Artificial intelligence-based county fire safety evaluation standard generation method

Abstract

This invention relates to the field of fire safety inspection technology and discloses a method for generating county-level fire safety evaluation standards based on artificial intelligence. The method includes receiving a raw county-level fire data set containing fire records from multiple dimensions; performing multiple rounds of cross-validation within the set to filter for a stable set of elements; inputting this set into a dynamic semantic network that simulates element dependencies and transmission relationships using nodes and associated edges; injecting real-time county-level environmental state parameters to trigger node state migration; and reconstructing the text-based fire safety evaluation standard clauses according to the final steady-state distribution. By obtaining stable elements through multiple rounds of cross-validation, and generating environmentally adaptable clauses through the dynamic semantic network combined with real-time parameters, the method achieves stable standard framework elements and dynamic response clauses to the environment.

Description

Technical Field

[0001] This invention relates to the field of fire safety inspection technology, specifically to a method for generating county-level fire safety evaluation standards based on artificial intelligence. Background Technology

[0002] The current county-level fire safety evaluation standards rely heavily on manual review of legal provisions and historical cases, or directly select frequently occurring fire hazard records as core elements. Conventional practices often involve statistically analyzing data based on a single dimension before selecting elements, neglecting the consistency and stability within multi-dimensional data. This makes the selected elements susceptible to data noise or biases in certain scenarios, hindering the formation of a universally applicable core framework.

[0003] Conventional solutions also suffer from static standard generation logic, often relying on fixed rules to map elements and clauses. The relationships between elements are solidified by pre-defined logic, failing to respond to real-time environmental changes. When the county-level environment changes, standard requirements struggle to adapt promptly, and the dependencies and transmission relationships between elements cannot be dynamically adjusted, resulting in a disconnect between generated clauses and actual scenario needs. Existing methods lack both a verification and screening mechanism for stable elements within multi-dimensional data and a technical path for generating standards by dynamically associating elements with the real-time environment. A standard generation method that can handle the stability of multi-dimensional data and respond to real-time environmental changes is needed. Summary of the Invention

[0004] The purpose of this invention is to provide a method for generating county-level fire safety evaluation standards based on artificial intelligence, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, this invention provides a method for generating county-level fire safety evaluation standards based on artificial intelligence, the method comprising: Receive a raw county-level fire protection data set, which contains fire-related records of various dimensions; Multiple rounds of cross-validation were performed within the original county-level fire protection data set to select a stable set of elements that constitute the core standard framework from fire protection-related records from various dimensions. The set of stable elements is input into a dynamic semantic network, which simulates the dependency and transmission relationships between fire protection elements through nodes and associated edges; Real-time acquired county-level environmental status parameters are injected into the dynamic semantic network to trigger state transitions of nodes within the network; Based on the final steady-state distribution of node state transitions, the fire safety evaluation standard clauses in text form are parsed and reconstructed.

[0006] Preferably, performing multiple rounds of cross-validation within the original county-level fire protection data set includes the following steps: A subset of data was randomly selected from the original county-level fire protection data set as the verification benchmark; The remaining fire-related records were used to perform a consistency comparison of the verification benchmark, and data records that maintained a high degree of consistency in multiple comparisons were marked. Change the verification benchmark and repeat the consistency comparison and marking process until all data subsets have been traversed; The set of highly consistent data records obtained from all the marking processes is defined as the stable element set.

[0007] Preferably, the construction and initialization of the dynamic semantic network are achieved in the following ways: Map each independent element in the stable element set to an independent node in the network; Analyze the patterns of simultaneous or successive occurrence of different fire protection elements in historical data, and quantify the frequent co-occurrence or successive relationships as the weights of the associated edges; Based on the quantized associated edge weights, each independent node is connected to form an initial structure of a dynamic semantic network with weighted connections.

[0008] Preferably, the county-level environmental status parameters include real-time meteorological monitoring data, registration information of major events in the region, and real-time on-duty status data of fire-fighting forces; The injected real-time acquired county-level environmental status parameters refer to converting meteorological monitoring data, major event registration information, and real-time on-duty status data of fire fighting forces into numerical vectors respectively. The transformed numerical vector is used as an external stimulus signal and superimposed on the corresponding environment-aware node in the dynamic semantic network.

[0009] Preferably, the triggering of state transitions for nodes within the network includes an iterative propagation process: Starting from the environmental perception node that receives the external stimulus signal, the stimulus signal is transmitted proportionally to the next layer node directly connected to it according to the weight of the associated edge. The next layer node integrates all the received incoming signals and combines them with its own historical state to calculate the node's new state value; The node with the new state value is used as a new signal source to continue propagating the state change to its neighboring nodes, and this process is repeated cyclically. When the change in the state value of all nodes in the entire dynamic semantic network is lower than a preset threshold, the network is determined to have entered the final steady-state distribution.

