Smart park-oriented intelligent three-dimensional security and protection monitoring method and system

Abstract

The invention relates to a smart park-oriented intelligent three-dimensional security and protection monitoring method and system, and relates to the technical field of security and protection monitoring, and the method comprises the steps: constructing a multi-dimensional three-dimensional sensing network, obtaining multi-modal heterogeneous data, and carrying out the preprocessing of the multi-modal heterogeneous data; performing parallel post-processing on the time-space diagram neural network and the Bayesian network, performing adaptive fusion through a gating fusion mechanism at a decision-making layer, and constructing a security and protection monitoring model; inputting the preprocessed data into the security and protection monitoring model, and respectively generating a dynamic feature vector and a rule confidence vector; and carrying out weighted fusion on the two vectors through a decision level adaptive gating fusion mechanism, and outputting a real-time alarm level, event type probability distribution and an evidence abstract.

Description

Technical Field

[0001] The present invention relates to the field of security monitoring technology, and in particular to an intelligent three-dimensional security monitoring method and system for a smart park. Background Art

[0002] Existing industrial park security systems generally face multiple technical bottlenecks, which are severely restricting the effectiveness of smart park security operations. Recent reports on the security monitoring industry indicate that over half of security incidents in my country's industrial parks are directly related to delayed security system responses, while nearly half of property losses in science and technology parks are closely related to the waste of disposal resources caused by false alarms. This data not only reveals the failure of security systems at critical moments but also highlights the deep structural flaws of the existing technology system, which is unable to meet the increasingly complex security needs of industrial parks.

[0003] A thorough analysis of the existing security system reveals that the monolithic nature of its technical architecture is the key to the aforementioned problems. Statistics show that as many as 89% of existing campus security systems overly rely on two-dimensional video surveillance. This reliance presents numerous drawbacks. The effective monitoring height of traditional cameras is typically limited to 3 to 10 meters, making the system incapable of detecting high-altitude risks such as drone intrusions and leaks from overhead pipelines, creating a "vertical gap." Furthermore, only a very small number of campuses have deployed underground sensing equipment, preventing real-time monitoring of underground hazards such as cable trench theft and hazardous chemical pipeline leaks, creating a "underground gap." Even more worryingly, existing security systems often provide less than half of the coverage of hidden areas such as building mezzanines and equipment rooms. These areas, in turn, become hotbeds for illegal intrusion, creating "hidden blind spots." This coexistence of vertical gaps, underground gaps, and hidden blind spots creates a "triple defense gap."

[0004] Single-layer security monitoring, whether high-altitude, on the ground, or underground, cannot provide comprehensive security. This flat, fragmented monitoring model creates multiple weak links within the entire security system; failure in any one link can lead to the collapse of the entire system. Therefore, existing security systems not only present significant security risks but also struggle to meet the urgent demands of smart parks for security, efficiency, and intelligence. A more comprehensive, multi-dimensional, and advanced technological solution is urgently needed. Summary of the Invention

[0005] In order to overcome the shortcomings of the above-mentioned prior art, the main purpose of the present invention is to provide an intelligent three-dimensional security monitoring method and system for smart parks. By constructing a multi-dimensional perception network, innovatively integrating spatiotemporal graph neural networks and Bayesian networks, and designing an adaptive decision-making mechanism, efficient, accurate and explainable park security monitoring is achieved.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: an intelligent three-dimensional security monitoring method for a smart park, comprising the following steps: A multi-dimensional stereoscopic perception network is constructed using a drone perception network, a ground fixed sensor network, and a mobile IoT perception network. Based on the stereoscopic perception network, multimodal heterogeneous data is acquired and preprocessed. The spatiotemporal graph neural network and the Bayesian network are parallelized and adaptively integrated at the decision layer to build a security monitoring model; Inputting the preprocessed multimodal heterogeneous data into the security monitoring model, processing the dynamic time series data in the preprocessed multimodal heterogeneous data through a spatiotemporal graph neural network to obtain a dynamic feature vector, processing the state event data through a differentiable conditional probability model, performing rule reasoning based on the conditional probability table generated by the expert library rules, and outputting a rule confidence vector; A decision-level adaptive gated fusion mechanism is designed to perform adaptive weighted fusion on the dynamic feature vector and the rule confidence vector to generate multi-level decision outputs, including real-time alarm level, event type probability distribution and evidence summary, as security monitoring results.

[0007] The construction of the multi-dimensional stereoscopic perception network includes: The drone perception network adopts a heterogeneous cluster collaborative architecture, including visible light drones, infrared temperature measurement drones and laser scanning drones; The ground fixed sensor network deploys multispectral video surveillance nodes; The mobile IoT sensing network tracks personnel through UWB / BLE dual-mode positioning tags.

