A method for constructing a spatiotemporal dynamic intelligent agent for factory personnel
Patent Information
- Application Number
- CN202611093574.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-22
- Publication Date
- 2026-09-11
AI Technical Summary
第一，系统仅能被动响应预设规则，无法理解业务场景
Smart Images
Figure CN122736102A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method for constructing a spatiotemporal dynamic intelligent agent for personnel in a factory area, belonging to the technical field of intelligent industrial safety production. Background Technology
[0002] With increasingly stringent requirements for industrial safety supervision, enterprises in high-risk industries such as energy and chemicals have widely deployed risk monitoring and early warning systems, dual-prevention information systems, electronic systems for special operations, information systems for training, and personnel positioning systems to achieve digitalization of factory safety operations and real-time monitoring of personnel locations. However, existing personnel positioning systems mainly remain at the level of data collection and rule triggering, and have the following shortcomings: First, the system can only passively respond to preset rules and cannot understand business scenarios. When an alarm occurs, the system can only indicate "who entered the danger zone and when", but cannot answer deeper questions such as "why did they enter?", "was there any intention to violate the rules?", and "is there any connection with surrounding work activities", resulting in a gap of "having data but no insight".
[0003] Second, there is a serious problem of data silos. Although the various modules are on the same platform, there is insufficient data linkage and application. The business data between different modules has not achieved the effect of linkage and integration, and it is difficult to effectively coordinate with other systems (such as work ticket system, video surveillance system, environmental monitoring system, etc.).
[0004] Third, safety management personnel are overwhelmed by a deluge of alarm information. Traditional systems lack the ability to intelligently filter alarm information and understand its context, so signals with real risk value are drowned out by a large number of false alarms, failing to achieve the transformation from "humans searching for reports" to "intelligent agents proactively presenting value".
[0005] Fourth, it lacks natural language interaction capabilities. Managers cannot directly ask questions to the system using natural language; they must obtain information through fixed report interfaces or complex query conditions, resulting in high operational barriers and low efficiency.
[0006] Therefore, there is an urgent need for a method to construct a spatiotemporal dynamic intelligent agent for factory personnel that can integrate multi-source spatiotemporal data, understand business scenarios, support natural language interaction, and proactively generate security insights, in order to solve the above-mentioned deficiencies of existing technologies. Summary of the Invention
[0007] To address the aforementioned problems, this invention discloses a method for constructing a spatiotemporally dynamic intelligent agent for factory personnel, the specific technical solution of which is as follows: A method for constructing a spatiotemporal dynamic intelligent agent for factory personnel includes the following steps: Step S1: Construct a spatiotemporal data acquisition and fusion layer for the factory area: Real-time acquisition of multi-source heterogeneous sensor data within the factory area, including personnel positioning sensors, video surveillance equipment, work permit management system, environmental monitoring sensors, and equipment operation status sensors. Spatiotemporal alignment and semantic fusion of the acquired multi-source heterogeneous data are performed to form a unified spatiotemporal data model. Step S2: Construct a spatiotemporal knowledge graph of the factory area: Based on the unified spatiotemporal data model, extract spatial entities, temporal events and relationships between entities within the factory area, construct a spatiotemporal knowledge graph including five types of nodes: personnel, areas, equipment, work activities and risk factors, and establish spatiotemporal relationships between various types of nodes. Step S3: Construct a scene understanding and intent recognition engine: Based on the spatiotemporal knowledge graph, use a pre-trained large language model to perform semantic parsing of natural language management requests, transform management requests into spatiotemporal data query and reasoning tasks, and identify the current management scenario type and corresponding risk level; Step S4: Construct an intelligent decision generation module: Based on the output of the scene understanding and intent recognition engine, combined with the industry regulatory knowledge base and business rule base, cross-compare and deeply reason the spatiotemporal data to generate personalized security insights and decision suggestions, and push them to the manager after prioritizing them according to risk level. Step S5: Construct a multi-agent collaborative mechanism: Deploy multiple sub-agents with different professional capabilities, including a safety monitoring agent, an operation compliance agent, an emergency response agent, and a risk warning agent. Each sub-agent shares information and collaborates on tasks through a message bus to achieve intelligent safety management and control across the entire factory area.
