Intelligent campus safety management early warning method and system based on big data

Abstract

The invention belongs to the technical field of smart campus, and relates to a smart campus safety management early warning method and system based on big data. The method comprises the following steps: collecting registration information, real-time behavior data and state information of an environment of a foreign vehicle; constructing a dynamic space-time diagram, and generating an edge attribute reflecting a relationship between deployment nodes in the dynamic space-time diagram; calculating a reasonable path set of the foreign vehicle, and analyzing trajectory deviation data of the foreign vehicle; abnormal scoring is carried out on the external vehicle; generating a corresponding intervention control strategy; performing causal influence analysis and generating a decision basis; and training the local graph neural network model, and carrying out centralized aggregation and distribution updating on shared model parameters in a training result of the local graph neural network model. According to the method, a dynamic behavior modeling strategy based on a heterogeneous graph neural network is introduced, structural modeling and time sequence behavior evolution expression of campus moving target multi-source data are achieved, the system can capture tiny anomalies in individual trajectory evolution, and the abnormal behavior recognition sensitivity is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of smart campuses, and specifically relates to a smart campus safety management early warning method and system based on big data. Background Art

[0002] With the continuous advancement of smart campus construction, various activities such as the flow of people and vehicle traffic on campus are highly complex, and campus safety management faces multiple challenges: on the one hand, the frequency of external vehicles entering the campus has increased, and the uncertainty of driving trajectories, stopping areas and behavior patterns may cause traffic congestion, collision accidents or safety hazards; on the other hand, the road network on campus is complex, and there are many areas with dense crowds (such as teaching buildings, cafeterias and playgrounds). The coupling of environmental factors (such as construction areas, weather changes and lighting conditions) with the behavior of traffic participants further aggravates the difficulty of safety management.

[0003] Existing campus security management systems use video surveillance equipment to achieve visual monitoring of key areas, combine facial recognition technology to verify personal identities, and rely on access control systems to record vehicle and personnel entry and exit information; some systems introduce rule engines or simple machine learning models (such as decision trees and support vector machines) to judge abnormal behavior based on preset thresholds (such as speeding monitoring and detection of entering restricted areas) and trigger alarm prompts.

[0004] However, existing systems mostly rely on fixed rules or static features (such as a single speed threshold and fixed no-entry zones) to determine anomalies, making it difficult to adapt to the dynamic changes in the campus environment. In the campus transportation system, there are complex spatial correlations and interactions between vehicles, roads, and people. For example, vehicle speed is restricted by pedestrian density and road construction affects the travel path. Existing technologies lack structured modeling of these relationships. Some systems use models (such as complex neural network models) to output anomaly judgment results, but the logical basis for the result generation is not clear, making it difficult for managers to trace the causes of abnormal behavior and unable to verify the effectiveness of intervention strategies. Traditional methods have weak ability to capture minor anomalies (such as vehicles slowly deviating from reasonable paths and speed gradient anomalies), and often trigger alarms only when the risks accumulate to a significant level, missing the opportunity for intervention.

[0005] Therefore, there is an urgent need for a smart campus safety management and early warning system that can integrate multi-source data, dynamically model entity relationships, be interpretable, and support cross-regional collaboration, so as to improve the recognition accuracy and intervention efficiency of abnormal behaviors in complex scenarios and ensure campus safety. Summary of the Invention

[0006] In order to solve the above technical problems, the present invention provides a smart campus safety management early warning method and system based on big data.

[0007] In a first aspect, the present invention provides a smart campus safety management early warning method based on big data, comprising: Collect registration information, real-time behavior data and status information of the environment of foreign vehicles; Using the collected data, a dynamic spatiotemporal graph containing deployment nodes is constructed, and edge attributes reflecting the relationships between deployment nodes are generated in the dynamic spatiotemporal graph; deployment nodes include vehicles, roads, and people; Calculate the reasonable path set of the foreign vehicle based on the dynamic space-time graph, and analyze the trajectory deviation data of the foreign vehicle based on the spatial relationship between the actual driving trajectory of the foreign vehicle and the reasonable path set; Based on the trajectory deviation data and the real-time behavior data changes of foreign vehicles, the foreign vehicles are scored as abnormal; Generate corresponding intervention control strategies based on the anomaly score and the current state information of the foreign vehicle; Conduct causal impact analysis on intervention and control strategies and generate decision-making basis; The local graph neural network model is trained separately in multiple deployment nodes, and the shared model parameters in the local graph neural network model training results are centrally aggregated and distributed for update.

[0008] In a second aspect, the present invention provides a smart campus safety management and early warning system based on big data, including an external vehicle information collection unit, a dynamic spatiotemporal graph construction unit, a path modeling and deviation analysis unit, an abnormal behavior measurement unit, an intervention strategy decision unit, an analysis and generation unit, and a federated learning collaboration unit; The external vehicle information collection unit is used to collect the registration information, real-time behavior data and status information of the environment in which the external vehicle is located; A dynamic spatiotemporal graph construction unit is used to construct a dynamic spatiotemporal graph containing deployment nodes using the collected data, and to generate edge attributes reflecting the relationships between the deployment nodes in the dynamic spatiotemporal graph; the deployment nodes include vehicles, roads, and people; A path modeling and deviation analysis unit is used to calculate a reasonable path set for an incoming vehicle based on a dynamic spatiotemporal graph, and to analyze the trajectory deviation data of the incoming vehicle based on the spatial relationship between the actual driving trajectory of the incoming vehicle and the reasonable path set; Abnormal behavior measurement unit, used to score the abnormality of foreign vehicles based on trajectory deviation data and real-time behavior data changes of foreign vehicles; An intervention strategy decision unit, configured to generate a corresponding intervention control strategy based on the anomaly score and the current state information of the foreign vehicle; Analysis and generation unit, used to perform causal impact analysis on intervention control strategies and generate decision basis; The federated learning collaborative unit is used to train the local graph neural network model in multiple deployment nodes separately, and to centrally aggregate and distribute the updates of the shared model parameters in the training results of the local graph neural network model.

[0009] On the basis of the above technical solution, the present invention can also be improved as follows.

[0010] Furthermore, the registration information of external vehicles is collected, including the license plate number, vehicle type, time of entry, location of entry and destination of the external vehicle; Collecting real-time behavior data of external vehicles includes collecting GPS tracks, speeds, and accelerations of external vehicles on campus; Collecting the state information of the environment in which the external vehicle is located includes collecting the road state, pedestrian density, light intensity and construction area distribution information of the location of the external vehicle.

[0011] Furthermore, the edge-to-edge attributes of the dynamic spatiotemporal graph are generated according to the spatial distance, speed difference and risk value between the deployed nodes.