[0008] The method of processing the dynamic time series data in the preprocessed multimodal heterogeneous data through a spatiotemporal graph neural network to obtain a dynamic feature vector includes the following steps: Taking key points of the monitoring area as nodes, taking the spatiotemporal associations between nodes as edges, and constructing the nodes, edges and adjacency matrix into a spatiotemporal graph based on the adjacency matrix obtained from the dynamic time series data; The spatiotemporal graph convolution layer in the spatiotemporal graph neural network extracts the spatial features of the spatiotemporal graph and obtains the spatiotemporal features through the gated temporal convolution layer; the spatiotemporal features are encoded using the encoder of the spatiotemporal graph neural network to generate a dynamic feature vector; The process of processing state event data through a differentiable conditional probability model, performing rule reasoning based on a conditional probability table generated based on expert library rules, and outputting a rule confidence vector includes the following steps: The security-related events are used as nodes of the Bayesian network, and the probability dependency relationships defined by the expert library rules are used as edges to construct the Bayesian network topology. Based on the network topology and initial conditional probability table defined by the expert library rules, a differentiable conditional probability model is constructed, in which the conditional probability table of each non-root node is obtained through neural network parameterization; Mapping the events and state record data in the state event data to corresponding leaf nodes in the Bayesian network topology, encoding the observation data of the events and states into embedding vectors of the corresponding leaf nodes; and using the corresponding leaf nodes and the embedding vectors of the corresponding leaf nodes as evidence vectors; Performing probabilistic reasoning using the evidence vector and the differentiable conditional probability model to obtain a posterior probability distribution of the target event node; The confidence of the target event is extracted according to the posterior probability distribution to generate a rule confidence vector.

[0009] Each non-root node adopts a hierarchical neural network structure to model its conditional probability table, including: The first layer of sub-network groups the parent node states according to logical correlation and outputs the intermediate representation of each group of parent nodes; The second layer fusion network concatenates the intermediate representations and inputs them, outputting the state probability distribution of the node; The probabilistic reasoning adopts a differentiable approximate reasoning algorithm.

[0010] The rule confidence vector is generated as follows: The target event node is a binary state, and its positive state probability is taken as the confidence level; The target event node is a multivariate state, and the confidence of its complete state probability vector is taken.

[0011] The gated fusion mechanism includes: Dimensionally aligning and concatenating the dynamic feature vector and the rule confidence vector to generate a fusion input vector; Generate adaptive weights for the dynamic feature vector and the rule confidence vector through a gated neural network; wherein the gated neural network structure includes two fully connected layers, the first fully connected layer uses a ReLU activation function, and the second fully connected layer uses a Softmax function; the adaptive weights include a dynamic feature vector weight α and a rule confidence vector weight β, and satisfy α + β = 1; Based on the adaptive weight, performing weighted fusion on the dynamic feature vector and the rule confidence vector to generate a fused feature vector; A multi-task decision head is introduced, and the fused feature vector is input into the multi-task decision head to generate a multi-level decision output.

[0012] The multi-task decision head includes: Output the continuous alarm level value through the fully connected layer as the alarm level generation branch; The fully connected layer and the Softmax function are used to output the probability distribution of event types as the event classification branch: Extracting key evidence from the dynamic feature vector and the rule confidence vector based on the attention weight, and generating a structured text summary as an evidence summary generation branch; The generation of the evidence summary includes: Obtaining the attention weights of the fused feature vector, the dynamic feature vector, and the rule confidence vector; Select the top N features with the highest attention weights as key evidence; Key evidence is filled in according to a predefined template to generate a structured text summary; wherein the predefined template includes an event type field, a confidence field, and an evidence source field.

[0013] Compared with existing technologies, the present invention achieves the following beneficial effects: By integrating a multi-dimensional sensing network consisting of drones, ground-based fixed systems, and mobile IoT, it solves the "triple protection gap" problem inherent in traditional security systems. This enables comprehensive coverage and real-time monitoring of the campus's high-altitude, ground-based, underground, and hidden spaces, significantly enhancing the breadth and depth of security protection and effectively preventing various security incidents. By employing a spatiotemporal graph neural network to process dynamic time-series data, it accurately captures the temporal and spatial evolution and correlations of events, extracting high-dimensional dynamic features. Furthermore, by combining a differentiable conditional probability model with a Bayesian network for rule-based reasoning, expert experience is effectively incorporated into the model, enabling parallel processing and deep fusion of data-driven and rule-driven approaches, improving understanding of complex scenarios and decision-making accuracy. Furthermore, the designed decision-level adaptive gating fusion mechanism dynamically adjusts the weights of data-driven features and rule confidence based on the varying characteristics of the input data, generating more accurate and reliable multi-level decision outputs. This ensures optimal responses under varying security scenarios and enhances the intelligence and adaptability of decision-making.

