Campus hidden danger identification method based on double nested agent architecture and related equipment

Abstract

The invention discloses a campus hidden danger identification method based on a double nested intelligent agent architecture and related equipment, and aims to focus on a hidden danger identification level by a double nested intelligent agent and correspondingly improve an intelligent agent model based on the application field of campus hidden danger identification. According to the campus hidden danger recognition system, the multi-level dynamic interaction of individual-local-global can be carried out, real-time interaction between personnel behaviors and local environments and real-time interaction between different environment units are recognized in real time, and therefore the whole campus hidden danger recognition is completed.

Description

Technical Field

[0001] This invention relates to the field of campus hazard identification technology, specifically to a campus hazard identification method and related equipment based on a dual-nested intelligent agent architecture. Background Technology

[0002] Currently, in the field of campus hazard identification based on dynamic data-driven approaches, existing technologies generally adopt a centralized data processing and analysis framework. Specifically, these technologies construct a campus Internet of Things (IoT) by deploying various sensors and cameras, aggregating multi-source data such as video streams, equipment status, and personnel trajectories to a central server, and using machine learning algorithms for anomaly detection and pattern recognition. Its main implementation path can be summarized as "data acquisition - centralized fusion - model recognition - early warning output," with its core relying on building a unified feature engineering and statically or periodically updated analysis model to achieve automated monitoring of visible hazards such as fires, crowd gatherings, and equipment malfunctions.

[0003] While existing technologies have achieved basic data-driven approaches, they have inherent limitations when dealing with the complex and dynamic system of a campus. First, their centralized, hierarchical processing architecture is inherently static, failing to effectively model and characterize the dynamic evolution of potential risks. In particular, they neglect the shaping effect of real-time interactions and game theory between "personnel behavior" and the "local environment," as well as between "different environmental units," on system risk. Second, existing methods primarily understand "dynamism" at the level of temporal updates of data flows, lacking in-depth descriptions of the nonlinear interactions between various elements within the system and the mechanisms of risk transmission across domains. This makes it difficult to predict systemic risks that gradually evolve from the coupling of multiple factors. Finally, their model updates are typically lagging and passive, unable to adaptively adjust and learn online based on real-time interactive feedback during operation, resulting in limited generalization ability when facing unknown scenarios or changes in behavioral patterns.

[0004] In summary, existing technologies fail to reflect the nature of campuses as complex adaptive systems in terms of architecture, and lack the ability to dynamically characterize the mechanisms of risk emergence in terms of methodology. Therefore, how to construct an adaptive intelligent system that can accurately simulate multi-level dynamic interactions of "individual-local-global" and identify and predict the emergence of systemic hidden dangers in real time has become the core challenge facing current technologies. Summary of the Invention

[0005] Based on the problems raised in the background technology, the purpose of this invention is to provide a campus hazard identification method and related equipment based on a dual nested intelligent agent architecture, which solves the problems that the existing technology fails to reflect the nature of the campus as a complex adaptive system in terms of architecture and lacks the ability to dynamically characterize the risk emergence mechanism in terms of method.

[0006] This invention is achieved through the following technical solution:

[0007] The first aspect of this invention provides a method for identifying campus safety hazards based on a dual-nested intelligent agent architecture, comprising the following steps:

[0008] Step S1: Receive the campus multi-source heterogeneous data stream, and construct several inner-layer intelligent agents based on the campus multi-source heterogeneous data stream; wherein, each inner-layer intelligent agent is configured to perform hazard analysis on the corresponding local campus scene and output building hazard data stream;

[0009] Step S2: Integrate the building hazard data streams output by each inner-layer intelligent agent and fuse them with the campus multi-source heterogeneous data streams; construct a nested observation environment based on the fused data.

[0010] Step S3: Construct a nested outer intelligent agent. The nested outer intelligent agent performs hazard analysis on the global campus scene by observing the nested observation environment and outputs campus hazard data.

[0011] In the aforementioned technical solution, firstly, a dedicated analysis process is established at the local level. Specifically, a corresponding inner-layer intelligent agent is created for each local campus scenario (e.g., teaching buildings, laboratory buildings, playgrounds, etc.). This agent continuously receives real-time monitoring information from specific areas and analyzes the interaction between the behavioral characteristics of active subjects and the conditions of the location. These processes transform the correlation between visible behavioral patterns and the functional state of the location, as well as the mutual influence between different elements, into a measurable risk development trajectory. This achieves a complete micro-dynamic depiction of the risk formation process, realizing dynamic interaction between the "individual" and the "local."

[0012] Then, by integrating the building hazard data streams output by each inner-layer agent, a global correlation analysis process is established. The building hazard data streams obtained from each local analysis are integrated and merged with the overall campus operation status. This constructs a nested observation environment for nesting outer-layer agents. This nested observation environment can be used to identify potential correlation paths between different areas, explore the ripple effects of local anomalies on adjacent areas, and explore the chain reaction patterns between multiple areas, thereby providing a basis for realizing dynamic interaction between "local and global".

[0013] Finally, a nested outer agent is constructed. This outer agent is a comprehensive agent nested within several inner agents. It can observe not only the potential risks identified by the inner agents but also the entire campus environment (e.g., main roads, paths connecting different buildings, and trails) outside the inner agents, enabling global identification of the entire campus and allowing local conclusions to evolve synergistically with global assessments. When new risk characteristics are discovered, the global analysis results can guide local processes to adjust their observation priorities; simultaneously, the experience of local processes in handling unknown situations can provide a reference for global assessments. This continuous adaptive optimization allows the entire identification process to continuously improve as site conditions change, effectively enhancing the ability to identify new risks.