[0012] Furthermore, a reasonable path set for external vehicles is calculated based on the dynamic space-time graph, including: calculating the shortest path between the external vehicle's entry node and the access target node as the basic path in the dynamic space-time graph; setting the tolerance parameter of the preset basic path; and constructing a reasonable path set according to the path length of the basic path and the preset tolerance parameter.

[0013] Furthermore, the trajectory deviation data of the foreign vehicle is analyzed based on the spatial relationship between the actual driving trajectory of the foreign vehicle and the reasonable path set, including: Calculate the path deviation of the foreign vehicle based on its actual driving trajectory and the path on the selected road section in the reasonable path set; Constructing a historical speed distribution model corresponding to each foreign vehicle based on the historical data of each foreign vehicle; comparing the current speed distribution of the foreign vehicle with the historical speed distribution of the foreign vehicle to obtain a speed distribution difference of the foreign vehicle; According to the path deviation and speed distribution difference of the foreign vehicle, the track deviation data of the foreign vehicle is obtained.

[0014] Further, according to the path deviation and speed distribution difference of the foreign vehicle, obtaining the trajectory deviation data of the foreign vehicle includes: calculating the abnormality score of the foreign vehicle based on the path deviation and speed distribution difference; Assume that the anomaly score is , the average spatial deviation between the actual driving trajectory of the external vehicle and the reasonable path in the reasonable path set is , and is the preset non-negative real weighting coefficient, and the Kullback-Leibler divergence is , the current speed distribution of the external vehicle is , the historical speed distribution of foreign vehicles is , then the anomaly score is: .

[0015] Furthermore, the location information of the foreign vehicle, the graph embedding representation, the anomaly score and the environment state information are constructed into a state vector; Generate intervention control strategy based on state vector using reinforcement learning model; Intervention control strategies include voice prompts, route replanning, safety passes and road restrictions.

[0016] Furthermore, a causal impact analysis is conducted on the intervention control strategy and a basis for decision making is generated, including: Construct a causal graph structure between intervention behavior variables, vehicle response behavior variables, and environmental variables; Based on the causal graph structure, the expectation of the vehicle response behavior variable under the intervention behavior variable is calculated as the causal reasoning result; Generate intervention decision description information based on causal reasoning results as the basis for making decisions.

[0017] Furthermore, the local graph neural network model is trained separately in multiple deployment nodes, and the shared model parameters in the local graph neural network model training results are centrally aggregated and distributed and updated, including: Train the graph neural network model based on local data in each deployment node to obtain a local model; Upload the shared model parameters in the local model to the central server; Perform weighted average calculation on the shared model parameters uploaded by multiple deployment nodes and generate global shared model parameters; The global shared model parameters are distributed to each deployment node to update the local model.

[0018] The beneficial effects of the present invention are: (1) This paper introduces a dynamic behavior modeling strategy based on heterogeneous graph neural networks, which realizes the structured modeling and temporal behavior evolution expression of multi-source data of mobile targets on campus. This enables the system to capture small anomalies in the evolution of individual trajectories, improves the sensitivity of abnormal behavior recognition, and avoids the problem that traditional methods mostly rely on static rule matching or simple sequence statistics and are difficult to adapt to the diversity and evolution of behavior patterns in complex environments; (2) By integrating causal reasoning calculations, the present invention can quantitatively analyze the effectiveness of intervention measures and trace the causal logic behind policy triggering, achieving the dual goals of credible control interpretation and reasonable policy optimization. Compared with the existing solutions that generally use black box models to output control instructions but lack reasoning transparency, this method effectively solves the system trust problem; (3) Without transmitting the original data, the present invention constructs a federated learning collaborative mechanism, which enables different campus nodes to share graph neural network knowledge, ensuring the consistency of model updates and the collaborative learning ability of the system. This mechanism overcomes the data leakage risks that are prone to occur in existing centralized training and the bottleneck of insufficient cross-regional model generalization ability, and significantly enhances the adaptability of the system to cross-school deployment. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 Schematic diagram of the smart campus safety management and early warning method based on big data provided in Example 1 of the present invention; Figure 2 This is a flowchart of a specific implementation method of Example 1 of the present invention; Figure 3 This is a schematic diagram of the big data-based smart campus safety management and early warning system provided in Example 2 of the present invention. DETAILED DESCRIPTION

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings herein can be arranged and designed in various different configurations.

[0021] Example 1 As an example, Figure 1 As shown, to solve the above technical problems, this embodiment provides a smart campus safety management and early warning method based on big data, including: Collect registration information, real-time behavior data and status information of the environment of foreign vehicles; Using the collected data, a dynamic spatiotemporal graph containing deployment nodes is constructed, and edge attributes reflecting the relationships between deployment nodes are generated in the dynamic spatiotemporal graph; deployment nodes include vehicles, roads, and people; Calculate the reasonable path set of the foreign vehicle based on the dynamic space-time graph, and analyze the trajectory deviation data of the foreign vehicle based on the spatial relationship between the actual driving trajectory of the foreign vehicle and the reasonable path set; Based on the trajectory deviation data and the real-time behavior data changes of foreign vehicles, the foreign vehicles are scored as abnormal; Generate corresponding intervention control strategies based on the anomaly score and the current state information of the foreign vehicle; Conduct causal impact analysis on intervention and control strategies and generate decision-making basis; The local graph neural network model is trained separately in multiple deployment nodes, and the shared model parameters in the local graph neural network model training results are centrally aggregated and distributed for update.

[0022] As attached Figure 2 The flowchart shown is a specific implementation method of the smart campus safety management and early warning method based on big data.

[0023] Optionally, collect registration information of external vehicles, including license plate number, vehicle type, time of entry, location of entry, and destination of visit; Collecting real-time behavior data of external vehicles includes collecting GPS tracks, speeds, and accelerations of external vehicles on campus; Collecting the state information of the environment in which the external vehicle is located includes collecting the road state, pedestrian density, light intensity and construction area distribution information of the location of the external vehicle.

[0024] By collecting multi-dimensional and multi-modal data on external vehicles entering the campus area, and closely integrating the data flow and logic flow with the subsequent dynamic spatiotemporal graph, the system generally initiates the collection of external vehicle data upon receiving a new external vehicle entry event, including registration information, real-time behavior data, and environmental status information.

[0025] As a typical application process, when an external vehicle enters the controlled area of ​​the campus, the system will initiate real-time detection of the vehicle's registration, trajectory, and environmental status through the multi-source sensing equipment deployed at the school gate. These data will be uniformly timestamped and stored in a structured cache queue as the basic data fragment for subsequent dynamic graph construction.

[0026] The registration information of external vehicles is connected to the campus access control platform and linked to the visitor reservation system through a high-definition license plate recognition camera. The system can simultaneously obtain information including but not limited to: The license plate number of the external vehicle; the vehicle type, such as a small car, a commercial van, etc.; the actual entry timestamp, in the format of a Unix timestamp; the geographic location of the entry, which can be represented by a static point code; the target location to be visited, usually provided by the campus reservation system.