[0014] In order to enhance the interpretability and credibility of security monitoring, the present invention uses the rule reasoning process of the Bayesian network and the dynamic adjustment of weights in the gated fusion mechanism, so that the final decision is not only based on the data pattern, but also integrates expert rules, making the decision-making process more transparent. In particular, the evidence summary generation module can extract key evidence from massive data and present it in a structured manner, providing security personnel with an intuitive and reliable decision-making basis, significantly improving the interpretability of the security system and user trust. The present invention not only outputs traditional alarm signals, but also generates continuous alarm levels, fine-grained event type probability distributions, and structured evidence summaries. The refined output method enables security management personnel to grasp the event situation more comprehensively and accurately, providing more specific and powerful support for subsequent resource scheduling and emergency response, and improving the overall security operation efficiency.

[0015] In summary, through technological innovation, the present invention effectively solves many pain points of existing smart park security monitoring, realizes the transition from passive response to active defense, and from single monitoring to intelligent decision-making, providing strong technical guarantee for the safe and stable operation of smart parks. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The drawings described herein are used to provide further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute improper limitations on the present application.

[0017] Figure 1 It is a schematic flow chart of the present invention.

[0018] Figure 2 It is a schematic diagram of the flow chart of the architecture of the present invention.

[0019] Figure 3 It is a schematic diagram of the stereoscopic perception network of the present invention.

[0020] Figure 4 It is a schematic diagram of the spatiotemporal graph neural network processing flow of the present invention.

[0021] Figure 5 Schematic diagram of the Bayesian network processing flow of the present invention.

[0022] Figure 6 It is a schematic diagram of the structure of the hierarchical conditional probability table of the present invention.

[0023] Figure 7 Schematic diagram of the gated fusion process of the present invention.

[0024] Figure 8 It is a schematic diagram of the multi-task decision-making process of the present invention. DETAILED DESCRIPTION

[0025] Smart campuses integrate a vast array of advanced facilities and complex activities, posing unprecedented challenges to security and monitoring. Traditional security systems often rely on a single monitoring method, such as fixed cameras, which can lead to blind spots, data silos, and delayed analysis. Single-layer security monitoring, whether at high altitude, on the ground, or underground, struggles to provide comprehensive security. This flat, fragmented monitoring model creates multiple weak links throughout the security system, and failure in any one link can lead to the collapse of the entire system. Therefore, existing security systems not only present significant security risks but also struggle to meet the urgent demands of smart campuses for safety, efficiency, and intelligence. A more comprehensive, multi-dimensional, and advanced technological solution is urgently needed.

[0026] To address these challenges, this paper proposes an intelligent, three-dimensional security monitoring method for smart campuses. This method breaks down data barriers and achieves full, multi-dimensional monitoring coverage by building a multi-dimensional, three-dimensional perception network. By deeply fusing spatiotemporal graph neural networks with Bayesian networks, this method leverages the data-driven deep learning capabilities of surveillance data to capture complex patterns while also integrating rule-driven expert knowledge to improve decision reliability. Through adaptive gating fusion and multi-task decision head design, this method generates rich, reliable, and easy-to-understand security monitoring results. Figures 1 to 8 , it can be seen that the specific steps include: Using the drone perception network, ground fixed sensor network and mobile IoT perception network, a multi-dimensional three-dimensional perception network is constructed. The specific steps include: The drone perception network adopts a heterogeneous cluster collaborative architecture, including visible light drones, infrared temperature measurement drones, and laser scanning drones; Deploy multispectral video surveillance nodes in ground-based fixed sensor networks; The mobile IoT sensing network tracks people through UWB and BLE dual-mode positioning tags.

[0027] According to the constructed multi-dimensional stereoscopic perception network, multimodal heterogeneous data is acquired and preprocessed to obtain preprocessed multimodal heterogeneous data; The spatiotemporal graph neural network and the Bayesian network are parallelized and adaptively integrated at the decision layer to build a security monitoring model; The preprocessed multimodal heterogeneous data is input into the security monitoring model. The dynamic time series data in the preprocessed multimodal heterogeneous data is processed by the spatiotemporal graph neural network to obtain the dynamic feature vector, which is specifically expressed as follows: Taking key points of the monitoring area as nodes, taking the spatiotemporal associations between nodes as edges, and constructing the nodes, edges and adjacency matrix into a spatiotemporal graph based on the adjacency matrix obtained from the dynamic time series data; The spatiotemporal graph convolution layer in the spatiotemporal graph neural network extracts the spatial features of the spatiotemporal graph and obtains the spatiotemporal features through the gated temporal convolution layer; the spatiotemporal features are encoded using the encoder of the spatiotemporal graph neural network to generate a dynamic feature vector; The state event data is then processed using a differentiable conditional probability model. Rule reasoning is performed based on the conditional probability table generated by the expert base rules, and the rule confidence vector is output. Specifically, the Bayesian network topology is constructed by treating security-related events as nodes and the probabilistic dependencies defined by the expert base rules as edges. Based on the network topology and initial conditional probability table defined by the expert library rules, a differentiable conditional probability model is constructed, wherein the conditional probability table of each non-root node is obtained through neural network parameterization; for each non-root node, a hierarchical neural network structure is used to model its conditional probability table, including the following steps: The first layer of sub-network groups the parent node states according to logical correlation and outputs the intermediate representation of each group of parent nodes; The second layer fusion network concatenates the intermediate representations and inputs them, outputting the state probability distribution of the node; Probabilistic reasoning uses a differentiable approximate reasoning algorithm.