[0010] Preferably, the step of parsing and reconstructing the fire safety evaluation standard clauses in text form based on the final steady-state distribution of node state transitions specifically includes: Under the final steady-state distribution, the state values ​​of all nodes in the dynamic semantic network are sorted, and the set of active nodes whose state values ​​exceed the critical level is selected. From the set of active nodes, trace the original fire protection element corresponding to each active node in the set of stable elements; Based on the weights of the associated edges between active nodes, determine the logical order and subordinate relationships among the original fire protection elements; According to the established logical order and subordinate relationships, the original fire protection elements are organized into a structured natural language description, forming the final fire safety evaluation standard clauses.

[0011] Preferably, determining the logical order and subordinate relationship between the original fire protection elements based on the association edge weights between active nodes is achieved through the following steps: Set a threshold for the weight of associated edges, and filter out all associated edges that connect two active nodes and whose weights exceed the threshold from the dynamic semantic network to form a set of strongly associated edges; Analyze the two active nodes connected by each edge in the strongly correlated edge set. Based on the direction of the directed edge, mark the original fire protection element corresponding to the node as the starting point as the preceding element or the superior element, and mark the original fire protection element corresponding to the node as the ending point as the following element or the subordinate element. Traverse the set of strongly related edges to construct a logical relationship graph with active nodes as vertices and directed strongly related edges as connections, and identify one or more nodes with the highest in-degree or centrality in the logical relationship graph as core nodes. Using the core node as the root node, and based on the direction of the associated edges, the logical relationship graph is traversed using a breadth-first or depth-first strategy to generate a tree-like or hierarchical logical structure model. Based on the hierarchical position and connection path of each node in the logical structure model, the logical order and subordinate relationship between the original fire protection elements corresponding to all active nodes are finally determined.

[0012] Preferably, the method further includes a confidence assessment of the generated fire safety evaluation standard clauses: Extract the set of active nodes used to generate the standard terms, as well as the weights of the associated edges between these active nodes; Calculate the average state value of the active node set and the average strength of the associated edge weights; The average state value and average intensity are input into a pre-trained evaluation function, and the output of the evaluation function is the confidence score of the currently generated fire safety evaluation standard clause.

[0013] Preferably, when the confidence score is lower than the acceptance threshold, the method initiates a feedback optimization process: Record the set of active nodes that lead to low confidence scores and their associated edge weights; The associated edge weight information is matched with high-confidence cases in a standard clause case library to identify differences. Based on the identified discrepancies, the weight parameters of the corresponding related edges in the dynamic semantic network are adjusted in reverse to complete an adaptive optimization of the network parameters.

[0014] Preferably, the method further includes performing a logical conflict detection after generating the fire safety evaluation standard clauses: Extract all key phrases containing mandatory requirements from the generated textual fire safety evaluation standard clauses; In a predefined fire protection code conflict rule base, retrieve other code entries that may be mutually exclusive with the key phrase; If mutually exclusive entries are found, a logical conflict warning will be marked in the generation record of the fire safety evaluation standard clause.

[0015] Compared with the prior art, the beneficial effects of the present invention are: Multiple rounds of cross-validation were performed within the original county-level fire protection dataset to select a stable set of elements constituting the core standard framework from fire-related records across various dimensions. The multiple rounds of cross-validation divided the original data into several subsets for iterative verification. Each round tested whether the elements consistently served as core support under different subset combinations. Only those elements retained in all or most rounds were included in the stable element set. This cross-subset consistency check freed the element set from dependence on single-dimensional data or specific samples, enabling it to maintain the same core position across multi-source heterogeneous data. This ensured that the elements upon which the core standard framework is based possess a common foundation across scenarios, reducing element omissions and misselections caused by data sampling bias or local anomalies, and making the framework more closely aligned with the complex characteristics of actual county-level fire protection.

[0016] A dynamic semantic network simulating the dependencies and transmission relationships between fire safety elements is input into a stable set of elements. This network uses nodes and associated edges to simulate these dependencies and transmission relationships. Real-time acquired county-level environmental state parameters trigger state transitions within the network. Based on the final steady-state distribution of these state transitions, the textual form of the fire safety evaluation standard clauses is analyzed and reconstructed. The dynamic semantic network uses nodes to carry elements and associated edges to express the interactions between elements. Upon the input of real-time environmental parameters, nodes undergo state changes according to dependency and transmission rules, which are then transmitted to associated nodes, forming a state transition process across the entire network. The steady-state distribution presents the activation degree and priority order of each node under environmental influence. This distribution is analyzed, and clause elements are organized according to intensity and order. This allows clause generation to be based on the linkage results between elements driven by the real-time environment, enabling the standard clauses to present a focus and order that matches the current environment. The relationships and presentation forms between elements are automatically adjusted as the environmental state changes, ensuring that the clause content corresponds to the dynamic conditions of the county-level fire scene. Attached Figure Description

