Dynamically changing campus environment safety monitoring system and method

Abstract

The invention discloses a dynamically changing campus environment safety monitoring system and method, and relates to the technical field of intelligent campus monitoring, and the method comprises the steps: fusing a campus function map and a real-time data flow of a multi-region sensor through a graph neural network, and generating a dynamic weight instruction driven by a scene adaptation degree; constructing a structural causal model based on the dynamic weight instruction to determine causal association strength between the environmental parameters and the equipment state; the causal association strength is used as a condition to be input into the generation model, and an anti-fact risk trajectory conforming to the campus physical law is output; and fusing the anti-fact risk trajectory and the real-time data stream, and constructing a virtual task set to drive the cross-cycle migration optimization of the security monitoring model.

Description

Technical Field

[0001] This application relates to the field of smart campus monitoring technology, and more specifically, to a dynamic campus environment security monitoring system and method. Background Technology

[0002] Against the backdrop of the booming development of IoT technology, the construction of smart campuses has become an important direction for improving the efficiency of education management. Among them, campus security monitoring, as a core link in ensuring the personal and property safety of teachers and students, has received widespread attention. In existing technologies, IoT-based campus security monitoring systems typically achieve real-time monitoring of the campus environment by deploying video surveillance equipment, which to some extent improves the level of automation in security monitoring.

[0003] However, traditional monitoring methods have relatively simple processing for multi-source data fusion, often using fixed weights to integrate sensor data streams, making it difficult to flexibly cope with the complex and ever-changing environment and equipment operating status within the campus. Data collected by each sensor is often fragmented from information about functional areas of the campus, failing to achieve deep integration. This results in limited comprehensive analysis capabilities for security risks, easily leading to misjudgments or omissions, and failing to meet the accuracy and real-time requirements of smart campus security monitoring.

[0004] There is currently no effective solution to the above problems. Summary of the Invention

[0005] This application provides a dynamic campus environment security monitoring system and method to solve the above-mentioned technical problems.

[0006] This application provides a dynamically changing campus environment security monitoring system, including: The weight instruction generation module is used to generate dynamic weight instructions driven by scene adaptability by fusing the campus functional map and real-time data streams from multiple area sensors through a graph neural network. The correlation determination module is used to construct a structural causal model based on the dynamic weight instructions in order to determine the strength of the causal correlation between environmental parameters and equipment status. The risk trajectory output module is used to generate a model by taking the strength of the causal relationship as a conditional input and outputting a counterfactual risk trajectory that conforms to the physical laws of the campus. The migration optimization module is used to integrate the counterfactual risk trajectory with real-time data streams from multiple regional sensors to construct a virtual task set, thereby driving cross-cycle migration optimization of the safety monitoring model.

[0007] This application provides a method for dynamically changing campus environment security monitoring, including: By fusing real-time data streams from campus functional maps and multi-area sensors using graph neural networks, dynamic weight instructions driven by scene adaptability are generated. A structural causal model is constructed based on the dynamic weighting instructions to determine the strength of the causal relationship between environmental parameters and equipment status. The causal correlation strength is used as a conditional input to generate a model, which outputs a counterfactual risk trajectory that conforms to the physical laws of the campus. By integrating the counterfactual risk trajectory with the real-time data stream, a virtual task set is constructed to drive the cross-cycle migration optimization of the security monitoring model.

[0008] Furthermore, the generation of the dynamic weight instruction includes: The campus functional map is modeled in multiple regions using a spatiotemporal graph neural network. The laboratory is divided into a chemical storage area sub-map and an experimental operation area sub-map, and the dormitory is divided into an electricity monitoring area sub-map and a public activity area sub-map. For the chemical storage area subgraph, gas sensor concentration distribution data and ventilation equipment operation status data are extracted, and the correlation weight between gas diffusion and ventilation efficiency is calculated through the subgraph attention mechanism of the spatiotemporal graph neural network. For the experimental operation sub-graph, current harmonic data and temperature and humidity sensor data of the experimental equipment are extracted to construct a weighted evaluation model of the equipment operating status. For the dormitory electricity monitoring area sub-map, the air conditioner usage time period and circuit current characteristics are extracted, and the weight features of illegal electricity use patterns are identified through the spatiotemporal graph neural network; The weight matrices of each subgraph are input into the cross-regional fusion module. Combined with the spatial correlation features of the campus functional map, the global dynamic weight allocation scheme is determined through the graph pooling operation of the spatiotemporal graph neural network to generate dynamic weight instructions driven by scene adaptability.

[0009] Furthermore, the construction of the structural causal model based on the dynamic weight instructions includes: The dynamic weight instruction is used as a regional prior input to the structural causal model to construct a cross-regional causal graph. The cross-regional causal graph includes nodes and causal edges of the chemical storage area subgraph, the experimental operation area subgraph, and the electricity monitoring area subgraph. The spatiotemporal graph neural network is used to extract causal features from each subgraph, including: for the chemical storage area subgraph, extracting the spatiotemporal features of temperature, humidity, and gas concentration, and learning the causal mapping between chemical volatilization and ventilation efficiency through a multi-scale residual module; for the experimental operation area subgraph, extracting the correlation features between equipment current harmonics and temperature and humidity, and constructing a causal inference chain between equipment overheating and current mutations; for the electricity monitoring area subgraph, extracting the causal correlation model between load change time-series features and illegal electricity use patterns. When the dynamic weight of a subgraph exceeds the risk weight threshold, the intervention inference channel is activated in the cross-regional causal graph to generate a causal effect intensity matrix.

[0010] Furthermore, the step of using the strength of the causal relationship as a conditional input to generate a model and outputting a counterfactual risk trajectory that conforms to the physical laws of the campus includes: The causal relationship between environmental parameters and equipment status is encoded as a constraint and input into the generative model that integrates campus building physical parameters. Collect 3D point cloud data of the teaching building staircase and structural parameters of the roof corridor to construct a physical constraint library, and establish a joint constraint module of geometric safety and causal logic in the generated model; Output a counterfactual risk trajectory that conforms to the physical laws of the campus. The counterfactual risk trajectory includes hazard level labeling information. The spatial coordinates and time series of the counterfactual risk trajectory are jointly defined by the causal correlation strength and the physical parameters of the campus buildings.

[0011] Furthermore, the verification and dynamic correction of the counterfactual risk trajectory includes: A spatial collision constraint field is constructed based on the three-dimensional point cloud of the teaching building staircase. When the angle between the diffusion vector of the counterfactual risk trajectory and the normal vector of the load-bearing structure exceeds the safety threshold, the trajectory reconstruction mechanism is triggered. The propagation rate of the counterfactual risk trajectory is dynamically damped and corrected using the building structure elastic modulus parameters in the physical constraint library. Output counterfactual risk trajectories with building structure conflict markers, where the hazard level marking information is positively correlated with the intensity of building structure conflict.

[0012] Furthermore, the construction of the physical constraint library for the generated model also includes: A mapping table between crowd density and steel structure deformation rate is established in the rooftop corridor area. When the counterfactual risk trajectory involves the rooftop corridor area, the deformation rate threshold is used as a hard constraint condition for the spatial distribution of the trajectory. An embedded laboratory equipment current safety envelope is used to generate a trajectory bifurcation path for the equipment protection mechanism when the simulated current value in the counterfactual risk trajectory exceeds the laboratory equipment current safety envelope.