[0027] In an optional implementation, the registration information may also record the vehicle's estimated departure time and reservation contact information for subsequent trajectory deviation and stop anomaly analysis.

[0028] The real-time behavior data of vehicles on campus is continuously tracked. Generally, the system deploys several global navigation satellite system differential positioning base stations on campus and combines the inertial navigation module with the vehicle-mounted GPS terminal to achieve high-precision sampling. In some embodiments, the sampling frequency of the track points is set to 1Hz, and the collected data includes: location information, set to , recorded in WGS-84 or projected coordinate system; instantaneous velocity ; Acceleration is calculated from the velocity difference between two adjacent time points. For the moment Vehicle acceleration in m / s 2 , The vehicle speed at time , The vehicle speed at the time is , is the sampling time interval, which is generally 1 second, and the vehicle acceleration is expressed as: .

[0029] It is worth noting that in order to suppress the influence of GPS noise, a Kalman filter is often introduced in actual deployment to smooth the trajectory.

[0030] When collecting status information of the environment in which the external vehicle is located, scene perception and context modeling around the vehicle are performed.

[0031] In one possible design, the environment status is obtained from the following data sources: Road slipperiness: Real-time rainfall and humidity are obtained through micro-meteorological monitoring equipment deployed on key roads. Crowd density: The instantaneous number of people in hot spots is perceived using facial counting cameras or Wi-Fi probes. Light intensity: Detected using ambient light sensors, measured in Lux, is often used to judge nighttime driving behavior. Regional construction information: Typically provided with map-based coding by a campus management system or building information modeling platform.

[0032] As an option, in order to improve the accuracy of scene expression, the system performs block mapping processing on the above environmental information, that is, dividing the campus into multiple perception grids, each grid maintains a set of environmental state vectors; let the environmental state vector be , is the road slipperiness index, is the crowd density, in persons / square meter, is the light level in Lux, Code for construction signs. If there is construction, set to 1, otherwise 0. Road slippery index Normalized to [0,1], the environment state vector is: .

[0033] These structured environmental vectors are mapped to the attributes of road and crowd nodes during the dynamic graph construction phase. In some embodiments, to reduce sensor load and communication pressure, the system adopts an edge computing node deployment strategy. Each collection point performs data preprocessing, compression, and encoding locally, and only uploads key feature values ​​to the central node. For example, a trajectory point sequence is compressed into a five-dimensional state summary, including the starting point, end point, maximum speed, average acceleration, and dwell time. Furthermore, in practical applications, to enhance adaptability to special types of vehicles (such as delivery trucks and construction vehicles), the system supports preliminary classification and labeling of collected data using a rule base. For example, if a vehicle's trajectory is clearly concentrated between the cafeteria and the teaching building, and the dwell time is short, it is initially labeled as a "short-term delivery type"; if the trajectory wanders to a non-booked destination, the system will label it as "high deviation potential" to trigger an early warning condition.

[0034] Optionally, edge-to-edge attributes of the dynamic spatiotemporal graph are generated based on the spatial distance, speed difference, and risk value between the deployed nodes.

[0035] Based on the collected external vehicle information, a dynamic spatiotemporal graph consisting of vehicle, road, and crowd nodes is constructed, and edge attributes reflecting the relationships between nodes are generated. Specifically, the heterogeneous data of external vehicle information is converted into a unified graph data structure and then input into the dynamic spatiotemporal graph for subsequent graph computing tasks such as path modeling and behavior analysis. The core components of the overall system architecture, data structure abstraction and graph encoding, are utilized to organize multimodal raw information into a dynamic spatiotemporal graph that adapts to the graph neural network processing process.

[0036] This paper introduces a dynamic behavior modeling strategy based on heterogeneous graph neural networks, realizes the structured modeling and temporal behavior evolution expression of multi-source data of campus mobile targets, enables the system to capture subtle anomalies in the evolution of individual trajectories, improves the sensitivity of abnormal behavior identification, and avoids the problem that traditional methods mostly rely on static rule matching or simple sequence statistics and are difficult to adapt to the diversity and evolution of behavior patterns in complex environments.

[0037] Typically, after the system completes data collection for incoming vehicles, it immediately triggers graph encoding of the spatial, behavioral, and environmental data involved in the current vehicle behavior cycle. During the dynamic graph construction process, the graph nodes, edge attributes, and their temporal embeddings are dynamically updated to ensure that the constructed graph effectively reflects changes in the vehicle's spatiotemporal environment and potential risks.

[0038] The construction of a dynamic spatiotemporal graph includes node construction, edge attribute generation, and graph embedding extraction. The specific implementation is as follows: Node construction is used to model elements related to vehicle activity on campus as nodes in the graph. There are at least three types of nodes: external vehicle nodes, which represent individual vehicles entering the campus. Their attributes may include vehicle identification, current trajectory coordinates, speed, and acceleration; campus road nodes, which represent the spatial location, directionality, and traffic status of on-campus roads; and crowded area nodes, which represent specific areas on campus with high traffic flow, such as cafeterias, teaching buildings, and libraries.

[0039] In some embodiments, to enhance the node expression capability, the system defines each node as a set of vehicle node attribute vectors. For vehicles at time The horizontal coordinate information, For vehicles at time The vertical coordinate information is in meters, using the WGS-84 coordinate system or projection coordinate system. is the vehicle speed in meters per second (m / s), is the vehicle acceleration in meters per second squared (m / s²), is the sampling timestamp, the format is Unix timestamp, the vehicle node attribute vector is , expressed as: .

[0040] As an option, a node category encoding can be attached to each node in the graph to identify the semantic role played by the node in the graph to support subsequent embedding learning.

[0041] Edge attributes are used to define the connection relationships and attributes between different types of nodes. Edges are used to model the structural relationships and spatiotemporal interactions between nodes. The main edge types include: navigation edges between vehicles and road nodes, indicating whether a vehicle is traveling on a specific road; proximity edges between vehicles and crowd area nodes, indicating the spatial proximity between vehicles and crowd hotspots; and topological edges between road nodes, forming the basic skeleton of the transportation network.

[0042] Specifically, For nodes and The spatial Euclidean distance between them, in meters, If the vehicle and the road node are connected, it represents the speed difference in m / s. is the risk weight value, which reflects the potential safety risk of the connection path. It can be obtained by comprehensive evaluation of historical accident records, crowd density and lighting conditions. It is normalized to [0,1]. The attribute vector of each edge is ,but Expressed as: .

[0043] In one possible implementation, in order to enhance the model's ability to model complex relationships, the system can also adopt a multi-dimensional edge attribute structure, in which the risk weight value is further refined into multiple dimensions, such as congestion index, slippery index and lighting confidence.