[0017] Figure 1 This is a schematic diagram illustrating the working principle of the county-level fire safety evaluation standard generation method based on artificial intelligence as described in this invention. Figure 2 The flowchart for multi-round cross-validation; Figure 3 A flowchart for the construction and initialization of a dynamic semantic network; Figure 4 A graph showing the correlation between conflict resolution rate and number of revisions; Figure 5 This is a statistical distribution chart showing the types of conflicts and their handling. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1This invention provides a method for generating county-level fire safety evaluation standards based on artificial intelligence. The method includes: receiving a raw county-level fire data set containing fire-related records from various dimensions; performing multiple rounds of cross-validation within the raw county-level fire data set to select a stable set of elements constituting the core standard framework from the fire-related records from various dimensions; inputting the stable set of elements into a dynamic semantic network, which simulates the dependency and transmission relationships between fire elements through nodes and associated edges; injecting real-time acquired county-level environmental state parameters into the dynamic semantic network to trigger state transitions of nodes within the network; and parsing and reconstructing the text-based fire safety evaluation standard clauses based on the final steady-state distribution of the node state transitions.

[0020] In one embodiment of the present invention, see [reference] Figure 2 A subset of data is randomly selected from the original county-level fire protection data set as a verification benchmark. The remaining fire-related records are then used to perform a consistency comparison against this benchmark, marking data records that maintain a high degree of consistency across multiple comparisons. The verification benchmark is changed, and the consistency comparison and marking process is repeated until all data subsets have been traversed. The highly consistent data records obtained from all marking processes are then summarized, and the set of these highly consistent data records is defined as the stable element set.

[0021] In practice, the original county-level fire protection dataset contains various fire-related records across different dimensions, including building fire protection facility records, fire history records, fire inspection records, and safety assessment reports. These records are stored in a database in structured or semi-structured form. A subset of data is randomly selected from the original county-level fire protection dataset as the validation benchmark. The selection of the validation benchmark is achieved using a uniform random sampling algorithm to ensure that each data record has an equal probability of being selected. The remaining fire-related records are used to perform a consistency comparison against the validation benchmark. The consistency comparison process involves calculating the similarity between each data record in the validation benchmark and the corresponding dimension records in the remaining records. Data records that maintain a high degree of consistency across multiple comparisons are marked. High consistency is defined as a similarity score exceeding a preset consistency threshold.

[0022] In some embodiments, the similarity score is calculated using the following formula: in: Indicating the first in the verification benchmark The first record and the remaining records The similarity score of the records. This indicates the total number of dimensions in fire-related records. It is the first The weighting coefficients for each dimension are pre-set based on the dimension's importance in historical fire inspections. It is the first The matching degree between two records in each dimension is calculated using a standardized distance function, with a value ranging from 0 to 1. During the consistency comparison process, for each record in the verification benchmark, all corresponding records in the remaining records are traversed, the similarity score is calculated, and the number of times the consistency threshold is exceeded is counted. When the number of times exceeds the preset requirement for repeated comparison rounds, the record is marked as a high consistency data record.

[0023] The validation benchmark is changed, and the consistency comparison and labeling process is repeated. Changing the validation benchmark is achieved by randomly selecting a different subset of data from the original county-level fire protection dataset. This process is repeated until all subsets of data in the original county-level fire protection dataset have been traversed. The subsets are divided according to a fixed size or proportion to ensure that the selected validation benchmarks do not overlap and cover all data each time. Optionally, all highly consistent data records obtained from the labeling process are aggregated. The aggregation operation is completed through database queries and aggregation functions, collecting all record identifiers that have been marked as highly consistent in at least one labeling. The set of these highly consistent data records is defined as the stable feature set, which is stored in an independent data table containing record content, original dimensional information, and consistency labeling history. It can be understood that the multi-round cross-validation process is scheduled by an automated script, which controls the cyclical execution of random sampling, consistency comparison, labeling, and aggregation. The entire implementation relies on a computing server. The server loads the original county-level fire protection dataset and executes the algorithm, ultimately outputting the stable feature set as input for subsequent steps.

[0024] In one embodiment of the present invention, see [reference] Figure 3 This approach maps each independent element in the stable element set to an independent node in the network. It analyzes the patterns of simultaneous or successive occurrences of different fire-fighting elements in historical data, quantifying frequent co-occurrence or successive relationships as weights of the associated edges. Based on these quantified edge weights, the independent nodes are connected to form an initial structure of a dynamic semantic network with weighted connections.

[0025] In practice, the stable element set is derived from the output of multiple rounds of cross-validation. Each data record in the stable element set represents a validated and highly consistent fire protection element. Each independent element in the stable element set is mapped to an independent node in the network. This mapping process is achieved by assigning a unique node identifier to each entry in the stable element set, and the node identifier is bound to the unique code of the fire protection element. The analysis focuses on patterns of simultaneous or sequential occurrence of different fire protection elements in historical data, which includes several years of county-level fire case reports, fire inspection checklists, and emergency response plan documents. The analysis is performed on any combination of two fire protection elements in the stable element set.