[0013] Furthermore, the process of fusing the counterfactual risk trajectory with the real-time data stream to construct a virtual task set to drive cross-cycle migration optimization of the security monitoring model includes: A bio-inspired counterfactual network is constructed, whose input layer receives the real-time data stream and whose output layer is connected to a spatiotemporal capsule network. When multi-region sensors detect the spatiotemporal coupling between a sudden change in population density in the target area and an abnormal signal from equipment in adjacent areas, the causal neuron module of the bio-inspired counterfactual network activates the matching risk patterns in the historical context memory bank. The real-time data stream is decoupled into entity capsules and relational capsules using the spatiotemporal capsule network, wherein the entity capsules encode the state characteristics of physical devices, and the relational capsules encode the interactions between devices; A set of virtual tasks with local intervention conditions is generated through a capsule routing protocol to drive cross-cycle migration optimization of the security monitoring model.

[0014] Furthermore, the construction and retrieval of the historical context memory bank includes: Establish a two-dimensional index architecture, where the first dimension is classified into scenario tags according to fire, structural collapse, and equipment cascading failure, and the second dimension is divided into time tags according to teaching cycle periods and holiday periods. Spatiotemporal encoding of historical risk scenarios is performed through neural radiation fields, including: spatial encoding binding to building structure point cloud coordinates, and temporal encoding associating with school calendar event timestamps; When the causal neuron module of the bio-inspired counterfactual network is activated, a matching memory fragment is retrieved in the two-dimensional indexing architecture based on the spatiotemporal coupling characteristics of the current sensor data stream. The retrieved memory fragments are input into the spatiotemporal capsule network for feature decoupling and recombination, and a set of memory retrieval instructions adapted to the current dynamic weights is output.

[0015] Furthermore, the updating of the historical context memory bank includes: After the security monitoring model completes cross-cycle transfer training, the error matrix between the virtual task prediction results and the real scene monitoring data is compared. When the trajectory localization error under a specific scene label exceeds the error threshold for three consecutive cycles, the capsule network feature reconstruction mechanism is triggered. The solid capsule is subjected to device state characteristic distillation to compress redundant current harmonic characteristics; Spatial association strength recalibration is performed on the relationship capsule, and device interaction weights are corrected based on the causal association strength of the structural causal model; The reconstructed capsule features are re-encoded into a spatiotemporal tensor using a neural radiation field, and the storage priority of the memory is updated according to a two-dimensional indexing architecture.

[0016] Based on the embodiments provided in this application, a graph neural network is used to fuse a campus functional map with real-time data streams from multiple sensor areas and generate dynamic weight instructions driven by scene adaptability. This changes the simple processing method of integrating sensor data streams with fixed weights in traditional methods. The campus functional map covers the functional characteristics of different areas within the campus. After fusion with the real-time sensor data streams, the weights can be dynamically adjusted according to the complex and ever-changing environment within the campus, making the system more flexible in adapting to the security monitoring needs of different scenarios. This effectively solves the problems of insufficient multi-source data fusion and difficulty in dealing with complex environments in traditional methods. A structural causal model is constructed based on the dynamic weight instructions to determine the strength of the causal relationship between environmental parameters and equipment status, achieving deep fusion of multi-source data. In traditional methods, the data from each sensor and the information of campus functional areas are in a fragmented state. However, this method, by mining the causal relationships between data, can more accurately analyze the intrinsic connection between changes in environmental parameters and equipment status, improve the comprehensive analysis capability of security risks, and reduce risk misjudgments or omissions caused by data fragmentation. The strength of causal relationships is used as a conditional input to generate the model, outputting counterfactual risk trajectories that conform to the physical laws of the campus, providing more forward-looking risk predictions for security monitoring. This counterfactual risk trajectory, generated based on the actual physical laws of the campus, can simulate the potential development paths of risks under different conditions. This allows the monitoring system to not only monitor the current state in real time but also predict potential security risks in advance, providing more comprehensive decision-making basis for security management. By integrating the counterfactual risk trajectory with real-time data streams to construct a virtual task set, the security monitoring model is driven to migrate and optimize across cycles, improving the model's long-term adaptability to campus security monitoring scenarios. In this way, the security monitoring model can continuously learn and optimize in different cycles, constantly adapting to the dynamic changes in the campus environment, thereby meeting the higher requirements of smart campuses for the accuracy and real-time performance of security monitoring. Attached Figure Description

[0017] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings: Figure 1 This is a structural diagram of an optional dynamically changing campus environment security monitoring system according to an embodiment of this application; Figure 2 This is a flowchart of an optional dynamically changing campus environment security monitoring method according to an embodiment of this application; Figure 3 This is a flowchart of an optional method for constructing a virtual task set according to an embodiment of this application.

[0018] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0020] Optionally, such as Figure 1 As shown, this application provides a dynamically changing campus environment security monitoring system, including: The weight instruction generation module 101 is used to generate dynamic weight instructions driven by scene adaptability by fusing the campus functional map and real-time data streams from multiple area sensors through a graph neural network. The correlation determination module 102 is used to construct a structural causal model based on dynamic weight instructions in order to determine the strength of the causal correlation between environmental parameters and equipment status. The risk trajectory output module 103 is used to generate a model by taking the strength of causal relationship as a conditional input and outputting a counterfactual risk trajectory that conforms to the physical laws of the campus. The migration optimization module 104 is used to integrate counterfactual risk trajectories and real-time data streams from multiple regional sensors to construct a virtual task set, thereby driving cross-cycle migration optimization of the safety monitoring model.

[0021] Optionally, such as Figure 2 As shown, this application provides a method for dynamically changing campus environmental security monitoring, including: S201 uses a graph neural network to fuse the campus functional map with real-time data streams from multiple area sensors to generate dynamic weight instructions driven by scene adaptability. The sensors include, but are not limited to, environmental monitoring, equipment status, and spatial structure sensors. Environmental monitoring sensors include: gas concentration sensors (for laboratory chemical storage areas, detecting ethanol, formaldehyde, etc., with an accuracy of ±5ppm); temperature and humidity sensors (for teaching building corridors, with a temperature measurement range of 0-50℃ and an accuracy of ±0.5℃) and rooftop corridors, with a humidity accuracy of ±2%); and crowd density sensors (for playgrounds and canteen entrances, based on infrared arrays, with a detection range of 0-5 people / m²).

[0022] Equipment status sensors include current harmonic sensors: experimental equipment (sampling frequency 10kHz, detecting harmonic distortion rate); circuit current sensors: dormitory power distribution boxes (range 0-100A, accuracy ±0.5%); ventilation equipment sensors: fan speed (0-3000rpm), valve opening (0-100%).

[0023] Spatial structure sensors include a 3D laser scanner for the teaching building staircase (point cloud density 10 points / cm², constructing a collision constraint field); steel structure strain gauges for the roof corridor (measuring deformation rate, resolution 0.01mm / s); and a building settlement monitoring instrument for the laboratory building foundation (accuracy ±0.1mm).

[0024] S202, constructs a structural causal model based on dynamic weight instructions to determine the strength of the causal relationship between environmental parameters and equipment status; S203 uses the strength of causal relationship as a conditional input to generate a model and outputs a counterfactual risk trajectory that conforms to the physical laws of the campus. S204 integrates counterfactual risk trajectories with real-time data streams to construct a set of virtual tasks, driving the cross-cycle migration and optimization of the security monitoring model.