[0008] Furthermore, in step S1, the specific method for spatiotemporal alignment and semantic fusion of multi-source heterogeneous data is as follows: Step S11: Segment the personnel positioning data into trajectory segments and identify stop points according to time windows, and extract the spatiotemporal trajectory features of the personnel; Step S12: Use target detection algorithms to detect and track people in the video surveillance data, and associate video targets with location targets for identification; Step S13: Perform spatial and temporal matching between the work area, work time, and personnel information of the work ticket data and the spatiotemporal trajectory data; Step S14: Perform spatial interpolation and temporal alignment on environmental monitoring data and equipment operating status data to establish a three-dimensional correlation matrix of environment-equipment-personnel; Step S15: Use a spatiotemporal graph neural network to extract features from the above fused data and generate a unified spatiotemporal data model representation.
[0009] Furthermore, in step S2, the specific method for constructing the spatiotemporal knowledge graph is as follows: Step S21: Define the plant area ontology model, which includes five sub-ontologies: spatial ontology, temporal ontology, personnel ontology, operational ontology, and risk ontology. Each sub-ontology is associated with other ontologies through spatiotemporal attributes. Step S22: Extract entities and relationships between entities from the fused spatiotemporal data. The entities include individual personnel, functional areas, production equipment, work activities, and safety risk events. The relationships include spatiotemporal co-occurrence relationships, work authorization relationships, regional access relationships, and risk transmission relationships. Step S23: Represent the extracted entities and relations using a four-tuple structure <subject, relation, object, spatiotemporal context> to store the knowledge, where the spatiotemporal context includes time interval and spatial range information; Step S24: Establish an incremental update mechanism for the knowledge graph. When new spatiotemporal data arrives, update entity attributes and relationships between entities in real time and trigger a re-evaluation of relevant reasoning rules.
[0010] Furthermore, in step S3, the scene understanding and intent recognition engine operates as follows: Step S31: Receive management requests input by managers in natural language form, and use a pre-trained large language model to perform word segmentation, entity recognition, and intent classification on the input text; Step S32: Map the identified intent to a predefined scenario template library, which includes cluster risk identification scenario, unlicensed operation detection scenario, regional access compliance scenario, operation timeout warning scenario, and emergency evacuation guidance scenario; Step S33: Based on the matched scene template, automatically generate a spatiotemporal data query plan, which includes the required data source, the spatiotemporal range of the query, the associated knowledge graph nodes, and the reasoning path; Step S34: Optimize the query plan based on the current spatiotemporal context, and adjust the query priority and inference depth by combining historical query records and real-time risk situation.
[0011] Furthermore, in step S4, the working method of the intelligent decision generation module is as follows: Step S41: Based on the query plan generated by the scene understanding and intent recognition engine, retrieve relevant entities and relationships from the spatiotemporal knowledge graph to obtain scene context information; Step S42: Match the scenario context information with the industry regulatory knowledge base according to rules. The regulatory knowledge base includes safety production standards, work specifications, and regional access rules. Step S43: Perform multi-dimensional cross-comparison analysis on spatiotemporal data, including regional-time cross-comparison of personnel trajectories and work tickets, personnel-time cross-comparison of aggregation behavior and shift schedule, and equipment-personnel cross-comparison of equipment status and work activities. Step S44: Based on the comparative analysis results, a security insight report in natural language form is generated using a large language model. The report includes a risk description, risk cause analysis, impact scope assessment, and recommended remedial measures. Step S45: Prioritize the decision recommendations according to risk level, and present the results to the manager through a visual interface and message push.
[0012] Furthermore, in step S5, the specific implementation method of the multi-agent cooperative mechanism is as follows: Step S51: The safety monitoring intelligent agent is responsible for continuously monitoring the real-time location and activity status of personnel in the factory area. When abnormal gathering, boundary crossing behavior or overcrowding in an area is detected, it sends an early warning message to other intelligent agents. Step S52: The work compliance intelligent agent is responsible for automatically reviewing the compliance of work activities, including verifying the validity of work tickets, verifying the qualifications of workers, and checking the time range of work. When it detects suspected work without a ticket, it triggers an alarm. Step S53: The emergency response agent is responsible for emergency dispatch when an emergency occurs, including evacuation route planning, emergency resource allocation and rescue force guidance, and coordinating with other agents to update the risk situation. Step S54: The risk warning intelligent agent is responsible for assessing and predicting the risk situation of the entire factory area, establishing a risk propagation model based on historical and real-time data, identifying potential risk chains in advance and issuing warnings. Step S55: Each sub-agent communicates through a message bus based on a publish-subscribe model. The message format adopts a unified spatiotemporal event description language, which supports semantic interoperability across agents.
