; S44. Combine the path attention score and the fuzzy responsibility score to calculate the global attribution score of each node in all paths. Set the total responsibility score of node v to Γ(v). The calculation formula is as follows: Among them, β i Represents path p i The attention score, μ(v j ,p i ) is the node v j Fuzzy scoring in the path; S45, output attribution score vector Γ={Γ(v1),Γ(v2),…,Γ(v k )}, each element in the score vector corresponds to a construction unit node or an operation behavior node in the HSE map.

7. The method for tracing the safety responsibility of construction units based on HSE graph according to claim 1 is characterized in that: The fuzzy responsibility scoring function constructed in S42 is a weighted combination of node centrality, risk level and path position factor. The weight parameters are trainable and meet normalization constraints, thereby realizing attribution scoring under multi-dimensional features.

8. The method for tracing the safety responsibility of construction units based on HSE graph according to claim 1 is characterized in that: The S5 specifically includes: S51. Define the model optimization objective function J, which is used to measure the error between the responsibility score result generated by the model and the known responsibility label. The specific calculation formula is: Among them, U represents the set of nodes with labeled responsibility levels, w x represents the importance weight of node x, φ x Represents the responsibility scoring result output by the model, Indicates the true responsibility level of node x; S52. Set the parameter vector to be optimized Θ = [L, H, α1, α2, α3], where L represents the number of graph convolution layers of the logical attention path network, H represents the number of attention sub-heads, and α1, α2, and α3 represent the node centrality weight, behavior risk level weight, and path position weight in the fuzzy responsibility scoring function, respectively, satisfying the normalization constraint: α1+α2+α3＝1； S53, based on the cuckoo search evolution mechanism, construct the initial parameter population, and use the Levy flight strategy to generate each current solution vector Θ i New candidate solution Θ′ i , and perform integer projection on the layer number parameter L and sub-head number parameter H in the candidate solution, and perform normalization correction on the weight parameters α1, α2, and α3; S54, the candidate solution vector Θ′ i Applied to the responsibility identification model, the aforementioned graph embedding encoding, path attention calculation and fuzzy responsibility scoring processes are executed to obtain the model output responsibility scoring result φ x , and calculate the optimization objective function value J(Θ′ i ); S55. Evaluate the fitness of the current population based on the objective function value, adopt a winner-replacement strategy to retain low-error candidate solutions, and introduce some random individuals to enhance population diversity to form the next generation solution set; S56, repeat S53 to S55 until the maximum number of iterations or the objective function convergence condition is reached, and output the optimal parameter vector The responsibility identification model after final structural optimization is configured and applied to the calculation and output of node responsibility scores.

9. The method for tracing the safety responsibility of construction units based on HSE graph according to claim 1 is characterized in that: The S6 specifically includes: Applied to the logical attention path network model and fuzzy responsibility scoring function, combined with the aforementioned graph embedding representation, path attention score and node attribution score results, the global responsibility score φ of each construction unit node is calculated. x ; According to the responsibility score, interval mapping is performed within the set level threshold range, divided into multi-level responsibility level labels, forming a responsibility node set and level labeling results; the responsibility level results are associated with the original HSE map, and a responsibility traceability table containing node identification, responsibility value and level category is output for construction unit accountability decision-making and risk control management.

Images