[0028] Mapping the events and state record data in the state event data to corresponding leaf nodes in the Bayesian network topology, encoding the observation data of the events and states into embedding vectors of the corresponding leaf nodes; and using the corresponding leaf nodes and the embedding vectors of the corresponding leaf nodes as evidence vectors; Performing probabilistic reasoning using the evidence vector and the differentiable conditional probability model to obtain a posterior probability distribution of the target event node; The confidence of the target event is extracted according to the posterior probability distribution to generate a rule confidence vector. The rule confidence vector is generated as follows: The target event node is a binary state, and its positive state probability is taken as the confidence level; The target event node is a multivariate state, and the confidence of its complete state probability vector is taken.

[0029] In the final decision-making phase, it is necessary to design a gated fusion mechanism that is adaptive at the decision-making level. This mechanism performs adaptive weighted fusion on the dynamic feature vector and the rule confidence vector to generate multi-level decision outputs, including real-time alarm levels, event type probability distributions, and evidence summaries, as security monitoring results. The design of the gated fusion mechanism includes: Dimensionally aligning and concatenating the dynamic feature vector and the rule confidence vector to generate a fusion input vector; Generate adaptive weights for the dynamic feature vector and the rule confidence vector through a gated neural network; wherein the gated neural network structure includes two fully connected layers, the first fully connected layer uses the ReLU activation function, and the second fully connected layer uses the Softmax function; the adaptive weights include the dynamic feature vector weight α and the rule confidence vector weight β, and satisfy α + β = 1; Based on the adaptive weight, the dynamic feature vector and the rule confidence vector are weightedly fused to generate a fused feature vector; A multi-task decision head is introduced, and the fused feature vector is input into the multi-task decision head to generate a multi-level decision output.

[0030] The introduced multi-task decision head includes: Output the continuous alarm level value through the fully connected layer as the alarm level generation branch; The fully connected layer and the Softmax function are used to output the probability distribution of event types as the event classification branch: Based on the attention weight, extract key evidence from the dynamic feature vector and the rule confidence vector, and generate a structured text summary as an evidence summary generation branch; Evidence summary generation includes: Obtain the attention weights of the fusion feature vector, dynamic feature vector, and rule confidence vector; Select the top N features with the highest attention weights as key evidence; Key evidence is filled in according to a predefined template to generate a structured text summary; wherein the predefined template includes an event type field, a confidence field, and an evidence source field.

[0031] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0032] Example 1: The following is an example of an intelligent three-dimensional security monitoring method for a smart park, combined with detailed technical details to better understand the implementation methods and application scenarios of each link. In the security monitoring system of the smart park, firstly, a multi-dimensional three-dimensional perception network is used to achieve comprehensive monitoring of the park environment. Figures 1 to 8 It can be seen that the perception network consists of the following parts: The first step is to build a three-dimensional perception network, which includes: The drone perception network is built by integrating multiple types of drones to form a heterogeneous swarm perception network. Visible light drones, infrared temperature measurement drones, and laser scanning drones are responsible for different data collection tasks.

[0033] Visible light drones are responsible for acquiring real-time image data and conducting real-time monitoring.

[0034] Infrared temperature measurement drones are responsible for monitoring temperature changes, especially in areas with large temperature fluctuations within the park.

[0035] Laser scanning drones are used for high-precision mapping to generate three-dimensional models of the park, providing detailed data support for spatial layout and safety protection.

[0036] A ground-based fixed sensor network deploys multispectral video surveillance nodes to cover key areas within the campus. These sensors perform intelligent analysis by fusing video and infrared data.

[0037] The mobile IoT sensing network uses UWB / BLE dual-mode positioning tags for personnel tracking, which can accurately identify the movement trajectories of people within the park.

[0038] The collected multimodal data, including images, infrared videos, and temperature data, are preprocessed. Through noise removal and data standardization, the quality and accuracy of the data input into the subsequent model are ensured.