[0014] In one optional embodiment, each inner-layer agent is configured to perform hazard analysis on a corresponding local campus scene, including the following steps:

[0015] Extract the corresponding local campus scene from the multi-source heterogeneous data stream of the campus as the inner observation environment;

[0016] The inner-layer agent obtains personnel status datasets and facility status datasets by observing the inner-layer observation environment, and establishes an inner-layer state space using these datasets;

[0017] Historical campus hazard data is acquired, and an action space is constructed based on the historical campus hazard data. The inner layer intelligent agent determines the action behavior from the action space based on the inner layer state space.

[0018] Execute the aforementioned action, and obtain the first and second campus hazard statuses before and after execution. Calculate the reward value based on the first and second campus hazard statuses and historical campus hazard data.

[0019] An experience pool for training the inner agent is established based on the inner state space, the action space, and the reward value.

[0020] In one optional embodiment, the personnel status dataset includes: personnel spatial distribution data, movement trajectory data, and behavioral pattern data; the facility status dataset includes: facility structural safety data, electrical facility safety data, fire protection facility safety data, and specific facility safety data.

[0021] The inner-layer agent acquires behavioral pattern data by observing the inner-layer observation environment, including the following steps:

[0022] The inner-layer intelligent agent acquires video surveillance data and audio surveillance data through data acquisition devices;

[0023] The video surveillance data is subjected to human posture recognition. If the recognized human posture involves interactive actions, the audio surveillance data is subjected to audio recognition. The recognized audio is then spliced ​​with the recognized human posture to generate behavior pattern data.

[0024] In one optional embodiment, the reward value is calculated based on the first campus hazard status, the second campus hazard status, and historical campus hazard data, including the following steps:

[0025] The historical campus hazard data is used to construct reward weights for accuracy of identification, timeliness of identification, progress, and punishment.

[0026] The confidence level of the second campus hazard status is calculated to obtain the credibility. The credibility is then integrated with the recognition accuracy reward weight to serve as the recognition accuracy reward.

[0027] Determine the state time of the second campus hazard state, and concatenate the state time with the second campus hazard state to obtain state-time data;

[0028] Obtain the timeline of hazard development, compare the state-time data with the hazard data line to obtain the timeliness value, and integrate the timeliness value with the identification timeliness reward weight as the identification timeliness reward;

[0029] The scenario fitness value is determined by the first campus hazard status and the second campus hazard status, and the scenario fitness value is fused with the progress weight as a progress reward.

[0030] The error cost is determined based on the second campus hazard status, and the error cost is integrated with the penalty weight as the error penalty;

[0031] The reward value is calculated by taking the recognition accuracy reward, the recognition timeliness reward, and the progress reward as positive reward values ​​and the error penalty as a negative reward value.

[0032] In one optional embodiment, the building hazard data streams output by each inner-layer intelligent agent are integrated and fused with the multi-source heterogeneous data streams of the campus. A nested observation environment is constructed based on the fused data, including the following steps:

[0033] Step S21: Obtain the building hazard data stream output by each inner-layer agent, and based on the preset campus map data, map each inner-layer agent to a spatial node in the map data to generate a dynamic graph structure containing spatial association attributes; wherein, the node attributes of the dynamic graph structure at least include the building hazard data output by the corresponding inner-layer agent.

[0034] Step S22: Spatiotemporally align and fuse the dynamic graph structure with the campus multi-source heterogeneous data stream to construct a unified digital environmental space that integrates the physical environment and the status of potential hazards.

[0035] In one optional embodiment, the nested outer agent performs a hazard analysis of the global campus scene by observing the nested observation environment, including the following steps:

[0036] Step S31: The nested outer agent obtains the hidden danger state dataset of each inner agent by observing the nested observation environment, and establishes the nested outer state space using the dataset.

[0037] Step S32: Obtain historical campus hazard data, construct a collaborative action space based on the historical campus hazard data, and determine the collaborative action behavior from the collaborative action space based on the outer state space of the nested outer layer intelligent agent;

[0038] Step S33: Execute the collaborative action behavior, obtain the hidden danger status of multiple inner-layer intelligent agents after execution, and associate the hidden danger status of multiple inner-layer intelligent agents based on the campus space topology to generate an associated campus hidden danger status map.

[0039] Step S34: Calculate the associated reward value based on the associated campus hazard status map;

[0040] Step S35: Establish an experience pool for training the nested outer agent based on the nested outer state space, the cooperative action behavior, and the associated reward value.

[0041] In one optional embodiment, the potential hazard states of multiple inner-layer intelligent agents are associated based on the campus spatial topology to generate associated campus potential hazard states, including the following steps:

[0042] Based on the physical connectivity in the campus spatial topology and the campus functional dependencies, the potential danger states of the multiple inner-layer intelligent agents are physically and functionally associated to obtain associated potential dangers;

[0043] An interaction analysis is performed on the associated hazards, and the associated hazards are mapped into space based on the results of the interaction analysis to form a status map of associated campus hazards.