[0044] Graph embedding extraction encodes the constructed dynamic graph through the graph neural network method and extracts the low-dimensional embedding representation of each node to support the input requirements of subsequent path modeling and behavior discrimination model. Generally, this module uses the spectral graph convolution method to encode the graph structure. Assume that the current graph ,set up For the The layer graph embedding input, the adjacency matrix of the graph is , the unit matrix is , Add the identity matrix to the adjacency matrix of the graph; for The degree matrix of , the elements on the diagonal are the node degrees; For the Layer trainable weight matrices, As the activation function, ReLU activation function or ELU activation function is often used, and the calculation formula for single-layer graph convolution is as follows: .

[0045] As an option, in the case of considering dynamic graph changes, this embodiment supports sliding updates of the graph based on a time window, that is, at fixed time intervals Reconstruct the current dynamic graph to adapt to the dynamic evolution characteristics of vehicle behavior.

[0046] In some embodiments, to enhance the temporal modeling capability of dynamic graphs, the system further introduces a temporal encoding mechanism to map the time attributes of nodes to time vectors and splice them into the node features as input. , the time embedding vector is , represents the dimension index in the vector, The total number of time embedding vector dimensions, the sine and cosine position encoding of the time embedding vector is expressed as , the cosine position encoding of the time embedding vector is expressed as ,but: ; .

[0047] The final node feature input is represented by the concatenation of the original feature and the time embedding, namely: .

[0048] As an extended implementation, the system also supports a graph layering mechanism. This constructs different types of nodes into a heterogeneous graph model and uses a graph attention mechanism to apply different attention weights to different types of nodes and edges, thereby improving the model's ability to express multivariate relationships.

[0049] Optionally, a reasonable path set for external vehicles is calculated based on a dynamic space-time graph, including: calculating the shortest path between the external vehicle's entry node and the access target node as a basic path in the dynamic space-time graph; setting a tolerance parameter for a preset basic path; and constructing a reasonable path set based on the path length of the basic path and the preset tolerance parameter.

[0050] Specifically, path modeling and deviation analysis follow the construction of the dynamic space-time graph, inheriting the dynamic graph structure generated by the latter, which includes road nodes, vehicle nodes, and crowd nodes. Its main function is to establish a reasonable path model for external vehicles from the entry point to the target area visited, and calculate the path deviation degree based on their real-time trajectory information, so as to provide the input basis of the spatial deviation dimension for the abnormal behavior measurement module.

[0051] Generally speaking, after completing the preliminary graph modeling, the system will call this module to perform path modeling and real-time comparative analysis based on the school entrance and target node accessed provided by the external vehicle in the registration information. This can be used for preliminary screening of behavioral trends and for determining the rationality of trajectories.

[0052] In this embodiment, path modeling and deviation analysis at least include path generation, reasonable path screening and deviation analysis.

[0053] Path generation is mainly used to calculate the shortest path from the vehicle's entry location to the target destination as the basic path based on the road node topology structure in the dynamic graph. Specifically, the system first calculates the entry location provided by the registration information collection unit in the external vehicle information collection module. To access the target location, determine the corresponding starting node and end node in the graph structure.

[0054] The system uses Dijkstra algorithm or A* heuristic search algorithm. The minimum cost path is calculated on the basis of the vehicle. is the path length, and the path with the minimum cost is , Nodes in the path With node The spatial Euclidean distance between them, the path length Obtained by summing the distance attributes of all edges on the path: .

[0055] The reasonable path screening process constructs a reasonable path set based on the basic path, and the reasonable path set is , Indicates the candidate paths; Indicates the path Total length; Indicates the maximum acceptable path length floating range in meters. The preset value can be dynamically set based on road density or traffic complexity. Generally, the system sets an adjustable path tolerance parameter to represent the maximum acceptable path gain. During the path screening phase, the system traverses all feasible paths and retains those that meet the following constraints: .

[0056] As an option, to avoid affecting computational efficiency due to an excessive number of paths, path structure constraints can be introduced, such as the number of node visits or the number of path loops, to further reduce the size of the path set.

[0057] Optionally, analyzing the trajectory deviation data of the foreign vehicle based on the spatial relationship between the actual driving trajectory of the foreign vehicle and the set of reasonable paths includes: Calculate the path deviation of the foreign vehicle based on its actual driving trajectory and the path on the selected road section in the reasonable path set; Constructing a historical speed distribution model corresponding to each foreign vehicle based on the historical data of each foreign vehicle; comparing the current speed distribution of the foreign vehicle with the historical speed distribution of the foreign vehicle to obtain a speed distribution difference of the foreign vehicle; According to the path deviation and speed distribution difference of the foreign vehicle, the track deviation data of the foreign vehicle is obtained.

[0058] Deviation analysis is used to compare the actual driving trajectory of foreign vehicles in real time and reasonable path set Calculate the path deviation index .

[0059] In one possible implementation, the system discretizes the actual driving trajectory of the foreign vehicle into a sequence of trajectory points: .

[0060] The system selects the path from the set of reasonable paths. The most similar path , Hausdorff distance or weighted DTW (Dynamic Time Warping) algorithm can be used for matching.

[0061] set up is the average spatial deviation index, in meters. is the number of actual trajectory points, The actual trajectory Point coordinates, Indicates a reasonable path The coordinates of each path point in the path, the path deviation index is defined as: .

[0062] The higher the path deviation index value, the more obvious the vehicle deviates from the expected path. This index is usually passed to the abnormal behavior measurement module to participate in the abnormality score calculation.

[0063] In some embodiments, to accommodate the uncertainty of multi-source location data, the system can set confidence intervals for trajectory points and incorporate the deviation index calculation process into a probabilistic modeling mechanism. For example, each trajectory point can be treated as the center of a two-dimensional Gaussian distribution, and the expected distance can be used to calculate the expected deviation value, enhancing the algorithm's robustness in the presence of position drift.

[0064] In some embodiments, to accommodate the uncertainty of multi-source location data, the system can set confidence intervals for trajectory points and incorporate the deviation index calculation process into a probabilistic modeling mechanism. For example, each trajectory point can be treated as the center of a two-dimensional Gaussian distribution, and the expected distance can be used to calculate the expected deviation value, enhancing the algorithm's robustness in the presence of position drift.

[0065] As an extended implementation method, the system can also classify deviation behaviors into categories, such as "short-term path deviation", "long-term offset and detention" and "reverse traffic" and other behavioral labels to assist subsequent intervention strategy selection logic.

[0066] Optionally, obtaining trajectory deviation data of the foreign vehicle according to the path deviation and speed distribution difference of the foreign vehicle includes: calculating an abnormality score of the foreign vehicle based on the path deviation and speed distribution difference; Assume that the anomaly score is , the average spatial deviation between the actual driving trajectory of the external vehicle and the reasonable path in the reasonable path set is , and is the preset non-negative real weighting coefficient, and the Kullback-Leibler divergence is , the current speed distribution of the external vehicle is , the historical speed distribution of foreign vehicles is , then the anomaly score is: .