[0013] Furthermore, in step S12, the method for associating the identity of the video target and the location target is as follows: The system extracts appearance feature vectors from the targets detected in the video, extracts the location coordinates of the corresponding time from the location data according to the timestamp, searches for matching targets in the video with the location as a constraint, and confirms that they are the same person when the appearance feature similarity exceeds a preset threshold and the spatial distance is less than a preset threshold, and establishes a mapping relationship between video ID and location ID.
[0014] Furthermore, the working method for the cluster risk identification scenario is as follows: When the number of people in a certain area exceeds the threshold and the duration exceeds the set time, the system automatically retrieves the shift schedule, shift handover records and emergency drill plan for that area. By querying the work attributes and relationships of the gathered personnel through the spatiotemporal knowledge graph, it determines whether the gathering behavior is a normal shift handover activity, a planned drill activity or an unplanned abnormal gathering. For unplanned abnormal gatherings, a risk alarm is generated and disposal suggestions are pushed out.
[0015] Furthermore, the working method for the no-ticket operation detection scenario is as follows: Managers submit a query request to the system in natural language to check for suspected unlicensed work. The system automatically retrieves spatiotemporal trajectory data within a specified time range, cross-compares the personnel's stay records in the work area with the work ticket database, filters out personnel whose stay time in the work area exceeds a preset threshold and who do not have a corresponding valid work ticket, generates a list of suspected unlicensed work, and pushes an alarm.
[0016] Furthermore, the method for constructing the industry regulatory knowledge base is as follows: Collect relevant laws, regulations, industry standards, and internal safety management systems for safe production. Structuralize the regulatory texts, extract the applicable conditions, behavioral requirements, and consequences of violations from the regulatory clauses, and transform the regulatory clauses into computable rule expressions. The rule expressions are stored in the form of a four-tuple of <triggering condition, constraint condition, violation judgment, and handling suggestion>, and establish a mapping relationship between the rules and spatiotemporal knowledge graph entities.
[0017] The beneficial effects of this invention are: This invention achieves an intelligent leap from "passive response" to "active understanding" by constructing a spatiotemporal data acquisition and fusion layer for personnel, a spatiotemporal knowledge graph, a scene understanding and intent recognition engine, an intelligent decision generation module, and a multi-agent collaborative mechanism. This enables the factory safety management system to possess four core capabilities: data acquisition and fusion, knowledge base fusion, demand understanding, and decision generation.
[0018] This invention enables intelligent agents to leap from "seeing" to "understanding." By constructing a spatiotemporal knowledge graph and a scene understanding engine, this invention endows the system with the ability to understand business scenarios. It can identify the attribute status of personnel in dangerous areas and provide monitoring and early warning, such as whether they have the necessary permissions to enter a certain area, whether they have been trained on the corresponding two forms and three cards, whether there is equipment malfunction prohibiting entry, whether they are under maintenance prohibiting entry, and whether personnel gathering is within permitted timeframes. It also optimizes the causal and temporal relationships of alarms by modeling them, pushing only root cause alarms and corresponding handling measures, thus optimizing alarm processing.
[0019] This invention supports interactive decision-making using natural language. Managers can directly ask questions to the system in natural language, and the system automatically completes cross-system data comparison and analysis without the need for complex query condition settings. Combining an abnormal operating condition database and knowledge graph, it provides concrete feedback and guidance, reducing training barriers and improving learning and response efficiency.
[0020] This invention realizes a data-driven decision-making closed loop. From multi-source data collection to spatiotemporal knowledge graph construction to scene understanding to intelligent decision generation to multi-agent collaborative execution, a complete data-driven decision-making closed loop is formed, truly realizing "intelligent agents proactively presenting value".
[0021] This invention possesses multi-agent collaborative capabilities. By deploying four specialized sub-agents—safety monitoring, operational compliance, emergency response, and risk warning—each sub-agent collaborates and shares information, achieving intelligent safety management and control across the entire factory area, both in real-time and across all scenarios.
[0022] This invention boasts strong adaptability and flexible deployment. Based on an industry-specific abnormal operating condition database, a GB standard rule engine, a risk factor model, a scenario template library, and rule expressions, the system can quickly adapt to the management needs of different enterprises without requiring extensive customization, thus reducing adaptation costs. Attached Figure Description
[0023] Figure 1 This is a schematic diagram of the system architecture of the present invention; Figure 2 This is a flowchart of the spatiotemporal data acquisition and fusion layer in this invention; Figure 3 This is a schematic diagram of the spatiotemporal knowledge graph in this invention; Figure 4 This is a flowchart of the scene understanding and intent recognition engine in this invention; Figure 5 This is a flowchart of the intelligent decision generation module in this invention; Figure 6 This is a schematic diagram of information interaction in the multi-agent collaborative mechanism of this invention. Detailed Implementation
[0024] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.