[0025] In one embodiment, data fusion and dynamic weight generation are performed by fusing a campus functional map (marking the spatial topology of laboratories, corridors, and emergency exits) with multi-area sensor data via a graph neural network: a gas concentration sensor (Room 302, Building A) collects real-time ethanol concentration data (current value 180 ppm); a ventilation equipment status sensor provides feedback on fan speed (1200 rpm); and a corridor temperature and humidity sensor detects a temperature of 28°C and humidity of 65%. Dynamic weight instructions are generated for chemical spill scenarios: the gas concentration data weight is set to 0.7 (exceeding the conventional monitoring weight of 0.5), the ventilation status weight is 0.3, and a high-risk area monitoring mode is triggered.

[0026] Structural causal model construction: A causal model is constructed based on dynamic weights to determine the correlation strength between environmental parameters and equipment status: for every 1°C increase in temperature, the ethanol evaporation rate increases by 8% (causal strength coefficient 0.08); for every 10% decrease in ventilation volume, the gas concentration retention rate increases by 15% (causal strength coefficient 0.15). A causal chain of "increased temperature → accelerated evaporation → insufficient ventilation → excessive concentration" is generated, providing a logical basis for risk prediction.

[0027] Counterfactual risk trajectory generation: The strength of causal correlation is input into the generation model integrating building parameters: 3D point cloud data of the laboratory is collected (wall thickness 240mm, door and window location coordinates); physical parameters for ethanol diffusion are set (diffusion coefficient 0.15m² / s). The counterfactual risk trajectory is output: If the ventilation equipment malfunctions, the ethanol concentration will diffuse along the corridor to the stairwell after 30 minutes, reaching 20% ​​of the lower explosive limit at a distance of 10 meters from the leak source (hazard level marked in yellow).

[0028] Virtual task set-driven model optimization: By integrating trajectory and real-time data stream (current fan speed suddenly drops to 800 rpm), a virtual task set is constructed: simulating the scenario of "complete shutdown of ventilation equipment", generating 10 sets of diffusion trajectories at different time periods; driving the safety monitoring model to learn emergency ventilation strategies during the peak period of the final experiment (across cycles), increasing the alarm response weight from 0.6 to 0.8.

[0029] Based on the embodiments provided in this application, a graph neural network is used to fuse a campus functional map with real-time data streams from multiple sensor areas and generate dynamic weight instructions driven by scene adaptability. This changes the simple processing method of integrating sensor data streams with fixed weights in traditional methods. The campus functional map covers the functional characteristics of different areas within the campus. After fusion with the real-time sensor data streams, the weights can be dynamically adjusted according to the complex and ever-changing environment within the campus, making the system more flexible in adapting to the security monitoring needs of different scenarios. This effectively solves the problems of insufficient multi-source data fusion and difficulty in dealing with complex environments in traditional methods. A structural causal model is constructed based on the dynamic weight instructions to determine the strength of the causal relationship between environmental parameters and equipment status, achieving deep fusion of multi-source data. In traditional methods, the data from each sensor and the information of campus functional areas are in a fragmented state. However, this method, by mining the causal relationships between data, can more accurately analyze the intrinsic connection between changes in environmental parameters and equipment status, improve the comprehensive analysis capability of security risks, and reduce risk misjudgments or omissions caused by data fragmentation. The strength of causal relationships is used as a conditional input to generate the model, outputting counterfactual risk trajectories that conform to the physical laws of the campus, providing more forward-looking risk predictions for security monitoring. This counterfactual risk trajectory, generated based on the actual physical laws of the campus, can simulate the potential development paths of risks under different conditions. This allows the monitoring system to not only monitor the current state in real time but also predict potential security risks in advance, providing more comprehensive decision-making basis for security management. By integrating the counterfactual risk trajectory with real-time data streams to construct a virtual task set, the security monitoring model is driven to migrate and optimize across cycles, improving the model's long-term adaptability to campus security monitoring scenarios. In this way, the security monitoring model can continuously learn and optimize in different cycles, constantly adapting to the dynamic changes in the campus environment, thereby meeting the higher requirements of smart campuses for the accuracy and real-time performance of security monitoring.

[0030] Furthermore, the generation of dynamic weight instructions includes: Multi-region topological modeling of the campus functional map was carried out using a spatiotemporal graph neural network. The laboratory was divided into a chemical storage area sub-map and an experimental operation area sub-map, and the dormitory was divided into an electricity monitoring area sub-map and a public activity area sub-map. For the chemical storage area subgraph, gas sensor concentration distribution data and ventilation equipment operation status data are extracted, and the correlation weight between gas diffusion and ventilation efficiency is calculated through the subgraph attention mechanism of the spatiotemporal graph neural network. In this embodiment, the subgraph attention mechanism refers to assigning dynamic association weights to different subgraph data features in a spatiotemporal graph neural network to suppress interference from irrelevant information and strengthen the association of key risks.

[0031] For example, for the chemical storage area sub-map, a threshold angle between the ventilation vent direction and the gas diffusion direction is set: when the angle is <30°, the correlation weight between ventilation efficiency and gas diffusion increases from 0.5 to 0.7 (an increase of 40%); when the angle is ≥60°, the weight decreases to 0.3 (a decrease of 40%).

[0032] For the experimental operation subplot, the correlation weight threshold between equipment current harmonics and temperature and humidity is set to 0.6. When the harmonic distortion rate exceeds 15% and the temperature is >35℃, the weight is automatically increased to 0.8 (exceeding the threshold by 33%) to focus on the risk of equipment overheating.

[0033] For the experimental operation sub-graph, current harmonic data and temperature and humidity sensor data of the experimental equipment are extracted to construct a weighted evaluation model of the equipment operating status. For the dormitory electricity monitoring area submap, the air conditioner usage time period and circuit current characteristics are extracted, and the weight features of illegal electricity use patterns are identified through spatiotemporal graph neural network; In this embodiment, the identification of weight features of illegal electricity use patterns includes: analyzing the characteristics of dormitory electricity use periods and current waveforms through a spatiotemporal graph neural network to extract weight features of electricity use patterns that exceed safety regulations.

[0034] For example, time-based thresholds and current characteristics: After 11 PM is designated as a high-risk period for illegal electricity use (threshold time). If the current in a dormitory during this period exceeds 3A (rated current 2A, exceeding the threshold by 50%), the weight of the electricity use pattern increases from 0.4 to 0.7. If a continuous spike in the current waveform during air conditioner use is detected (exceeding the average by 1.5 times, the threshold multiple), it is determined to be an illegal appliance connection, and the weight increases to 0.8. Pattern feature extraction: A joint weighted feature of "air conditioner use period + current harmonic distortion rate > 10%" is constructed. When a dormitory simultaneously meets the conditions of current > 3A and harmonic distortion rate reaching 12% after 11 PM, the system marks it as "illegal use of electric kettle" with a weight of 0.9, triggering an alert.

[0035] The weight matrices of each subgraph are input into the cross-regional fusion module. Combined with the spatial correlation features of the campus functional map, the global dynamic weight allocation scheme is determined through graph pooling operation of the spatiotemporal graph neural network to generate dynamic weight instructions driven by scene adaptability.

[0036] In this embodiment, the graph pooling operation includes: fusing the local weight matrices of each subgraph with the spatial correlation features of the campus functional map, and extracting the global dynamic weight allocation scheme through dimensionality reduction.