[0044] A second aspect of this invention provides a campus hazard identification system based on a dual-nested intelligent agent architecture, comprising:

[0045] The inner-layer intelligent agent module is used to receive multi-source heterogeneous data streams from the campus and construct several inner-layer intelligent agents based on the multi-source heterogeneous data streams from the campus; wherein, each inner-layer intelligent agent is configured to perform hazard analysis on the corresponding local campus scene and output building hazard data streams;

[0046] The nested observation environment module is used to integrate the building hazard data streams output by each inner-layer intelligent agent and fuse them with the multi-source heterogeneous data streams of the campus, and construct the nested observation environment based on the fused data;

[0047] The nested outer agent module is used to construct a nested outer agent, which performs hazard analysis on the global campus scene by observing the nested observation environment and outputs campus hazard data.

[0048] A third aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a campus hazard identification method based on a dual-nested intelligent agent architecture.

[0049] The fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a campus hazard identification method based on a dual-nested intelligent agent architecture.

[0050] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0051] Compared to existing multi-agent applications, this invention focuses on the identification of hidden dangers in the application field of campus hazard identification. It improves the agent model accordingly, enabling it to perform multi-level dynamic interaction of "individual-local-global". It can identify the real-time interaction between "personnel behavior" and "local environment" as well as between "different environmental units", thereby completing the identification of hidden dangers in the whole campus. Attached Figure Description

[0052] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:

[0053] Figure 1 This is a flowchart illustrating the campus hazard identification method based on a dual-nested intelligent agent architecture provided in Embodiment 1 of the present invention.

[0054] Figure 2 This is a schematic diagram of the campus hazard identification system based on a dual-nested intelligent agent architecture provided in Embodiment 2 of the present invention;

[0055] Figure 3 This is a schematic diagram of the structure of an electronic device provided in Embodiment 2 of the present invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0057] Example 1

[0058] This embodiment relates to the field of campus hazard identification technology, specifically including multi-agent systems, deep Q networks, motion capture and other technical means.

[0059] Figure 1 This is a flowchart illustrating the campus hazard identification method based on a dual-nested intelligent agent architecture provided in Embodiment 1 of the present invention. Figure 1 As shown, the campus hazard identification method based on a dual-nested intelligent agent architecture includes the following steps:

[0060] Step S1: Receive the campus multi-source heterogeneous data stream, and construct several inner-layer intelligent agents based on the campus multi-source heterogeneous data stream; wherein, each inner-layer intelligent agent is configured to perform hazard analysis on the corresponding local campus scene and output building hazard data stream;

[0061] Step S2: Integrate the building hazard data streams output by each inner-layer intelligent agent and fuse them with the campus multi-source heterogeneous data streams; construct a nested observation environment based on the fused data.

[0062] Step S3: Construct a nested outer intelligent agent. The nested outer intelligent agent performs hazard analysis on the global campus scene by observing the nested observation environment and outputs campus hazard data.

[0063] It should be noted that, firstly, a dedicated analysis process is set up at the local level. That is, corresponding inner-layer intelligent agents are established for each local campus scene (e.g., teaching buildings, laboratory buildings, playgrounds, etc.). These agents are used to continuously receive real-time monitoring information of specific areas and analyze the interaction between the behavioral characteristics of the active subjects and the conditions of the site. These processes transform the correlation between visible behavioral patterns and the functional state of the site, as well as the mutual influence between different elements, into a measurable risk development trajectory, thereby achieving a complete micro-dynamic depiction of the risk formation process. This step realizes the dynamic interaction between "individual" and "local".

[0064] Then, by integrating the building hazard data streams output by each inner-layer agent, a global correlation analysis process is established. The building hazard data streams obtained from each local analysis are integrated and merged with the overall campus operation status. This constructs a nested observation environment for nesting outer-layer agents. This nested observation environment can be used to identify potential correlation paths between different areas, explore the ripple effects of local anomalies on adjacent areas, and explore the chain reaction patterns between multiple areas, thereby providing a basis for realizing dynamic interaction between "local and global".

[0065] Finally, a nested outer agent is constructed. This outer agent is a comprehensive agent nested within several inner agents. It can observe not only the potential risks identified by the inner agents but also the entire campus environment (e.g., main roads, paths connecting different buildings, and trails) outside the inner agents, enabling global identification of the entire campus and allowing local conclusions to evolve synergistically with global assessments. When new risk characteristics are discovered, the global analysis results can guide local processes to adjust their observation priorities; simultaneously, the experience of local processes in handling unknown situations can provide a reference for global assessments. This continuous adaptive optimization allows the entire identification process to continuously improve as site conditions change, effectively enhancing the ability to identify new risks.

[0066] Compared to existing multi-agent applications, this embodiment focuses on the identification of hidden dangers in the application field of campus hazard identification. The dual-nested agent is improved to enable multi-level dynamic interaction of "individual-local-global". It can identify the real-time interaction between "personnel behavior" and "local environment" and between "different environmental units" in real time, thereby completing the identification of hidden dangers in the whole campus.

[0067] It should be emphasized that all data in this method was obtained legally and compliantly, and the obtained data will be used for the purpose of maintaining campus public safety.

