A safety production management method based on multi-level hidden danger investigation

Abstract

The present invention discloses a safe production management method based on multi-level hidden danger investigation, which specifically relates to the field of safe production management and is used to solve the problems that existing investigation tasks are mostly issued layer by layer by superior units, with low allocation efficiency, mismatched resource utilization and poor real-time performance; the method constructs a task structure trajectory library by extracting historical investigation task execution logs, identifies the structural position of each unit in the task transmission chain, and establishes a standardized task management hierarchical architecture; further identifies samples with task allocation differences, extracts resource configuration information of non-execution layer units and identifies task execution structure changes, and constructs a joint parameter set; trains a policy model for execution layer task allocation based on the joint parameter set, and constructs a current joint parameter set input model in combination with real-time resource and structure data, generates investigation task allocation results for execution layer units, and realizes accurate push and automatic execution of task allocation.

Description

Technical Field

[0001] The present invention relates to the technical field of safe production management, and more specifically, to a safe production management method based on multi-level hidden danger detection. Background Art

[0002] In current workplace safety management practices, hazard investigation tasks are often assigned by higher-level departments through a series of administrative tiers, ultimately reaching specific implementation units. However, this layered approach to task allocation often results in complex processes and slow response times for periodic tasks. This is particularly true within large enterprises or regional organizations, where multiple units collaborate and tasks are meticulously divided. On the one hand, the long task transmission path leads to information delays, often delaying the optimal response window for implementation units by the time they receive the task. On the other hand, because higher-level units lack accurate information on the real-time resource allocation of their subordinate units, resource shortages or redundant investments often occur, leading to the dual problems of reduced task execution efficiency and resource waste. The existing task allocation mechanism relies heavily on human judgment. Units at all levels may allocate resources and assign tasks based on their own experience, relationships, or established practices during the task decomposition process. This subjective decision-making model not only lacks traceability but also struggles to adapt to the demand for rapid and accurate dispatch during sudden safety risk events. During multiple rounds of periodic inspections, the task allocation structure remains stable on the surface, but the internal task transfer logic and the resource-sharing relationship between units may have undergone subtle but critical changes. These changes are difficult to be effectively utilized in subsequent inspection tasks because they have not been systematically identified and modeled, which in turn affects the scientific nature and timeliness of the inspection strategy.

[0003] Therefore, it is urgent to build a task allocation mechanism for hidden danger investigation and management scenarios, which can automatically identify the management hierarchy structure and resource allocation behavior in historical task communication, extract key influencing factors and build a strategy model, and realize the automated decision-making process of investigation tasks from topic determination to execution unit assignment, so as to solve the problems of response lag, structural redundancy and resource waste caused by manual communication. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a production safety management method based on multi-level hidden danger investigation to solve the problems raised in the above-mentioned background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A production safety management method based on multi-level hidden danger investigation includes the following steps:

[0007] S1: Based on the theme of production safety hazard investigation, historical investigation task execution logs are extracted to establish a task structure trajectory library;

[0008] S2: Extract task nodes and restore the transmission chain from the task structure trajectory library, and build a task management hierarchical structure containing units at all levels based on the task transmission chain;

[0009] S3: In different hidden danger investigation cycles, screen for difference samples that differ from the task allocation of the execution layer of the task management hierarchy, extract the resource configuration parameters of the non-execution layer units in the difference samples, and integrate them to form a resource parameter set;

[0010] S4: Deconstruct the task hierarchy, perform relationship change identification on the task management hierarchy corresponding to the difference samples, and establish a joint parameter set in combination with the resource parameter set;

[0011] S5: Construct the attribution mapping relationship between the joint parameter set and the execution layer task allocation, and train the output execution layer unit task allocation policy model;

[0012] S6: Obtain the joint parameter set built based on real-time data and input it into the policy model, and directly assign troubleshooting tasks to the execution layer based on the model output.

[0013] In a preferred embodiment, in S1, extracting historical inspection task execution logs based on the production safety hazard inspection theme and establishing a task structure trajectory library specifically includes:

[0014] Perform keyword decomposition on all issued hidden danger investigation topics, establish task topic search phrases, and filter task execution record sets from historical investigation task logs;

[0015] Extract the initiating unit, undertaking unit and task assignment content from the task execution record set, and organize the task content into a task structure track based on the master-slave relationship between the units;

[0016] Collect the task structure trajectories of multiple historical hidden danger inspection cycles, centrally integrate all task structure trajectories belonging to the same hidden danger inspection cycle, and establish a task structure trajectory library based on the hidden danger inspection theme.

[0017] In a preferred embodiment, in S2, the task structure trajectory library is subjected to task node extraction and transmission chain restoration, and a task management hierarchical structure including various hierarchical units is constructed based on the task transmission chain, specifically including:

[0018] Connect each trajectory in the task structure trajectory library according to the master-slave relationship between units to form a multi-hop task transfer path between units, record the transfer direction, unit identification and task assignment content, and build a node-based task transfer chain;

[0019] Based on the position differences of units in the task communication chain during the communication process, all unit nodes are divided into three communication levels: initiation layer, intermediate layer and execution layer;

[0020] Among them, the initiation layer and the execution layer correspond to the head and tail nodes of the task transmission chain respectively. The middle layer is a multi-layer structure and the corresponding layer number is marked;

[0021] Construct a task management transfer diagram by matching all task transmission chains with the same unit identifier;

[0022] Bind each unit in each task transmission chain to its corresponding transmission level, and fill the level label back into the task management transmission diagram;

[0023] Structural aggregation is performed on nodes with hierarchical annotations to extract the arrangement patterns that convey the hierarchical structure and construct the task management hierarchy architecture for each hidden danger investigation topic.

