Public health service resource optimization decision support system for industry and trade enterprises

Abstract

The invention discloses an industry and trade enterprise public health service resource optimization decision support system, and relates to the technical field of public health management, and the system comprises a data perception and access layer, a semantic fusion and modeling layer, a knowledge graph and risk calculation layer, an intelligent early warning and decision layer, and a feedback optimization closed loop. Multi-source heterogeneous data are collected through a protocol adapter, a unified data view is constructed through industry and trade public health field ontology semantic fusion, a space-time knowledge graph is dynamically generated, risks are deduced in combination with an improved SEIR space-time propagation model and a Monte Carlo method, visual early warning and interactive decision making are achieved based on digital twinborn bodies, and the method is applied to the field of industry and trade public health. According to the method, data islands can be broken, accurate advanced early warning of public health risks is realized, resource allocation is optimized, and the method is suitable for industry and trade enterprises with dense personnel such as machine manufacturing and electronic processing.

Description

Technical Field

[0001] This invention relates to the field of public health management technology, specifically to a decision support system for optimizing public health service resources in industrial and commercial enterprises. Background Technology

[0002] As a vital component of the economy, industrial and commercial enterprises face significant challenges in public health management due to their densely populated and highly mobile production environments. Traditional public health management methods rely heavily on manual recording and decentralized systems, making it difficult to optimize resource allocation and accurately predict risks. While the development of the Internet of Things (IoT) and big data technologies has led to some degree of information technology application, the multi-source heterogeneous data within industrial and commercial enterprises (such as production execution systems, environmental sensors, personnel location data, and third-party disease control data) has not been effectively integrated, resulting in severe data silos. This leads to a lack of scientific basis for the allocation of public health service resources and hinders effective prediction and rapid response to public health emergencies.

[0003] Currently, there are some technical solutions related to public health management. For example, the "Decision Support System for Public Health Safety Supervision and Control Based on WebGIS" (Jiangsu University, 2005) proposes combining WebGIS technology with a decision support system for public health safety supervision and control. This achieves a certain degree of integration of spatial information and business data, but it mainly focuses on macro-level resource allocation and information display, failing to address the semantic fusion and real-time risk projection problems of multi-source heterogeneous data in the micro-level scenarios of industrial and commercial enterprises. Another related paper, "A 10-Year Comparative Analysis of the Health Protection Effects of Foreign-Invested Enterprises" (2004), focuses on a retrospective comparative analysis of the health protection effects of enterprises, emphasizing the importance of scientific design and management maintenance for protection effectiveness. However, this method lacks real-time data-driven and dynamic early warning capabilities, failing to meet the proactive prevention and control needs of modern industrial and commercial enterprises. Existing technologies generally suffer from poor data sharing, outdated early warning models, and insufficient integration with actual production.

[0004] To address the aforementioned problems, this invention aims to solve the following issue: how to break down data silos in public health management within industrial and commercial enterprises and achieve accurate and proactive risk warnings. Specifically, it requires designing an intelligent early warning architecture capable of automatically integrating multi-source heterogeneous data, dynamically constructing a micro-spatiotemporal knowledge graph of the enterprise, and simulating and extrapolating epidemic transmission based on this graph. This will enable accurate perception, proactive prediction, and intelligent early warning of public health risks, providing real-time and scientific decision support for the optimal allocation of public health service resources within enterprises.

[0005] Therefore, there is an urgent need for a public health service resource optimization decision support system that is geared towards industrial and commercial enterprises, can deeply integrate data, and can achieve accurate early warning, in order to make up for the shortcomings of existing technologies. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a decision support system for optimizing public health service resources in industrial and commercial enterprises. Through a containerized protocol adapter in the data perception and access layer, it is compatible with collecting heterogeneous data from multiple sources, including production execution systems, environmental sensors, personnel positioning, and third-party disease control. Then, through a semantic fusion and modeling layer based on the ontology of public health in the industrial and commercial sector, it completes data cleaning, semantic alignment, and constructs a unified data view. Relying on a knowledge graph and risk calculation layer, it dynamically constructs a spatiotemporal knowledge graph, integrates an improved SEIR spatiotemporal propagation model and Monte Carlo methods to conduct risk simulation, and combines intelligent early warning with the digital twin visualization early warning and interactive decision-making functions of the decision-making layer. Furthermore, it continuously optimizes the model through feedback optimization loops and reinforcement learning. This system can break down data silos, achieve accurate and proactive early warning of public health risks, and provide scientific decision support for the optimized allocation of public health service resources in industrial and commercial enterprises, overcoming the shortcomings of existing technologies such as poor data sharing, delayed early warning, and weak integration with production.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In one aspect, a decision support system for optimizing public health service resources in industrial and commercial enterprises, comprising: The data sensing and access layer is used to collect multi-source heterogeneous data within industrial and trading enterprises through a protocol adapter. The multi-source heterogeneous data includes production execution system data, environmental sensor data, personnel positioning data, and third-party disease control data. The semantic fusion and modeling layer connects to the data perception and access layer. It includes an adaptive parsing engine based on the ontology of the industrial, trade and public health fields, which is used to clean, transform and semantically align the multi-source heterogeneous data, and assign a globally unique semantic ID to each data point to build a unified data view. The knowledge graph and risk calculation layer, connected to the semantic fusion and modeling layer, is used to construct and dynamically update the spatiotemporal knowledge graph of public health for industrial and commercial enterprises based on the unified data view, and to integrate the spatiotemporal propagation model for risk calculation and simulation. The intelligent early warning and decision-making layer connects the knowledge graph and the risk calculation layer, and is used to receive risk calculation results and issue graded early warning information; The ontology definition in the field of industrial and commercial public health includes core concepts such as personnel, equipment, workshops, processes, raw materials, public health events, risk indicators, and prevention and control resources, as well as core relationships such as location, belonging, operation, contact, subordination, initiation, and consumption.

[0008] Furthermore, the data perception and access layer includes: The protocol adapter framework adopts a containerized deployment approach and comes pre-installed with various industrial protocol parsing plugins. The metadata registration module is used to define metadata for each data point, including aspects, attributes, characteristics, and units. The semantic ID allocation module uses a unified resource naming convention to assign a globally unique identifier to each entity and data point. The protocol adapter framework supports the automatic generation of parsing scripts for unknown and private protocols using machine learning data samples, enabling flexible extension of the protocols. The metadata registration module defines rich contextual information for data points according to the industrial data semantic modeling framework; The identifiers generated by the semantic ID allocation module adopt a hierarchical structure to explicitly represent the organizational structure and physical location of the entity.

[0009] Furthermore, the semantic fusion and modeling layer includes: The ontology construction module uses a top-down classification design method to construct an ontology for the industrial, commercial, and public health domains. The domain ontology includes clearly defined core concepts and core relationships. The data mapping module uses the semantic ID to automatically map the parsed multi-source data to the corresponding concepts and instances of the domain ontology; The semantic association module automatically constructs an association network between people, machines, materials, methods, and environment based on the relationships defined in the domain ontology; The public health events defined by the ontology construction module include cluster outbreaks and public health emergencies; The data mapping module maps the attendance system data to the concepts of personnel and workshop, and the sensor data to the concepts of equipment and risk indicators. The production work order data is mapped to the concepts of process and raw materials. The semantic association module constructs a spatial layout network based on location relationships, a personnel interaction network based on contact relationships, and a human-machine relationship network based on operation relationships.