[0068] In one optional embodiment, each inner-layer agent is configured to perform hazard analysis on a corresponding local campus scene, including the following steps:

[0069] Extract the corresponding local campus scene from the multi-source heterogeneous data stream of the campus as the inner observation environment;

[0070] The inner-layer agent obtains personnel status datasets and facility status datasets by observing the inner-layer observation environment, and establishes an inner-layer state space using these datasets;

[0071] Historical campus hazard data is acquired, and an action space is constructed based on the historical campus hazard data. The inner layer intelligent agent determines the action behavior from the action space based on the inner layer state space.

[0072] Execute the aforementioned action, and obtain the first and second campus hazard statuses before and after execution. Calculate the reward value based on the first and second campus hazard statuses and historical campus hazard data.

[0073] An experience pool for training the inner agent is established based on the inner state space, the action space, and the reward value.

[0074] The campus multi-source heterogeneous data stream includes various types of data such as video data obtained through surveillance, audio data obtained through microphones, and campus map modeling.

[0075] Extracting corresponding local campus scenes from multi-source heterogeneous data streams on campus as inner-layer observation environments involves extracting map modeling data and using the corresponding local campus scenes as inner-layer observation environments. For example, for an inner-layer agent in a teaching building, the building's corridor can be used as its corresponding inner-layer observation environment, enabling the identification of potential campus hazards in the corridor through the inner-layer agent.

[0076] In one optional embodiment, the personnel status dataset includes: personnel spatial distribution data, movement trajectory data, and behavioral pattern data; the facility status dataset includes: facility structural safety data, electrical facility safety data, fire protection facility safety data, and specific facility safety data.

[0077] The inner-layer agent acquires behavioral pattern data by observing the inner-layer observation environment, including the following steps:

[0078] The inner-layer intelligent agent acquires video surveillance data and audio surveillance data through data acquisition devices;

[0079] Human posture recognition is performed on the video surveillance data. If the recognized human posture involves interactive actions, audio recognition is performed on the sound surveillance data. The recognized audio is then spliced ​​with the recognized human posture to generate behavior pattern data.

[0080] The personnel status dataset includes: spatial distribution data, movement trajectory data, and behavioral pattern data. The personnel status dataset is used to identify potential hazards among personnel on campus, such as: overcrowding hazards, abnormal movement trajectory hazards (e.g., moving away from their assigned classroom during class), and pushing hazards.

[0081] The facility status dataset includes: structural safety data, electrical safety data, fire safety data, and specific facility safety data. The facility status dataset is used to identify potential campus hazards, such as deformed facilities posing a risk of punctures or cuts, and electrical malfunctions causing fire hazards.

[0082] It should be noted that the process by which an intelligent agent acquires its state by observing the environment is a perception-representation-update loop, and its core objective is to construct an internal representation of the environment to support decision-making.

[0083] The observation process can be either fully observable or partially observable. Fully observable refers to information that can be directly obtained through images, while partially observable refers to information that cannot be directly obtained through images or audio data.

[0084] Therefore, in this embodiment, the above-mentioned state dataset is divided into fully observable and partially observable data. The partially observable data needs to be further processed so that it can be used as the state for subsequent actions.

[0085] Behavioral pattern data is a portion of the observable data collected in this embodiment. In this embodiment, video surveillance data and audio data need to be fused to obtain behavioral patterns. Specifically, this includes: performing human posture recognition on video surveillance data; if the recognized human posture involves interactive actions, performing audio recognition on sound surveillance data; and splicing the recognized audio with the recognized human posture to generate behavioral pattern data.

[0086] Specifically, in this embodiment, the observation behavior of the inner agent is implemented by a visual branch and a visual-audio fusion branch. The visual branch is based on a hybrid CNN-Transformer architecture, which includes a convolutional network and a Transformer encoder. The convolutional network is responsible for local feature extraction, and the Transformer encoder is responsible for global relation modeling. The two are connected through a feature map-to-sequence transformation layer. The implementation process begins with the original image input, and after multiple convolutional operations, visual features from edges to objects are abstracted step by step. These spatial features are segmented into patch sequences and positional encodings are added. Then, they are fed into a multi-head self-attention layer to capture the long-range dependencies between different regions of the image. Then, important regions are dynamically weighted and highlighted through a spatial attention mechanism. Finally, the data is compressed into a representation vector through global pooling and output through a fully connected layer.

[0087] The visual-audio fusion branch begins with the parallel preprocessing and feature extraction of the raw multimodal data. Images are abstracted into spatial semantics through cascaded convolutional layers, while audio is converted into a spectrogram and its temporal dynamics are captured using a temporal model. Subsequently, a bidirectional cross-modal attention layer first expands the global audio features to the spatial dimension of the visual feature map, laying the foundation for position-by-position interaction. Then, a gating mechanism adaptively adjusts the contribution weights of the two modalities at each spatial position. Next, the bidirectional attention mechanism extracts cross-modal association information through query-key-value matching. Finally, global pooling converges the data into a unified and robust multimodal joint representation. This representation is then fed into a multi-task output head, simultaneously driving classification decisions, regression predictions, and generating an interpretable attention heatmap.

[0088] The inner state space at time t based on the above dataset The structure is as follows:

[0089] ;

[0090] In the above formula, This represents the spatial distribution of people at time t. This represents the movement status of people at time t. Indicates the status of personnel behavior. Indicates facility status. Indicates the status of electrical facilities. Indicates the status of fire protection facilities.