[0039] The integration of a spatiotemporal graph neural network and a Bayesian network begins with the construction of the spatiotemporal graph neural network and the extraction of dynamic features. A spatiotemporal graph is constructed using key locations within the park, such as gates and monitoring points, as nodes. The connections between nodes reflect their temporal and spatial correlations. A spatiotemporal graph convolutional layer extracts the spatial features of the nodes, while a gated temporal convolutional layer extracts the temporal features between nodes. An encoder encodes the spatiotemporal features to generate a dynamic feature vector. This dynamic feature vector contains key information from the time series data, reflecting real-time dynamic changes within the park.

[0040] Security events such as fire and intrusion are used as nodes in the Bayesian network. The dependencies between nodes are determined by the rules of the expert library to construct the topological structure of the Bayesian network.

[0041] Based on the Bayesian network, the conditional probability table of each non-root node is parameterized using a neural network, the posterior probability of the event is calculated through inference, and the rule confidence vector is output.

[0042] The dynamic feature vector and rule confidence vector are dimensionally aligned and concatenated to generate a fused input vector.

[0043] Adaptive weights are assigned to each vector through a gated neural network. The gated neural network consists of two fully connected layers. The first layer uses the ReLU activation function, and the second layer uses the Softmax function. Ultimately, the dynamic feature vector weight (α) and the rule confidence vector weight (β) are derived, satisfying α + β = 1.

[0044] Based on the adaptive weights α and β, the dynamic feature vector and the rule confidence vector are weighted fused to generate a fused feature vector.

[0045] The fused feature vector is input into the multi-task decision head for subsequent decision making.

[0046] Alarm level generation: Outputs continuous alarm level values through a fully connected layer.

[0047] Event type classification: Output the probability distribution of event types through the fully connected layer and the Softmax function.

[0048] Evidence summary generation: Based on the attention mechanism, key evidence is extracted from the fused feature vector to generate a structured text summary.

[0049] To generate an evidence summary, we first obtain the attention weights of the dynamic feature vector and the rule confidence vector. We select the top N features with the highest weights as key evidence. Using predefined templates such as event type, confidence level, and evidence source, we populate the key evidence and generate a structured text summary for further analysis and decision-making.

[0050] Based on the alarm level, event type probability distribution and evidence summary output by the decision head, real-time security monitoring results are generated.

[0051] Real-time alert level: Based on dynamically changing environmental data, the alert level is given in real time, such as high risk and low risk.

[0052] Event type probability distribution, determines the type of event and gives the corresponding probability distribution such as fire, intrusion, etc.

[0053] Evidence summary: Outputs a structured evidence summary containing information such as event type, confidence level, and evidence source to facilitate decision-making by security personnel.

[0054] Through the above embodiments, a powerful intelligent security monitoring solution can be provided for smart parks. Combining the multi-dimensional perception of drones and ground sensors with efficient data processing technology, it can provide security managers with real-time decision support and early warning information.

[0055] Example 2: This example illustrates the application of the present invention to the deployment of an intelligent, three-dimensional security monitoring system in the XX Science and Technology Park. The park, covering approximately 2 square kilometers, includes offices, R&D buildings, production workshops, warehouses, an energy center, and public green space. The park has complex personnel flows and high-value assets, placing stringent security requirements.

[0056] First, a multi-dimensional stereoscopic perception network is built in the science and technology park: The drone perception network is deployed at three fixed takeoff and landing points and employs a heterogeneous fleet of six drones from three different categories. Two visible light drones, equipped with 30x optical zoom cameras, are used for wide-area patrols and detailed identification. They have a cruising speed of 8 m / s, a flight time of 30 minutes, and a single-shot coverage area of approximately 0.5 square kilometers. One infrared temperature measurement drone, equipped with an 80mm infrared lens and a temperature measurement accuracy of ±2°C, is used for night patrols and equipment thermal anomaly detection. It has a cruising speed of 6 m / s and a flight time of 25 minutes. One laser scanning drone, equipped with a 250-line laser radar and a point cloud density of 0.1 points / cm², is used for perimeter terrain modeling and illegal structure detection. It has a flight speed of 5 m / s and a flight time of 20 minutes.

[0057] Heterogeneous drones collaborate in real-time via 5G networks with ground control centers, dynamically deploying patrols based on pre-set patrol routes and real-time alerts. When an infrared drone detects an abnormal temperature in a transformer area, it immediately notifies a visible light drone to conduct image verification. Meanwhile, a laser drone simultaneously scans the surrounding environment.