[0010] Furthermore, the knowledge graph and risk calculation layer includes: The spatiotemporal graph construction module extends spatiotemporal semantics on the basis of the domain ontology, adding time validity and spatial coordinate attributes to entities and relations; The knowledge graph update module adopts a TransH-based incremental construction method for time-series knowledge graphs, continuously generating the optimal set of triples from the real-time data stream and incrementally updating it to the knowledge graph; The knowledge extraction module uses a Bi-LSTM-CRF model that incorporates trigger words to extract named entities and relations from unstructured text data; The spatiotemporal map construction module adds a time validity attribute to personnel entities to indicate their on-duty time and adds a three-dimensional coordinate attribute to spatial entities to indicate their specific location. The graph update module generates an optimal set of triples from the new data using an optimal subset extraction algorithm. The optimal set of triples is calculated using the following formula:

[0011] in, Denotes the set of candidate triples. This represents all triples generated from the new data. This represents the triplet scoring function. This indicates a time-limited penalty term for the triple. Indicates the balance parameters; The knowledge extraction module extracts structured information and semantic relationships using the ReLU activation function.

[0012] Furthermore, the knowledge graph and risk calculation layer also includes: The parameter extraction module extracts dynamic parameters required for risk deduction from the spatiotemporal knowledge graph in real time, including node attribute parameters and edge weight parameters. The propagation model construction module introduces spatiotemporal constraints on the basis of the SEIR model to establish an improved spatiotemporal propagation model; The simulation module uses the Monte Carlo method to simulate the spread of the epidemic. The node attributes extracted by the parameter extraction module include the number of people, health status, immunity status, spatial geometric layout, volume, ventilation rate, and functional area division. The edge weight parameters include the frequency and distance of contact between people, the movement trajectory of people in space, and the dwell time. The spatial constraints introduced by the propagation model construction module are defined by the personnel movement matrix, and the temporal constraints are achieved by dynamically adjusting the contact rate parameter.

[0013] Furthermore, the improved spatiotemporal propagation model is achieved through the following formula:

[0014] in, Indicates the first Number of susceptible individuals in each region Indicates the number of lurking individuals. Indicates the number of infected persons. Indicates the number of recovered patients. This represents a time-related contact rate parameter. Indicates from the region To the area Personnel movement matrix elements, Indicates the first Total population of each region This indicates the conversion rate from the incubation period to the infectious period. This indicates the conversion rate from the infection period to the recovery period. Indicates the rate of immune system decay; The personnel movement matrix The contact rate parameter is calculated directly from the spatiotemporal knowledge graph. Adjustments will be made dynamically based on production shifts and work schedules.

[0015] Furthermore, the intelligent early warning and decision-making layer includes: The digital twin construction module constructs a virtual mapping that runs synchronously with the enterprise based on the spatiotemporal knowledge graph and propagation model. The visualization module provides a three-dimensional view of the company's epidemic heat map, personnel movement lines, and resource distribution on a 3D map. The interactive decision-making module allows users to try different prevention and control plans by dragging and dropping and immediately see the simulation effect comparison; The early warning release module disseminates tiered early warning information through multiple channels. The virtual mapping implemented by the digital twin construction module not only reflects the current state, but can also predict the future state through simulation; The visualization module uses the color depth of a heatmap to represent the risk level and the thickness of streamlines to represent the intensity of personnel movement. The interactive decision-making module supports prevention and control plans including entrance blockade, shift adjustment, and disinfection frequency modification. The information released by the early warning release module includes specific risk areas, risk levels, and recommended prevention and control measures.

[0016] Furthermore, the system also includes: The feedback optimization loop records all manual corrections and subsequent effects, using them as feedback signals for reinforcement learning to continuously optimize the model algorithm. The feedback optimization closed loop achieves policy updates through the following formula:

[0017] in Indicates model parameters, Indicates the learning rate. Represents the policy gradient; The policy gradient is calculated using the following formula:

[0018] in Indicates the decision-making trajectory. Represents the policy function. Indicates time The action, Indicates time state, Indicates time Value function estimation; The value function estimation is calculated based on the difference between the actual effect after manual correction and the system's predicted effect.

[0019] Compared with existing technologies, this decision support system for optimizing public health service resources in industrial and commercial enterprises has the following beneficial effects: I. This invention achieves compatible collection of multi-source heterogeneous data from production execution systems, environmental sensors, personnel positioning, and third-party disease control through a protocol adapter at the data perception and access layer. Combined with semantic fusion and modeling layers, it constructs a unified data view by defining the industrial and commercial public health domain, data mapping, and semantic associations, effectively breaking down data silos in public health management within industrial and commercial enterprises. Furthermore, relying on a dynamically constructed spatiotemporal knowledge graph based on a knowledge graph and risk calculation layer, and integrating an improved SEIR spatiotemporal propagation model and Monte Carlo simulation methods, it accurately extracts dynamic risk parameters such as node attributes and edge weights. This enables the prediction of epidemic spread trends, achieving precise perception and early warning of public health risks, providing real-time decision-making support for the scientific allocation of public health service resources within enterprises. This addresses the shortcomings of existing technologies, such as poor data sharing, outdated early warning models, and weak integration with actual production.

[0020] Second, this invention utilizes a digital twin construction module for intelligent early warning and decision-making to achieve real-time virtual mapping and future state prediction of enterprise public health status. Combined with a visualization module, it presents epidemic heat maps, personnel movement flows, and resource distribution, allowing administrators to intuitively grasp risk dynamics. The interactive decision-making module supports drag-and-drop operation to test different prevention and control plans, enabling real-time comparison of effectiveness indicators such as the number of infections and resource consumption, quickly selecting the optimal solution. Simultaneously, it provides feedback and optimization loop recording of manual corrections and effect differences to reinforce learning and update model parameters, continuously improving the system's risk calculation accuracy and decision rationality. This enhances the system's adaptability to scenarios such as production shifts and spatial layouts in industrial and commercial enterprises, balancing public health prevention and control with production efficiency, and avoiding over- or under-prevention issues.

[0021] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0023] Figure 1 This is a system hierarchy architecture diagram of the present invention; Figure 2 This is a flowchart illustrating the data fusion and risk simulation process of the present invention. Figure 3 This is a flowchart of the process of the present invention. Detailed Implementation

[0024] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0025] Example 1

[0026] like Figure 1 and Figure 2 As shown in the illustration, this embodiment discloses a specific implementation of a decision support system for optimizing public health service resources in industrial and commercial enterprises. This embodiment uses a medium-sized machinery manufacturing enterprise as the application scenario. It constructs a full-process technical architecture encompassing data perception, semantic fusion, knowledge modeling, risk calculation, intelligent decision-making, and feedback optimization. This architecture enables automatic access and semantic alignment of multi-source heterogeneous data; dynamically constructs a spatiotemporal knowledge graph of public health for industrial and commercial enterprises; integrates an improved SEIR spatiotemporal propagation model to conduct risk simulation; provides visualized early warning and interactive decision support based on digital twins; and forms a feedback optimization closed loop through reinforcement learning. This embodiment can effectively improve the accurate perception and early warning capabilities of public health risks in industrial and commercial enterprises, providing a scientific basis for the optimal allocation of public health service resources. It is suitable for industrial and commercial enterprises with dense personnel and complex production processes, such as machinery manufacturing, electronics processing, and food production.