[0091] Obtain historical campus hazard data, and construct the action space at time t based on the historical campus hazard data. :

[0092] ;

[0093] in, This indicates the highest level of potential hazard to personnel. This indicates the highest level of potential hazard to the facility.

[0094] It should be noted that constructing the action space based on historical campus hazard data refers to classifying campus hazards into m and n levels, respectively, based on past campus hazard information, categorizing them into personnel hazards and facility hazards. The inner-layer agent determines its actions from the action space based on the inner-layer state space. This means determining the current level of personnel hazard by analyzing the spatial distribution, movement, and behavior of personnel within the inner-layer state space; and determining the current level of facility hazard by analyzing the status of facilities, electrical facilities, and fire protection facilities within the inner-layer state space.

[0095] The execution action refers to adjusting the campus hazard level to the corresponding hazard level and obtaining the corresponding historical campus hazard data under that hazard level. For example, adjusting the hazard level to... This indicates that there are many people in the corridor and that the facilities have slight deformation; therefore, historical campus hazard data that has occurred in the past should be retrieved.

[0096] In one optional embodiment, the reward value is calculated based on the first campus hazard status, the second campus hazard status, and historical campus hazard data, including the following steps:

[0097] The historical campus hazard data is used to construct reward weights for accuracy of identification, timeliness of identification, progress, and punishment.

[0098] The confidence level of the second campus hazard status is calculated to obtain the credibility. The credibility is then integrated with the recognition accuracy reward weight to serve as the recognition accuracy reward.

[0099] Determine the state time of the second campus hazard state, and concatenate the state time with the second campus hazard state to obtain state-time data;

[0100] Obtain the timeline of hazard development, compare the state-time data with the hazard data line to obtain the timeliness value, and integrate the timeliness value with the identification timeliness reward weight as the identification timeliness reward;

[0101] The scenario fitness value is determined by the first campus hazard status and the second campus hazard status, and the scenario fitness value is fused with the progress weight as a progress reward.

[0102] The error cost is determined based on the second campus hazard status, and the error cost is integrated with the penalty weight as the error penalty;

[0103] The reward value is calculated by taking the recognition accuracy reward, the recognition timeliness reward, and the progress reward as positive reward values ​​and the error penalty as a negative reward value.

[0104] In this embodiment, the reward value is determined based on identification accuracy, identification timeliness, scenario adaptability, and error cost. Specifically, firstly, weight benchmarks are set for different aspects based on historical data to provide a basis for subsequent evaluation. Regarding identification accuracy, the credibility calculation result of the current hazard status is combined with the accuracy weight constructed from historical data to form a quantitative reward for the degree of identification accuracy.

[0105] When assessing the timeliness of identification, the status of a hazard is combined with its corresponding discovery time to form status-time data. This data is then compared with the historical timeline of hazard development to derive a quantitative value for timeliness. This timeliness value is subsequently combined with a timeliness weight derived from historical experience to form a reward component for the speed of early warning.

[0106] Next, the adaptability to the new scenario is assessed by comparing the old and new potential risks, and this adaptability value is integrated with the progress weight supported by historical data as a reward for learning and adaptability.

[0107] Finally, the costs that may result from identification errors are estimated based on the current status of the potential hazards, and combined with historical penalty weights to form a penalty item.

[0108] By using the above-mentioned accuracy, timeliness, and progress rewards as positive incentives and error penalties as negative constraints, a balanced overall reward value that reflects multiple aspects of performance is calculated.

[0109] Specifically, the reward function at time t used to calculate the reward value. The structure is as follows:

[0110] ;

[0111] In the above formula, , , and These are the reward weights for recognition accuracy, recognition timeliness, progress, and penalty weights at time t, respectively. Let be the confidence level at time t. Let be the time-effect value at time t. Let be the scene fitness value at time t. Let t be the error cost at time t.

[0112] Furthermore, credibility The calculation is as follows:

[0113] ;

[0114] In the above formula, For video confidence, For audio confidence, For data integrity (the ratio of the number of key monitoring data items acquired to the total number of key data items that should be acquired). For environmental interference, the value range is [0,1). , , , These are the video coefficient, audio coefficient, integrity coefficient, and interference coefficient, respectively.

[0115] The reliability calculation integrates the quality and consistency of multi-source real-time data and eliminates the impact of environmental interference, thus providing a quantitative comprehensive evaluation of the reliability of the current identification results.

[0116] Timeliness The calculation is as follows:

[0117] ;

[0118] In the above formula, Rewards are based on historical baseline response times. This is the critical time point. To identify time, For the total development time of potential hazards, This is the time-dependent gain coefficient.

[0119] The scenario fitness value refers to the cosine similarity between the multi-dimensional feature vector of the current hidden danger and the historical case library.

[0120] Error cost refers to the penalty value that should be borne for identification errors by scaling the penalty base for underreporting and the penalty base for false reporting according to the relative proportion of the actual severity of the consequences and the scale of historical impact.

[0121] Finally, , , and the state at time t+1 The experience is integrated and placed into an experience pool. The inner-layer agent is trained using the experience in the experience pool. The trained inner-layer agent then outputs the corresponding data stream of building hazards.