[0025] Combined with appendix Figure 1As can be seen, this invention comprises five core components: a spatiotemporal data acquisition and fusion layer, a spatiotemporal knowledge graph, a scene understanding and intent recognition engine, an intelligent decision generation module, and a multi-agent collaborative mechanism. The spatiotemporal dynamic intelligent agent system architecture for factory personnel is divided into three layers: the top layer is the manager interaction layer, supporting natural language questioning, intelligent insight presentation, and decision suggestion push; the middle layer is the AI intelligent agent engine, including an intelligent decision generation module (regulatory matching / cross-comparison / insight generation), a scene understanding and intent recognition engine (semantic parsing / scene matching / query planning), a spatiotemporal knowledge graph (five types of nodes: personnel / area / equipment / operation / risk), and a multi-agent collaborative mechanism (four sub-agents: safety monitoring / operational compliance / emergency response / risk warning); the bottom layer is the spatiotemporal data acquisition and fusion layer, including five data sources: personnel positioning (UWB / Bluetooth), video surveillance (target detection and tracking), work ticket system (authorization / qualification / time), environmental monitoring (temperature / pressure / gas), and equipment status (vibration / current / electricity). The three layers are connected through bidirectional data flow, achieving a complete closed loop from data acquisition to intelligent decision-making.
[0026] Example 1. The following are examples of the execution of each step, with reference to the specific execution content of each step:
[0027] like Figure 2 As shown, the workflow of the spatiotemporal data acquisition and fusion layer is as follows: Step S11: Segment the personnel positioning data into trajectory segments and identify stop points according to time windows, and extract the spatiotemporal trajectory features of the personnel; Step S12: Use target detection algorithms to detect and track people in the video surveillance data, and associate video targets with location targets to identify them; Step S13: Perform spatial and temporal matching between the work area, work time, and personnel information of the work ticket data and the spatiotemporal trajectory data; Step S14: Perform spatial interpolation and temporal alignment on environmental monitoring data and equipment operating status data to establish a three-dimensional correlation matrix of environment-equipment-personnel; Step S15: Use a spatiotemporal graph neural network to extract features from the above fused data and generate a unified spatiotemporal data model representation.
[0028] The five steps are executed sequentially, ultimately outputting a unified spatiotemporal data model.
[0029] Within the factory area, personnel positioning sensors (UWB positioning base stations, Bluetooth beacons, etc.), video surveillance equipment, a work permit management system, environmental monitoring sensors (temperature, pressure, gas concentration, etc.), and equipment operation status sensors (vibration, current, etc.) are deployed to collect the aforementioned multi-source heterogeneous data in real time. The acquisition frequency is set according to the sensor type: 1Hz for personnel positioning data, 25fps for video surveillance, 5 seconds for environmental monitoring data, and 1 second for equipment operation status data. The collected multi-source heterogeneous data undergoes spatiotemporal alignment and semantic fusion: First, time synchronization is performed on each data source using a unified timestamp server as a reference; second, spatial alignment is performed using the factory's three-dimensional coordinate system as a reference; finally, ST-GNN spatiotemporal graph neural network is used to extract features from the fused data, generating a unified spatiotemporal data model representation.
[0030] ST-GNN (Spatiotemporal Graph Neural Network) is a GNN structure specifically designed for jointly processing time-varying network data based on a spatiotemporal topology. In general convolution operators... In the diagram, *D represents the convolution operator applied to a signal with an underlying architecture of D. Representation function With signal In the domain The result of the convolution operation on, It is the kernel function of the Fourier transform. The homogeneous diffusion equation is defined as follows: ,in It is the Laplacian operator. Based on this, we obtain temporal convolution and graph convolution: , in It is a translation operator. It is about variables Continuous functions, Indicates the position after translation. express and Convolution operation, This represents the impulse response sequence of a discrete system. This represents the discrete translation operator.
[0031] The ST-GNN spatiotemporal graph neural network consists of L layers, each containing a set of spatiotemporal graph filters and a pointwise nonlinear activation function. The input of the l-th layer is the output of the previous layer, and the output of the l-th layer is defined as follows:
[0032] in Indicates the first Layer, First Output tensors of each feature channel These are learnable spatiotemporal filter coefficients. By constructing the above spatiotemporal network, stable transmission of spatiotemporal data and a unified representation of three-dimensional coordinates can be achieved.