[0067] Abnormal behavior measurement is based on the path deviation index calculated by the path modeling and deviation analysis module and the changes in the behavior patterns of foreign vehicles, and is used to score foreign vehicles for abnormality. Specifically, abnormal behavior measurement is located after path modeling and deviation analysis in the system architecture of the present invention. It is mainly used to comprehensively measure the spatial behavior deviation and speed behavior changes of vehicles, and to construct a quantifiable behavior anomaly scoring system as an important input state source for the subsequent intervention strategy decision module.

[0068] The speed behavior deviation analysis model based on Kullback-Leibler divergence can accurately compare and evaluate the vehicle's actual operating status with the historical behavior baseline, thereby achieving more quantitative anomaly score calculation. Compared with traditional solutions that use a single threshold judgment or manually set deviation rules, this solution is more scalable and has more stable judgment, making it particularly suitable for risk screening needs in high-density traffic scenarios.

[0069] Typically, after obtaining spatial trajectory deviation indicators for incoming vehicles, the system combines these with historical and current speed pattern changes to determine whether their behavior exhibits abnormal trends. This module combines structured path deviation data with time-series speed data to form a unified scoring model.

[0070] In this embodiment, the abnormal behavior measurement module mainly includes a behavior pattern modeling unit, a behavior change comparison unit, and an abnormality score generation unit.

[0071] The behavior pattern modeling unit is used to establish a speed behavior reference model of foreign vehicles.

[0072] Specifically, the system first extracts the speed sequence data of the vehicle during the past several times of entering the school. The speed value in each trajectory is normalized, and statistical modeling is performed on all samples to form the historical speed distribution of the vehicle. .

[0073] In one possible implementation, the system uses kernel density estimation or histogram estimation to perform probability density modeling on the velocity data and outputs a continuous probability distribution function: ; in: is the velocity variable, in meters per second (m / s), Indicates that the speed value is The historical distribution probability density value at time , Indicates the historical speed distribution in speed The probability density function value at is usually obtained by histogram estimation or kernel density smoothing estimation of historical sampling data. This speed distribution is used as a baseline behavior feature to reflect the speed habits of vehicles during normal traffic.

[0074] The behavior change comparison process calculates the difference between current and historical behavior. During the vehicle's current training session, the system records its latest speed series in real time and uses this data to generate a current speed distribution model. This distribution uses a modeling method consistent with the historical speed distribution to ensure comparability.

[0075] In order to measure the degree of variation between distributions, the system uses the Kullback-Leibler divergence to calculate the difference, which is calculated as follows: ; in: represents the KL divergence between the actual velocity distribution and the historical velocity distribution; is the total number of discrete speed intervals, indicating the number of segments divided in the speed space; The actual velocity distribution In the The probability value in the speed interval is is the historical velocity distribution In the Probability value in the speed interval; is a logarithmic function, usually with the natural logarithm as the base.

[0076] Generally, to ensure the stability of divergence calculations, the system smooths all interval probabilities, for example, by adding 1 or setting a minimum threshold to avoid division by zero. Anomaly score generation is used to jointly model the path deviation indicator and the degree of speed behavior deviation to form a unified anomaly score.

[0077] In a standard implementation, the system introduces a linear weighted combination function as a scoring model. represents the abnormality score of foreign vehicles, and is the preset non-negative real weighting coefficient, then: .

[0078] Generally, the system determines the security incidents through supervised learning or expert calibration based on past security incident samples. and The system can also select a reasonable value for to ensure that the scoring results are sensitive to abnormal behavior. Alternatively, the system can introduce a dynamic adjustment mechanism to adaptively adjust the weight coefficient during peak periods or complex weather conditions, making the scoring system more adaptable to the environment.

[0079] In some embodiments, the system supports setting multi-level thresholds for abnormal scoring results. Based on different scoring intervals, labels such as "low risk," "medium risk," and "high risk" are generated and fed back to the intervention strategy decision module through confidence scores, forming a graded response mechanism. Furthermore, to enhance the traceability and interpretability of scoring, the system can record the original input data, model parameters, and intermediate calculation results used for each scoring process, supporting subsequent causal reasoning module calls.

[0080] Optionally, construct a state vector by combining the location information of the foreign vehicle, the graph embedding representation, the anomaly score, and the environment state information; Generate intervention control strategy based on state vector using reinforcement learning model; Intervention control strategies include voice prompts, route replanning, safety passes and road restrictions.

[0081] The intervention control strategy is trained by setting the reward function: Indicates the intervention strategy at time The reward value under Indicates at time Safety scores corresponding to foreign vehicles; Indicates at time The false alarm loss value; represents the delay time of intervention response; 、 and is a non-negative real number parameter used for weighting in the reward function, then the reward function is expressed as: .

[0082] The intervention strategy decision-making process generates a corresponding intervention control strategy based on the abnormal score of the abnormal behavior measurement and the current state information of the foreign vehicle; specifically, the intervention strategy decision is a control execution process in the system structure of the present invention, which follows the abnormal behavior measurement and takes the vehicle abnormal score as the core trigger factor. Combined with its current state vector, with the support of the spatiotemporal features provided by the graph neural network, it dynamically generates the optimal intervention response strategy, which is a key logical node for decision generation, strategy evaluation and intervention instruction output to achieve active safety control. Generally, after receiving the abnormal score of a certain foreign vehicle, the system will start the strategy generation process of this module. On the basis of the completion of the state space encoding, the system calls the reinforcement learning model for strategy selection, and sends the selected control action to the control system or security personnel in real time.

[0083] The intervention strategy decision-making process includes state vector construction, strategy generation and control output, as follows: The state vector construction process is used to integrate the multi-source state information of the external vehicle at the current moment and convert it into an input format acceptable to the reinforcement learning algorithm. Specifically, the state vector includes at least the following elements: the vehicle's current position information , in meters; the graph embedding representation vector is , which comes from the node embedding results in the dynamic spatiotemporal graph construction module; the anomaly score is , from the abnormal behavior measurement module; the set of environmental state parameters is , including crowd density, road slipperiness index, light level and construction status; time label is , represents the timestamp of state sampling. Therefore, the state vector can be written as: .

[0084] As an option, to improve the contextual consistency of policy generation, the system can also introduce state sequences from several past moments to form a state trajectory stack for time series dependency modeling.

[0085] The strategy generation process uses a reinforcement learning model to perform decision calculations for intervention strategies. In a standard implementation, the system uses the Deep Q-Network (DQN) framework to construct a strategy function and achieve the mapping between the state vector and the action space. The system defines a discrete action space Including: voice prompts, route re-planning, security notifications and road closure suggestions. To build reinforcement learning objectives, the system defines the following reward function: ; in: Indicates the strategy at time The instant reward value under Indicates the safety score improvement value of vehicle behavior under the current strategy intervention, It represents the false positive loss metric generated after the strategy is executed. Indicates the time delay between policy execution and response generation, in seconds. 、 and is a non-negative weighting coefficient used to control the contribution of the three indicators to the reward function. Generally, the system trains the weights of the reinforcement learning network by replaying data from historical event samples and uses the experience replay mechanism to stabilize the training process.