[0058] A ground-based fixed sensor network has been deployed along the park perimeter, along major roads, at key entrances and exits, and on building rooftops. A total of 50 multispectral video surveillance nodes are deployed, including 40 high-definition visible light cameras with a resolution of 4MP and 10 thermal imaging cameras with a temperature measurement range of -40°C to 650°C. Some cameras feature starlight-level night vision. Coverage ensures no blind spots, and critical areas such as warehouses and data centers have 200% surveillance redundancy. Auxiliary sensors are also deployed along the perimeter wire fence and along key pipelines, including vibration sensors to detect shearing and climbing, and water sensors to detect leaks.

[0059] For the mobile IoT sensing network, all 25 security personnel within the campus were equipped with UWB / BLE dual-mode positioning tags. UWB, with an average error of 0.8 meters, provides sub-meter high-precision positioning, used to pinpoint personnel locations during emergency calls. BLE is used for medium-range (10-30 meters) tracking of personnel activity areas and regional alerts. The tags report location data every 2 seconds.

[0060] All data generated by sensing devices, including video streams, infrared images, point cloud data, vibration signals, and personnel location coordinates, first enters the edge computing node for preliminary processing. The video stream is used for moving target detection (such as people and vehicles), and metadata such as target trajectory and speed are extracted to reduce transmission bandwidth requirements. Infrared data is used for temperature anomaly area segmentation. Point cloud data undergoes dimensionality reduction and feature extraction. Sensor data is filtered and denoised. The pre-processed multimodal heterogeneous data, including target trajectory, temperature anomaly coordinates, point cloud features, vibration intensity, and personnel location, is encapsulated into a unified data package and sent to the central security monitoring platform.

[0061] 100 key monitoring points within the campus, such as camera locations, access control points, and sensor locations, are considered nodes. Edge weights between nodes are normalized based on historical event correlations, such as the frequency of simultaneous events between two points and the physical distance between them. A dynamic adjacency matrix is used to represent the current node connectivity.

[0062] In this example, the preprocessed dynamic time series data for each node over the past five minutes, including target trajectories, personnel flow speed, and equipment vibration frequency, is selected as input. A variant of the spatiotemporal graph convolutional layer (GCN) extracts spatial features and event coordination patterns in adjacent regions. A variant of the gated temporal convolutional layer (LSTM) processes the time series, capturing the temporal dependencies of event development and generating a continuous action sequence of intrusion behavior. A two-layer Transformer encoder is then used to encode the fused spatiotemporal features, generating a 128-dimensional dynamic feature vector that represents the overall dynamic state of the park.

[0063] To construct the Bayesian network, key security event nodes were first defined, including "intrusion," "fire," "equipment failure," and "crowd gathering," as well as related status nodes, including "access control anomaly," "temperature exceeding standard," "violent vibration," and "number of people in area exceeding threshold." Based on expert knowledge, including security regulations and equipment maintenance manuals, dependencies between nodes were defined. For example, "access control anomaly" and "number of people in area exceeding threshold" both point to an "intrusion" event. A Bayesian network topology consisting of 15 nodes was constructed. A hierarchical neural network structure was used to model the conditional probability table (CPT) for each non-root node. The first-layer subnetwork grouped parent node states into 2-3 groups based on logical dependencies, such as temporal correlation and spatial proximity. Each group of parent node states was encoded, and an intermediate representation (32 dimensions) was output. The second-layer fusion network concatenated the intermediate representations, resulting in a total dimension of 96. This representation was then fed into a fully connected layer, which output a probability distribution for the node's state, such as a binary distribution of "normal" / "abnormal" or a multivariate distribution of specific fault types.

[0064] The preprocessed state event data, namely the number of consecutive access control failures, the average infrared temperature in a certain area, the reading of a pipeline vibration sensor, and the occupant density in a certain area as indicated by UWB positioning, are mapped to the corresponding leaf nodes in the Bayesian network and encoded as embedding vectors with a dimension of 16. The leaf node identifier and the embedding vector are used as the evidence vector. A differentiable variational inference algorithm is used. In this embodiment, KL divergence minimization is used for approximate inference to calculate the posterior probability distribution of the target event node.

[0065] For binary state events, such as "intrusion," the probability value of the "abnormal" state in its posterior probability distribution is used as the confidence level. For multi-state events, such as the specific type of "equipment failure," the complete state probability vector is output as the confidence level vector. Ultimately, a rule confidence level vector of length 20 is generated, corresponding to 20 predefined security events and their confidence levels.

[0066] The 128-dimensional dynamic feature vector and the 20-dimensional rule confidence vector are dimensionally aligned, the rule confidence vector is replicated 6 times, or the dimension is increased to 128 through linear transformation, and then spliced into a 256-dimensional fusion input vector.

[0067] The fused input vector is fed into a gated neural network. The network consists of two fully connected layers: First layer: 256 input -> 64 output, activation function ReLU.

[0068] Second layer: 64 inputs -> 2 outputs (α, β), with Softmax activation function to ensure α + β = 1. For example, the current input causes the network to output α = 0.7 and β = 0.3, indicating that the dynamic feature vector contributes more.