[0027] This embodiment uses a medium-sized machinery manufacturing enterprise as an application scenario. This enterprise includes multiple production workshops, raw material warehouses, office buildings, and an employee canteen, operating on a two-shift system (day shift: 8:00-20:00, night shift: 20:00-8:00 the next day). This presents challenges in public health management due to frequent personnel movement and relatively enclosed production areas. The specific implementation methods for each level of this system are described in detail below.

[0028] Implementation of Data Awareness and Access Layer: The core function of the data perception and access layer is to achieve comprehensive collection and standardized access of multi-source heterogeneous data within the application scenario. This is achieved through the collaborative implementation of a protocol adapter framework, a metadata registration module, and a semantic ID allocation module, as detailed below: Deployment and operation of the protocol adapter framework: The protocol adapter framework is deployed using Docker containers on edge computing nodes within the enterprise intranet (configured to meet industrial-grade data processing requirements, ensuring low latency and high stability). The framework comes pre-installed with various industrial protocol parsing plugins, including Modbus-RTU, OPCUA, MQTT, and HTTP, and can directly interface with existing enterprise production execution systems and environmental sensors (temperature and humidity sensors, PM2.5 sensors, etc.). Sensors), personnel positioning systems, and third-party disease control data interfaces (local disease control center epidemic early warning API).

[0029] For proprietary protocols used by older devices in application scenarios, the protocol adapter framework automatically generates parsing scripts through machine learning data samples: First, it collects sufficient historical data samples of the proprietary protocol (including key information such as device operating status codes and parameter values), and extracts features from the data format using unsupervised learning algorithms (such as K-means clustering) to identify the start identifier, field length, and checksum rules of the data frame; second, it constructs a parsing rule model based on the extracted features to generate a parsing script in Python; finally, it encapsulates the script as a containerized plugin and integrates it into the protocol adapter framework to achieve real-time parsing of the proprietary protocol data, with parsing latency controlled within 100ms to meet real-time requirements.

[0030] Implementation of the metadata registration module: The metadata registration module, following the industrial data semantic modeling framework (referencing GB / T35362-2023 "Industrial Data Semantic Description Specification"), defines metadata for each collected data point, including "aspect, attribute, feature, and unit". For example, the metadata for temperature and humidity sensor data points in the workshop is defined as follows: aspect is "environmental parameters", attribute is "temperature / humidity", feature is "real-time monitoring value / daily average value / maximum value", and unit is "℃ / %RH"; the metadata for personnel positioning data points is defined as follows: aspect is "personnel dynamics", attribute is "location coordinates / movement speed / stay time", feature is "real-time coordinates / historical trajectory / number of times staying in the area", and unit is "m / m·". ¹ / min.

[0031] The metadata registration module provides metadata entry and management functions through a web interface. Enterprise administrators can select a preset metadata template according to the data point type, add personalized information, and complete the registration. The registered metadata is stored in a relational database and associated with the semantic ID of the data point, supporting contextual queries in the subsequent semantic fusion process to ensure data traceability and semantic consistency.

[0032] Implementation of the semantic ID allocation module: The semantic ID allocation module adopts the Uniform Resource Naming (URN) format to assign globally unique identifiers to all entities (personnel, equipment, workshops, processes, etc.) and data points within the application scenario. The identifiers adopt a hierarchical structure, which clearly indicates the organizational structure and physical location to which the entity belongs. The structure format is "urn:industrial:enterprise:[enterprise identifier]:[level 1]:[level 2]:[entity / data point identifier]".

[0033] For example, the semantic ID corresponding to a certain machine tool in the workshop is: "urn:industrial:enterprise:Mach001:Workshop:A:Equipment:M01", The semantic ID of the machine tool temperature sensor data point is: "urn:industrial:enterprise:Mach001:Workshop:A:Equipment:M01:Sensor:Temp01".

[0034] Semantic ID allocation is achieved through automated scripts: the module extracts basic entity information from the enterprise HR system, equipment management system, and building information model (BIM) system, automatically generates semantic IDs according to preset hierarchical rules, and stores them in the semantic registry; when a new entity or data point is added, the module automatically triggers the semantic ID generation process by listening for new events in the system, ensuring that all data points have unique and traceable semantic identifiers, laying the foundation for subsequent semantic fusion.

[0035] Implementation of semantic fusion and modeling layer: The semantic fusion and modeling layer, based on the ontology of the industrial, trade, and public health fields, realizes the cleaning, transformation, and semantic alignment of multi-source heterogeneous data, and constructs a unified data view. The specific implementation method is as follows: Implementation of the ontology building module: The ontology construction module adopts a top-down classification design method to construct an ontology in the field of industrial and commercial public health. This ontology clearly defines eight core concepts: "personnel, equipment, workshop, process, raw materials, public health event, risk indicator, and prevention and control resources," as well as seven core relationships: "located, belong to, operate, contact, subordinate to, cause, and consume."

[0036] The core concepts are defined as follows: "Personnel" includes attributes such as employee number, health status, and department; "Equipment" includes attributes such as equipment number, type, operating status, and location; "Public Health Event" includes attributes such as event type (cluster outbreak, public health emergency), occurrence time, and scope of impact; "Risk Indicators" include attributes such as temperature, humidity, personnel density, and contact frequency; and "Prevention and Control Resources" includes attributes such as the type and inventory of disinfectants, protective clothing, and thermometers.

[0037] The core relationships are defined as follows: "Located" indicates the spatial relationship between personnel / equipment and the workshop; "Operation" indicates the operational relationship between personnel and equipment; "Contact" indicates the interactive relationship between personnel; "Initiate" indicates the causal relationship between risk indicators and public health events; and "Consumption" indicates the relationship between public health events and the use of prevention and control resources.

[0038] The ontology is described using the Web Ontology Language (OWL) and stored in an ontology knowledge base (such as an OWL file generated by Protégé). It is also validated for consistency using an ontology inference engine (such as Pellet) to ensure that there are no logical conflicts between concepts and relations. At the same time, the ontology supports dynamic expansion. When new core concepts (such as "logistics vehicles") or relations (such as "transportation") are added, the administrator can add definitions and update the knowledge base through the ontology editing interface to adapt to changes in enterprise business.