[0086] In some embodiments, the strategy generation network may adopt a dual network structure, that is, a main Q network and a target Q network are used to update parameters alternately to avoid the problem of Q value overestimation.

[0087] The intervention action is executed according to the decision output generated by the strategy. Based on the strategy results, the system calls different execution interface modules, such as the voice broadcast module, the vehicle navigation instruction module, the security alarm interface or the traffic management API interface, to complete the external effect of the intervention strategy. As a possible deployment method, some intervention strategies can be implemented in real time at the vehicle entry control terminal (such as the gate system and the road crossing broadcast), while some operations involving management authority need to be uploaded to the central platform and confirmed by security personnel before execution. In some embodiments, the system generates a unique control number for each intervention, and records the action issued, execution time, response result and strategy score for causal reasoning and tracing in the subsequent explainability analysis module.

[0088] Optionally, conduct a causal impact analysis of the intervention control strategy and generate a basis for decision making, including: Construct a causal graph structure between intervention behavior variables, vehicle response behavior variables, and environmental variables; Based on the causal graph structure, the expectation of the vehicle response behavior variable under the intervention behavior variable is calculated as the causal reasoning result; Generate intervention decision description information based on causal reasoning results as the basis for making decisions.

[0089] The explainability analysis module is used to perform causal influence reasoning on the control strategy output by the intervention strategy decision module and generate decision-making basis; Specifically, the explainability analysis module is located in the decision-making feedback loop of the system architecture, following the intervention strategy decision module. Its primary function is to perform a causal analysis of executed or recommended intervention control strategies, clarifying whether the intervention measures have truly led to improved vehicle behavior. Through a structured reasoning process, it provides a verifiable and traceable logical basis for system control decisions.

[0090] Typically, once an intervention strategy has been executed or submitted for confirmation, the interpretability analysis module initiates the modeling task. By constructing a causal relationship diagram between the intervention variables, vehicle response behavior variables, and environmental covariates, the system estimates and explains the actual effectiveness of the behavioral intervention.

[0091] In this embodiment, the interpretability analysis process mainly includes causal structure modeling, causal reasoning calculation, and decision explanation generation, as follows.

[0092] The causal structure modeling unit is used to define and construct the causal graph structure of intervention behavior variables, vehicle response behavior variables and key environmental factors.

[0093] Specifically, the system first divides the input variables into three categories: Intervention variables : Refers to the specific control strategies executed or recommended by the system, such as "voice warning", "route replanning" and "triggering security linkage", etc.; vehicle response behavior variables : refers to the observable behavioral changes after vehicle intervention, often including the decrease in abnormality score, the degree of path deviation relief, speed stability, etc.; environmental covariates Factors such as the vehicle's current road type, pedestrian density, time of day, and weather conditions influence the response but are not directly affected by the control strategy. Typically, the system trains a structured causal graph based on historical decision samples. The relationships between each variable node are modeled using a probabilistic graph structure, with directions and conditional dependencies marked. In one possible implementation, the system uses a structural learning algorithm (such as the Peter-Clark algorithm or the acyclic constraint algorithm) to perform data-driven determination of the existence of edges between variables, thereby constructing a complete directed acyclic graph.

[0094] The causal inference computation process is responsible for estimating the expected change in the vehicle response behavior variables under given intervention conditions, such as using The intervention effect is modeled using the algorithm. Assume: Indicates mandatory intervention Post-vehicle response behavior variables The conditional probability of Represents a set of environmental covariates with a range of ; Indicates the current distribution of system environment variables probability; Indicates that in a given and Under the premise of Conditional probability, calculated as follows: .

[0095] In general, the system builds a joint distribution model based on training samples , and perform integration or summation reasoning according to the above formula to output the expected effect of a certain control strategy in the current environment. In some embodiments, the device selection strategy is , the system can also compare and The numerical value of the evaluation strategy and alternative strategies The relative advantages and disadvantages under the same environmental conditions support strategy optimization and multi-strategy evaluation. The decision explanation generation unit is used to format and output the above causal reasoning results, and generate explanation labels and auxiliary explanation content for the intervention behavior. Specifically, the system organizes the following structural information into a decision report output: the selected intervention strategy Description and number of actual environment status Sampling values, such as crowd density, crowd risk level, time label, etc.; vehicle response behavior variables The direction of change and estimated value, such as a decrease in the anomaly score; causal effect explanation statements, such as "In the current high-flow environment, triggering a voice prompt can reduce the anomaly score by about 18%." As an option, the system can archive each decision statement together with the original state vector, model parameters, and graph structure information for subsequent review or review by security personnel. In some embodiments, to facilitate visual analysis, the system supports the display of reasoning links between variables in the form of a causal path diagram, such as: "voice prompt", "driving behavior deceleration", "deviation reduction" and "anomaly score reduction". In addition, it also supports the feedback of causal graph results to the intervention strategy module, which is used to modify the reward function structure in the reinforcement learning process and build a causal-strategy closed-loop feedback chain.

[0096] By integrating causal reasoning calculations, the present invention can quantitatively analyze the effectiveness of intervention measures and trace the causal logic behind policy triggering, achieving the dual goals of credible control interpretation and reasonable strategy optimization. Compared with existing solutions that generally use black box models to output control instructions but lack reasoning transparency, this method effectively solves the system trust problem.

[0097] Optionally, the local graph neural network model is trained separately in multiple deployment nodes, and the shared model parameters in the local graph neural network model training results are centrally aggregated and distributed and updated, including: Train the graph neural network model based on local data in each deployment node to obtain a local model; Upload the shared model parameters in the local model to the central server; Perform weighted average calculation on the shared model parameters uploaded by multiple deployment nodes and generate global shared model parameters; The global shared model parameters are distributed to each deployment node to update the local model.

[0098] The federated learning collaborative process is used to train the local graph neural network model separately in multiple deployment nodes, and to centrally aggregate and distribute the updates of the shared model parameters in the training results of the local graph neural network model.

[0099] Specifically, a federated learning collaboration module is set up in the model training layer of the smart campus safety management and early warning system. Its main function is to realize distributed graph neural network model training between nodes deployed on different campuses, thereby achieving knowledge sharing and model generalization without exchanging original data.