[0024] In a preferred embodiment, performing structural aggregation on nodes with hierarchical annotations, extracting the arrangement pattern of the conveyed hierarchical levels, and constructing the task management hierarchical structure of each hidden danger investigation theme specifically includes:

[0025] Group the task management transfer diagram nodes belonging to different hidden danger investigation cycles, and align the unit nodes of different groups according to their hierarchical execution structure;

[0026] Convert the hierarchical numbers of the aligned nodes in different communication chains into relative hierarchical position sequences, and construct a communication hierarchical position matrix, where the matrix dimension is consistent with the maximum depth of the communication hierarchical level;

[0027] Identify the repeated and skipped distribution of levels in the level position matrix, and extract the starting and ending spans of the levels, the average number of levels, and the distribution density of unit nodes in each level in each chain as structural feature vectors;

[0028] The hierarchical arrangement of nodes is clustered and identified using the structural feature vector as input to form a cluster of node arrangement patterns that convey the hierarchy;

[0029] Integrate the clusters of communication level node arrangement patterns into a node-based tree-like task management hierarchy.

[0030] In a preferred embodiment, in S3, in different hidden danger investigation cycles, difference samples that differ from the task allocation of the execution layer of the task management hierarchy are screened, resource configuration parameters of non-execution layer units in the difference samples are extracted, and integrated to form a resource parameter set, specifically including:

[0031] Compare the task management transfer diagrams constructed in different hidden danger investigation cycles with the execution layer of the corresponding hidden danger investigation theme task management hierarchy architecture to extract difference samples with different task assignment contents;

[0032] Extract all intermediate unit nodes in the non-execution layer from the difference samples and collect the hidden danger investigation resource configuration data of the corresponding nodes within the investigation cycle, including the number of investigation personnel and the number of investigation equipment in the subordinate units of the node corresponding to the unit, as well as the available deployment time;

[0033] The resource configuration data of the intermediate layer unit nodes corresponding to all difference samples are converted into a unified resource parameter expression and integrated into a resource parameter set according to the unit identifier.

[0034] In a preferred embodiment, in S4, deconstructing the task hierarchy architecture, performing relationship change identification on the task management hierarchy corresponding to the difference samples, and establishing a joint parameter set in combination with the resource parameter set specifically includes:

[0035] Compare the task management transfer diagram corresponding to the difference sample with the task management hierarchy structure to determine the task transfer path and transmission direction, and identify and mark the structural changes of addition, reduction and rearrangement in the task management hierarchy structure;

[0036] Extract the intermediate unit nodes where the structural relationship changes occur, and generate a structural change parameter set including the change type, unit identifier, and level number;

[0037] The structural change parameter set is integrated with the resource parameter set of the intermediate unit nodes in the difference sample to generate a unified joint parameter set.

[0038] In a preferred embodiment, in S5, constructing an attribution mapping relationship between the joint parameter set and the execution layer task allocation, and training and outputting a strategy model for the execution layer unit task allocation specifically includes:

[0039] The joint parameter set of the intermediate layer unit in each difference sample is associated with the task allocation result of its corresponding execution layer unit to construct a training sample set;

[0040] Train a multi-output regression model in the training sample set with joint parameters as input and execution layer unit tasks as output;

[0041] At the same time, the task allocation error is used as the loss function, the model structure parameters are iteratively updated, and the trained multi-output regression model is used as the strategy model for task allocation.

[0042] In a preferred embodiment, in S6, obtaining a joint parameter set constructed based on real-time data and inputting it into a policy model, and directly assigning troubleshooting tasks to the execution layer according to the model output specifically includes:

[0043] Obtain the current hidden danger investigation theme, analyze the resource status and task management hierarchy of each unit under the current hidden danger investigation theme, and extract the corresponding resource parameter set and structure change parameter set in real time;

[0044] Integrate the resource parameter set and the structure change parameter set into a real-time joint parameter set consistent with the structure of the policy model training phase;

[0045] Input the real-time joint parameter set into the strategy model, record the model's results of allocating inspection tasks to each execution layer unit, and establish a task assignment list.

[0046] The technical effects and advantages of the present invention's production safety management method based on multi-level hidden danger investigation are as follows:

[0047] By constructing a task structure trajectory library and restoring the chain of historical inspection task transmission, a hierarchical analysis of the hidden danger inspection task management architecture is achieved. This clearly reflects the role of each unit in the task transmission process and the transmission sequence, effectively resolving the structural ambiguity and untraceable path issues inherent in traditional multi-level manual transmission. Furthermore, by screening samples with discrepancies in task assignment, extracting and standardizing resource configuration parameters for non-execution-level units, and constructing a joint parameter set that incorporates structural change factors, a structured representation of the key drivers behind task assignment behavior is achieved. The task assignment strategy model trained based on this joint parameter set accurately maps resource and structural conditions to execution-level task configurations, demonstrating excellent generalization and responsiveness. Ultimately, during real-time inspection task delivery, the model output is driven by the real-time constructed joint parameter set, directly completing task assignments for execution-level units. This significantly reduces intermediate transmission links and manual intervention, improves the automation of task dispatching, and aligns response time with resource scheduling, thereby enhancing the overall intelligence of hidden danger inspection management. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is a schematic diagram of a production safety management method based on multi-level hidden danger detection according to the present invention. DETAILED DESCRIPTION

[0049] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0050] Example 1

[0051] Figure 1 The present invention provides a production safety management method based on multi-level hidden danger investigation, which includes the following steps:

[0052] S1: Based on the theme of production safety hazard investigation, historical investigation task execution logs are extracted to establish a task structure trajectory library;

[0053] S2: Extract task nodes and restore the transmission chain from the task structure trajectory library, and build a task management hierarchical structure containing units at all levels based on the task transmission chain;

[0054] S3: In different hidden danger investigation cycles, screen for difference samples that differ from the task allocation of the execution layer of the task management hierarchy, extract the resource configuration parameters of the non-execution layer units in the difference samples, and integrate them to form a resource parameter set;

[0055] S4: Deconstruct the task hierarchy, perform relationship change identification on the task management hierarchy corresponding to the difference samples, and establish a joint parameter set in combination with the resource parameter set;

[0056] S5: Construct the attribution mapping relationship between the joint parameter set and the execution layer task allocation, and train the output execution layer unit task allocation policy model;

[0057] S6: Obtain the joint parameter set built based on real-time data and input it into the policy model, and directly assign troubleshooting tasks to the execution layer based on the model output.