[0039] Implementation of the data mapping module: The data mapping module uses semantic IDs as association keys to automatically map the parsed multi-source data to the corresponding concepts and instances of the domain ontology. The specific mapping logic is as follows: The attendance system data maps employee attendance records (including employee ID, attendance time, and attendance location) to "personnel" instances through the "employee ID" attribute of the semantic ID "personnel" concept, and associates them with "workshop" instances through the "location" relationship (attendance location corresponds to the workshop). Environmental sensor data: The monitoring values ​​of temperature, humidity and PM2.5 sensors are mapped to "risk indicator" instances (such as temperature sensor data corresponding to the "temperature" risk indicator) through the "sensor identifier" attribute of the semantic ID "device" concept, and associated with the "workshop" instance through the "belonging" relationship; Production work order data: Information such as process number, raw material type, and equipment number in the work order is mapped to the corresponding instance through the corresponding attributes of the semantic ID concepts of "process", "raw material" and "equipment". "Personnel" and "equipment" are associated through "operation" relationship, and "process" and "raw material" are associated through "consumption" relationship.

[0040] The data mapping module uses ETL (Extract-Transform-Load) tools to achieve automated mapping. When configuring mapping rules, the administrator selects the data source, target ontology concept, and associated semantic ID field through the interface, and the system automatically generates the mapping script. During the mapping process, the module cleans (e.g., removes outliers in attendance times) and transforms (e.g., converts temperature units from °F to °C) to ensure that the data format is consistent with the ontology attributes. Finally, it is loaded into the storage engine of the unified data view (e.g., PostgreSQL) to form a structured unified data resource.

[0041] Implementation of the semantic association module: The semantic association module automatically constructs an association network between "human, machine, material, method, and environment" based on the core relationships defined in the domain ontology. Specifically, it includes a spatial layout network, a human interaction network, and a human-machine relationship network. The construction logic of each network is as follows: Spatial Layout Network: Based on the "located" relationship, the module extracts the correlation data of "personnel-workshop" and "equipment-workshop" from the unified data view, and forms a hierarchical spatial layout network with "workshop" as the node and "personnel / equipment" as the child node; the network stores the specific location coordinates of personnel / equipment in the workshop, supports spatial query, and provides the foundation for the subsequent propagation model of spatial constraints; Personnel Interaction Network: Based on "contact" relationships, the module extracts close-range interaction records between personnel from personnel location data (e.g., distance ≤ 1m and duration ≥ 5s), and forms a personnel interaction network with "personnel" as nodes, "contact" as edges, and edge weights as contact frequency (number of contacts per unit time). This network can be used to analyze potential paths of epidemic transmission and identify high-risk interaction nodes. Human-Machine Relationship Network: Based on the "operation" relationship, the module extracts records of personnel operating equipment from production work order data and equipment operation logs, and forms a human-machine relationship network with "personnel" and "equipment" as nodes, "operation" as edges, and edge weights as operation duration (number of operation minutes per unit time). This network can be used to analyze the risk of personnel infection caused by equipment contamination and assist in the precise allocation of prevention and control resources.

[0042] The semantic association module stores the association network through a graph database (such as Neo4j), supporting real-time querying and visualization. When multi-source data is updated (such as when a new personnel location record is added), the module automatically triggers the update of the association relationship to ensure that the network is consistent with the actual state of the enterprise and maintain the timeliness of the association data.

[0043] Implementation of Knowledge Graph and Risk Computation Layer: The knowledge graph and risk calculation layer constructs and dynamically updates a spatiotemporal knowledge graph of public health for industrial and commercial enterprises based on a unified data view. It integrates a spatiotemporal propagation model to perform risk calculation and simulation. The specific implementation method is as follows: Implementation of the spatiotemporal graph construction module: The spatiotemporal graph construction module extends spatiotemporal semantics on the basis of the domain ontology, adding "time validity" and "spatial coordinates" attributes to entities and relations, thereby enhancing the spatiotemporal dimension of the knowledge graph. The specific extension methods are as follows: Extended spatiotemporal attributes of entities: For the "Personnel" entity, a "On-Duty Hours" time validity attribute is added. This attribute value is extracted from the attendance system and location data, and is in the format of "Start Time - End Time". It is used to identify the valid activity time of personnel within the enterprise. For spatial entities such as "Workshop" and "Equipment", a "3D Coordinates" attribute is added. This attribute value is collected from the enterprise's BIM system and UWB positioning system, and is in the format of "X,Y,Z". It is used to identify the specific spatial location of the entity. Spatiotemporal attribute extension of relationships: For relationships such as "contact" and "operation", "occurrence time" and "occurrence location" attributes are added. "Occurrence time" is extracted from location data and operation logs, and "occurrence location" is obtained by association from spatial coordinate attributes. These attributes are used to identify the spatiotemporal scenario in which the relationship occurs, providing key parameters for subsequent spatiotemporal propagation models.

[0044] The spatiotemporal graph is stored in a graph database (Neo4j), and the spatiotemporal attributes of entities and relationships are used as attribute fields for graph nodes and edges. The module extracts spatiotemporal data from a unified data view through scheduled tasks (such as every 5 minutes) and updates the spatiotemporal attributes of the graph to ensure that the graph is synchronized with the real-time status of the enterprise and to meet the real-time requirements of risk calculation.

[0045] Implementation of the map update module: The knowledge graph update module adopts a TransH-based incremental construction method for time-series knowledge graphs. It continuously generates the optimal set of triples from the real-time data stream and incrementally updates it to the spatiotemporal knowledge graph, avoiding the resource consumption caused by full updates. The specific implementation steps are as follows: Triple generation: The module extracts entities and relationships from real-time data streams (such as location data, sensor data, and disease control data) to generate a set of candidate triples. The triple format is "(head entity) ,relation Tail entity For example, “(Employee 1, in contact with, Employee 2)” or “(Workshop A, located at, temperature sensor 1)”; Optimal Triple Selection: The module calculates and selects the optimal set of triples using the following formula. To ensure the effectiveness and timeliness of newly added triples:

[0046] in, The triplet scoring function is calculated using the TransH model. , , Semantic similarity in low-dimensional vector space (the higher the similarity, the higher the score). This is a time-sensitive penalty for triples. If the time elapsed between the occurrence of the event corresponding to the triple and the current time exceeds a preset threshold (e.g., 24 hours), the penalty will be increased, reducing the probability of the triple being selected. This is a balancing parameter, ranging from 0.1 to 0.5, used to balance the weights of score and timeliness (in this embodiment). (Set to 0.3); Incremental update: The module will filter out... The triples in the graph are incrementally added to the spatiotemporal knowledge graph. If the entity or relation in the triple already exists in the graph, its attributes are updated (such as updating the occurrence time of the "contact" relation); if it does not exist, a new node or edge is added to ensure the integrity and timeliness of the graph.

[0047] The update frequency of the graph update module can be dynamically adjusted according to the amount of data. During peak data periods (such as the day shift), the update interval is set to 1 minute, and during off-peak periods (such as the night shift), it is set to 5 minutes. During the update process, distributed locks are used to avoid data conflicts and ensure data consistency.