[0033] Step S2: Construct a spatiotemporal knowledge graph of the factory area like Figure 3 As shown, the spatiotemporal knowledge graph contains five types of nodes and four types of relationships: the five types of nodes are personnel nodes (employees / contractors / visitors), area nodes (factory area / equipment / room / passageway), equipment nodes (production equipment / safety facilities), operation nodes (hot work / confined space / work at height), and risk nodes (fire / explosion / poisoning / leakage); the four types of relationships are spatiotemporal co-occurrence relationships (the temporal-spatial association between personnel and areas), operation authorization relationships (authorization binding between personnel and operations), area access relationships (access rules between personnel / operations and areas), and risk transmission relationships (the transmission link between equipment / operations and risk events). All relationships are stored using a quadruple <subject, relationship, object, spatiotemporal context>.
[0034] Based on the unified spatiotemporal data model obtained in step S1, a plant ontology model is defined, including five sub-ontologies: spatial ontology (plant area, equipment, rooms, passageways, etc.), temporal ontology (shifts, work periods, inspection cycles, etc.), personnel ontology (employees, contractors, visitors, etc.), operational ontology (hot work, confined space work, work at height, etc.), and risk ontology (fire, explosion, poisoning, leakage, etc.). Entities and relationships between entities are extracted from the fused spatiotemporal data. Entities include individual personnel, functional areas, production equipment, work activities, and safety risk events. Relationships include spatiotemporal co-occurrence relationships, work authorization relationships, regional access relationships, and risk transmission relationships. A four-tuple structure <subject, relation, object, spatiotemporal context> is used to store knowledge, where the spatiotemporal context includes time interval and spatial range information. An incremental update mechanism for the knowledge graph is established; when new spatiotemporal data arrives, a CRF random conditional field is used.
[0035] in In a given sequence Under the given conditions, output the label sequence The conditional probability, These are the weight parameters corresponding to the transition feature function. It is a state transition activation function. These are the weight parameters of the corresponding state feature function. Entity relationships are determined probabilistically based on contextual information, entity attributes and relationships between entities are updated in real time, and relevant inference rules are re-evaluated.
[0036] like Figure 4 As shown, the workflow of the scene understanding and intent recognition engine is as follows: Step S31: Receive natural language input from the administrator and use a large language model for word segmentation, entity recognition, and intent classification; Step S32: Map the identified intent to the scenario template library (aggregation risk / work without a ticket / area access / work timeout / emergency evacuation); Step S33: Automatically generate a spatiotemporal data query plan based on the matched scene template; Step S34: Optimize the query plan based on the current spatiotemporal context and output the optimized query plan.
[0037] The four steps are executed sequentially to achieve the conversion from natural language to structured queries.
[0038] The system receives management requests input by administrators in natural language, such as "What unlicensed operations have been suspected in the past week?" or "Are there any abnormal clusters in Unit 3?". A pre-trained Large Language Model (LLM) is used to segment the input text, perform entity recognition, and intent classification. The BiLSTM-CRF sequence labeling algorithm is then used to... , It is the score corresponding to the actual label. The process involves calculating the loss value from the index scores of all candidate options, extracting key entities such as the time and location with the minimum loss value, and mapping them to a predefined scenario template library based on the identified intent. This library includes scenarios for cluster risk identification, unlicensed operation detection, regional access compliance, operation timeout warnings, and emergency evacuation guidance. Based on the matched scenario templates, a spatiotemporal data query plan is automatically generated. This plan includes the required data sources, the spatiotemporal scope of the query, associated knowledge graph nodes, and the inference path. The query plan is then optimized based on the current spatiotemporal context, incorporating historical query records and real-time risk assessment to adjust query priority and inference depth.
[0039] like Figure 5 As shown, the workflow of the intelligent decision generation module is as follows: Step S41: Retrieve relevant entities and relationships from the spatiotemporal knowledge graph according to the query plan to obtain the scene context; Step S42: Match the scenario context with the legal knowledge base according to the rules; Step S43: Perform multi-dimensional cross-comparison analysis on the spatiotemporal data; Step S44: Generate a natural language security insight report based on the comparison results using a large language model; Step S45: Sort by risk level and push to the manager.
[0040] The five steps are executed sequentially to achieve a complete process from data query to decision push.