[0058] In S1, based on the safety production hazard investigation theme, historical investigation task execution logs are extracted to establish a task structure trajectory library.

[0059] In actual hazard investigation work, regulatory agencies and businesses at all levels regularly issue inspection tasks with clear themes, based on annual, quarterly, or special rectification schedules. These task themes include specific terms such as "Special Project on Risks of Working at Heights," "Remediation of Electrical Fire Hazards," and "Inspection of Hot Work Procedures in Chemical Plants." To conduct structured search and traceability analysis of these tasks, a keyword-based decomposition of all previously issued hazard investigation task themes is performed. This process involves semantically segmenting each investigation topic text field, identifying semantic keywords such as task object (e.g., "chemical equipment"), task behavior (e.g., "inspection"), and risk type (e.g., "fire"), to form a set of search phrases with categorical attributes. Using this generated set of search phrases as query criteria, a full-text match is performed against the task description fields in the historical investigation task log database, filtering out task records that have semantic overlap or inclusion relationships with any of the term phrases. These matching criteria must cover the task log's subject field, task description field, and task notes to enhance comprehensiveness and accuracy. To eliminate the interference caused by synonymous expressions, a task semantic expansion vocabulary should also be introduced to map some non-standard or colloquial expressions to standard task phrases. For example, although "electricity safety inspection" and "electrical fire prevention and control" use different terms, they should belong to the same investigation topic. Therefore, when establishing a search phrase set, a semantic similarity table needs to be preset to expand the coverage of the keyword phrase. The retrieved historical task logs are matched one by one, and all investigation task records that meet the current topic keyword matching rules are screened out. They are then archived according to the issuance time, task initiator, and record number to form a set of task execution records under the hidden danger investigation topic.

[0060] Core structural information is extracted from each task record to form a task structure trajectory. Each record typically contains the name of the initiating unit, the name of the receiving unit, a description of the task content, and a timestamp. In this step, structural field extraction is first performed on each record to obtain the two key nodes, the initiating unit and the receiving unit. A one-time assignment relationship is confirmed between the two, establishing a master-slave task transfer structure. This extraction excludes batch broadcasts and non-hierarchical notifications. For example, if the same initiating unit issues a horizontal task to multiple receiving units, each task should be constructed as a separate unit pair, and their master-slave relationships recorded separately. Furthermore, keywords in the task content field (e.g., "inspecting the safety of construction work at subcontracting units") are used to assist in confirming whether the receiving unit is performing a transfer task, thereby determining whether the record is an intermediate node in the transfer chain. Unit-part task transfer segments with a master-slave relationship are marked as task structure trajectories, and information such as the task transfer direction, content summary, and timestamp is recorded.

[0061] Using the investigation topic as the primary index dimension, task structure trajectories are categorized according to their corresponding investigation topic tags. The investigation topic tags are derived from the search phrase matching results described above. Clear associations have already been established during the record screening phase, eliminating the need for re-matching. Within the same investigation topic, aggregation is further performed using the investigation period as the secondary index dimension. Investigation periods can be categorized by the year, quarter, or specific remediation period in which tasks are issued, such as "2022 Q3 Chemical Equipment Inspection" or "2023 Spring Festival Fire Protection Special Remediation." Task structure trajectories within each period are considered the transmission chain for a complete investigation. For each period, all trajectories are aggregated by initiating unit and numbered and archived. Statistics such as the period identifier, the number of task structure trajectories, and the number of units covered are recorded. A task structure trajectory library is constructed, dually indexed by investigation topic and period. Each entry in the trajectory library contains several task structure trajectories, each consisting of a unit, task content, and connections between upstream and downstream units. Once the trajectory library is constructed, it is directly used for subsequent analysis of the transmission chain structure and the construction of the task management hierarchy.

[0062] In S2, the task structure trajectory library is used to extract task nodes and restore the communication chain, and a task management hierarchical architecture containing units at all levels is constructed based on the task communication chain.

[0063] After the task structure trajectory library is constructed, the structural information within the trajectory is further organized to form a node-based task transfer chain structure. Each task structure trajectory consists of several unit nodes connected in chronological order and with task transfer logic, exhibiting a clear master-slave relationship. In this step, the unit nodes within the trajectory are sequentially connected according to their master-slave assignment logic to establish a complete multi-hop task transfer path. Each hop in this path represents a task transfer from a superior unit to a subordinate unit, essentially a hierarchical process of task management responsibility. When constructing the path, three key elements must be clearly recorded for each hop of the transfer: the task transfer direction, i.e., the unit that initiates the task assignment; the unit identifier, including the unique identifiers of the superior unit issuing the instruction and the subordinate unit receiving the task; and the task assignment content, i.e., the task field information contained in the task for that hop, such as the operational task description, coverage, or required time limit. This information should be recorded one by one in the connection relationships between the nodes, forming a chain structure with parsable semantics. Each unit node must have a unique identifier and retain its contextual position within the entire task structure trajectory. Repeated unit nodes within the same trajectory are numbered based on the chronological order of tasks and assigned content, and logically numbered using an index. The resulting task delivery chain is a linear, directed structure composed of multiple unit nodes and directional relationships, linked sequentially according to the order of task delivery. Each chain can be traced back to its corresponding original task structure trajectory, ensuring traceability and structural integrity.