[0026] In some embodiments, the weight quantization of associated edges is accomplished through a multi-dimensional metric function. This function comprehensively evaluates the frequency of co-occurrence, temporal density, and semantic strength of the association. Frequently occurring co-occurrence or successive relationships are quantified into the weights of associated edges. The quantization process derives specific values ​​based on statistical calculations, and the weight values ​​are stored in an adjacency matrix or edge list data structure. It can be understood that for directed relationships, weight quantization distinguishes the direction of the relationship. For example, a successive occurrence pattern from "aging electrical wiring" to "short circuit fire" would generate a directed edge and its weight.

[0027] Optionally, the weights of associated edges are calculated using the following formula: in: This represents the weight of the associated edge from node A to node B. This represents the conditional probability of fire element B occurring after fire element A has occurred in historical data. This represents the reciprocal of the average time interval between the occurrence of fire element A and the occurrence of fire element B in historical data, used to measure the strength of the correlation in the time series. This represents the semantic contextual similarity between fire safety element A and fire safety element B in their textual descriptions, calculated using a natural language processing (NLP) model. The NLP model first loads a pre-trained semantic understanding module to parse the textual descriptions of fire safety elements A and B, extracting key semantic features and converting them into numerical representations. Then, the model calculates the association strength between these numerical representations through its internal hierarchical structure, weighting the association strength based on the co-occurrence patterns of semantic features and contextual dependencies. Finally, the model outputs a normalized similarity score, fixed between 0 and 1, to reflect the semantic closeness between fire safety elements A and B. Yes , , The function performs weighted normalization on the three parameters and outputs a final weight value between 0 and 1.

[0028] Based on the quantified edge weights, the individual nodes are connected. This connection operation is performed using an application programming interface (API) of a graph database or graph computing library, and the nodes and edges with weight attributes are persistently stored. This forms an initial structure of a dynamic semantic network with weighted connections. The topology of this initial structure can be visualized, where nodes represent fire safety elements and edge thickness represents weight. In some embodiments, after the initial structure of the dynamic semantic network is constructed, a connectivity check is performed to remove edges with weights below a very low threshold and their associated isolated nodes, ensuring the compactness and effectiveness of the network structure.

[0029] In one embodiment of the present invention, the county-level environmental state parameters include real-time meteorological monitoring data, major event registration information within the region, and real-time on-duty status data of fire-fighting forces. The meteorological monitoring data, major event registration information, and real-time on-duty status data of fire-fighting forces are converted into numerical vectors. These converted numerical vectors are used as external stimulus signals and superimposed onto the corresponding environmental perception nodes in the dynamic semantic network. Starting from the environmental perception node that receives the external stimulus signal, the stimulus signal is proportionally transmitted to the next layer node directly connected to it according to the weight of the associated edges. The next layer node integrates all received incoming signals and, combined with its own historical state, calculates the node's new state value. The node with the new state value is used as a new signal source to continue propagating the state change to its neighboring nodes; this process is repeated cyclically. When the state value changes of all nodes in the entire dynamic semantic network are below a preset threshold, the network is determined to have entered the final steady-state distribution.

[0030] In implementation, a dynamic semantic network has been pre-built and loaded into memory. Some nodes in the network are pre-defined as environmental perception nodes, such as "wind force level perception node," "major event crowd density perception node," and "fire station attendance rate perception node." County-level environmental status parameters are obtained in real time through data interfaces. Meteorological monitoring data comes from the meteorological department's application programming interface and includes fields such as wind speed, wind direction, temperature, and humidity. Major event registration information within the region comes from the government's event management platform and includes event type, estimated number of participants, and venue type. Real-time fire service on-duty status data comes from the fire department's duty management system and includes fields such as the number of personnel on duty at each fire station and vehicle and equipment status. Meteorological monitoring data, major event registration information, and real-time fire service on-duty status data are converted into numerical vectors. The conversion process uses predefined encoding rules; for example, a wind force level of "5" is mapped to the value 5, an event type of "large gathering" is mapped to a category code, and a fire station status of "full attendance" is mapped to the value 1.

[0031] In some embodiments, the dimension of the numerical vector matches the input dimension of the corresponding environment-aware node in the dynamic semantic network, and each element in the vector represents a specific environmental state index. The transformed numerical vector is used as an external stimulus signal and superimposed on the corresponding environment-aware node in the dynamic semantic network. The superposition operation is manifested as adding or modifying the current state value of the target node. Triggering the state transition of nodes within the network includes an iterative propagation process, which is simulated in discrete time steps. Starting from the environment-aware node that receives the external stimulus signal, the stimulus signal is proportionally transmitted to the next layer node directly connected to it according to the weight of the associated edge. The signal transmission amount is equal to the product of the change in the state value of the source node and the weight of the associated edge.