[0041] Based on the query plan generated in step S3, relevant entities and relationships are retrieved from the spatiotemporal knowledge graph to obtain scenario context information. This scenario context information is then matched against industry regulatory knowledge bases, which include safety production standards such as AQ 3063, operational specifications, and regional access rules. Multi-dimensional cross-analysis of the spatiotemporal data is performed, including regional-time cross-comparison of personnel trajectories and work tickets, personnel-time cross-comparison of clustering behavior and shift schedules, and equipment-personnel cross-comparison of equipment status and work activities. Based on this logic, the Analytic Hierarchy Process (AHP) is used, combined with the industry regulatory knowledge base, to construct the criteria layer A:
[0042] Among them Representing the logic of regulations, It is the number of elements to compare; the elements of A satisfy... Risk levels are quantified and scored, and a ranking algorithm is used to determine the order of push notifications. A large language model is then used to generate a security insight report in natural language, which includes a risk description, risk causal analysis, impact assessment, and recommended remedial measures. Decision recommendations are prioritized according to risk level, and the results are presented to managers through a visual interface and push notifications.
[0043] like Figure 6 As shown, in the multi-agent collaborative mechanism, four sub-agents interact through a message bus: Safety Monitoring Agent (S51): Monitors personnel location and activity status, detects abnormal gatherings / boundary crossings / overstaffing, and sends alerts; Operation Compliance Agent (S52): Reviews operation compliance, verifies work permits / personnel qualifications / time ranges, and detects unlicensed operations; Emergency Response Agent (S53): Handles emergency dispatch, plans evacuation routes / allocates resources / guides rescue efforts; Risk Warning Agent (S54): Assesses and predicts the overall risk situation, establishes a risk propagation model, and issues warnings. The message bus, based on a publish-subscribe model, uses a unified spatiotemporal event description language and supports cross-agent semantic interoperability. The four sub-agents communicate bidirectionally through the message bus, forming a collaborative linkage.
[0044] Four sub-agents with different professional capabilities are deployed: (1) a safety monitoring agent, which is responsible for continuously monitoring the real-time location and activity status of personnel in the factory area. When abnormal gathering, boundary crossing, or overcrowding in an area is detected, it sends a warning message to other agents; (2) an operation compliance agent, which is responsible for automatically reviewing the compliance of operation activities, including verifying the validity of work tickets, verifying the qualifications of operators, and verifying the time range of operations. When an unlicensed operation is suspected, an alarm is triggered; (3) an emergency response agent, which is responsible for emergency dispatch when an emergency occurs, including evacuation route planning, emergency resource allocation, and rescue force guidance; (4) a risk warning agent, which is responsible for assessing and predicting the risk situation of the entire factory area, establishing a risk propagation model based on historical and real-time data, identifying potential risk chains in advance, and issuing warnings. Each sub-agent communicates through a message bus based on a publish-subscribe pattern. The message format adopts a unified spatiotemporal event description language, which supports semantic interoperability across agents.
[0045] Taking a cluster alarm scenario in a chemical plant as an example, the specific application of this invention is illustrated. When the safety monitoring agent detects that six people have gathered in Unit 3 for more than 10 minutes, the traditional system can only issue an alarm stating "Six people have gathered in Unit 3 for more than 10 minutes." However, the spatiotemporal agent of this invention processes the information as follows: First, the safety monitoring agent publishes the cluster event information to other agents via a message bus; after receiving the message, the work compliance agent queries the work attributes and relationships of the six individuals in the spatiotemporal knowledge graph, finding that five are on-duty operators and one is an inspector; simultaneously, it queries the shift schedule and handover records to confirm that the current time period is a normal shift handover time; the emergency response agent confirms that there is no emergency drill plan. Based on the analysis results of all agents, the intelligent decision generation module determines that this cluster is a normal shift handover activity, generating the insight conclusion "The cluster in Unit 3 is a normal shift handover and requires no action," thus avoiding false alarms.
[0046] Take the scenario of detecting unlicensed operations at a petrochemical company as an example. The manager inputs a natural language query into the system: "What unlicensed operations are suspected in the past week?" The scenario understanding and intent recognition engine interprets this request as an unlicensed operation detection scenario and generates a query plan: the time range is the last 7 days, the data source is personnel location trajectories and a work ticket database, and the query condition is "staying in the work area for more than 10 minutes without a valid work ticket." The intelligent decision generation module executes the query, cross-compares the spatiotemporal trajectory data and work ticket data, filters out 3 personnel who stayed in confined spaces for more than 10 minutes but did not have corresponding valid work tickets, generates a list of suspected unlicensed operations including personnel name, stay time, stay area, and stay duration, and pushes an alert to the manager.
[0047] Those skilled in the art will understand that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as herein.
[0048] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.