[0064] After the node-based task transmission chain is constructed, a unified hierarchical classification process is performed based on the position of each unit node in the transmission path. This process uses the sequence information in the transmission path as the basis for classification, dividing all unit nodes into three transmission layers: the initiating layer, the intermediate layer, and the executing layer. The initiating layer corresponds to the starting unit node in the task transmission chain, that is, the unit that first initiates the task transfer, typically a higher-level management unit at the municipal or head office level. The executing layer is the terminal unit node in the transmission chain, that is, the unit that ultimately assumes responsibility for task execution, such as grassroots grids, project departments, and equipment maintenance teams. The intermediate layer comprises all transfer units between initiation and execution during the transmission process, typically intermediate control or scheduling units in a multi-level organizational structure. The intermediate layer unit nodes should be assigned a hierarchical number within each task transmission chain based on their sequential position in the path. The initiating layer is uniformly labeled Layer 1, and the number is incremented by 1 with each subsequent hop, until the executing layer node. The intermediate layers may have multiple layers, for example, with tasks being assigned from the provincial company to the municipal company, and then to the county company, with levels numbered 2 and 3, respectively. Numbering should strictly follow the path sequence, with no skipping or reversed numbering. All task chains should be uniformly numbered in this manner, forming a structurally consistent hierarchical identification system. Each unit node not only has a unique unit identifier, task content, and information about connections between superiors and subordinates, but also a clear transmission level number, forming a three-dimensional hierarchical information structure.

[0065] All chains are integrated into a global task management transfer graph. This graph uses unit nodes as vertices and edges connecting task transfer relationships between units as connecting paths. It integrates all historical task structure trajectories, forming a panoramic task transfer structure map. The construction process uses unit identification as the core index dimension, merging unit nodes appearing in all chains. Repeated occurrences of the same unit in different task chains are identified and mapped to the same graph node. During the construction process, upstream and downstream path information for all task transfers is preserved, allowing unit nodes in the graph to have multiple upstream and downstream connections, forming a dynamic many-to-many task path network structure. For unit nodes with multiple positions in different tasks, their historical upstream and downstream unit information is recorded as node attributes in the graph to facilitate path dependency analysis. The transfer level number and role category (initiating layer, intermediate layer, execution layer) corresponding to each unit node are written into the node attribute fields, and the full graph is backfilled with these level attributes. Specifically, for all unit nodes from the task delivery chain, their level numbers in the chain are extracted and written uniformly into their corresponding nodes in the management delivery graph, forming a corresponding relationship between the graph node structure and the task delivery chain.

[0066] Structural aggregation is performed on nodes with hierarchical annotations, extracting the permutation patterns of the transmission hierarchy to construct the task management hierarchy for each hidden danger investigation theme. The transfer graphs are grouped by hidden danger investigation cycle, with each group corresponding to a complete task execution cycle for the investigation theme, ensuring structural closure. The graph structure within each cycle includes the task transfer path, the connections between units, and the hierarchical attributes of the nodes. By traversing all transfer graphs, each unit node and its hierarchical annotation information are extracted, and the hierarchical position of the unit in each graph is recorded. After uniformly numbering the unit nodes, cross-cycle alignment is performed based on the unit identification. This addresses the problem of hierarchical drift that may occur between units in different cycles. This means that the same unit node may assume different task roles in different cycles, resulting in different hierarchical numbers in the structure. Therefore, the alignment operation does not simply rely on hierarchical numbers. Instead, it uses the unit identification as the primary anchor point and performs refined matching based on the relative position of the unit in the task path. For example, a unit serves as a middle-level node (e.g., level 3) in cycle A and sinks to the execution level (e.g., level 5) in cycle B. Empty nodes are added at the missing middle-level levels as an alignment control aid (e.g., inserting two empty nodes before level 3 in cycle A to convert level 3 to level 5), and dynamic alignment relationship maintenance is performed by establishing a unified unit level table across cycles.

[0067] After aligning all unit nodes in the transfer graph, the hierarchical information of each unit across multiple cycles is further standardized to form a structural representation that facilitates statistical analysis. Specifically, a transmission hierarchical position matrix is ​​constructed. The hierarchical numbers of all unit nodes in each task transmission chain are converted into relative hierarchical positions and mapped into a fixed-dimensional matrix. Each row of the matrix corresponds to a chain, and each column corresponds to a standardized hierarchical position, starting from level 1 and proceeding to the highest level number corresponding to the maximum transmission depth in the data set. After constructing the position matrix, the distribution of unit nodes within each column of the matrix is ​​analyzed to identify several key structural features. First, patterns of dense recurrence of unit nodes in certain hierarchical columns (identified by statistical node frequency) are identified to capture the repetitive distribution characteristics of the structure. Second, if certain chains continuously skip levels, such as jumping directly from level 3 to level 6, the structure exhibits jump distribution characteristics (including empty nodes). Structural indicators are extracted for each chain, including but not limited to the span of the hierarchy (the difference between the first and last hierarchy levels), the average number of levels in the chain (the average level to which nodes belong), and the number and proportion of unit nodes in each level. These structural indicators are converted into multidimensional numerical features to form the structural feature vector of each chain.