[0048] Implementation of the knowledge extraction module: The knowledge extraction module employs a Bi-LSTM-CRF model that incorporates trigger words to extract named entities (such as "cluster outbreak" and "disinfectant") and relationships (such as "trigger" and "consume") from unstructured text data (such as epidemic reports from local disease control centers, enterprise equipment maintenance records, and employee health declaration texts). The specific implementation steps are as follows: Text preprocessing: The module cleans unstructured text (removing special characters and redundant spaces), segments it (using the jieba segmentation tool), tags it with parts of speech, and constructs word vectors (using the Word2Vec model, trained on a corpus of industrial, trade and public health fields) to provide structured data for model input; Trigger word recognition: The module has a pre-set dictionary of trigger words in the industrial and commercial public health fields, which includes trigger words such as "epidemic", "infection", "disinfection" and "abnormal body temperature". By matching trigger words, it can locate sentence fragments related to public health in the text, narrow the extraction range and improve extraction efficiency. Bi-LSTM-CRF Model Training and Inference: The model's input consists of preprocessed word vectors and trigger word identifiers (trigger word positions are marked as 1, other positions as 0). The Bi-LSTM layer extracts contextual features of the text through a bidirectional recurrent neural network, and the CRF layer optimizes the labeling accuracy of entity boundaries through a conditional random field model. The model uses the ReLU activation function to perform non-linear transformation on the extracted features, enhancing the ability to express structured information and semantic relationships. During inference, the model outputs entity labels (e.g., "employees with abnormal body temperature" is labeled as "personnel-abnormal") and relationship labels (e.g., "occurrence" is labeled as "trigger"). Extraction Result Verification: The module matches and verifies the extracted entities and relations against the core concepts and relations in the domain ontology. If a conflict exists (e.g., the extracted "device type" is not within the scope of the ontology definition), it is marked as a suspicious result and manually reviewed by the administrator. Results that pass the review are converted into triples and added to the candidate triples set. They participate in map updates to ensure the accuracy of extracted data.

[0049] The model training corpus for the knowledge extraction module comes from publicly available public health literature and corporate historical text data. The training set is constructed through manual annotation (the accuracy of annotating entities and relationships is ≥95%). The model is incrementally trained every month using newly added text data to continuously improve the extraction accuracy and adapt to changes in text data.

[0050] Implementation of the parameter extraction module: The parameter extraction module extracts dynamic parameters required for risk inference in real time from the spatiotemporal knowledge graph, including node attribute parameters and edge weight parameters, to provide input for the subsequent propagation model. The specific extraction method is as follows: Node attribute parameter extraction: The module extracts parameters from the attributes of entities such as "Personnel," "Workshop," and "Prevention and Control Resources," including the number of personnel (obtained from the total count of the "Personnel" entity) and health status (obtained from the "Health Status" attribute of the "Personnel" entity, categorized as susceptible). , infiltration ,Infect Rehabilitation The parameters are categorized into four types: immunization status (obtained from the "vaccination record" attribute of the "personnel" entity), spatial geometry (calculated from the "3D coordinates" attribute of the "workshop" entity, such as area and volume), ventilation rate (obtained from the "ventilation equipment operation status" attribute of the "workshop" entity), and functional area division (obtained from the "area type" attribute of the "workshop" entity, such as production area and rest area). The parameter extraction frequency is consistent with the map update frequency to ensure the real-time nature of the parameters. Edge weight parameter extraction: The module extracts parameters from the attributes of relationships such as "contact", "location", and "operation", including the contact frequency between people (calculated from the "number of occurrences" attribute of the "contact" relationship, the number of contacts per unit time), the contact distance between people (calculated from the "location" attribute of the "contact" relationship, the Euclidean distance between the spatial coordinates of the two people), the movement trajectory of people in space (obtained from the "location" sequence of the "location" relationship, forming a coordinate change path), and the dwell time of people in space (calculated from the "time of occurrence" of the "location" relationship, the total effective time in a certain spatial entity). The edge weight parameters are calculated using a sliding window (such as a 1-hour window) to avoid the impact of instantaneous data fluctuations and ensure parameter stability.

[0051] The parameter extraction module stores the extracted parameters in a time series database (such as InfluxDB), supports querying historical parameters by time dimension, provides data support for parameter tuning of the propagation model, and facilitates tracing the parameter change process.

[0052] Implementation of the propagation model building module: The propagation model construction module introduces spatiotemporal constraints based on the classic SEIR model to establish an improved spatiotemporal propagation model, thereby enhancing the accuracy of epidemic propagation simulation. The specific construction method is as follows: Model assumptions: It is assumed that employees within the company are divided into susceptible individuals. Infiltrators infected persons Recovered patients Four categories: The movement of people between different areas can lead to the cross-regional spread of the epidemic; the contact rate changes dynamically with production shifts and work and rest times (day shift staff are densely packed and have a high contact rate; night shift staff are sparsely packed and have a low contact rate). Spatiotemporal constraints are introduced: Spatial constraints are addressed through the "personnel movement matrix". "definition, Indicates from the region To the area The proportion of personnel movement, this matrix is ​​calculated from the "location" relationship data of the spatiotemporal knowledge graph, that is... area To the area Personnel movement frequency / area Total number of personnel; time constraints are achieved through the "contact rate parameter". "accomplish, According to the production shift adjustment, the day shift is from 8:00 to 20:00. Use a value of 0.3-0.5, and 0.1-0.2 for night shifts (20:00-8:00 the next day). The specific value should be calibrated based on historical infection data. Model formulation construction: The improved spatiotemporal propagation model describes the changes in the state of people in each region through the following set of differential equations:

[0053] in, For the first The number of susceptible individuals in each region For the number of lurkers, For the number of infected people, For the number of recovered patients, For the first Total population of each region The conversion rate from the incubation period to the infectious period (values ​​range from 0.1 to 0.2, and 0.15 in this example). The conversion rate from infection to recovery (values ​​range from 0.05 to 0.1, and 0.08 in this example) The immune decay rate is 0.01-0.02, and 0.015 is used in this example.

[0054] The propagation model uses Python's SciPy library to solve differential equations, supporting input of initial parameters for different regions (such as the initial number of infected individuals) and outputting the changes in each region over time. , , , The number of people curves provides quantitative model support for risk projection.

[0055] Implementation of the simulation and deduction module: The simulation module uses the Monte Carlo method to simulate the spread of the epidemic. Through multiple random simulations, it predicts the trend of epidemic spread and provides a basis for risk assessment. The specific implementation steps are as follows: Initial parameter settings: The module obtains the initial parameters for each region from the parameter extraction module. , , , Quantity (e.g., initial) (i.e., there is one infected person in a certain area), and , , , , Set the simulation time step (e.g., 1 hour) and total simulation duration (e.g., 24 hours, 48 ​​hours) to the specified parameter values. Stochastic simulation process: For each time step, the module performs a stochastic simulation according to the following steps: Calculate the propagation probability: Based on the improved spatiotemporal propagation model, calculate the region... Inner susceptible individuals Probability of infection ; State transition: For each susceptible individual, generate a random number between 0 and 1. If the random number... Then the susceptible person becomes a latent individual. For each infiltrator, generate a random number if Then they become infected. For each infected person, generate a random number; if Then they become recovered patients. For each recovered patient, generate a random number if... Then, weakened immunity makes one susceptible to infection. ; Cross-regional transmission: based on personnel movement matrix Random selection The proportion of people from the region Move to area The status of moving personnel is updated synchronously (e.g., after an infected person moves to a new area, they participate in the spread of the virus in that area). Results statistics: The above simulation process was repeated N times (N≥1000, 2000 times in this embodiment), and the statistics of each region were analyzed at the end of each simulation. Quantity, calculation The mean, standard deviation, and probability distribution (such as the range of infected people within the 95% confidence interval) provide a quantitative basis for risk level determination.