[0037] For example, the weight matrix of the laboratory subgraph (gas concentration weight 0.7, ventilation status 0.3) and the weight matrix of the dormitory subgraph (electrical safety weight 0.6, personnel activity 0.4) are input into the cross-regional fusion module; combined with the spatial association feature of "the laboratory and dormitory are 50 meters apart and share the same ventilation system" in the functional graph, the weight of the impact of laboratory gas leakage on the dormitory is set to 0.2 through graph pooling operation (the original local weight was 0 when cross-regional association was not considered), and a global weight scheme is generated: laboratory building monitoring weight 0.8, dormitory area 0.2 (during class break).

[0038] Based on the embodiments provided in this application, the dynamic weight instruction generation process is refined. A spatiotemporal graph neural network is used to perform multi-regional topological modeling of the campus functional map, and subgraphs are constructed for different functional areas. This solves the problem that traditional monitoring methods cannot perform refined monitoring tailored to the characteristics of different functional areas on campus. By dividing different functional areas such as laboratories and dormitories into subgraphs and extracting specific data for each subgraph (such as gas sensor concentration distribution data and ventilation equipment operating status data in chemical storage areas), combined with the subgraph attention mechanism and graph pooling operation of the spatiotemporal graph neural network, a more accurate global dynamic weight allocation scheme can be generated. This allows the system to dynamically adjust the monitoring focus and weight according to the actual situation of different areas, improving its adaptability to the complex and ever-changing environment on campus, achieving deeper integration of multi-source data, and enhancing the system's ability to accurately identify safety risks.

[0039] Furthermore, a structural causal model is constructed based on dynamic weight instructions, including: Using dynamic weight instructions as regional prior input structural causal models, a cross-regional causal graph is constructed. The cross-regional causal graph includes nodes and causal edges of the chemical storage area subgraph, the experimental operation area subgraph, and the electricity monitoring area subgraph. The spatiotemporal graph neural network is used to extract causal features of each subgraph, including: for the chemical storage area subgraph, extracting the spatiotemporal features of temperature, humidity and gas concentration, and learning the causal mapping between chemical volatilization and ventilation efficiency through multi-scale residual modules; for the experimental operation area subgraph, extracting the correlation features between equipment current harmonics and temperature and humidity, and constructing a causal inference chain between equipment overheating and current mutation; for the electricity monitoring area subgraph, extracting the causal correlation model between load change time series features and illegal electricity use patterns. When the dynamic weight of a subgraph exceeds the risk weight threshold, the intervention inference channel is activated in the cross-regional causal graph to generate a causal effect intensity matrix.

[0040] In one specific implementation, the causal effect strength of the structural causal model is calculated based on the following formula:

[0041] in, This represents the causal effect strength between subgraph i and subgraph j, and is also an element of the causal effect strength matrix. It is used to represent the degree of influence of the state change of subgraph i on subgraph j, such as the risk transmission strength of abnormal gas concentration in the chemical storage area to the experimental operation area, through the Sigmoid function (i.e., ...). After normalization, the value is in the range [0,1]; i, It is a subgraph index, for example, i=1 represents the chemical storage area subgraph, j=2 represents the experimental operation area subgraph, used to uniquely identify subgraphs of different functional areas and to construct cross-regional causal graph nodes; These are the dynamic weights of subgraphs i and j, generated by the subgraph attention mechanism of the spatiotemporal graph neural network, reflecting the risk monitoring priority of the subgraph in the current scenario, such as when the gas concentration in a chemical storage area exceeds the standard. ; , These are the causal feature strengths of subgraphs i and j, extracted through a spatiotemporal graph neural network, used to determine the causal relationship between environmental parameters and equipment status within the subgraph, such as in a chemical storage area. This represents the strength of the causal mapping between temperature and humidity changes and the gas evaporation rate, and its value ranges from [0,1]. It is the k-th type of region association factor, representing the degree of spatial or functional logical association between subgraphs. Spatially adjacent subgraphs Sub-graphs sharing the same ventilation system ; This represents the number of categories of regional association factors; This is the quantified value of the k-th association between subgraphs i and j, ranging from [0,1]. It comes from the spatial association features of the campus functional map, such as the association degree of the ventilation system between the laboratory subgraph and the corridor subgraph. .

[0042] In this embodiment, the risk weight threshold is used to quantify the priority of different safety risks in campus buildings. By setting numerical boundaries, risk characteristics are transformed into decision-making level indicators. For example, weights are assigned based on causal features extracted by a spatiotemporal graph neural network (such as illegal electricity use and structural deformation), and an early warning is triggered if the threshold is exceeded.

[0043] For example, risk weight thresholds include: Electricity load risk threshold: When the cumulative power weight of electrical equipment in a certain area exceeds 0.8 (maximum value 1.0), it is judged as high risk. For example, when three 1500W appliances are connected to a single socket in the dormitory area at the same time, the weight is calculated to be 0.85, triggering an overload warning. Structural deformation risk threshold: The displacement weight threshold of the load-bearing structure of the connecting corridor is set to 0.6. If a node is monitored to have a settlement of 0.3 mm in 24 hours (corresponding to a weight model calculation value of 0.65), it is marked as a structural hazard.

[0044] Based on the embodiments provided in this application, this application further improves the process of constructing a structural causal model based on dynamic weight instructions. It uses dynamic weight instructions as regional prior inputs to the structural causal model and constructs a cross-regional causal graph, changing the traditional method's independent analysis of data from each region and realizing cross-regional data correlation analysis. By extracting the causal features of each subgraph, it can uncover the inherent causal relationships between different environmental parameters and equipment states, such as the causal mapping between chemical volatilization and ventilation efficiency, and the causal inference chain between equipment overheating and current surges. When the dynamic weight of a subgraph exceeds the risk weight threshold, the intervention inference channel is activated and a causal effect intensity matrix is ​​generated, enabling the system to intervene and issue early warnings in the early stages of risk occurrence, improving the system's risk prediction and response capabilities, and reducing the probability of safety accidents.

[0045] Furthermore, by using the strength of causal correlation as a conditional input to generate a model, the output of counterfactual risk trajectories that conform to the physical laws of the campus include: The causal relationship between environmental parameters and equipment status is encoded as a constraint and input into the generative model that integrates campus building physical parameters. Collect 3D point cloud data of the teaching building staircase and structural parameters of the roof corridor to construct a physical constraint library, and establish a joint constraint module of geometric safety and causal logic in the generative model; In this embodiment, the 3D point cloud of the teaching building staircase is acquired through non-contact laser scanning or UAV photogrammetry technology to obtain data such as the spatial coordinates and material texture of the connecting corridor; the structural parameters of the roof connecting corridor (such as span and load-bearing beam cross-sectional dimensions) are retrieved through actual measurement or Building Information Modeling (BIM); the hazard level labeling is based on the acquired data, using numbers or color codes to classify the risk levels. The structural parameters of the roof connecting corridor include geometric data such as step height difference, roof slope, and drainage outlet distribution, as well as parameters such as shoe sole friction coefficient and equipment material properties.

[0046] For example, the acquisition methods are as follows: Laser scanning: using non-contact equipment to perform a surround scan of the roof truss of the connecting corridor with millimeter-level precision, acquiring millions of point cloud data per second to generate a 3D model; UAV photogrammetry: using a UAV to photograph the facade of the connecting corridor at a preset flight altitude and overlap rate, and converting the images into point clouds.