[0068] After constructing the structural feature vectors, clustering was performed to identify hierarchical patterns within the task management structure based on numerical representations. During implementation, a clustering algorithm was used to cluster the structural feature vectors of all chains, forming multiple node arrangement pattern clusters. Each pattern cluster represents a set of chains with significant similarities in the hierarchical structure of the transmission path. Its members exhibit similar or identical hierarchical trends within the task transmission structure (identified through matrix similarity). After cluster analysis of the structural feature vectors, multiple structural pattern clusters were obtained. Each cluster consists of several task transmission chains that exhibit high similarity in terms of structural profile, hierarchical distribution, and task node density. To extract representative task management architectures from these multiple structural pattern clusters that can be used for standardized modeling, cluster representativeness assessment was necessary. Specifically, clusters were screened based on the highest frequency of occurrence of their corresponding unit node identifiers. Structural pattern clusters that recur across multiple inspection cycles, demonstrate high structural stability, and reflect typical transmission logic were prioritized. The identifiers and hierarchical distribution characteristics of all unit nodes in the pattern clusters were extracted to create a structural template, and empty nodes in the template were removed. Based on the hierarchical position of unit nodes in the structure, the direction of communication connections, and the logic of task assignment, a standardized tree-like task management hierarchy is integrated. This tree-like structure takes the initiation layer as the root node and sequentially establishes unit nodes at the intermediate and execution layers downward, maintaining the characteristics of directed acyclicity and structural closure. Each node contains the unit identification, the direction of the task transfer path, and the communication layer to which it belongs, forming a complete multi-level task management organizational structure chart. The final output tree-like structure architecture has three characteristics: first, it is derived from the real task execution structure in multiple actual cycles and has engineering feasibility; second, the structural pattern cluster formed by clustering structural feature vectors has statistical significance and wide adaptability; third, the tree-like structure provides a unified alignment benchmark for subsequent task resource allocation modeling and structural change identification, becoming the basic structural expression for the intelligent evolution of task scheduling. This process effectively overcomes the problems of hierarchical inconsistency and severe structural drift in historical task structure data, and outputs a stable structural input that can be directly used by subsequent models.

[0069] In S3, during different hidden danger investigation cycles, difference samples that differ from the task allocation of the execution layer of the task management hierarchy are screened, and the resource configuration parameters of the non-execution layer units in the difference samples are extracted and integrated to form a resource parameter set.

[0070] Based on archived historical hazard inspection records, task management transfer graphs from different time periods are grouped by hazard inspection theme, with each group corresponding to an independent hazard inspection cycle. For each cycle, a task management transfer graph is constructed in the previous step and contains information about the task assignment paths, transmission directions, and structural hierarchies between unit nodes. For the current hazard inspection theme, a standard task management hierarchy architecture established for that theme is selected as a benchmark template for structural alignment and discrepancy identification. The set of execution-level nodes in the task management transfer graph for each inspection cycle is compared one-to-one with the unit nodes in the standard architecture. This comparison not only matches unit identification but also, if the task assignment content is not standardized, performs semantic analysis of the task assignment content to extract elements such as task type, workload estimate, and dispatch frequency from the task description. Task content vectors are then constructed for similarity analysis. If the type or number of tasks received by a unit node in a given cycle differs significantly from that in the standard architecture, the node and the inspection cycle it belongs to are marked as discrepant samples. The discrepancy is determined based on mismatches in task type and content. The final output is a set of difference samples that includes the inspection cycle identifier, unit node identifier, and task difference description.

[0071] Taking the generated set of difference samples as input, we focus on all intermediate-level unit nodes at the non-execution level. Intermediate-level nodes perform intermediate management responsibilities such as task transmission, coordination, aggregation, or secondary dispatch. Based on the aforementioned transmission chain and structural annotations, we extract all unit nodes that meet the intermediate-level judgment criteria from each difference sample to form a set of intermediate-level unit nodes. For each unit node in this set, we collect its hidden danger investigation resource allocation data for the investigation cycle. Specifically, we collect data on the following three categories: the total number of subordinate units corresponding to the node's unit and the number of investigation personnel available; the types and quantities of investigation equipment available for the unit; and the effective dispatchable time for the unit to be assigned to investigation tasks during the investigation period. Resource data is collected from original task dispatch records, equipment scheduling logs, and personnel schedule data, standardized and aggregated using a unified field to avoid resource statistics errors caused by inconsistent data caliber. All data is organized using unit identifiers as the index dimension to form a raw resource configuration dataset for intermediate-level nodes. A standardized resource parameter representation format is constructed for the collected resource configuration data for intermediate-level unit nodes. To facilitate subsequent model training and policy attribution, this format must meet the following requirements: fixed parameter items, consistent dimensions, and easy batch comparison. The resource configuration data of each unit node will be converted into a multidimensional vector form, specifically including: the number of subordinate inspection personnel, the total number of available inspection equipment, the number of types of unit inspection equipment, the unit's effective inspection and deployment time, the unit's inspection task acceptance frequency, and other dimensions. Numerical normalization and dimensional unification operations will be performed to form a resource parameter vector with a consistent structure. After the parameterized expression is completed, the resource parameter vectors of the intermediate unit nodes in all difference samples are integrated according to the unit's unique identifier, duplicate records are eliminated, and the resource change track of each unit in each inspection cycle is retained. The final output result is a resource parameter set.

[0072] In S4, the task hierarchy architecture is deconstructed, relationship change identification is performed on the task management hierarchy corresponding to the difference samples, and a joint parameter set is established in combination with the resource parameter set.

[0073] Based on the set of differential samples screened in the previous step, the task management transfer diagram corresponding to each differential sample is compared against the standard task management hierarchy structure for the current hidden danger investigation topic. The key elements of this comparison include the task transfer paths and transmission directions between unit nodes. In this process, all paths in the transfer diagram are first parsed through unit node sequences to extract the upstream and downstream transmission relationships between consecutive units. A set of directed paths is then constructed based on unit identifiers and hierarchical numbers. Simultaneously, a set of directed paths in the standard hierarchy structure is constructed in the same format for comparison. During this comparison, any path that appears in the differential sample but not in the standard structure is marked as a newly added structure. Paths that exist in the original standard structure but are missing from the differential sample are marked as a newly added structure. If a unit node exists in both structures but its connections to upper and lower-level units have changed, such as a shift in task transfer order or relative hierarchical position, this is marked as a structural rearrangement. All annotations are mapped to the unit node level and attached as structural change labels to the corresponding intermediate-level unit nodes, forming a unit-level structural change record. At the same time, a standard structural change parameter format is generated for the above-mentioned change annotation content. This parameter set is constructed for each intermediate unit node that undergoes structural changes. The parameter items include: change type (addition, reduction or rearrangement), unit unique identifier, original level number and changed level number (if applicable), and a list of upstream and downstream connected units in the task communication chain. All structural change parameters are uniformly coded in a structured form.