[0100] In general, each campus or subsystem will independently run a local version of the system architecture, receive local vehicle information, and complete the abnormality identification and intervention process. This module uses periodic training tasks to enable local systems to share global model parameters under the premise of safety, thereby improving the generalization ability and convergence efficiency of each model. In this embodiment, the federated learning collaborative process includes at least local model training, model parameter uploading, global model aggregation, and model update distribution. Specifically: the local model training unit is used to perform local training on the graph neural network model within a single campus node. Specifically, at each deployment node On top of that, the system builds local map data , and train a graph neural network model with local feature extraction capabilities. The internal weights of the model are divided into two parts: private parameters Only updated locally, not shared; shared parameters Participate in the global synchronization process. During local training, the system uses local data to perform forward propagation and backward updates, and the optimization objective is usually the anomaly score prediction error or node classification loss.

[0101] The model parameter upload process is used to submit the shared weights to the central coordination server; generally, the system uploads the model parameters at fixed time intervals. Initiate a synchronization request. After completing several rounds of local training, each node will Upload to the server. During the upload process, the system will also send the sample quantity information , used for subsequent aggregation weight allocation calculation.

[0102] The global model aggregation process is deployed on the central server to perform weighted fusion of shared parameters from each node. In a standard implementation, the federated averaging algorithm is used to complete the model aggregation. Assume: is the shared weight matrix after global aggregation, For the Shared parameters uploaded by each node; For the The number of training samples per node, is the total number of training samples for all nodes, is the total number of deployment nodes participating in this round of aggregation, then: .

[0103] As an option, the system can also set an aggregation weight adjustment mechanism based on factors such as node performance and data representativeness to dynamically adjust the parameter fusion contribution ratio.

[0104] The model update distribution process is responsible for sending the aggregated shared model parameters to all participating nodes. Specifically, after the system completes the aggregation calculation, Synchronously distributed to each campus subsystem, each node replaces the local old version of the shared parameters , forming a new round of hybrid model. The local private parameters remain unchanged, so as to ensure that the models in each region are still adaptable to the local scene characteristics. In some embodiments, in order to improve the convergence speed and stability, the system can use the sliding average method for parameter fusion setting: is the fusion weight coefficient, ranging from [0,1], It is the local shared parameter before updating. is the updated local shared parameter, namely: .

[0105] In some embodiments, the system supports multiple rounds of validation of global aggregation results. By deploying a validation graph dataset on the central side, the system can evaluate the performance of the current global model in typical scenarios, such as anomaly recognition accuracy, recall rate, and scoring error, thereby assisting in optimizing subsequent aggregation strategies. Furthermore, to protect the data privacy of each node, the system incorporates parameter encryption mechanisms and differential privacy processing during federated communication to prevent data content from being leaked through parameter reasoning.

[0106] Example 2 Based on the same principle as the method shown in Example 1 of the present invention, as shown in the attached Figure 3 As shown, an embodiment of the present invention further provides a smart campus safety management and early warning system based on big data, including an external vehicle information collection unit, a dynamic spatiotemporal graph construction unit, a path modeling and deviation analysis unit, an abnormal behavior measurement unit, an intervention strategy decision unit, an analysis and generation unit, and a federated learning collaboration unit; The external vehicle information collection unit is used to collect the registration information, real-time behavior data and status information of the environment in which the external vehicle is located; A dynamic spatiotemporal graph construction unit is used to construct a dynamic spatiotemporal graph containing deployment nodes using the collected data, and to generate edge attributes reflecting the relationships between the deployment nodes in the dynamic spatiotemporal graph; the deployment nodes include vehicles, roads, and people; A path modeling and deviation analysis unit is used to calculate a reasonable path set for an incoming vehicle based on a dynamic spatiotemporal graph, and to analyze the trajectory deviation data of the incoming vehicle based on the spatial relationship between the actual driving trajectory of the incoming vehicle and the reasonable path set; Abnormal behavior measurement unit, used to score the abnormality of foreign vehicles based on trajectory deviation data and real-time behavior data changes of foreign vehicles; An intervention strategy decision unit, configured to generate a corresponding intervention control strategy based on the anomaly score and the current state information of the foreign vehicle; Analysis and generation unit, used to perform causal impact analysis on intervention control strategies and generate decision basis; The federated learning collaborative unit is used to train the local graph neural network model in multiple deployment nodes separately, and to centrally aggregate and distribute the updates of the shared model parameters in the training results of the local graph neural network model.

[0107] Optionally, collect registration information of external vehicles, including license plate number, vehicle type, time of entry, location of entry, and destination of visit; Collecting real-time behavior data of external vehicles includes collecting GPS tracks, speeds, and accelerations of external vehicles on campus; Collecting the state information of the environment in which the external vehicle is located includes collecting the road state, pedestrian density, light intensity and construction area distribution information of the location of the external vehicle.

[0108] Optionally, edge-to-edge attributes of the dynamic spatiotemporal graph are generated based on the spatial distance, speed difference, and risk value between the deployed nodes.

[0109] Optionally, a reasonable path set for external vehicles is calculated based on a dynamic space-time graph, including: calculating the shortest path between the external vehicle's entry node and the access target node as a basic path in the dynamic space-time graph; setting a tolerance parameter for a preset basic path; and constructing a reasonable path set based on the path length of the basic path and the preset tolerance parameter.

[0110] Optionally, analyzing the trajectory deviation data of the foreign vehicle based on the spatial relationship between the actual driving trajectory of the foreign vehicle and the set of reasonable paths includes: Calculate the path deviation of the foreign vehicle based on its actual driving trajectory and the path on the selected road section in the reasonable path set; Constructing a historical speed distribution model corresponding to each foreign vehicle based on the historical data of each foreign vehicle; comparing the current speed distribution of the foreign vehicle with the historical speed distribution of the foreign vehicle to obtain a speed distribution difference of the foreign vehicle; According to the path deviation and speed distribution difference of the foreign vehicle, the track deviation data of the foreign vehicle is obtained.

[0111] Optionally, obtaining trajectory deviation data of the foreign vehicle according to the path deviation and speed distribution difference of the foreign vehicle includes: calculating an abnormality score of the foreign vehicle based on the path deviation and speed distribution difference; Assume that the anomaly score is , the average spatial deviation between the actual driving trajectory of the external vehicle and the reasonable path in the reasonable path set is , and is the preset non-negative real weighting coefficient, and the Kullback-Leibler divergence is , the current speed distribution of the external vehicle is , the historical speed distribution of foreign vehicles is , then the anomaly score is: .

[0112] Optionally, construct a state vector by combining the location information of the foreign vehicle, the graph embedding representation, the anomaly score, and the environment state information; Generate intervention control strategy based on state vector using reinforcement learning model; Intervention control strategies include voice prompts, route replanning, safety passes and road restrictions.

[0113] Optionally, conduct a causal impact analysis of the intervention control strategy and generate a basis for decision making, including: Construct a causal graph structure between intervention behavior variables, vehicle response behavior variables, and environmental variables; Based on the causal graph structure, the expectation of the vehicle response behavior variable under the intervention behavior variable is calculated as the causal reasoning result; Generate intervention decision description information based on causal reasoning results as the basis for making decisions.