[0032] The calculation of the new state value of a node follows the formula: in: Represents a node At time step The state value, Represents a node In the next time step The state value. Represents all pointer nodes The set of source nodes, Indicates from node Pointing to node The weight of the associated edges, Represents a node At time step The change in the state value compared to the previous time step. Function It is a state update function that synthesizes nodes Its own historical state value The node is calculated by weighting and summing all incoming signals. The new state value.

[0033] The next layer node synthesizes all received incoming signals and, combined with its own historical state, calculates a new state value. The node with the new state value is used as a new signal source to propagate the state change to its neighboring nodes. This process is repeated cyclically, with each cycle corresponding to one time step. When the state value changes of all nodes in the entire dynamic semantic network are below a preset threshold, the network is considered to have entered a final steady-state distribution. The preset threshold is a near-zero positive decimal used to determine the convergence of the state update. Optionally, the iterative propagation process has a maximum iteration limit to prevent convergence failure and infinite loops in extreme cases. In some embodiments, the state update function... The specific form can be an accumulation function with decay to ensure the stability of the network state.

[0034] In one embodiment of the present invention, under the final steady-state distribution, the state values ​​of all nodes in the dynamic semantic network are sorted, and a set of active nodes whose state values ​​exceed a critical level is selected. From the set of active nodes, the original fire protection element corresponding to each active node in the stable element set is traced. A threshold for the weight of associated edges is set, and all associated edges connecting two active nodes with weights exceeding the threshold are selected from the dynamic semantic network to form a set of strongly associated edges. The two active nodes connected by each edge in the set of strongly associated edges are analyzed. Based on the direction of the directed edges, the original fire protection element corresponding to the node as the starting point is marked as a predecessor element or a superior element, and the original fire protection element corresponding to the node as the ending point is marked as a successor element or a subordinate element. The set of strongly associated edges is traversed to construct a logical relationship graph with active nodes as vertices and directed strongly associated edges as connections, and one or more nodes with the highest in-degree or centrality in the logical relationship graph are identified as core nodes. With the core nodes as the root nodes, the logical relationship graph is traversed using a breadth-first or depth-first strategy based on the direction of the associated edges to generate a tree-like or hierarchical logical structure model. Based on the hierarchical position and connection path of each node in the logical structure model, the logical order and subordinate relationships among the original fire protection elements corresponding to all active nodes are finally determined. According to the determined logical order and subordinate relationships, the original fire protection elements are organized into a structured natural language description, forming the final fire safety evaluation standard clauses.

[0035] In practical implementation, the dynamic semantic network has reached a final steady-state distribution through iterative propagation. Each node in the network has a stable state value, which is stored in a floating-point array. Under the final steady-state distribution, the state values ​​of all nodes in the dynamic semantic network are sorted. The sorting operation uses the quicksort algorithm to arrange the node identifiers and their corresponding state values ​​in descending order. A set of active nodes whose state values ​​exceed a critical level is selected. The critical level is a predefined score threshold used to distinguish between ordinary nodes and active nodes sensitive to the current environmental state. From the set of active nodes, the original fire protection element corresponding to each active node in the stable element set is traced. The tracing process is completed by querying the mapping relationship table between node identifiers and fire protection element codes, ensuring that each active node can be mapped back to a specific fire inspection item or fire protection facility attribute. In some embodiments, refer to Table 1, which shows the set of active nodes and their corresponding state values.

[0036] Table 1: Active Node Status Values ​​Table A threshold for the weight of associated edges is set, based on the average edge strength from several historically generated high-confidence standard cases. All associated edges connecting two active nodes with weights exceeding the threshold are selected from the dynamic semantic network, forming a set of strongly associated edges. This set is stored in memory as a list of edges. The two active nodes connected by each edge in the set are analyzed. Based on the direction of the directed edge, the original fire protection element corresponding to the starting node is marked as a predecessor or superior element, and the original fire protection element corresponding to the ending node is marked as a successor or subordinate element.

[0037] Traverse the set of strongly related edges to construct a logical relationship graph with active nodes as vertices and directed strongly related edges as connections. The graph is constructed using a graph data structure, where vertices are active node identifiers and edges are selected strongly related edges. Identify one or more nodes with the highest in-degree or centrality as core nodes. In-degree is calculated by counting the number of directed edges pointing to a given node, and centrality is calculated using the betweenness centrality algorithm. Using the core nodes as the root node, traverse the logical relationship graph using a breadth-first or depth-first strategy based on the directions of the related edges, generating a tree-like or hierarchical logical structure model. This model records the parent-child or hierarchical relationships between nodes. The logical structure model is constructed by starting with the core node as the root node and traversing the logical relationship graph using either a breadth-first or depth-first strategy based on the direction of the associated edges. During the traversal, the system visits adjacent nodes layer by layer starting from the root node. The breadth-first strategy expands horizontally according to the hierarchical order, while the depth-first strategy delves vertically along the path until it reaches the leaf nodes and then backtracks. Each node is assigned a specific hierarchical value during traversal. The root node's level is initialized to zero, and the level of a child node is calculated by adding the parent node's level to the standard level span. The standard level span is usually set to one to represent a direct subordinate relationship. After traversal, a tree-like or hierarchical logical structure model is generated. This model uses nodes to represent fire protection elements and edges to represent the logical dependencies between elements. The model records the parent-child or hierarchical relationships between nodes, where the parent node corresponds to the superior element and the child node corresponds to the subordinate element. Based on the hierarchical position and connection path of each node in the logical structure model, the system ultimately determines the logical order and subordinate relationships between the original fire protection elements corresponding to all active nodes, providing a framework for organizing the original fire protection elements into a structured natural language description.