[0074] From the set of identified intermediate-layer unit nodes with structural relationship changes, extract all units with clear structural change labels and establish a unique index based on the unit identifier. In the structural change record of each unit node, retain the change type information corresponding to the unit (such as addition, reduction, or rearrangement), and extract the level numbers before and after the change for subsequent structural change trend analysis. If a unit node undergoes repeated changes of different types in multiple investigation cycles, record them separately and introduce a timestamp dimension to distinguish them. After the structural change parameter set is constructed, it is fused with the resource parameter set. The fusion logic matches based on the unit identifier, and the structural change parameters of the same unit in the same investigation cycle are spliced ​​with the corresponding resource parameter vector to generate a joint parameter vector for a single unit cycle. Each joint parameter vector consists of two main components: the first is the structural change parameter segment, which includes a discrete code for the change type, the change level difference (e.g., a change from level 3 to level 2 is marked as -1), and whether multiple transmission chains are involved, among other structural evolution characteristics; the second is the resource parameter segment, which directly inherits the resource dimension vector defined in the previous step and includes the number of personnel, equipment, and task allocation duration. An example is shown below:

[0075] Taking the unit node PA as an example, during the spring special equipment hidden danger inspection cycle, the unit was originally affiliated with the middle-level unit OA, but now it is directly affiliated with the upper-level unit OB after the structural adjustment, and the level number remains unchanged at the third level. The resource status corresponding to this node is 2 subordinate units, 8 inspection personnel, 3 detection equipment, and the available working time in this cycle is 360 hours. To construct a joint parameter vector, the structural change and resource configuration data are uniformly coded using the following paradigm: [PA, OA, OB, 3, 2, -1, 02, 0, 3, 8, 3, 360], where the meanings of the fields in this vector are: current unit ID, original upper-level unit, current upper-level unit, level number, structural change type (2 indicates "affiliation change"), level position change flag (-1 indicates the level remains unchanged), unit type code (02 indicates a middle-level unit), redundant flag bit (0), number of subordinate units (3), number of inspection personnel (8), number of equipment (3), and available working time (360 hours). This paradigm combines structural change information with resource allocation feature expressions. The resulting joint parameter vector is written as a unified sample unit into a joint parameter set. This parameter set uses unit identification and troubleshooting period as dual indexes in data organization, ensuring high-precision joint attribution of structure and resources during subsequent policy modeling. This parameter set not only provides a structural basis for changes in task allocation logic but also incorporates resource-level constraint information, serving as the core input for the intelligent evolution of task scheduling strategies.

[0076] In S5, the attribution mapping relationship between the joint parameter set and the execution layer task allocation is constructed, and the policy model for the output execution layer unit task allocation is trained.

[0077] After constructing the joint parameter set, a clear input-output correspondence is established between it and the actual task execution results, thereby constructing a training sample set. For each difference sample, an association mapping is established based on the task transfer path between the intermediate-level units and the execution-level units in the task management transfer diagram. The joint parameter vector is associated with the task category and amount actually completed by the execution unit during the current inspection cycle. The task allocation results include multiple dimensions, such as task type, workload estimate (which can be the number of tasks or operation time), and allocation frequency. These multi-dimensional outputs are mapped one-to-one with the input vectors to form a complete training sample. During the sample organization process, to ensure consistent representation and modeling efficiency, all input vectors are standardized as fixed-length numerical sequences, with each field position corresponding to a clear resource or structural parameter. The output results are concatenated according to the execution-level unit order and normalized to eliminate numerical deviations caused by differences in unit number or task density. The resulting training sample set consists of a large number of matching pairs of joint parameter input vectors and corresponding task allocation output results.

[0078] After the sample set is constructed, the task allocation strategy model is trained based on this dataset. Considering the multi-objective and multi-dimensional nature of task allocation results, and the mutual constraints between the task loads of multiple execution units, this embodiment uses a multi-output regression modeling approach for training. The multi-output regression model essentially simultaneously learns the mapping relationship between multiple related target variables, enabling a holistic prediction of the task allocation results of execution-level units within a unified model structure. During the model training phase, the input data is the joint parameter vector constructed in the previous step, and the output is the execution-level task allocation vector bound to it. Through repeated iterative training, the model structure gradually captures the nonlinear mapping between input features and output results. In the model structure design, a branching structure is adopted to accommodate multiple output channels, ensuring that the prediction subspaces of different execution units are independent of each other while also having a certain degree of parameter sharing, thereby improving the model's generalization and prediction accuracy. During the training process, training samples are loaded in batches, and forward propagation, loss calculation, and gradient backpropagation steps are performed. The internal weight parameters are continuously adjusted, allowing the model to gradually converge to the optimal mapping function between the joint parameter vector and the task allocation results. After training convergence, the model accepts any formatted joint parameter input vector and outputs the corresponding multi-objective task allocation prediction results, forming the core logic of the policy model. To improve the reliability of prediction results, task allocation error is introduced as the primary loss function during model training. This error metric is based on the absolute deviation or squared difference between the predicted task value and the actual task execution value for each execution layer unit. A total loss function is constructed by batch averaging over all samples, and this loss is optimized in each iteration. Specifically, for each execution unit in each sample, the difference between the task amount output by the model and the true value in the log record is calculated. The differences across all units are then weighted and aggregated to form the loss value for each sample. During full-sample batch training, the average of all sample losses is then taken as the total loss for this iteration, which drives the inverse adjustment of model weights. To prevent individual extreme task values ​​from excessively affecting the overall loss, a truncation mechanism is introduced in the loss calculation, which applies a weight reduction strategy for deviations exceeding a set upper limit, thereby improving training robustness. After training, the model has the ability to convert any legal joint parameter vector input into execution-layer unit task predictions, and can adapt to complex task environment changes such as intermediate-layer structural adjustments and resource allocation fluctuations. Therefore, it can be directly deployed as a policy model in the automated task allocation process, replacing the traditional manual experience configuration process, and realizing intelligent and real-time inspection task assignment.