[0069] Weighted fusion: Calculate fusion feature vector = 0.7 Dynamic eigenvector +0.3 Rule confidence vector (dimension aligned).

[0070] The fused feature vector with a dimension of 128 is input into the multi-task decision head: After passing through a fully connected layer (128->1), it outputs a continuous alarm level value between 0 and 10. The alarm level generation branch outputs 7.2, indicating that the current security risk is high.

[0071] A fully connected layer (128->20) is connected to a Softmax function to output a 20-dimensional probability distribution vector for the event type. The event classification branch generates [0.01, 0.02, 0.85, 0.05, ..., 0.01], indicating that the probability of an "intrusion" event is the highest (85%).

[0072] The evidence summary generation branch needs to obtain attention weights and extract key evidence. An attention mechanism is designed. It takes the fused feature vector as input and calculates its similarity with the original dynamic feature vector and the original rule confidence vector to generate the respective attention weight vectors (dimensions 128 and 20). The attention weights for the dynamic feature vector are then used to select the features corresponding to the top five dimensions with the highest weights; the attention weights for the rule confidence vector are used to select the confidence values corresponding to the top three events with the highest weights. These features / events are considered key evidence.

[0073] Generate text based on a predefined template: Event Type: Intrusion, Confidence: 85%.

[0074] Key evidence: Source of evidence: visible light from a drone, content of evidence: suspicious persons were monitored climbing over a wall.

[0075] Evidence source: UWB positioning. Evidence content: The population density in the walled area has increased abnormally.

[0076] Evidence source: access control system, evidence content: access control on a certain side failed 3 times in a row.

[0077] Evidence source: spatiotemporal pattern, evidence content: There has been a high incidence of recent invasions in this area.

[0078] Evidence source: thermal imaging, evidence content: a mobile heat source appeared outside the wall.

[0079] The summary visually presents the cause and basis of the alert.

[0080] System performance: After being deployed at the XX Science and Technology Park, the system successfully issued warnings for five real intrusion incidents during a three-month test, shortening the average response time to 45 seconds. The false alarm rate was reduced by 60% compared to traditional systems, effectively reducing ineffective deployments of security personnel. The generated structured evidence summary helped security supervisors quickly understand the full picture of the incident, improving decision-making efficiency by 40%.

[0081] It should be noted that, in the present invention, relational terms such as first and second, etc. are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "include," "comprise," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0082] The above embodiments are merely examples of the present invention and do not limit the scope of protection of the present invention. Any designs that are identical or similar to the present invention fall within the scope of protection of the present invention.

Claims

1. An intelligent three-dimensional security monitoring method for smart parks, characterized in that: The following steps are involved: Utilize the drone perception network, the ground fixed sensor network, and the mobile IoT perception network to build a multi-dimensional stereoscopic perception network, and acquire and pre-process multimodal heterogeneous data based on the stereoscopic perception network; The spatiotemporal graph neural network and the Bayesian network are parallelized and adaptively integrated at the decision layer to build a security monitoring model; Inputting the preprocessed multimodal heterogeneous data into the security monitoring model, processing the dynamic time series data in the preprocessed multimodal heterogeneous data through a spatiotemporal graph neural network to obtain a dynamic feature vector, processing the state event data through a differentiable conditional probability model, performing rule reasoning based on the conditional probability table generated by the expert library rules, and outputting a rule confidence vector; A decision-level adaptive gated fusion mechanism is designed to perform adaptive weighted fusion on the dynamic feature vector and the rule confidence vector to generate multi-level decision outputs, including real-time alarm level, event type probability distribution and evidence summary, as security monitoring results.

2. The intelligent three-dimensional security monitoring method for a smart park according to claim 1, characterized in that: The multi-dimensional stereoscopic perception network is constructed, comprising the following steps: The drone perception network adopts a heterogeneous cluster collaborative architecture, including visible light drones, infrared temperature measurement drones and laser scanning drones; The ground fixed sensor network deploys multispectral video surveillance nodes; The mobile IoT sensing network tracks people through UWB and BLE dual-mode positioning tags.