[0122] In one optional embodiment, the building hazard data streams output by each inner-layer intelligent agent are integrated and fused with the multi-source heterogeneous data streams of the campus. A nested observation environment is constructed based on the fused data, including the following steps:

[0123] Step S21: Obtain the building hazard data stream output by each inner-layer agent, and based on the preset campus map data, map each inner-layer agent to a spatial node in the map data to generate a dynamic graph structure containing spatial association attributes; wherein, the node attributes of the dynamic graph structure at least include the building hazard data output by the corresponding inner-layer agent.

[0124] Step S22: Spatiotemporally align and fuse the dynamic graph structure with the campus multi-source heterogeneous data stream to construct a unified digital environmental space that integrates the physical environment and the status of potential hazards.

[0125] First, based on a pre-set campus map, each inner-layer intelligent agent is mapped as a node with spatial location attributes, forming a dynamic graph structure containing spatial relationships. Each node not only carries real-time hazard data for the corresponding building area, but also expresses the physical connectivity and functional correlation between areas through edge relationships in the graph structure. Subsequently, this dynamic graph is spatiotemporally aligned with multi-source data streams from the campus, and personnel flow patterns, environmental monitoring indicators, and hazard status are uniformly encoded to construct a digital environmental space that fully corresponds to the physical environment and risk situation, providing an observation foundation with complete spatiotemporal semantics for the upper-layer intelligent agents.

[0126] The dynamic graph structure visually presents the mutual influence paths of hazard states in different building areas; while the fusion of multi-source data allows the environmental space to simultaneously encompass objective physical states and subject behavioral characteristics, achieving a leap from static environmental modeling to dynamic interactive scenarios. The resulting nested observation environment not only reflects the current hazard distribution but also predicts the cross-regional transmission trend of risks through the evolution of the graph structure, providing a spatiotemporal analysis dimension for systemic hazard identification.

[0127] In one optional embodiment, the nested outer agent performs a hazard analysis of the global campus scene by observing the nested observation environment, including the following steps:

[0128] Step S31: The nested outer agent obtains the hidden danger state dataset of each inner agent by observing the nested observation environment, and establishes the nested outer state space using the dataset.

[0129] Step S32: Obtain historical campus hazard data, construct a collaborative action space based on the historical campus hazard data, and determine the collaborative action behavior from the collaborative action space based on the outer state space of the nested outer layer intelligent agent;

[0130] Step S33: Execute the collaborative action behavior, obtain the hidden danger status of multiple inner-layer intelligent agents after execution, and associate the hidden danger status of multiple inner-layer intelligent agents based on the campus space topology to generate an associated campus hidden danger status map.

[0131] Step S34: Calculate the associated reward value based on the associated campus hazard status map;

[0132] Step S35: Establish an experience pool for training the nested outer agent based on the nested outer state space, the cooperative action behavior, and the associated reward value.

[0133] Among them, the model building technology of nested outer agents is similar to that of inner agents, the difference lies in the state space, action space and reward.

[0134] Specifically, the nested outer state space at time t The structure is as follows:

[0135] ;

[0136] In the above formula, This represents the building hazard status of the k-th innermost agent. This represents the potential vulnerability status on the channel from the i-th inner agent to the k-th inner agent.

[0137] The hazard status on the channel from the i-th inner layer agent to the k-th inner layer agent is used to determine the possibility of hazard spread between local scenarios. For example, if any one of the i-th inner layer agent or the k-th inner layer agent has a fire hazard, and the hazard status on the channel from the i-th inner layer agent to the k-th inner layer agent is a spreadable state (i.e., there are combustibles on the channel), then there is a possibility of hazard spread. That is, the hazard status on the channel from the i-th inner layer agent to the k-th inner layer agent is obtained by extracting spreadable features from video data and audio data, including but not limited to: crowding level and combustibles.

[0138] Cooperative action behavior at time t The structure is as follows:

[0139] ;

[0140] In the above formula, Let M represent the level of the potential hazard spreading from the i-th inner layer agent to the k-th inner layer agent, and N represent the level of the potential hazard spreading from the personnel.

[0141] Implementing collaborative actions refers to determining the extent to which a potential hazard can spread.

[0142] In one optional embodiment, the potential hazard states of multiple inner-layer intelligent agents are associated based on the campus spatial topology to generate associated campus potential hazard states, including the following steps:

[0143] Based on the physical connectivity in the campus spatial topology and the campus functional dependencies, the potential danger states of the multiple inner-layer intelligent agents are physically and functionally associated to obtain associated potential dangers;

[0144] An interaction analysis is performed on the associated hazards, and the associated hazards are mapped into space based on the results of the interaction analysis to form a status map of associated campus hazards.

[0145] It should be noted that physical connectivity includes: the connectivity of passageways between buildings and the connection of pipelines and electrical networks; functional associations include: dependence on teaching functions and dependence on infrastructure supply. Based on the physical connectivity (such as corridors and pipelines between buildings) and functional dependencies (such as resource sharing between teaching buildings and laboratories and the temporal association of course arrangements) in the campus spatial topology, the independent hazard states output by multiple inner-layer agents are physically and functionally associated to identify cross-regional associated hazards. Subsequently, by analyzing the interactions between these associated hazards—including the coupling effects of enhancement, inhibition, and triggering among hazards—and mapping the analysis results onto the campus spatial structure, a map of associated campus hazard states is finally formed that can intuitively reflect the dynamic relationships, spatial distribution characteristics, and potential transmission paths among hazards. This map not only presents the cluster characteristics of current hazards but also reveals the networked structure and evolution trend of systemic risks.