[0047] Hazard level labeling: A numerical rating system of 1-5 is used. For example, if the rusted area of ​​the corridor railing accounts for 30%, it is labeled as level 3 (yellow); if the rusted area exceeds 50%, it is labeled as level 4 (orange). The specific location is marked with a symbol in the BIM model.

[0048] Output counterfactual risk trajectories that conform to the physical laws of the campus. The counterfactual risk trajectories include hazard level labeling information. The spatial coordinates and time series of the counterfactual risk trajectories are jointly defined by the causal relationship strength and the physical parameters of the campus buildings.

[0049] Based on the embodiments provided in this application, this application improves the generation process of counterfactual risk trajectories by encoding causal correlation strength as a constraint condition and inputting it into a generation model that integrates campus building physical parameters. Simultaneously, it collects 3D point clouds of teaching building staircases and structural parameters of rooftop corridors to construct a physical constraint library, establishing a joint constraint module for geometric safety and causal logic. This improvement ensures that the generated counterfactual risk trajectory not only considers the causal correlation between environmental parameters and equipment status but also incorporates the physical structural parameters of campus buildings, such as 3D point clouds of staircases and structural parameters of rooftop corridors, ensuring that the counterfactual risk trajectory conforms to the physical laws of the campus. The output counterfactual risk trajectory includes hazard level labeling information, and its spatial coordinates and time series are jointly defined by the causal correlation strength and campus building physical parameters, providing more accurate risk predictions that better reflect the actual physical environment for safety monitoring. This allows safety management personnel to more intuitively understand the development trend and potential impact range of risks.

[0050] Furthermore, the verification and dynamic correction of counterfactual risk trajectories include: A spatial collision constraint field is constructed based on the 3D point cloud of the teaching building staircase. When the angle between the diffusion vector of the counterfactual risk trajectory and the normal vector of the load-bearing structure exceeds the safety threshold, the trajectory reconstruction mechanism is triggered. In this embodiment, the safety threshold is used to define the safety boundary of the building structure or equipment operating status. When the monitoring data exceeds the threshold, a risk assessment or emergency response is initiated.

[0051] For example, the safety threshold includes a load-bearing structure pressure threshold and a deformation rate threshold. Specifically, the load-bearing structure pressure threshold is set at 500 kg / m² for the main beam of the connecting corridor. A level one alarm is triggered when the real-time monitored pressure reaches 550 kg / m² (exceeding the threshold by 10%), and a level two alarm is triggered when it reaches 600 kg / m². Deformation rate threshold: The safe threshold for the deformation rate of the connecting corridor node is 0.5 mm / h. If the deformation reaches 0.7 mm within 1 hour (exceeding the threshold by 40%), the system will automatically mark it as an emergency risk.

[0052] In this embodiment, the included angle is determined through three-dimensional spatial vector operations. The diffusion vector represents the direction of risk propagation, and the normal vector of the load-bearing structure represents the perpendicular direction of the structural stress surface. The size of the included angle reflects the degree of impact of the risk on the structure. For example, when a crack occurs in a column of the connecting corridor, if the angle between the risk diffusion direction simulated in the counterfactual simulation and the perpendicular direction of the column's load-bearing surface is greater than 90°, it indicates that the risk diffusion direction and the structural stress surface are acting in opposite directions, and reinforcement should be prioritized.

[0053] The propagation rate of counterfactual risk trajectories is dynamically damped and corrected using the building structure elastic modulus parameters in the physical constraint library. Dynamic damping correction involves adjusting the structural damping coefficient based on the material's elastic modulus parameter to optimize vibration response and reduce the risk of resonance. For example, when a steel structure of a connecting corridor vibrates due to dense pedestrian traffic, the system dynamically adjusts the damping ratio according to material properties. The original damping ratio was 0.02, which was increased to 0.04 after correction, thus improving energy dissipation.

[0054] Output counterfactual risk trajectories with building structure conflict markers, where the hazard level marking information is positively correlated with the intensity of building structure conflict.

[0055] Building structural conflict marking involves using algorithms to identify spatial interference between structural components during design or construction and visually marking conflict points. For example, the designed distance between a connecting corridor staircase handrail and a fire pipe should be 30cm. When the actual distance is detected to be only 15cm, the system automatically marks the conflict location in red and generates a report indicating "insufficient horizontal distance".

[0056] In one specific implementation, the propagation rate of the counterfactual risk trajectory is dynamically damped and corrected based on the following formula:

[0057] in, It is the risk trajectory propagation rate after the building structure is corrected, that is, the actual risk diffusion speed after dynamic damping correction, such as the propagation rate of fire smoke after being blocked by the wall, and the unit is m / s. It is the original propagation rate output by the generative model, which comes from the initial calculation results of the counterfactual risk trajectory, such as the theoretical diffusion rate of a chemical leak in the absence of obstruction, and can take the value v=0.8m / s; It is the angle between the risk diffusion vector and the normal vector of the load-bearing structure, used to determine whether the risk trajectory conflicts with the building structure. A 90-degree angle indicates that the diffusion direction is perpendicular to the load-bearing surface. When the angle is greater than 90 degrees, the trajectory reconstruction mechanism will be triggered, with the value range being [0°, 180°]. It is the normalized value of the elastic modulus of building materials, which comes from the physical constraint library. For example, E=2.1 for steel structure corresponds to the actual elastic modulus of 2.1×105MPa, reflecting the damping characteristics of the structure to risk propagation. It is a structural conflict type index, corresponding to different building structures or obstacles. m=1 represents a load-bearing wall, and m=2 represents a fire door. The number of structural conflict types; It is the influence coefficient of the m-th structural conflict, ranging from [0,1], representing the damping capacity of the structure to the propagation of risk. (Fire door) A value of 0.3 means that the diffusion rate can be reduced by 30%; The intensity of the m-th structural conflict, ranging from [0,1], is positively correlated with the hazard level label. The risk trajectory penetrates three fire doors. It is 0.7.

[0058] Based on the embodiments provided in this application, this application further improves the verification and dynamic correction process of counterfactual risk trajectories. A spatial collision constraint field is constructed based on the 3D point cloud of the teaching building staircase. When the angle between the diffusion vector of the counterfactual risk trajectory and the normal vector of the load-bearing structure exceeds a safety threshold, a trajectory reconstruction mechanism is triggered, ensuring that the predicted risk trajectory does not conflict with the load-bearing structure of the campus building, thus improving the accuracy of risk prediction. The propagation rate of the counterfactual risk trajectory is dynamically damped and corrected using the building structure elastic modulus parameters in the physical constraint library, making the propagation rate of the risk trajectory more consistent with actual physical laws. The output is a counterfactual risk trajectory with building structure conflict markers, and the hazard level labeling information is positively correlated with the building structure conflict intensity, enabling safety management personnel to more clearly understand the relationship between risks and building structures, providing a basis for formulating more reasonable safety response measures.

[0059] Furthermore, the construction of the physical constraint library for the generative model also includes: A mapping table between crowd density and steel structure deformation rate is established in the rooftop corridor area. When the counterfactual risk trajectory involves the rooftop corridor area, the deformation rate threshold is used as a hard constraint condition for the spatial distribution of the trajectory. The deformation rate mapping table is used to map structural deformation rate data to risk levels and corresponding countermeasures, establishing a "rate-risk-response" correspondence to facilitate rapid decision-making. For example, a deformation rate ≤ 0.1 mm / h is considered low risk, requiring routine monitoring; a deformation rate > 1.0 mm / h is considered an emergency risk, requiring immediate evacuation and reinforcement.