[0056] The simulation module outputs the simulation results in the form of charts (such as line graphs and box plots) to the intelligent early warning and decision-making layer, allowing administrators to view the risk of epidemic spread at different time points and intuitively grasp the trend of risk changes.

[0057] Intelligent early warning and its implementation at the decision-making level: The intelligent early warning and decision-making layer receives risk calculation results, issues tiered early warning information, and provides visual displays and interactive decision support. The specific implementation method is as follows: Implementation of digital twin building blocks: The digital twin construction module, based on spatiotemporal knowledge graphs and an improved SEIR propagation model, builds a virtual mapping that operates synchronously with the physical space of the application scenario, enabling real-time mapping and future prediction of the enterprise's public health status. The specific construction method is as follows: Physical entity modeling: The module obtains 3D models of physical entities such as workshops, equipment, and pipelines from the enterprise's BIM system and equipment management system, imports them into a digital twin platform (such as Unity or Unreal Engine), and associates entity attributes (such as equipment operating status and personnel location) in the spatiotemporal knowledge graph through semantic IDs to achieve synchronization of attributes between the virtual model and the physical entity, ensuring the accuracy of virtual mapping; Real-time status mapping: The module receives real-time data streams (such as personnel positioning, sensor data, and risk calculation results) via the WebSocket protocol and updates the status of the digital twin every 10 seconds. For example, the position of the personnel model is synchronized to the latest positioning coordinates, the color of the equipment model changes according to the operating status (green for normal and red for abnormal), and the color of the workshop model changes according to the risk level (refer to the warning level), thus realizing a real-time virtual representation of the physical status. Future State Prediction: The module calls the API of the simulation and deduction module, inputs the current state parameters (such as the current number of infected people and prevention and control measures), and predicts the digital twin state for the next 12 hours, 24 hours, and 48 hours. For example, if it is predicted that the number of infected people in a certain workshop will be 5 after 24 hours, the digital twin will display 5 red personnel models, and mark the infection time and transmission path, providing intuitive support for early warning.

[0058] Digital twins support multi-terminal access, allowing enterprise administrators to view the virtual mapping via web, mobile, and large-screen terminals, enabling them to monitor the enterprise's public health status in real time and improve management efficiency.

[0059] Implementation of the visualization module: The visualization module uses data visualization technology to display the epidemic heat map, personnel movement lines, and resource distribution on the enterprise's 3D map based on a digital twin. The specific display method is as follows: COVID-19 Heat Map: Using areas such as workshops and warehouses as units, the heat map renders the colors of the areas on a 3D map based on the simulated infection risk probability (e.g., 0-20% is green, 21-40% is yellow, 41-60% is orange, and 61-100% is red). The color depth is positively correlated with the risk level. The heat map supports playback along a time dimension (e.g., playing an animation of risk changes over the past 24 hours), helping administrators trace the risk spread path and identify the source of the risk. Personnel movement flow: Historical movement trajectories of personnel are extracted from a spatiotemporal knowledge graph and displayed as dynamic lines on a 3D map. The line color is associated with the health status of the personnel (blue for susceptible individuals, purple for asymptomatic individuals, red for infected individuals, and green for recovered individuals), and the line thickness is positively correlated with the number of people (e.g., a thick line for movement of more than 10 people, and a thin line for movement of less than 10 people). The flow lines support filtering (e.g., filtering for personnel movement in a certain time period or area), which facilitates the analysis of the impact of personnel movement on the spread of the epidemic. Distribution of epidemic prevention and control resources: The inventory and location information of epidemic prevention and control resources (disinfectant, protective clothing, thermometers) are extracted from the spatiotemporal knowledge graph and marked on the 3D map with icons (e.g., blue bottle icon for disinfectant, red instrument icon for thermometers). When the mouse hovers over the icon, the resource details are displayed (e.g., "Thermometer at the workshop entrance, 80% remaining power"). The resource distribution supports filtering by type, helping administrators to quickly locate resource locations and optimize resource allocation efficiency.

[0060] The visualization module is implemented using front-end frameworks such as ECharts and Three.js, and supports interactive operations such as zooming, panning, and rotating. Administrators can view detailed data by clicking on area, personnel, and resource icons, improving the convenience of data acquisition.

[0061] Implementation of the interactive decision-making module: The interactive decision-making module allows administrators to try different prevention and control plans by dragging and dropping. The system calculates and displays the simulation effects of the plans in real time, helping administrators select the optimal resource allocation scheme. The specific implementation method is as follows: Contingency Plan Component Design: The module provides a library of prevention and control contingency plan components on the visual interface, including components such as "Entrance Blockade", "Shift Adjustment", "Disinfection Frequency Modification" and "Resource Allocation". Each component corresponds to a prevention and control measure. For example, the "Entrance Blockade" component is a red blockade icon, and the "Disinfection Frequency Modification" component is a slider (0-4 times / day). The component design is simple, intuitive and easy to operate. Drag-and-drop operation and effect calculation: Administrators can drag components to the target location on the 3D map (e.g., drag the "Entrance Blockade" icon to the east gate of a workshop). The system will call the simulation module in real time to calculate the changes in the spread of the epidemic under the plan. For example, after the east gate of the workshop is blocked, the number of infected people drops from 5 to 2 within 24 hours. The system will immediately display "The number of infected people has decreased by 60%, and resource consumption has increased by 10% (2 new people are added to the blockade)" on the interface, providing a quantitative effect reference for decision-making. Contingency Plan Comparison and Ranking: The module supports loading multiple prevention and control contingency plans simultaneously. The interface displays the effectiveness indicators (number of infections, scope of infection, resource consumption, and implementation difficulty) of each plan in a table format. The plans are ranked according to a comprehensive weight of "least number of infections + lowest resource consumption," and the optimal plan is recommended. For example, Plan 1 has 2 infections and a resource consumption of 1000 yuan; Plan 2 has 3 infections and a resource consumption of 800 yuan. The system calculates a comprehensive score based on the weights and recommends Plan 1 as the optimal plan, helping administrators make quick decisions.

[0062] The interactive decision-making module supports saving and sharing of contingency plans. Administrators can save the best contingency plan as a template and call it directly in similar scenarios in the future, improving decision-making efficiency.