[0038] It is understandable that in the logical structure model, each node needs to be assigned a specific hierarchical value. The formula for calculating the hierarchical value is as follows: in: This represents the hierarchy value of the current child node. This represents the hierarchy value of its parent node. This represents the standard hierarchical span of the directed edge from the parent node to the current child node in the logical relationship graph. It is typically set to 1, indicating a direct descendant. The root node's hierarchical value is initialized to 0.

[0039] Based on the hierarchical position and connection path of each node in the logical structure model, the logical order and subordinate relationships among the original fire protection elements corresponding to all active nodes are finally determined. The logical order is reflected in the sequence of clauses, and the subordinate relationship is reflected in the hierarchical numbering structure of the clauses. According to the determined logical order and subordinate relationships, the original fire protection elements are organized into structured natural language descriptions. The organization process follows the text framework of "chapter-section-article-clause," and templates are used to fill in the specific descriptions of the fire protection elements in their corresponding positions. This forms the final fire safety evaluation standard clauses, which are output in plain text or structured document format. Optionally, when generating the natural language descriptions, the prefixes of normative terms such as "should," "appropriate," and "may" are automatically added based on the data type of the original fire protection elements.

[0040] See Figure 4 In the analysis of fire safety clause conflict detection results during the logical conflict detection phase, a two-dimensional indicator was used to present the conflict resolution effectiveness and revision cost of different fire safety system modules. Specifically, the figure uses green bars to represent the conflict resolution rate of each module and orange broken lines (including node markers) to represent the number of revisions for the corresponding module. The two axes correspond to the quantitative scales of "conflict resolution rate" and "number of revisions," respectively. Looking at the data distribution: the conflict resolution rate for fire hydrant systems reached 0.95, with a corresponding number of revisions of 0.35; the conflict resolution rate for evacuation routes decreased to 0.88, but the number of revisions increased to 1.0; the conflict resolution rate for automatic fire extinguishing systems rebounded to 1.0, while the number of revisions decreased to 0.0; the conflict resolution rate for alarm systems was 0.92, with a number of revisions of 0.7; and the conflict resolution rate for emergency lighting was 0.98, with a number of revisions of 0.35. This shows an inverse correlation between conflict resolution rate and the number of revisions: when the conflict resolution rate of a module is high, its number of revisions is usually low (e.g., automatic fire extinguishing systems); while when the conflict resolution rate decreases, the number of revisions often increases (e.g., evacuation routes). This correlation can serve as a reference for subsequent optimization of fire protection clauses, thereby improving the efficiency and stability of the overall testing process by reducing the probability of conflicts in modules with high revision frequency.

[0041] In one embodiment of the present invention, the set of active nodes on which the standard clauses are generated, and the weights of the associated edges between these active nodes are extracted. The average state value of the set of active nodes and the average strength of the associated edge weights are calculated. The average state value and average strength are input into a pre-trained evaluation function, and the output of the evaluation function is the confidence score of the currently generated fire safety evaluation standard clause. When the confidence score is lower than the acceptance threshold, the set of active nodes and their associated edge weights that caused the low confidence score are recorded. The associated edge weights are then matched with high-confidence cases in a standard clause case library to find discrepancies. Based on the found discrepancies, the weight parameters of the corresponding associated edges in the dynamic semantic network are adjusted in reverse to complete an adaptive optimization of the network parameters. From the generated textual fire safety evaluation standard clauses, all key phrases containing mandatory requirements are extracted. In a predefined fire code conflict rule library, other code entries that are mutually exclusive with the key phrases are retrieved. If a mutually exclusive entry is found, a logical conflict warning is marked in the generation record of the fire safety evaluation standard clause.

[0042] In practice, after the generation process of fire safety evaluation standard clauses is completed, a confidence assessment of the generated fire safety evaluation standard clauses is initiated. The set of active nodes upon which the standard clauses were based, and the weights of the associated edges between these active nodes, are extracted. This extraction is achieved by reading the metadata log of the standard clause generation process. The average state value of the active node set is calculated by summing the steady-state values ​​of all nodes in the set and dividing by the total number of nodes. The average strength of the associated edge weights is calculated by summing the weight values ​​of all associated edges within the active node set and dividing by the total number of edges.