Claims
1. A method for constructing a spatiotemporal dynamic intelligent agent for personnel in a factory area, characterized in that, Includes the following steps: Step S1: Construct a spatiotemporal data acquisition and fusion layer for the factory area: Real-time acquisition of multi-source heterogeneous sensor data within the factory area, including personnel positioning sensors, video surveillance equipment, work permit management system, environmental monitoring sensors, and equipment operation status sensors. Spatiotemporal alignment and semantic fusion of the acquired multi-source heterogeneous data are performed to form a unified spatiotemporal data model. Step S2: Construct a spatiotemporal knowledge graph of the factory area: Based on the unified spatiotemporal data model, extract spatial entities, temporal events and relationships between entities within the factory area, construct a spatiotemporal knowledge graph including five types of nodes: personnel, areas, equipment, work activities and risk factors, and establish spatiotemporal relationships between various types of nodes. Step S3: Construct a scene understanding and intent recognition engine: Based on the spatiotemporal knowledge graph, use a pre-trained large language model to perform semantic parsing of natural language management requests, transform management requests into spatiotemporal data query and reasoning tasks, and identify the current management scenario type and corresponding risk level; Step S4: Construct an intelligent decision generation module: Based on the output of the scene understanding and intent recognition engine, combined with the industry regulatory knowledge base and business rule base, cross-compare and deeply reason the spatiotemporal data to generate personalized security insights and decision suggestions, and push them to the manager after prioritizing them according to risk level. Step S5: Construct a multi-agent collaborative mechanism: Deploy multiple sub-agents with different professional capabilities, including a safety monitoring agent, an operation compliance agent, an emergency response agent, and a risk warning agent. Each sub-agent shares information and collaborates on tasks through a message bus to achieve intelligent safety management and control across the entire factory area.
2. The method for constructing a spatiotemporal dynamic intelligent agent for factory personnel according to claim 1, characterized in that, In step S1, the specific method for spatiotemporal alignment and semantic fusion of multi-source heterogeneous data is as follows: Step S11: Segment the personnel positioning data into trajectory segments and identify stop points according to time windows, and extract the spatiotemporal trajectory features of the personnel; Step S12: Use target detection algorithms to detect and track people in the video surveillance data, and associate video targets with location targets to identify them; Step S13: Perform spatial and temporal matching between the work area, work time, and personnel information of the work ticket data and the spatiotemporal trajectory data; Step S14: Perform spatial interpolation and temporal alignment on environmental monitoring data and equipment operating status data to establish a three-dimensional correlation matrix of environment-equipment-personnel; Step S15: Use a spatiotemporal graph neural network to extract features from the above fused data and generate a unified spatiotemporal data model representation.
3. The method for constructing a spatiotemporal dynamic intelligent agent for factory personnel according to claim 1, characterized in that, In step S2, the specific method for constructing the spatiotemporal knowledge graph is as follows: Step S21: Define the plant area ontology model, which includes five sub-ontologies: spatial ontology, temporal ontology, personnel ontology, operational ontology, and risk ontology. Each sub-ontology is associated with other ontologies through spatiotemporal attributes. Step S22: Extract entities and relationships between entities from the fused spatiotemporal data. The entities include individual personnel, functional areas, production equipment, work activities, and safety risk events. The relationships include spatiotemporal co-occurrence relationships, work authorization relationships, regional access relationships, and risk transmission relationships. Step S23: Represent the extracted entities and relations using a four-tuple structure <subject, relation, object, spatiotemporal context> to store the knowledge, where the spatiotemporal context includes time interval and spatial range information; Step S24: Establish an incremental update mechanism for the knowledge graph. When new spatiotemporal data arrives, update entity attributes and relationships between entities in real time and trigger a re-evaluation of relevant reasoning rules.
4. The method for constructing a spatiotemporal dynamic intelligent agent for factory personnel according to claim 1, characterized in that, In step S3, the scene understanding and intent recognition engine works as follows: Step S31: Receive management requests input by managers in natural language form, and use a pre-trained large language model to perform word segmentation, entity recognition, and intent classification on the input text; Step S32: Map the identified intent to a predefined scenario template library, which includes cluster risk identification scenario, unlicensed operation detection scenario, regional access compliance scenario, operation timeout warning scenario, and emergency evacuation guidance scenario; Step S33: Based on the matched scene template, automatically generate a spatiotemporal data query plan, which includes the required data source, the spatiotemporal range of the query, the associated knowledge graph nodes, and the reasoning path; Step S34: Optimize the query plan based on the current spatiotemporal context, and adjust the query priority and inference depth by combining historical query records and real-time risk situation.