[0060] Embedded with the laboratory equipment current safety envelope, when the simulated current value in the counterfactual risk trajectory exceeds the laboratory equipment current safety envelope, a trajectory bifurcation path for the equipment protection mechanism is generated.

[0061] The simulated current value refers to the current value of electrical equipment in the connecting corridor area, which is monitored and calculated in real time by sensors. When the simulated value exceeds the safety threshold, it is judged as a "breakthrough" and triggers a leakage or overload warning. For example, the total rated current of the lighting and monitoring equipment in the connecting corridor is 10A, and the safety threshold is set at 12A. When the simulated current value is calculated to be 15A during a certain period (exceeding the threshold by 25%), the system judges it as a current simulation breakthrough and immediately cuts off unnecessary power.

[0062] Based on the embodiments provided in this application, this application further improves the construction process of the physical constraint library for the generative model. A mapping table between crowd density and steel structure deformation rate is established in the rooftop corridor area. When the counterfactual risk trajectory involves the rooftop corridor area, the deformation rate threshold is used as a hard constraint condition for the spatial distribution of the trajectory, ensuring that the prediction of the risk trajectory considers the special physical characteristics and safety requirements of the rooftop corridor area. A laboratory equipment current safety envelope is embedded. When the simulated current value in the counterfactual risk trajectory exceeds the laboratory equipment current safety envelope, a trajectory bifurcation path for the equipment protection mechanism is generated. This allows the system to predict equipment failure risks in advance and generate corresponding protection measure paths, improving the safety and reliability of laboratory equipment. These improvements further enhance the physical rationality and practicality of the counterfactual risk trajectory, providing a more comprehensive and accurate risk prediction for campus security monitoring.

[0063] Furthermore, such as Figure 3 As shown, by integrating counterfactual risk trajectories with real-time data streams, a virtual task set is constructed to drive cross-cycle migration optimization of the security monitoring model, including: S301 constructs a bio-inspired counterfactual network, whose input layer receives real-time data streams and whose output layer connects to a spatiotemporal capsule network. S302, when the multi-area sensor detects the spatiotemporal coupling between the sudden change signal of personnel density in the target area and the abnormal signal of equipment in the adjacent area, the risk pattern matched in the historical scenario memory bank is activated through the causal neuron module of the bio-inspired counterfactual network. In this embodiment, adjacent areas refer to areas that are directly related to the target area in terms of spatial location or functional logic. The determination criteria include: spatial adjacency: straight-line distance ≤ 10 meters or vertical relationship between adjacent floors; functional adjacency: areas that share the same circuit, ventilation system or emergency passage.

[0064] For example, spatially adjacent: Room 302 of Building A (target area) is spatially adjacent to Room 301 and the corridor, with a straight-line distance of less than 10 meters; functionally adjacent: The electrical distribution box on a certain floor of the dormitory area (target area) is functionally adjacent to all dormitories on that floor because they share the same circuit system.

[0065] S303 uses a spatiotemporal capsule network to decouple real-time data streams into entity capsules and relational capsules. Entity capsules encode the state characteristics of physical devices, while relational capsules encode the interactions between devices. In this embodiment, the physical device status characteristics encoded by the physical capsule include real-time operating parameters and status attributes, specifically including: operating parameters: quantifiable indicators such as current, voltage, temperature, humidity, and rotation speed; status attributes: qualitative characteristics such as switch status, fault codes, and maintenance cycles. For example, for experimental equipment: encoded current value (e.g., 10A), harmonic distortion rate (12%), and equipment temperature (55℃); for ventilation systems: encoded fan speed (1200rpm), valve opening (70%), and operating mode (automatic / manual).

[0066] The interactions between devices in relational capsule coding include physical connections or logical associations between devices, specifically: physical connections: series / parallel circuit relationships, ventilation duct topology; logical associations: device linkage rules (such as fire alarm → sprinkler activation), risk transmission paths.

[0067] For example, in the electrical system: coding the parallel connection between the dormitory air conditioner and water heater (sharing the same socket); in the fire protection system: coding the linkage logic between the smoke sensor and emergency lighting (smoke concentration > 5% LEL → lighting turns on).

[0068] S304 generates a set of virtual tasks with local intervention conditions through the capsule routing protocol to drive the cross-cycle migration optimization of the security monitoring model.

[0069] Driving cross-cycle migration optimization can enable security monitoring models to adapt to cyclical changes on campus and improve the accuracy of long-term monitoring, including: time cycles: differences in time periods such as semester / holiday, weekday / weekend, and day / night; scenario cycles: changes in personnel and equipment patterns during exam weeks / teaching weeks and activity days / regular days.

[0070] In one specific implementation, the loss function for cross-cycle migration optimization of the security monitoring model is:

[0071] in, It is the total loss of cross-cycle migration optimization, used to evaluate the effect of cross-cycle migration optimization of the safety monitoring model. The smaller the value, the better the model adapts to the risk patterns of different cycles. It is a multi-objective optimization index. It is a time dimension loss, which quantifies the difference in risk patterns between the current cycle and historical cycles. For example, the prediction error caused by the difference in equipment operation patterns between exam week and regular week is obtained by comparing the mean square error of real-time data stream and data of the same period in historical scenario memory bank. It is the spatial dimension loss, which reflects the spatiotemporal coupling consistency error of multi-region sensor data. For example, when the population density in the target area suddenly increases, is the correlation of abnormal signals from adjacent equipment reasonable? , , These are balance coefficients, and their sum is 1. They are used to adjust the optimization priority of time dimension, space dimension and weight change terms. For example, λt is 0.6 during the semester transition, focusing on the adaptation of time cycle. It is a device index used to uniquely identify physical devices on campus. =1 can represent laboratory ventilation equipment; It is the first The interaction weights of a device in the current period t and the previous period t-1, such as the association weights between a dormitory air conditioner and its circuit, during the semester. ; It is a time period index, representing different cycles of campus operation. t=1 is the first week of the teaching cycle, and t=16 is the end of the semester cycle. It is used to distinguish model training and evaluation in different time scenarios.

[0072] Based on the embodiments provided in this application, this application improves the process of constructing a virtual task set by fusing counterfactual risk trajectories and real-time data streams. A bio-inspired counterfactual network is constructed, with its input layer receiving the real-time data stream and its output layer connected to a spatiotemporal capsule network. When multiple area sensors detect the spatiotemporal coupling between a sudden change in personnel density in the target area and an abnormal signal from equipment in adjacent areas, the causal neuron module of the bio-inspired counterfactual network activates the matching risk patterns in the historical scenario memory bank, achieving effective utilization of historical risk patterns. The spatiotemporal capsule network decouples the real-time data stream into entity capsules and relational capsules, respectively encoding the physical device state characteristics and inter-device interactions, thus more clearly representing the state and relationships of equipment and the environment within the campus. A virtual task set with local intervention conditions is generated through a capsule routing protocol to drive cross-cycle migration optimization of the security monitoring model. This allows the model to dynamically adjust and optimize according to the actual situation in different cycles, improving the model's adaptability and accuracy to campus security monitoring scenarios.