[0043] The average state value and average intensity are input into a pre-trained evaluation function, which exists in the form of a machine learning model, such as a support vector regression model or a deep neural network model. The output of the evaluation function is the confidence score of the currently generated fire safety evaluation standard clause. The confidence score is a floating-point number between 0 and 1, representing the reliability of the generated clause. In some embodiments, the confidence score... The following evaluation function is calculated: in: This represents the final confidence score. This represents the average state value of the set of active nodes. The function represents the average strength of the associated edge weights. The evaluation function model represents the pre-trained model, which takes the input two-dimensional features Mapped to a single confidence score .

[0044] When the confidence score falls below the acceptance threshold, the method initiates a feedback optimization process. The acceptance threshold is a pre-set value, such as 0.7. The set of active nodes that caused the low confidence score and their associated edge weights are recorded and saved to a dedicated optimization history database. The associated edge weights are then matched against high-confidence cases in a standard clause case library. This library stores historically generated, manually verified, high-confidence standard clauses and their corresponding complete network generation parameters. Differences are identified, including comparing the current generated case with the high-confidence case in terms of active node composition and corresponding associated edge weight values.

[0045] Based on the identified discrepancies, the weight parameters of the corresponding related edges in the dynamic semantic network are adjusted in reverse. The adjustment follows a gradient direction, fine-tuning towards patterns closer to high-confidence cases. This completes one adaptive optimization of the network parameters. The optimized dynamic semantic network parameters will be saved and used in subsequent standard generation tasks. Optionally, the weight parameter adjustment follows these rules: if the weight of a related edge in the current case is significantly lower than the weight of the corresponding edge in a high-confidence case, the weight of the current edge is appropriately increased; conversely, it is appropriately decreased.

[0046] After generating the fire safety evaluation standard clauses, a logical conflict detection is performed. Logical conflict detection is a separate post-processing step independent of the confidence assessment. From the generated textual fire safety evaluation standard clauses, all key phrases containing mandatory requirements are extracted. The extraction process is based on keyword matching and syntactic analysis, and the keyword database includes "strictly prohibited," "must," "should," and "must not." In a predefined fire code conflict rule base, other code entries that are mutually exclusive with the key phrases are retrieved. The fire code conflict rule base stores known, paired, mutually exclusive fire regulations in a graph format. If mutually exclusive entries are found, a logical conflict warning is marked in the fire safety evaluation standard clause generation record. The logical conflict warning includes the specific clause content of both conflicting parties and a description of the conflict type. In some embodiments, a conflict report is generated after the logical conflict detection is completed. This report is submitted to reviewers along with the final standard clause document.

[0047] See Figure 5The logical conflict detection process tracks and statistically analyzes different types of conflicts. The chart, presented as a combination of bar and line graphs, shows the detection and handling of four types of conflicts: prohibited conflicts, mandatory conflicts, permission-related conflicts, and time-sensitive conflicts. Purple bars represent the number of detected conflicts, cyan bars represent the number of resolved conflicts, and the orange line represents the trend of the cumulative number of conflicts. Specifically, mandatory conflicts had the highest number of detections (12) and a relatively high number of resolved conflicts; while time-sensitive conflicts had the lowest number of detections (3) and a correspondingly lower number of resolved conflicts. The cumulative number of conflicts gradually decreases as the conflict type progresses, reflecting the effectiveness of the conflict handling process in resolving various types of conflicts. This visualization directly reflects the scale and processing efficiency of different conflict types, providing data support for the subsequent optimization of the conflict detection and resolution mechanism.

[0048] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0049] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for generating county-level fire safety evaluation standards based on artificial intelligence, characterized in that, Includes the following steps: Receive a raw county-level fire protection data set, which contains fire-related records of various dimensions; Multiple rounds of cross-validation were performed within the original county-level fire protection data set to select a stable set of elements that constitute the core standard framework from fire protection-related records from various dimensions. The set of stable elements is input into a dynamic semantic network, which simulates the dependency and transmission relationships between fire protection elements through nodes and associated edges; Real-time acquired county-level environmental status parameters are injected into the dynamic semantic network to trigger state transitions of nodes within the network; Based on the final steady-state distribution of node state transitions, the fire safety evaluation standard clauses in text form are parsed and reconstructed.

2. The method for generating county-level fire safety evaluation standards based on artificial intelligence as described in claim 1, characterized in that, The process of performing multiple rounds of cross-validation within the original county-level fire protection dataset includes the following steps: A subset of data was randomly selected from the original county-level fire protection data set as the verification benchmark; The remaining fire-related records were used to perform a consistency comparison of the verification benchmark, and data records that maintained a high degree of consistency in multiple comparisons were marked. Change the verification benchmark and repeat the consistency comparison and marking process until all data subsets have been traversed; The set of highly consistent data records obtained from all the marking processes is defined as the stable element set.