[0079] In S6, the joint parameter set built based on real-time data is obtained and input into the policy model, and the troubleshooting tasks are directly assigned to the execution layer according to the model output.

[0080] When a hazard inspection task is about to begin, input data for the strategy model is dynamically generated based on the actual scenario's topic type, organizational structure, and resource status, ensuring realistic and timely task assignment. First, the current inspection topic number and task classification type (such as "Special Remediation of Confined Space Operations" or "Inspection of Temporary Hazardous Chemical Storage Sites") are identified. This determines the set of organizational units required to respond to the task and retrieves their actual resource status data for the current inspection cycle. Resource status data includes, but is not limited to, the number of subordinate units under each intermediate-level unit, the total number of qualified personnel, the number of on-site inspection tools and equipment and their current availability, and the frequency of actual inspection task participation over the past three or seven days. This data is standardized to form a real-time resource parameter set, with the field structure consistent with the resource parameters in previous training samples. Combined with the current task management hierarchy, the model identifies any structural changes since the last training cycle. By comparing the current task management transfer diagram with the standard task management hierarchy from the previous training phase, structural changes such as the reassignment of an intermediate-level unit from Unit C to Unit D, premature downsizing of a unit from the three-tier structure to an execution-level unit, or the disappearance of a unit from the structure due to merger or rescission can be identified. These structural changes must be converted into a standardized structural change parameter set. Each change parameter record contains fields such as the change type, the unit's unique identifier, the hierarchy number before and after the adjustment, and the expression of the change in the parent unit. After obtaining the resource parameter set and structural change parameter set, the field structures of the two parameter sets are aligned and concatenated to generate a unified real-time joint parameter set for the current hazard investigation topic. The field order, value type specifications, and number of fields in this parameter set must strictly inherit the joint parameter template used during the model training phase to avoid input mismatches due to structural inconsistencies. Each vector in the real-time joint parameter set corresponds to a node in the intermediate-level unit, forming a set of vectors that serves as input.

[0081] After the joint parameter set is constructed, it is loaded as input into the trained policy model, triggering the task assignment process. The policy model is a multi-output regression structure. Its input channel receives the joint parameter vector of the intermediate-layer unit node, and its output channel predicts the task assignment result for the execution-layer unit associated with the node's task. The model execution process is a batch prediction process. Each inference cycle, based on a set of input vectors, sequentially outputs the task target values ​​for multiple execution units. These include dimensions such as the inspection task type identifier, task magnitude estimate, planned execution time window, and task weight level. To meet the execution efficiency requirements of the current inspection and deployment cycle, the model output must clearly map the execution unit's identification field to the corresponding position in the output vector. For example, the first dimension of the model output represents the number of inspection tasks for the execution unit, the second dimension represents the inspection round, and the third dimension represents the temporary inspection response instruction. Based on this structure, all task metrics output by the model are aggregated by unit to construct a complete inspection task assignment list. Each execution unit occupies a row in the list, recording the unit code, the received task type, the estimated task quantity, the recommended execution order, and the annotated time window. To enhance the transparency and traceability of task assignments, the task assignment list includes information about the source of the intermediate-layer unit nodes, that is, which intermediate-layer node input triggers the model assignment result for the task, to facilitate subsequent evaluation of whether there are structural deviations in task assignments. Once the task assignment list is formed, it serves as the execution input for the hidden danger investigation task assignment module, or is transferred to the task scheduling process after review and confirmation by the dispatcher. This operation realizes the task binding between the model reasoning output and the actual execution unit, effectively improving the automation, real-time performance and resource matching accuracy of the investigation task assignment, eliminating the lag and subjectivity problems existing in the traditional manual negotiation assignment model, and providing technical support for comprehensively improving the organizational response efficiency of hidden danger investigation.

[0082] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters and thresholds in the formulas are set by technicians in this field according to actual conditions.

[0083] The above embodiments can be implemented in whole or in part via software, hardware, firmware, or any other combination. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product comprises one or more computer instructions or computer programs. When loaded or executed on a computer, the processes or functions described in the embodiments of this application are fully or partially performed. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired means (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium accessible by a computer or a data storage device such as a server or data center that contains a collection of one or more available media. The available medium can be magnetic media (e.g., floppy disks, hard disks, tapes), optical media (e.g., DVDs), or semiconductor media. The semiconductor media can be a solid-state drive.

[0084] Those skilled in the art will appreciate that the modules and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0085] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and modules described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0086] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the modules is only a logical function division. In actual implementation, there may be other division methods, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or modules, which can be electrical, mechanical or other forms.

[0087] The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules, and may be located in one place or distributed across multiple network modules. Some or all of the modules may be selected to achieve the purpose of this embodiment according to actual needs.

[0088] In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0089] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0090] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

[0091] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A production safety management method based on multi-level hidden danger investigation, characterized in that: The steps include: S1: Based on the theme of production safety hazard investigation, historical investigation task execution logs are extracted to establish a task structure trajectory library; S2: Extract task nodes and restore the transmission chain from the task structure trajectory library, and build a task management hierarchical structure containing units at all levels based on the task transmission chain; S3: In different hidden danger investigation cycles, screen for difference samples that differ from the task allocation of the execution layer of the task management hierarchy, extract the resource configuration parameters of the non-execution layer units in the difference samples, and integrate them to form a resource parameter set; S4: Deconstruct the task hierarchy, perform relationship change identification on the task management hierarchy corresponding to the difference samples, and establish a joint parameter set in combination with the resource parameter set; S5: Construct the attribution mapping relationship between the joint parameter set and the execution layer task allocation, and train the output execution layer unit task allocation policy model; S6: Obtain the joint parameter set built based on real-time data and input it into the policy model, and directly assign troubleshooting tasks to the execution layer based on the model output.