[0073] Furthermore, the construction and retrieval of historical context memory banks include: Establish a two-dimensional index architecture, where the first dimension is classified into scenario tags according to fire, structural collapse, and equipment cascading failure, and the second dimension is divided into time tags according to teaching cycle periods and holiday periods. Spatiotemporal encoding of historical risk scenarios is performed through neural radiation fields, including: spatial encoding binding to building structure point cloud coordinates, and temporal encoding associating with school calendar event timestamps; When the causal neuron module of the bio-inspired counterfactual network is activated, matching memory fragments are retrieved in a two-dimensional indexing architecture based on the spatiotemporal coupling characteristics of the current sensor data stream. Spatiotemporal coupling characteristics include the correlation characteristics of sensor data in the time and spatial dimensions, specifically including: temporal coupling: continuous anomalies in the time series of data within the same region (such as sudden changes in temperature and humidity within 10 minutes); spatial coupling: coordinated anomalies in data from adjacent regions at the same point in time (such as equipment failure in area A + sudden increase in population density in area B).

[0074] For example, in a laboratory fire scenario: time coupling: within 5 minutes, the temperature sensor in a laboratory rises from 25°C to 60°C and the humidity drops from 60% to 30%; spatial coupling: at the same time as the gas concentration sensor in the laboratory alarms, the smoke sensor in the adjacent corridor detects an increase in particle concentration.

[0075] The retrieved memory fragments are input into the spatiotemporal capsule network for feature decoupling and recombination, and the output memory retrieval instruction set is adapted to the current dynamic weights.

[0076] Based on the embodiments provided in this application, this application improves the construction and retrieval process of the historical scenario memory bank, establishing a dual-dimensional index architecture to classify and divide historical risk scenarios according to scenario tags and time tags, making the organization of the historical scenario memory bank more reasonable and efficient. Historical risk scenarios are spatiotemporally encoded using neural radiation fields, binding spatial codes to building structure point cloud coordinates and temporal codes to calendar event timestamps, achieving accurate representation and storage of historical risk scenarios. When the causal neuron module of the bio-inspired counterfactual network is activated, based on the spatiotemporal coupling characteristics of the current sensor data stream, matching memory fragments are retrieved in the dual-dimensional index architecture. The retrieved memory fragments are then input into a spatiotemporal capsule network for feature decoupling and recombination, outputting a memory retrieval instruction set adapted to the current dynamic weights. This improvement enables the system to quickly and accurately find historical risk patterns matching the current scenario and adjust and apply them according to the current dynamic weights, improving the system's ability to respond to similar risk scenarios and its decision-making efficiency.

[0077] Furthermore, the updating of the historical context memory bank includes: After the security monitoring model completes cross-cycle transfer training, the error matrix is ​​compared between the prediction results of the virtual task and the monitoring data of the real scene. When the trajectory localization error under a specific scene label exceeds the error threshold for three consecutive cycles, the capsule network feature reconstruction mechanism is triggered. In this embodiment, specific scenario tags include classification identifiers for campus safety risk types, used for quick retrieval of historical risk patterns. These specifically include: risk type: fire, equipment failure, structural collapse, illegal electricity use, etc.; scenario attributes: area (laboratory / dormitory), time period (teaching / holiday), environment (high temperature / humidity). For example, tag 1: "Laboratory - Teaching period - Equipment overload"; tag 2: "Dormitory - Holiday period - Illegal electricity use".

[0078] The length of each cycle is a time unit set according to the campus's operational patterns, used for model training and evaluation, including: long cycles: 1 semester (approximately 16 weeks), 1 academic year; short cycles: 1 day, 1 week, 1 teaching day (8:00-22:00). For example, long cycles: using semesters as units, comparing equipment failure modes between the spring and autumn semesters; short cycles: using weeks as units, analyzing the probability of abnormal laboratory equipment startup on Monday mornings.

[0079] Error thresholds are used to measure the critical value of the deviation between the model's prediction results and the actual data. When the threshold is exceeded, model optimization is triggered. Specifically, these include: trajectory positioning error: the deviation between the counterfactual trajectory prediction location and the actual risk location; risk level error: the difference between the risk level determined by the model and the actual level.

[0080] For example, trajectory positioning error thresholds: Teaching building scenario: allowable error ≤ 3 meters, if the trajectory prediction deviation of a fire reaches 4 meters (exceeding the threshold by 33%), optimization is triggered; Laboratory scenario: allowable error ≤ 1.5 meters, if the trajectory deviation of a chemical leak reaches 2 meters (exceeding the threshold by 33%), optimization is triggered.

[0081] Risk level error thresholds: Equipment failure scenario: Allowable level deviation ≤ 1 level (e.g., predicted level 2 risk, actual level 3), if the deviation reaches level 2 (predicted level 2, actual level 4), optimization is triggered; Structural risk scenario: Allowable level deviation ≤ 0.5 level, if the deviation reaches level 1 (predicted low risk, actual medium risk), optimization is triggered.

[0082] Distillation of the equipment status characteristics of solid capsules is performed to compress redundant current harmonic characteristics; Spatial association strength recalibration is performed on the relationship capsule, and device interaction weights are corrected based on the causal association strength of the structural causal model. The reconstructed capsule features are re-encoded into a spatiotemporal tensor using a neural radiation field, and the storage priority of the memory is updated according to a two-dimensional indexing architecture.

[0083] Based on the embodiments provided in this application, this application improves the update process of the historical scenario memory bank. After the security monitoring model completes cross-cycle transfer training, the error matrix between the virtual task prediction results and the real scene monitoring data is compared. When the trajectory positioning error under a specific scene label is higher than the error threshold for three consecutive cycles, the capsule network feature reconstruction mechanism is triggered. By performing device state feature distillation on the physical capsules, compressing redundant current harmonic features, and performing spatial correlation strength recalibration on the relational capsules, the device interaction weights are corrected based on the causal correlation strength of the structural causal model. This removes redundant information in the historical scenario memory bank and improves the quality and effectiveness of the memory bank. The reconstructed capsule features are re-encoded into a spatiotemporal tensor through a neural radiation field, and the storage priority of the memory bank is updated according to a two-dimensional index architecture. This enables the historical scenario memory bank to continuously adapt to changes in the campus environment, maintain the ability to learn and remember the latest risk patterns, and thus continuously improve the performance and accuracy of the security monitoring model.

[0084] It should be noted that the embodiments implemented on the side of the dynamically changing campus environment security monitoring system in this application can be referenced with the embodiments implemented on the side of the dynamically changing campus environment security monitoring method, and will not be described in detail here.

[0085] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A dynamically changing campus environment security monitoring system, characterized in that, include: The weight instruction generation module is used to generate dynamic weight instructions driven by scene adaptability by fusing the campus functional map and real-time data streams from multiple area sensors through a graph neural network. The correlation determination module is used to construct a structural causal model based on the dynamic weight instructions in order to determine the strength of the causal correlation between environmental parameters and equipment status. The risk trajectory output module is used to generate a model by taking the strength of the causal relationship as a conditional input and outputting a counterfactual risk trajectory that conforms to the physical laws of the campus. The migration optimization module is used to integrate the counterfactual risk trajectory with real-time data streams from multiple regional sensors to construct a virtual task set, thereby driving cross-cycle migration optimization of the safety monitoring model.