3. The method for generating county-level fire safety evaluation standards based on artificial intelligence as described in claim 1, characterized in that, The construction and initialization of the dynamic semantic network are achieved through the following methods: Map each independent element in the stable element set to an independent node in the network; Analyze the patterns of simultaneous or successive occurrence of different fire protection elements in historical data, and quantify the frequent co-occurrence or successive relationships as the weights of the associated edges; Based on the quantized associated edge weights, each independent node is connected to form an initial structure of a dynamic semantic network with weighted connections.

4. The method for generating county-level fire safety evaluation standards based on artificial intelligence as described in claim 1, characterized in that, The county-level environmental status parameters include real-time meteorological monitoring data, registration information of major events in the region, and real-time on-duty status data of fire-fighting forces; The injected real-time acquired county-level environmental status parameters refer to converting meteorological monitoring data, major event registration information, and real-time on-duty status data of fire fighting forces into numerical vectors respectively. The transformed numerical vector is used as an external stimulus signal and superimposed on the corresponding environment-aware node in the dynamic semantic network.

5. The method for generating county-level fire safety evaluation standards based on artificial intelligence as described in claim 4, characterized in that, The triggering of state transitions for nodes within the network includes an iterative propagation process: Starting from the environmental perception node that receives the external stimulus signal, the stimulus signal is transmitted proportionally to the next layer node directly connected to it according to the weight of the associated edge. The next layer node integrates all the received incoming signals and combines them with its own historical state to calculate the node's new state value; The node with the new state value is used as a new signal source to continue propagating the state change to its neighboring nodes, and this process is repeated cyclically. When the change in the state value of all nodes in the entire dynamic semantic network is lower than a preset threshold, the network is determined to have entered the final steady-state distribution.

6. The method for generating county-level fire safety evaluation standards based on artificial intelligence as described in claim 5, characterized in that, The process of parsing and reconstructing the text-based fire safety evaluation standard clauses based on the final steady-state distribution of node state transitions specifically includes: Under the final steady-state distribution, the state values ​​of all nodes in the dynamic semantic network are sorted, and the set of active nodes whose state values ​​exceed the critical level is selected. From the set of active nodes, trace the original fire protection element corresponding to each active node in the set of stable elements; Based on the weights of the associated edges between active nodes, determine the logical order and subordinate relationships among the original fire protection elements; According to the established logical order and subordinate relationships, the original fire protection elements are organized into a structured natural language description, forming the final fire safety evaluation standard clauses.

7. The method for generating county-level fire safety evaluation standards based on artificial intelligence as described in claim 6, characterized in that, The determination of the logical order and hierarchy among the original fire protection elements based on the association edge weights between active nodes is achieved through the following steps: Set a threshold for the weight of associated edges, and filter out all associated edges that connect two active nodes and whose weights exceed the threshold from the dynamic semantic network to form a set of strongly associated edges; Analyze the two active nodes connected by each edge in the strongly correlated edge set. Based on the direction of the directed edge, mark the original fire protection element corresponding to the node as the starting point as the preceding element or the superior element, and mark the original fire protection element corresponding to the node as the ending point as the following element or the subordinate element. Traverse the set of strongly related edges to construct a logical relationship graph with active nodes as vertices and directed strongly related edges as connections, and identify one or more nodes with the highest in-degree or centrality in the logical relationship graph as core nodes. Using the core node as the root node, and based on the direction of the associated edges, the logical relationship graph is traversed using a breadth-first or depth-first strategy to generate a tree-like or hierarchical logical structure model. Based on the hierarchical position and connection path of each node in the logical structure model, the logical order and subordinate relationship between the original fire protection elements corresponding to all active nodes are finally determined.

8. The method for generating county-level fire safety evaluation standards based on artificial intelligence as described in claim 1, characterized in that, The method also includes a confidence assessment of the generated fire safety evaluation standard clauses: Extract the set of active nodes used to generate the standard terms, as well as the weights of the associated edges between these active nodes; Calculate the average state value of the active node set and the average strength of the associated edge weights; The average state value and average intensity are input into a pre-trained evaluation function, and the output of the evaluation function is the confidence score of the currently generated fire safety evaluation standard clause.

9. The method for generating county-level fire safety evaluation standards based on artificial intelligence as described in claim 8, characterized in that, When the confidence score falls below the acceptance threshold, the method initiates a feedback optimization process: Record the set of active nodes that lead to low confidence scores and their associated edge weights; The associated edge weight information is matched with high-confidence cases in a standard clause case library to identify differences. Based on the identified discrepancies, the weight parameters of the corresponding related edges in the dynamic semantic network are adjusted in reverse to complete an adaptive optimization of the network parameters.

10. The method for generating county-level fire safety evaluation standards based on artificial intelligence as described in claim 1, characterized in that, The method also includes performing a logical conflict detection after generating the fire safety evaluation criteria clauses: Extract all key phrases containing mandatory requirements from the generated textual fire safety evaluation standard clauses; In a predefined fire protection code conflict rule base, retrieve other code entries that may be mutually exclusive with the key phrase; If mutually exclusive entries are found, a logical conflict warning will be marked in the generation record of the fire safety evaluation standard clause.

Images