2. A production safety management method based on multi-level hidden danger investigation according to claim 1, characterized in that: In S1, based on the safety production hidden danger investigation theme, historical investigation task execution logs are extracted to establish a task structure trajectory library, which specifically includes: Perform keyword decomposition on all issued hidden danger investigation topics, establish task topic search phrases, and filter task execution record sets from historical investigation task logs; Extract the initiating unit, undertaking unit and task assignment content from the task execution record set, and organize the task content into a task structure track based on the master-slave relationship between the units; Collect the task structure trajectories of multiple historical hidden danger inspection cycles, centrally integrate all task structure trajectories belonging to the same hidden danger inspection cycle, and establish a task structure trajectory library based on the hidden danger inspection theme.

3. A production safety management method based on multi-level hidden danger investigation according to claim 1, characterized in that: In S2, the task structure trajectory library is extracted and the transmission chain is restored. Based on the task transmission chain, a task management hierarchical structure containing units at all levels is constructed, specifically including: Connect each trajectory in the task structure trajectory library into multiple task transfer paths between units according to the master-slave relationship between units, record the transmission direction, unit identification and task assignment content, and build a node-based task transmission chain; Based on the position differences of units in the task communication chain during the communication process, all unit nodes are divided into three communication levels: initiation layer, intermediate layer and execution layer; Among them, the initiation layer and the execution layer correspond to the head and tail nodes of the task transmission chain respectively. The middle layer is a multi-layer structure and the corresponding layer number is marked; Construct a task management transfer diagram by matching all task transmission chains with the same unit identifier; Bind each unit in each task transmission chain to its corresponding transmission level, and fill the level label back into the task management transmission diagram; Structural aggregation is performed on nodes with hierarchical annotations to extract the arrangement patterns that convey the hierarchical structure and construct the task management hierarchy architecture for each hidden danger investigation topic.

4. A production safety management method based on multi-level hidden danger investigation according to claim 3, characterized in that: The structure aggregation of nodes with hierarchical annotations is performed to extract the arrangement pattern of the conveyed hierarchical levels and construct the task management hierarchical structure of each hidden danger investigation theme, specifically including: Group the task management transfer diagram nodes belonging to different hidden danger investigation cycles, and align the unit nodes of different groups according to their hierarchical execution structure; Convert the hierarchical numbers of the aligned nodes in different communication chains into relative hierarchical position sequences, and construct a communication hierarchical position matrix, where the matrix dimension is consistent with the maximum depth of the communication hierarchical level; Identify the repeated and skipped distribution of levels in the level position matrix, and extract the starting and ending spans of the levels, the average number of levels, and the distribution density of unit nodes in each level in each chain as structural feature vectors; The hierarchical arrangement of nodes is clustered and identified using the structural feature vector as input to form a cluster of node arrangement patterns that convey the hierarchy; Integrate the clusters of communication level node arrangement patterns into a node-based tree-like task management hierarchy.

5. The method for safe production management based on multi-level hidden danger investigation according to claim 1, characterized in that: In S3, during different hidden danger investigation cycles, we screen for difference samples that differ from the task allocation of the execution layer of the task management hierarchy, extract the resource configuration parameters of the non-execution layer units in the difference samples, and integrate them to form a resource parameter set, which specifically includes: Compare the task management transfer diagrams constructed in different hidden danger investigation cycles with the execution layer of the corresponding hidden danger investigation theme task management hierarchy architecture to extract difference samples with different task assignment contents; Extract all intermediate unit nodes in the non-execution layer from the difference samples and collect the hidden danger investigation resource configuration data of the corresponding nodes within the investigation cycle, including the number of investigation personnel and the number of investigation equipment in the subordinate units of the node corresponding to the unit, as well as the available deployment time; The resource configuration data of the intermediate layer unit nodes corresponding to all difference samples are converted into a unified resource parameter expression and integrated into a resource parameter set according to the unit identifier.

6. A production safety management method based on multi-level hidden danger investigation according to claim 1, characterized in that: In S4, the task hierarchy is deconstructed, and relationship changes are identified for the task management hierarchy corresponding to the difference samples. A joint parameter set is established in combination with the resource parameter set. Specifically, the following steps are performed: Compare the task management transfer diagram corresponding to the difference sample with the task management hierarchy structure to determine the task transfer path and transmission direction, and identify and mark the structural changes of addition, reduction and rearrangement in the task management hierarchy structure; Extract the intermediate unit nodes where the structural relationship changes occur, and generate a structural change parameter set including the change type, unit identifier, and level number; The structural change parameter set is integrated with the resource parameter set of the intermediate unit nodes in the difference sample to generate a unified joint parameter set.

7. A production safety management method based on multi-level hidden danger investigation according to claim 1, characterized in that: In S5, the attribution mapping relationship between the joint parameter set and the execution layer task allocation is constructed, and the policy model for outputting the execution layer unit task allocation is trained. Specifically, the following are included: The joint parameter set of the intermediate layer unit in each difference sample is associated with the task allocation result of its corresponding execution layer unit to construct a training sample set; Train a multi-output regression model in the training sample set with joint parameters as input and execution layer unit tasks as output; At the same time, the task allocation error is used as the loss function, the model structure parameters are iteratively updated, and the trained multi-output regression model is used as the strategy model for task allocation.

8. A production safety management method based on multi-level hidden danger investigation according to claim 1, characterized in that: In S6, a joint parameter set based on real-time data is obtained and input into the policy model. Based on the model output, the execution layer is directly assigned troubleshooting tasks, specifically including: Obtain the current hidden danger investigation theme, analyze the resource status and task management hierarchy of each unit under the current hidden danger investigation theme, and extract the corresponding resource parameter set and structure change parameter set in real time; Integrate the resource parameter set and the structure change parameter set into a real-time joint parameter set consistent with the structure of the policy model training phase; Input the real-time joint parameter set into the strategy model, record the model's results of allocating inspection tasks to each execution layer unit, and establish a task assignment list.

Images