5. The method for constructing a spatiotemporal dynamic intelligent agent for factory personnel according to claim 1, characterized in that, In step S4, the working method of the intelligent decision generation module is as follows: Step S41: Based on the query plan generated by the scene understanding and intent recognition engine, retrieve relevant entities and relationships from the spatiotemporal knowledge graph to obtain scene context information; Step S42: Match the scenario context information with the industry regulatory knowledge base according to rules. The regulatory knowledge base includes safety production standards, work specifications, and regional access rules. Step S43: Perform multi-dimensional cross-comparison analysis on spatiotemporal data, including regional-time cross-comparison of personnel trajectories and work tickets, personnel-time cross-comparison of aggregation behavior and shift schedule, and equipment-personnel cross-comparison of equipment status and work activities. Step S44: Based on the comparative analysis results, a security insight report in natural language form is generated using a large language model. The report includes a risk description, risk cause analysis, impact scope assessment, and recommended remedial measures. Step S45: Prioritize the decision recommendations according to risk level, and present the results to the manager through a visual interface and message push.
6. The method for constructing a spatiotemporal dynamic intelligent agent for factory personnel according to claim 1, characterized in that, In step S5, the specific implementation method of the multi-agent cooperative mechanism is as follows: Step S51: The safety monitoring intelligent agent is responsible for continuously monitoring the real-time location and activity status of personnel in the factory area. When abnormal gathering, boundary crossing behavior or overcrowding in an area is detected, it sends an early warning message to other intelligent agents. Step S52: The work compliance intelligent agent is responsible for automatically reviewing the compliance of work activities, including verifying the validity of work tickets, verifying the qualifications of workers, and checking the time range of work. When it detects suspected work without a ticket, it triggers an alarm. Step S53: The emergency response agent is responsible for emergency dispatch when an emergency occurs, including evacuation route planning, emergency resource allocation and rescue force guidance, and coordinating with other agents to update the risk situation. Step S54: The risk warning intelligent agent is responsible for assessing and predicting the risk situation of the entire factory area, establishing a risk propagation model based on historical and real-time data, identifying potential risk chains in advance and issuing warnings. Step S55: Each sub-agent communicates through a message bus based on a publish-subscribe model. The message format adopts a unified spatiotemporal event description language, which supports semantic interoperability across agents.
7. The method for constructing a spatiotemporal dynamic intelligent agent for factory personnel according to claim 2, characterized in that, In step S12, the method for associating the identity of the video target and the location target is as follows: The system extracts appearance feature vectors from the targets detected in the video, extracts the location coordinates of the corresponding time from the location data according to the timestamp, searches for matching targets in the video with the location as a constraint, and confirms that they are the same person when the appearance feature similarity exceeds a preset threshold and the spatial distance is less than a preset threshold, and establishes a mapping relationship between video ID and location ID.
8. The method for constructing a spatiotemporal dynamic intelligent agent for factory personnel according to claim 1, characterized in that, The working method for the cluster risk identification scenario is as follows: When the number of people in a certain area exceeds the threshold and the duration exceeds the set time, the system automatically retrieves the shift schedule, shift handover records and emergency drill plan for that area. By querying the work attributes and relationships of the gathered personnel through the spatiotemporal knowledge graph, it determines whether the gathering behavior is a normal shift handover activity, a planned drill activity or an unplanned abnormal gathering. For unplanned abnormal gatherings, a risk alarm is generated and disposal suggestions are pushed out.
9. The method for constructing a spatiotemporal dynamic intelligent agent for factory personnel according to claim 4, characterized in that, The working method for detecting unlicensed operations is as follows: Managers submit a query request to the system in natural language to check for suspected unlicensed work. The system automatically retrieves spatiotemporal trajectory data within a specified time range, cross-compares the personnel's stay records in the work area with the work ticket database, filters out personnel whose stay time in the work area exceeds a preset threshold and who do not have a corresponding valid work ticket, generates a list of suspected unlicensed work, and pushes an alarm.
10. The method for constructing a spatiotemporal dynamic intelligent agent for factory personnel according to claim 1, characterized in that, The method for constructing the industry regulatory knowledge base is as follows: Collect relevant laws, regulations, industry standards, and internal safety management systems for safe production. Structuralize the regulatory texts, extract the applicable conditions, behavioral requirements, and consequences of violations from the regulatory clauses, and transform the regulatory clauses into computable rule expressions. The rule expressions are stored in the form of a four-tuple of <triggering condition, constraint condition, violation judgment, and handling suggestion>, and establish a mapping relationship between the rules and spatiotemporal knowledge graph entities.