[0114] Optionally, the local graph neural network model is trained separately in multiple deployment nodes, and the shared model parameters in the local graph neural network model training results are centrally aggregated and distributed and updated, including: Train the graph neural network model based on local data in each deployment node to obtain a local model; Upload the shared model parameters in the local model to the central server; Perform weighted average calculation on the shared model parameters uploaded by multiple deployment nodes and generate global shared model parameters; The global shared model parameters are distributed to each deployment node to update the local model.

[0115] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A smart campus safety management and early warning method based on big data, characterized by: include: Collect registration information, real-time behavior data and status information of the environment of foreign vehicles; Using the collected data, a dynamic spatiotemporal graph containing deployment nodes is constructed, and edge attributes reflecting the relationships between deployment nodes are generated in the dynamic spatiotemporal graph; Deployment nodes include vehicles, roads, and crowds; Calculate the reasonable path set of the foreign vehicle based on the dynamic space-time graph, and analyze the trajectory deviation data of the foreign vehicle based on the spatial relationship between the actual driving trajectory of the foreign vehicle and the reasonable path set; Based on the trajectory deviation data and the real-time behavior data changes of foreign vehicles, the foreign vehicles are scored as abnormal; Generate corresponding intervention control strategies based on the anomaly score and the current state information of the foreign vehicle; Conduct causal impact analysis on intervention and control strategies and generate decision-making basis; The local graph neural network model is trained separately in multiple deployment nodes, and the shared model parameters in the local graph neural network model training results are centrally aggregated and distributed for update.

2. The smart campus safety management early warning method based on big data according to claim 1 is characterized in that: Collect registration information of external vehicles, including license plate number, vehicle type, time of entry, location of entry, and destination of visit; Collecting real-time behavior data of external vehicles includes collecting GPS tracks, speeds, and accelerations of external vehicles on campus; Collecting the state information of the environment in which the external vehicle is located includes collecting the road state, pedestrian density, light intensity and construction area distribution information of the location of the external vehicle.

3. The smart campus safety management early warning method based on big data according to claim 1 is characterized in that: The edge-to-edge attributes of the dynamic spatiotemporal graph are generated based on the spatial distance, speed difference and risk value between the deployed nodes.

4. The smart campus safety management early warning method based on big data according to claim 1 is characterized in that: A reasonable path set for external vehicles is calculated based on a dynamic space-time graph, including: calculating the shortest path between the external vehicle's entry node and the access target node as the basic path in the dynamic space-time graph; setting the tolerance parameter of the preset basic path; and constructing a reasonable path set based on the path length of the basic path and the preset tolerance parameter.

5. The smart campus safety management early warning method based on big data according to claim 1 is characterized in that: Analyze the trajectory deviation data of foreign vehicles based on the spatial relationship between their actual driving trajectories and the set of reasonable paths, including: Calculate the path deviation of the foreign vehicle based on its actual driving trajectory and the path on the selected road section in the reasonable path set; Constructing a historical speed distribution model corresponding to each foreign vehicle based on the historical data of each foreign vehicle; comparing the current speed distribution of the foreign vehicle with the historical speed distribution of the foreign vehicle to obtain a speed distribution difference of the foreign vehicle; According to the path deviation and speed distribution difference of the foreign vehicle, the track deviation data of the foreign vehicle is obtained.

6. The smart campus safety management early warning method based on big data according to claim 5 is characterized in that: Obtaining trajectory deviation data of the foreign vehicle based on the path deviation and speed distribution difference of the foreign vehicle, including: calculating an abnormality score of the foreign vehicle based on the path deviation and speed distribution difference; Assume that the anomaly score is , the average spatial deviation between the actual driving trajectory of the external vehicle and the reasonable path in the reasonable path set is , and is the preset non-negative real weighting coefficient, and the Kullback-Leibler divergence is , the current speed distribution of the external vehicle is , the historical speed distribution of foreign vehicles is , then the anomaly score is: 。 7. The smart campus safety management early warning method based on big data according to claim 1 is characterized in that: The location information of the foreign vehicle, graph embedding representation, anomaly score and environment state information are constructed into a state vector; Generate intervention control strategy based on state vector using reinforcement learning model; Intervention control strategies include voice prompts, route replanning, safety passes and road restrictions.

8. The smart campus safety management early warning method based on big data according to claim 1 is characterized in that: Conduct causal impact analysis of intervention control strategies and generate a basis for decision making, including: Construct a causal graph structure between intervention behavior variables, vehicle response behavior variables, and environmental variables; Based on the causal graph structure, the expectation of the vehicle response behavior variable under the intervention behavior variable is calculated as the causal reasoning result; Generate intervention decision description information based on causal reasoning results as the basis for making decisions.

9. The smart campus safety management early warning method based on big data according to claim 1 is characterized in that: The local graph neural network model is trained separately in multiple deployment nodes, and the shared model parameters in the local graph neural network model training results are centrally aggregated and distributed for update, including: Train the graph neural network model based on local data in each deployment node to obtain a local model; Upload the shared model parameters in the local model to the central server; Perform weighted average calculation on the shared model parameters uploaded by multiple deployment nodes and generate global shared model parameters; The global shared model parameters are distributed to each deployment node to update the local model.

10. The smart campus safety management and early warning system based on big data is characterized by: It includes an external vehicle information collection unit, a dynamic space-time graph construction unit, a path modeling and deviation analysis unit, an abnormal behavior measurement unit, an intervention strategy decision unit, an analysis and generation unit, and a federated learning collaboration unit; The external vehicle information collection unit is used to collect the registration information, real-time behavior data and status information of the environment in which the external vehicle is located; A dynamic spatiotemporal graph construction unit is used to construct a dynamic spatiotemporal graph containing deployment nodes using the collected data, and to generate edge attributes reflecting the relationships between the deployment nodes in the dynamic spatiotemporal graph; the deployment nodes include vehicles, roads, and people; A path modeling and deviation analysis unit is used to calculate a reasonable path set for an incoming vehicle based on a dynamic spatiotemporal graph, and to analyze the trajectory deviation data of the incoming vehicle based on the spatial relationship between the actual driving trajectory of the incoming vehicle and the reasonable path set; Abnormal behavior measurement unit, used to score the abnormality of foreign vehicles based on trajectory deviation data and real-time behavior data changes of foreign vehicles; An intervention strategy decision unit, configured to generate a corresponding intervention control strategy based on the anomaly score and the current state information of the foreign vehicle; Analysis and generation unit, used to perform causal impact analysis on intervention control strategies and generate decision basis; The federated learning collaborative unit is used to train the local graph neural network model in multiple deployment nodes separately, and to centrally aggregate and distribute the updates of the shared model parameters in the training results of the local graph neural network model.

Images