[0063] Implementation of the early warning release module: The early warning release module, based on the simulated risk level (referencing the levels corresponding to the colors on the heatmap: green for no warning, yellow for level four warning, orange for level three warning, and red for level two / one warning), releases tiered early warning information through multiple channels to ensure that relevant personnel receive it in a timely manner. The specific implementation method is as follows: Warning Level Determination: The module has preset risk level determination rules, such as: no warning if the probability of infection risk in a single area is ≤20%, 21-40% is a Level 4 warning (remind to pay attention), 41-60% is a Level 3 warning (enhanced monitoring), 61-80% is a Level 2 warning (restricted movement), and 81-100% is a Level 1 warning (complete lockdown); the determination rules can be adjusted by administrators according to the actual situation of the enterprise to adapt to the management needs of different enterprises; Warning Information Generation: The module automatically generates warning information, which includes the name of the risk area, the warning level, the risk description, and recommended prevention and control measures. For example: "[Level 3 Warning] The probability of infection in a certain workshop is 55%, and a cluster outbreak may occur. Recommended measures: disinfect once every 2 hours, restrict personnel entry and exit, and conduct temperature screening for personnel in the workshop." The information is concise, clear, and easy to implement. Multi-channel dissemination: The module disseminates early warning information through channels such as enterprise APP push notifications, SMS notifications, workshop broadcasts, and office system pop-ups. Specifically, enterprise APP and office system pop-ups push information in real time (≤1 minute), SMS notifications are sent to regional managers and enterprise management, and workshop broadcasts are broadcast once every 30 minutes (once every 10 minutes for Level 2 / Level 1 early warnings). Early warning information is also synchronized to the digital twin and visual interface, with red pop-up reminders to ensure the timeliness and comprehensiveness of information delivery.

[0064] The early warning release module supports the early warning cancellation function. When the risk level in the simulation is reduced to no warning or level four warning, the system automatically releases an early warning cancellation notice to avoid over-prevention and control and balance the prevention and control effect with production efficiency.

[0065] Implementation of feedback optimization closed loop: The feedback optimization closed loop records all manual corrections and their subsequent effects, using this as feedback signals for reinforcement learning to continuously optimize the model algorithm and improve the accuracy of system decisions. The specific implementation method is as follows: Manual correction record: The module sets up a "manual correction" entry in the interactive decision-making module and the early warning release module. If the administrator adjusts the system-recommended plan or early warning level (such as changing the system-recommended "disinfection every 2 hours" to "disinfection every 1 hour"), he / she must record the correction content and the reason for the correction and submit it to the feedback database to ensure that the correction process is traceable. Effectiveness evaluation: The module periodically (e.g., daily) extracts manually corrected actual effect data from the real-time data stream and compares it with the original prediction of the system to calculate the difference in effect. For example, if the original prediction of the number of infections in 24 hours is 5, and the actual number of infections after manual correction is 1, the effect difference is "the number of infections has decreased by 80%". The effectiveness evaluation indicators also include differences in resource consumption, early warning accuracy, etc., to comprehensively measure the effect of correction. Reinforcement learning optimization: The module uses the difference in performance as a reward signal for reinforcement learning, updating the model parameters using the following formula. Optimize the propagation model and contingency plan recommendation algorithm:

[0066] in, For model parameters (such as those of the SEIR model) , , (weight of the recommended contingency plan) The learning rate (values ​​0.01-0.1, initially 0.05, decreasing with the number of iterations) The policy gradient is calculated using the following formula:

[0067] in, For decision trajectory (including state) ,action ,award), This is the strategy function (system-recommended contingency plan / early warning). For a moment Manual correction actions, For a moment The system status, For a moment The value function is estimated (based on the difference between the actual effect and the predicted effect; the greater the difference, the higher the value). The larger (the larger).

[0068] The feedback optimization loop updates the model parameters every 7 days. The updated model is used for subsequent risk calculation and decision recommendation, forming a continuous improvement mechanism of "prediction-decision-feedback-optimization" to ensure that the system performance continues to improve with the application process.

[0069] In summary, this embodiment details the entire implementation process of a decision support system for optimizing public health service resources in industrial and commercial enterprises. It comprehensively presents the technical solutions from multi-source heterogeneous data access and semantic fusion modeling to spatiotemporal knowledge graph construction, risk calculation and deduction, and finally, intelligent early warning decision-making and feedback optimization. This embodiment solves the compatibility problem of multi-source data access through containerized protocol adapters, achieves accurate prediction of public health risks through an improved SEIR spatiotemporal propagation model and Monte Carlo simulation, provides scientific tools for resource optimization and allocation based on digital twins and interactive decision support, and finally ensures continuous improvement in system performance through a feedback loop of reinforcement learning.

[0070] The technical solution in this embodiment is closely integrated with the actual production of industrial and commercial enterprises (such as dynamically adjusting the contact rate according to shifts and building a network of connections based on the production process), avoiding the defects of macro-management and micro-scenario disconnect in the existing technology. It can effectively improve the ability of industrial and commercial enterprises to accurately perceive and provide early warning of public health risks, and provide a scientific basis for the optimal allocation of public health service resources.

[0071] Example 2

[0072] like Figure 3 As shown in Example 1, this example elaborates on the specific steps of a decision support system for optimizing public health service resources in industrial and commercial enterprises during its operation. The specific steps are as follows: 1. Multi-source heterogeneous data acquisition: The protocol adapter collects real-time data from the production execution system, environmental sensor data (temperature, humidity, PM2.5), personnel positioning data (UWB positioning coordinates), and third-party disease control data (epidemic early warning API). Automatically generate parsing scripts for proprietary protocol data to ensure compatibility; 2. Data cleaning and semantic alignment: Perform outlier cleaning on the raw data (e.g., remove invalid attendance records); Define attributes for data points (such as "Environmental Parameters - Temperature / Unit °C") through the metadata registration module. Assign a hierarchical semantic ID to each entity; 3. Dynamic construction of spatiotemporal knowledge graphs: Based on the ontology of industrial and commercial public health, the cleaned data is mapped to core concepts (personnel, equipment, workshop). The semantic association module automatically constructs a "personnel-equipment-workshop" relationship network (such as generating "contact" relationships based on location data). Incrementally update the spatiotemporal knowledge graph, integrating new triples every 5 minutes (such as adding "Employee A contacts Employee B"). 4. Real-time extraction of risk parameters: Extracting dynamic parameters from knowledge graphs: Node attributes: personnel health status, workshop ventilation rate, functional area division; Edge weights: frequency of personnel contact and dwell time on movement trajectories; The contact rate is dynamically adjusted according to the production shift (0.3-0.5 for day shift and 0.1-0.2 for night shift). 5. Simulation and simulation of epidemic transmission: Input parameters such as the current distribution of infected individuals and the personnel movement matrix; The Monte Carlo method was used to conduct 2000 random simulations to output the probability distribution of the number of infected people in each region in the next 24 hours; 6. Visualized early warning and decision support: Render a heat map of the epidemic on a 3D map (red = high-risk area); Generate tiered early warning information (e.g., Level 3 warning: 55% probability of infection, disinfection recommended every 2 hours); Supports dragging and dropping prevention and control plan components (such as blocking workshop entrances) and real-time comparison of plan effectiveness (reduction in the number of infections, resource consumption). 7. Feedback optimization closed loop: Record the administrator's manual modifications to the warning or contingency plan (such as changing the disinfection frequency to once per hour); The difference in the actual number of infections before and after the correction is used as a reinforcement learning signal. The model parameters are updated every 7 days to optimize the accuracy of risk calculation.