2. A method for dynamically changing campus environment security monitoring, wherein the method implements the system described in claim 1, characterized in that, include: By fusing real-time data streams from campus functional maps and multi-area sensors using graph neural networks, dynamic weight instructions driven by scene adaptability are generated. A structural causal model is constructed based on the dynamic weighting instructions to determine the strength of the causal relationship between environmental parameters and equipment status. The causal correlation strength is used as a conditional input to generate a model, which outputs a counterfactual risk trajectory that conforms to the physical laws of the campus. By integrating the counterfactual risk trajectory with the real-time data stream, a virtual task set is constructed to drive the cross-cycle migration optimization of the security monitoring model.

3. The method for dynamically changing campus environmental safety monitoring according to claim 2, characterized in that, The generation of the dynamic weight instruction includes: The campus functional map is modeled in multiple regions using a spatiotemporal graph neural network. The laboratory is divided into a chemical storage area sub-map and an experimental operation area sub-map, and the dormitory is divided into an electricity monitoring area sub-map and a public activity area sub-map. For the chemical storage area subgraph, gas sensor concentration distribution data and ventilation equipment operation status data are extracted, and the correlation weight between gas diffusion and ventilation efficiency is calculated through the subgraph attention mechanism of the spatiotemporal graph neural network. For the experimental operation sub-graph, current harmonic data and temperature and humidity sensor data of the experimental equipment are extracted to construct a weighted evaluation model of the equipment operating status. For the dormitory electricity monitoring area sub-map, the air conditioner usage time period and circuit current characteristics are extracted, and the weight features of illegal electricity use patterns are identified through the spatiotemporal graph neural network; The weight matrices of each subgraph are input into the cross-regional fusion module. Combined with the spatial correlation features of the campus functional map, the global dynamic weight allocation scheme is determined through the graph pooling operation of the spatiotemporal graph neural network to generate dynamic weight instructions driven by scene adaptability.

4. The method for dynamically changing campus environmental safety monitoring according to claim 3, characterized in that, The construction of the structural causal model based on the dynamic weight instructions includes: The dynamic weight instruction is used as a regional prior input to the structural causal model to construct a cross-regional causal graph. The cross-regional causal graph includes nodes and causal edges of the chemical storage area subgraph, the experimental operation area subgraph, and the electricity monitoring area subgraph. The spatiotemporal graph neural network is used to extract causal features from each subgraph, including: for the chemical storage area subgraph, extracting the spatiotemporal features of temperature, humidity, and gas concentration, and learning the causal mapping between chemical volatilization and ventilation efficiency through a multi-scale residual module; for the experimental operation area subgraph, extracting the correlation features between equipment current harmonics and temperature and humidity, and constructing a causal inference chain between equipment overheating and current mutations; for the electricity monitoring area subgraph, extracting the causal correlation model between load change time series features and illegal electricity use patterns. When the dynamic weight of a subgraph exceeds the risk weight threshold, the intervention inference channel is activated in the cross-regional causal graph to generate a causal effect intensity matrix.

5. The method for dynamically changing campus environmental safety monitoring according to claim 2, characterized in that, The process of using the strength of the causal relationship as a conditional input to generate a model, and outputting a counterfactual risk trajectory that conforms to the physical laws of the campus, includes: The causal relationship between environmental parameters and equipment status is encoded as a constraint and input into the generative model that integrates campus building physical parameters. Collect 3D point cloud data of the teaching building staircase and structural parameters of the roof corridor to construct a physical constraint library, and establish a joint constraint module of geometric safety and causal logic in the generated model; Output a counterfactual risk trajectory that conforms to the physical laws of the campus. The counterfactual risk trajectory includes hazard level labeling information. The spatial coordinates and time series of the counterfactual risk trajectory are jointly defined by the causal correlation strength and the physical parameters of the campus buildings.

6. The method for dynamically changing campus environmental safety monitoring according to claim 5, characterized in that, The verification and dynamic correction of the counterfactual risk trajectory includes: A spatial collision constraint field is constructed based on the three-dimensional point cloud of the teaching building staircase. When the angle between the diffusion vector of the counterfactual risk trajectory and the normal vector of the load-bearing structure exceeds the safety threshold, the trajectory reconstruction mechanism is triggered. The propagation rate of the counterfactual risk trajectory is dynamically damped and corrected using the building structure elastic modulus parameters in the physical constraint library. Output counterfactual risk trajectories with building structure conflict markers, where the hazard level marking information is positively correlated with the intensity of building structure conflict.

7. The method for dynamically changing campus environmental safety monitoring according to claim 5, characterized in that, The construction of the physical constraint library for the generated model also includes: A mapping table between crowd density and steel structure deformation rate is established in the rooftop corridor area. When the counterfactual risk trajectory involves the rooftop corridor area, the deformation rate threshold is used as a hard constraint condition for the spatial distribution of the trajectory. An embedded laboratory equipment current safety envelope is used to generate a trajectory bifurcation path for the equipment protection mechanism when the simulated current value in the counterfactual risk trajectory exceeds the laboratory equipment current safety envelope.

8. The method for dynamically changing campus environmental safety monitoring according to claim 2, characterized in that, The process of fusing the counterfactual risk trajectory with the real-time data stream to construct a virtual task set to drive cross-cycle migration optimization of the security monitoring model includes: A bio-inspired counterfactual network is constructed, whose input layer receives the real-time data stream and whose output layer is connected to a spatiotemporal capsule network. When multi-region sensors detect the spatiotemporal coupling between a sudden change in population density in the target area and an abnormal signal from equipment in adjacent areas, the causal neuron module of the bio-inspired counterfactual network activates the matching risk patterns in the historical context memory bank. The real-time data stream is decoupled into entity capsules and relational capsules using the spatiotemporal capsule network, wherein the entity capsules encode the state characteristics of physical devices, and the relational capsules encode the interactions between devices; A set of virtual tasks with local intervention conditions is generated through a capsule routing protocol to drive cross-cycle migration optimization of the security monitoring model.

9. The method for dynamically changing campus environmental safety monitoring according to claim 8, characterized in that, The construction and retrieval of the historical context memory bank includes: Establish a two-dimensional index architecture, where the first dimension is classified into scenario tags according to fire, structural collapse, and equipment cascading failure, and the second dimension is divided into time tags according to teaching cycle periods and holiday periods. Spatiotemporal encoding of historical risk scenarios is performed through neural radiation fields, including: spatial encoding binding to building structure point cloud coordinates, and temporal encoding associating with school calendar event timestamps; When the causal neuron module of the bio-inspired counterfactual network is activated, a matching memory fragment is retrieved in the two-dimensional indexing architecture based on the spatiotemporal coupling characteristics of the current sensor data stream. The retrieved memory fragments are input into the spatiotemporal capsule network for feature decoupling and recombination, and a set of memory retrieval instructions adapted to the current dynamic weights is output.

10. The method for dynamically changing campus environmental safety monitoring according to claim 9, characterized in that, The updating of the historical context memory bank includes: After the security monitoring model completes cross-cycle transfer training, the error matrix between the virtual task prediction results and the real scene monitoring data is compared. When the trajectory localization error under a specific scene label exceeds the error threshold for three consecutive cycles, the capsule network feature reconstruction mechanism is triggered. The solid capsule is subjected to device state characteristic distillation to compress redundant current harmonic characteristics; Spatial association strength recalibration is performed on the relationship capsule, and device interaction weights are corrected based on the causal association strength of the structural causal model; The reconstructed capsule features are re-encoded into a spatiotemporal tensor using a neural radiation field, and the storage priority of the memory is updated according to a two-dimensional indexing architecture.

Images