[0146] Furthermore, the reward function at time t is used to calculate the associated reward value based on the associated campus hazard status. for:

[0147] ;

[0148] In the above formula, These are the weighting coefficients. The base reward value calculated by the i-th inner agent based on its own state (i.e., the reward value of the inner agent during its own training). The weighted centrality of the i-th inner-layer agent in the associated campus hazard state graph reflects its topological importance in the campus topology network; Let E be the global weight coefficient for the associated interactive rewards, and let E be the set of all edges in the associated campus hazard state graph. Let be the interaction strength between the i-th inner layer agent and the j-th inner layer agent. , These represent the weighting coefficients for physical and functional connections, respectively. This represents the strength of the physical connectivity between the i-th inner agent and the j-th inner agent. This represents the strength of the functional dependency between the i-th inner agent and the j-th inner agent.

[0149] Example 2

[0150] Figure 2 This is a schematic diagram of the campus hazard identification system based on a dual-nested intelligent agent architecture provided in Embodiment 2 of the present invention, as shown below. Figure 2 As shown, the campus hazard identification system based on a dual-nested intelligent agent architecture includes:

[0151] The inner-layer intelligent agent module is used to receive multi-source heterogeneous data streams from the campus and construct several inner-layer intelligent agents based on the multi-source heterogeneous data streams from the campus; wherein, each inner-layer intelligent agent is configured to perform hazard analysis on the corresponding local campus scene and output building hazard data streams;

[0152] The nested observation environment module is used to integrate the building hazard data streams output by each inner-layer intelligent agent and fuse them with the multi-source heterogeneous data streams of the campus, and construct the nested observation environment based on the fused data;

[0153] The nested outer agent module is used to construct a nested outer agent, which performs hazard analysis on the global campus scene by observing the nested observation environment and outputs campus hazard data.

[0154] Example 3

[0155] Figure 3 This is a schematic diagram of the structure of an electronic device provided in Embodiment 3 of the present invention, as shown below. Figure 3 As shown, the electronic device includes a processor 21, a memory 22, an input device 23, and an output device 24; the number of processors 21 in the computer device can be one or more. Figure 3 Taking a processor 21 as an example; the processor 21, memory 22, input device 23, and output device 24 in an electronic device can be connected via a bus or other means. Figure 3 For example, the connection is made via a bus; where processor 21 can be a graphics processor.

[0156] The memory 22, as a computer-readable storage medium, can be used to store software programs, computer-executable programs, and modules. The processor 21 executes various functional applications and data processing of the electronic device by running the software programs, instructions, and modules stored in the memory 22, thereby implementing the campus hazard identification method based on a dual-nested intelligent agent architecture as described in Embodiment 1.

[0157] The memory 22 may primarily include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a given function; the data storage area may store data created based on terminal usage. Furthermore, the memory 22 may include high-speed random access memory and non-volatile memory, such as at least one disk storage device, flash memory, or other non-volatile solid-state storage device. In some instances, the memory 22 may further include memory remotely located relative to the processor 21, which can be connected to the electronic device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks and combinations thereof, and industrial internet platforms.

[0158] Input device 23 can be used to receive user input such as ID and password. Output device 24 is used to output the network configuration page.

[0159] Example 4

[0160] Embodiment 4 of the present invention also provides a computer-readable storage medium, wherein the computer-executable instructions, when executed by a computer processor, are used to implement the campus hazard identification method based on a dual nested intelligent agent architecture as provided in Embodiment 1.

[0161] The storage medium containing computer-executable instructions provided in the embodiments of the present invention is not limited to the method operation provided in Embodiment 1, but can also execute related operations in the campus hazard identification method based on a dual nested intelligent agent architecture provided in any embodiment of the present invention.

[0162] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A campus hazard identification method based on a dual-nested intelligent agent architecture, characterized in that, Includes the following steps: Step S1: Receive the campus multi-source heterogeneous data stream, and construct several inner-layer intelligent agents based on the campus multi-source heterogeneous data stream; wherein, each inner-layer intelligent agent is configured to perform hazard analysis on the corresponding local campus scene and output building hazard data stream; Step S2: Integrate the building hazard data streams output by each inner-layer intelligent agent and fuse them with the campus multi-source heterogeneous data streams; construct a nested observation environment based on the fused data. Step S3: Construct a nested outer intelligent agent. The nested outer intelligent agent performs hazard analysis on the global campus scene by observing the nested observation environment and outputs campus hazard data.

2. The campus hazard identification method based on a dual-nested intelligent agent architecture according to claim 1, characterized in that, Each inner-layer intelligent agent is configured to perform hazard analysis on the corresponding local campus scene, including the following steps: Extract the corresponding local campus scene from the multi-source heterogeneous data stream of the campus as the inner observation environment; The inner-layer agent obtains personnel status datasets and facility status datasets by observing the inner-layer observation environment, and establishes an inner-layer state space using these datasets; Historical campus hazard data is acquired, and an action space is constructed based on the historical campus hazard data. The inner layer intelligent agent determines the action behavior from the action space based on the inner layer state space. Execute the aforementioned action, and obtain the first and second campus hazard statuses before and after execution. Calculate the reward value based on the first and second campus hazard statuses and historical campus hazard data. An experience pool for training the inner agent is established based on the inner state space, the action space, and the reward value.