[0073] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A decision support system for optimizing public health service resources in industrial and commercial enterprises, characterized in that, include: The data sensing and access layer is used to collect multi-source heterogeneous data within industrial and trading enterprises through a protocol adapter. The multi-source heterogeneous data includes production execution system data, environmental sensor data, personnel positioning data, and third-party disease control data. The semantic fusion and modeling layer connects to the data perception and access layer. It includes an adaptive parsing engine based on the ontology of the industrial, trade and public health fields, which is used to clean, transform and semantically align the multi-source heterogeneous data, and assign a globally unique semantic ID to each data point to build a unified data view. The knowledge graph and risk calculation layer, connected to the semantic fusion and modeling layer, is used to construct and dynamically update the spatiotemporal knowledge graph of public health for industrial and commercial enterprises based on the unified data view, and to integrate the spatiotemporal propagation model for risk calculation and simulation. The intelligent early warning and decision-making layer connects the knowledge graph and the risk calculation layer, and is used to receive risk calculation results and issue graded early warning information; The ontology definition in the field of industrial and commercial public health includes core concepts such as personnel, equipment, workshops, processes, raw materials, public health events, risk indicators, and prevention and control resources, as well as core relationships such as location, belonging, operation, contact, subordination, initiation, and consumption.

2. The decision support system for optimizing public health service resources in industrial and commercial enterprises according to claim 1, characterized in that, The data perception and access layer includes: The protocol adapter framework adopts a containerized deployment approach and comes pre-loaded with various industrial protocol parsing plugins; The metadata registration module is used to define metadata for each data point, including aspects, attributes, characteristics, and units. The semantic ID allocation module uses a unified resource naming convention to assign a globally unique identifier to each entity and data point. The protocol adapter framework supports the automatic generation of parsing scripts for unknown and private protocols using machine learning data samples, enabling flexible extension of the protocols. The metadata registration module defines rich contextual information for data points according to the industrial data semantic modeling framework; The identifiers generated by the semantic ID allocation module adopt a hierarchical structure to explicitly represent the organizational structure and physical location of the entity.

3. The decision support system for optimizing public health service resources in industrial and commercial enterprises according to claim 1, characterized in that, The semantic fusion and modeling layer includes: The ontology construction module uses a top-down classification design method to construct an ontology for the industrial, commercial, and public health domains. The domain ontology includes clearly defined core concepts and core relationships. The data mapping module uses the semantic ID to automatically map the parsed multi-source data to the corresponding concepts and instances of the domain ontology; The semantic association module automatically constructs an association network between people, machines, materials, methods, and environment based on the relationships defined in the domain ontology; The public health events defined by the ontology construction module include cluster outbreaks and public health emergencies; The data mapping module maps the attendance system data to the concepts of personnel and workshop, and the sensor data to the concepts of equipment and risk indicators. The production work order data is mapped to the concepts of process and raw materials. The semantic association module constructs a spatial layout network based on location relationships, a personnel interaction network based on contact relationships, and a human-machine relationship network based on operation relationships.

4. The decision support system for optimizing public health service resources in industrial and commercial enterprises according to claim 1, characterized in that, The knowledge graph and risk calculation layer includes: The spatiotemporal graph construction module extends spatiotemporal semantics on the basis of the domain ontology, adding time validity and spatial coordinate attributes to entities and relations; The knowledge graph update module adopts a TransH-based incremental construction method for time-series knowledge graphs, continuously generating the optimal set of triples from the real-time data stream and incrementally updating it to the knowledge graph; The knowledge extraction module uses a Bi-LSTM-CRF model that incorporates trigger words to extract named entities and relations from unstructured text data; The spatiotemporal map construction module adds a time validity attribute to personnel entities to indicate their on-duty time and adds a three-dimensional coordinate attribute to spatial entities to indicate their specific location. The graph update module generates an optimal set of triples from the new data using an optimal subset extraction algorithm. The optimal set of triples is calculated using the following formula: in, Denotes the set of candidate triples. This represents all triples generated from the new data. This represents the triplet scoring function. This indicates a time-limited penalty term for the triple. Indicates the balance parameters; The knowledge extraction module extracts structured information and semantic relationships using the ReLU activation function.

5. The decision support system for optimizing public health service resources in industrial and commercial enterprises according to claim 1, characterized in that, The knowledge graph and risk calculation layer also includes: The parameter extraction module extracts dynamic parameters required for risk deduction from the spatiotemporal knowledge graph in real time, including node attribute parameters and edge weight parameters. The propagation model construction module introduces spatiotemporal constraints on the basis of the SEIR model to establish an improved spatiotemporal propagation model; The simulation module uses the Monte Carlo method to simulate the spread of the epidemic. The node attributes extracted by the parameter extraction module include the number of people, health status, immunity status, spatial geometric layout, volume, ventilation rate, and functional area division. The edge weight parameters include the frequency and distance of contact between people, the movement trajectory of people in space, and the dwell time. The spatial constraints introduced by the propagation model construction module are defined by the personnel movement matrix, and the temporal constraints are achieved by dynamically adjusting the contact rate parameter.

6. The decision support system for optimizing public health service resources in industrial and commercial enterprises according to claim 5, characterized in that, The improved spatiotemporal propagation model is achieved through the following formula: in, Indicates the first Number of susceptible individuals in each region Indicates the number of lurking individuals. Indicates the number of infected persons. Indicates the number of recovered patients. This represents a time-related contact rate parameter. Indicates from the region To the area Personnel movement matrix elements, Indicates the first Total population of each region This indicates the conversion rate from the incubation period to the infectious period. This indicates the conversion rate from the infection period to the recovery period. Indicates the rate of immune system decay; The personnel movement matrix The contact rate parameter is calculated directly from the spatiotemporal knowledge graph. Adjustments will be made dynamically based on production shifts and work schedules.

7. The decision support system for optimizing public health service resources in industrial and commercial enterprises according to claim 1, characterized in that, The intelligent early warning and decision-making layer includes: The digital twin construction module constructs a virtual mapping that runs synchronously with the enterprise based on the spatiotemporal knowledge graph and propagation model. The visualization module provides a three-dimensional view of the company's epidemic heat map, personnel movement lines, and resource distribution on a 3D map. The interactive decision-making module allows users to try different prevention and control plans by dragging and dropping and immediately see the simulation effect comparison; The early warning release module disseminates tiered early warning information through multiple channels. The virtual mapping implemented by the digital twin construction module not only reflects the current state, but can also predict the future state through simulation; The visualization module uses the color depth of a heatmap to represent the risk level and the thickness of streamlines to represent the intensity of personnel movement. The interactive decision-making module supports prevention and control plans including entrance blockade, shift adjustment, and disinfection frequency modification. The information released by the early warning release module includes specific risk areas, risk levels, and recommended prevention and control measures.

8. The decision support system for optimizing public health service resources in industrial and commercial enterprises according to claim 1, characterized in that, The system also includes: The feedback optimization loop records all manual corrections and subsequent effects, using them as feedback signals for reinforcement learning to continuously optimize the model algorithm. The feedback optimization closed loop achieves policy updates through the following formula: in Indicates model parameters, Indicates the learning rate. Represents the policy gradient; The policy gradient is calculated using the following formula: in Indicates the decision-making trajectory. Represents the policy function. Indicates time The action, Indicates time state, Indicates time Value function estimation; The value function estimation is calculated based on the difference between the actual effect after manual correction and the system's predicted effect.

Images