3. The intelligent three-dimensional security monitoring method for a smart park according to claim 1, characterized in that: The method of processing the dynamic time series data in the preprocessed multimodal heterogeneous data through a spatiotemporal graph neural network to obtain a dynamic feature vector includes the following steps: Taking key points of the monitoring area as nodes, taking the spatiotemporal associations between nodes as edges, and constructing the nodes, edges and adjacency matrix into a spatiotemporal graph based on the adjacency matrix obtained from the dynamic time series data; The spatiotemporal graph convolution layer in the spatiotemporal graph neural network extracts the spatial features of the spatiotemporal graph and obtains the spatiotemporal features through the gated temporal convolution layer; the spatiotemporal features are encoded using the encoder of the spatiotemporal graph neural network to generate a dynamic feature vector; The process of processing state event data through a differentiable conditional probability model, performing rule reasoning based on a conditional probability table generated based on expert library rules, and outputting a rule confidence vector includes the following steps: The security-related events are used as nodes of the Bayesian network, and the probability dependency relationships defined by the expert library rules are used as edges to construct the Bayesian network topology. Based on the network topology and initial conditional probability table defined by the expert library rules, a differentiable conditional probability model is constructed, in which the conditional probability table of each non-root node is obtained through neural network parameterization; Mapping the events and state record data in the state event data to corresponding leaf nodes in the Bayesian network topology, encoding the observation data of the events and states into embedding vectors of the corresponding leaf nodes; and using the corresponding leaf nodes and the embedding vectors of the corresponding leaf nodes as evidence vectors; Performing probabilistic reasoning using the evidence vector and the differentiable conditional probability model to obtain a posterior probability distribution of the target event node; The confidence of the target event is extracted according to the posterior probability distribution to generate a rule confidence vector.

4. The intelligent three-dimensional security monitoring method for a smart park according to claim 3, characterized in that: For each non-root node, a hierarchical neural network structure is used to model its conditional probability table, including the following steps: The first layer of sub-network groups the parent node states according to logical correlation and outputs the intermediate representation of each group of parent nodes; The second layer fusion network concatenates the intermediate representations and inputs them, outputting the state probability distribution of the node; The probabilistic reasoning adopts a differentiable approximate reasoning algorithm.

5. The intelligent three-dimensional security monitoring method for a smart park according to claim 3, characterized in that: The rule confidence vector is generated as follows: The target event node is a binary state, and its positive state probability is taken as the confidence level; The target event node is a multivariate state, and the confidence of its complete state probability vector is taken.

6. The intelligent three-dimensional security monitoring method for a smart park according to claim 1, characterized in that: The gated fusion mechanism includes: Dimensionally aligning and concatenating the dynamic feature vector and the rule confidence vector to generate a fusion input vector; Generate adaptive weights for the dynamic feature vector and the rule confidence vector through a gated neural network; wherein the gated neural network structure includes two fully connected layers, the first fully connected layer uses a ReLU activation function, and the second fully connected layer uses a Softmax function; the adaptive weights include a dynamic feature vector weight α and a rule confidence vector weight β, and satisfy α + β = 1; Based on the adaptive weight, performing weighted fusion on the dynamic feature vector and the rule confidence vector to generate a fused feature vector; A multi-task decision head is introduced, and the fused feature vector is input into the multi-task decision head to generate a multi-level decision output.

7. The intelligent three-dimensional security monitoring method for a smart park according to claim 6, characterized in that: The multi-task decision head includes: Output the continuous alarm level value through the fully connected layer as the alarm level generation branch; The fully connected layer and the Softmax function are used to output the probability distribution of event types as the event classification branch: Based on the attention weight, extract key evidence from the dynamic feature vector and the rule confidence vector, and generate a structured text summary as an evidence summary generation branch; The generation of the evidence summary includes: Obtaining the attention weights of the fused feature vector, the dynamic feature vector, and the rule confidence vector; Select the top N features with the highest attention weights as key evidence; Key evidence is filled in according to a predefined template to generate a structured text summary; wherein the predefined template includes an event type field, a confidence field, and an evidence source field.

8. An intelligent three-dimensional security monitoring system for smart parks, characterized by: include: A data feature acquisition module is used to build a multi-dimensional stereoscopic perception network using the drone perception network, the ground fixed sensor network, and the mobile Internet of Things perception network. Based on the stereoscopic perception network, multi-modal heterogeneous data is acquired and pre-processed; The monitoring data analysis module is used to parallelize and adaptively fuse the spatiotemporal graph neural network and the Bayesian network at the decision layer to build a security monitoring model; Inputting the preprocessed multimodal heterogeneous data into the security monitoring model, processing the dynamic time series data in the preprocessed multimodal heterogeneous data through a spatiotemporal graph neural network to obtain a dynamic feature vector, processing the state event data through a differentiable hierarchical conditional probability module, performing rule reasoning based on the conditional probability table generated by the expert library rules, and outputting a rule confidence vector; The security monitoring result acquisition module is used to design a gated fusion mechanism with adaptive decision-making levels, perform adaptive weighted fusion on the dynamic feature vector and the rule confidence vector, and generate multi-level decision outputs, including real-time alarm level, event type probability distribution and evidence summary, as security monitoring results.

9. A computer-readable storage medium, characterized in that The storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

10. A computer device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method according to any one of claims 1 to 7 is implemented.

Images