3. The campus hazard identification method based on a dual-nested intelligent agent architecture according to claim 1, characterized in that, The personnel status dataset includes: personnel spatial distribution data, movement trajectory data, and behavior pattern data; the facility status dataset includes: facility structural safety data, electrical facility safety data, fire protection facility safety data, and specific facility safety data. The inner-layer agent acquires behavioral pattern data by observing the inner-layer observation environment, including the following steps: The inner-layer intelligent agent acquires video surveillance data and audio surveillance data through data acquisition devices; Human posture recognition is performed on the video surveillance data. If the recognized human posture involves interactive actions, audio recognition is performed on the sound surveillance data. The recognized audio is then spliced ​​with the recognized human posture to generate behavior pattern data.

4. The campus hazard identification method based on a dual-nested intelligent agent architecture according to claim 1, characterized in that, The reward value is calculated based on the first campus hazard status, the second campus hazard status, and historical campus hazard data, including the following steps: The historical campus hazard data is used to construct reward weights for accuracy of identification, timeliness of identification, progress, and punishment. The confidence level of the second campus hazard status is calculated to obtain the credibility. The credibility is then integrated with the recognition accuracy reward weight to serve as the recognition accuracy reward. Determine the state time of the second campus hazard state, and concatenate the state time with the second campus hazard state to obtain state-time data; Obtain the timeline of hazard development, compare the state-time data with the hazard data line to obtain the timeliness value, and integrate the timeliness value with the identification timeliness reward weight as the identification timeliness reward; The scenario fitness value is determined by the first campus hazard status and the second campus hazard status, and the scenario fitness value is fused with the progress weight as a progress reward. The error cost is determined based on the second campus hazard status, and the error cost is integrated with the penalty weight as the error penalty; The reward value is calculated by taking the recognition accuracy reward, the recognition timeliness reward, and the progress reward as positive reward values ​​and the error penalty as a negative reward value.

5. The campus hazard identification method based on a dual-nested intelligent agent architecture according to claim 1, characterized in that, The process involves integrating the building hazard data streams output by each inner-layer intelligent agent and fusing them with the multi-source heterogeneous data streams from the campus. Based on the fused data, a nested observation environment is constructed, including the following steps: Step S21: Obtain the building hazard data stream output by each inner-layer agent, and based on the preset campus map data, map each inner-layer agent to a spatial node in the map data to generate a dynamic graph structure containing spatial association attributes; wherein, the node attributes of the dynamic graph structure at least include the building hazard data output by the corresponding inner-layer agent. Step S22: Spatiotemporally align and fuse the dynamic graph structure with the campus multi-source heterogeneous data stream to construct a unified digital environmental space that integrates the physical environment and the status of potential hazards.

6. The campus hazard identification method based on a dual-nested intelligent agent architecture according to claim 1, characterized in that, The nested outer agent performs a hazard analysis of the global campus scene by observing the nested observation environment, including the following steps: Step S31: The nested outer agent obtains the hidden danger state dataset of each inner agent by observing the nested observation environment, and establishes the nested outer state space using the dataset. Step S32: Obtain historical campus hazard data, construct a collaborative action space based on the historical campus hazard data, and determine the collaborative action behavior from the collaborative action space based on the outer state space of the nested outer layer intelligent agent; Step S33: Execute the collaborative action behavior, obtain the hidden danger status of multiple inner-layer intelligent agents after execution, and associate the hidden danger status of multiple inner-layer intelligent agents based on the campus space topology to generate an associated campus hidden danger status map. Step S34: Calculate the associated reward value based on the associated campus hazard status map; Step S35: Establish an experience pool for training the nested outer agent based on the nested outer state space, the cooperative action behavior, and the associated reward value.

7. The campus hazard identification method based on a dual-nested intelligent agent architecture according to claim 1, characterized in that, Based on the campus spatial topology, the potential hazard states of multiple inner-layer intelligent agents are associated to generate associated campus potential hazard states, including the following steps: Based on the physical connectivity in the campus spatial topology and the campus functional dependencies, the potential danger states of the multiple inner-layer intelligent agents are physically and functionally associated to obtain associated potential dangers. An interaction analysis is performed on the associated hazards, and the associated hazards are mapped into space based on the results of the interaction analysis to form a status map of associated campus hazards.

8. A campus hazard identification system based on a dual-nested intelligent agent architecture, used to implement the campus hazard identification method based on a dual-nested intelligent agent architecture as described in any one of claims 1 to 7, characterized in that, include: The inner-layer intelligent agent module is used to receive multi-source heterogeneous data streams from the campus and construct several inner-layer intelligent agents based on the multi-source heterogeneous data streams from the campus; wherein, each inner-layer intelligent agent is configured to perform hazard analysis on the corresponding local campus scene and output building hazard data streams; The nested observation environment module is used to integrate the building hazard data streams output by each inner-layer intelligent agent and fuse them with the multi-source heterogeneous data streams of the campus, and construct the nested observation environment based on the fused data; The nested outer agent module is used to construct a nested outer agent, which performs hazard analysis on the global campus scene by observing the nested observation environment and outputs campus hazard data.

9. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the campus hazard identification method based on a dual nested intelligent agent architecture as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the campus hazard identification method based on a dual-nested intelligent agent architecture as described in any one of claims 1 to 7.

Images