Self-adaptive dynamic risk monitoring and early warning method and system

Abstract

The invention discloses a self-adaptive dynamic risk monitoring and early warning method and system. The method comprises the following steps: acquiring static information and dynamic factors in a scene to be monitored; performing entity identification and risk factor identification on the initial data, obtaining potential risks of the to-be-monitored scene in combination with an existing knowledge base, and performing learning fusion by adopting a multi-head attention network to obtain potential risk vectorization representation of the to-be-monitored scene; acquiring real-time monitoring data, and identifying real-time dynamic factors; the real-time dynamic factors serve as input, and risk information features are calculated; determining a risk type, a specific risk point and a risk value according to the risk information features; and determining an early warning processing strategy according to the risk type, the specific risk point and the risk value, and outputting the early warning processing strategy. By implementing the method provided by the invention, adaptive end-to-end risk monitoring and early warning in multiple scenes can be realized, the risk identification capability in a complex dynamic scene is remarkably improved, and the generalization performance of the system in different environments is enhanced.

Description

Technical Field

[0001] This invention relates to a public safety risk monitoring method, and more specifically to an adaptive dynamic risk monitoring and early warning method and system. Background Technology

[0002] Public safety risk monitoring involves using various technologies and data sources to monitor and assess risk factors that may threaten public safety in real time, and to identify potential hazardous events in advance. It typically combines multiple data collection methods, including sensors, video surveillance, IoT devices, and environmental monitoring, and utilizes data analytics to identify, predict, and issue early warnings for various risks.

[0003] Many existing traditional risk monitoring and early warning methods are typically targeted at specific scenarios or objects, relying on fixed risk factors. They fail to dynamically correlate multiple risk factors across different time and spatial dimensions, and cannot adaptively monitor risks in various scenarios. For example, traditional methods like threshold-triggered algorithms struggle to effectively integrate different types of data, such as video streams, IoT data, sensor data, and location data, especially multimodal information such as RGB images, infrared heatmaps, and pressure sensor sequences. Existing models, such as those based on a combination of convolutional neural networks and long short-term memory networks, can capture time-series data to some extent, but their ability to capture the spatiotemporal correlation characteristics of sudden events remains limited, and they cannot automatically adjust the weights of changing risk factors in real time. Fixed-time-window statistical methods are ill-suited for modeling complex long-term series dependencies, particularly when dealing with dynamic changes in complex scenarios such as sudden shifts in crowd density or fire spread trajectories, lacking effective modeling capabilities.

[0004] Therefore, it is necessary to design a new method to achieve adaptive end-to-end risk monitoring and early warning in multiple scenarios, which significantly improves the ability to identify risks in complex and dynamic scenarios and enhances the system's generalization performance in different environments, in order to address the shortcomings of existing technologies. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide an adaptive dynamic risk monitoring and early warning method and system.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: an adaptive dynamic risk monitoring and early warning method, comprising:

[0007] Acquire static information and dynamic factors in the scene to be monitored to obtain initial data;

[0008] The initial data is used for entity recognition and risk factor identification. The potential risks of the scene to be monitored are obtained by combining the existing knowledge base. A multi-head attention network is used for learning and fusion to obtain the vectorized representation of the potential risks of the scene to be monitored.

[0009] Acquire real-time monitoring data and identify real-time dynamic factors;

[0010] Using real-time dynamic factors as input, the risk information features are calculated by superimposing the vectorized representation of the potential risk with real-time dynamic factors.

[0011] The risk type, specific risk point, and risk value are determined based on the risk information characteristics.

[0012] The early warning handling strategy is determined based on the risk type, the specific risk point, and the magnitude of the risk value.

[0013] Output the risk type, the specific risk point, the risk value, and the early warning handling strategy.

[0014] The further technical solution is as follows: the static information includes the nature, characteristics and load distribution of the area; the dynamic factors are key dynamic factors affecting the environment and decision-making, including activities, weather conditions, personnel conditions and time nodes; the real-time dynamic factors are factors reflecting the real-time changes in the environment and behavior, including crowd density, movement vectors, carried items, high-temperature areas and sound change information.

[0015] The further technical solution is as follows: The initial data is subjected to entity recognition and risk factor identification; the potential risks of the scene to be monitored are obtained by combining the existing knowledge base; and a multi-head attention network is used for learning and fusion to obtain a vectorized representation of the potential risks of the scene to be monitored, including:

[0016] The static information is subjected to entity recognition and vectorization to obtain a first vector;

[0017] The dynamic factors are identified and vectorized to obtain a second vector;

[0018] The potential risks of the scene to be monitored are obtained by combining the existing knowledge base, the first vector, and the second vector. A multi-head attention network is then used to learn and fuse them to obtain the vectorized representation of the potential risks of the scene to be monitored.

[0019] The further technical solution is as follows: the potential risks of the scene to be monitored, obtained by combining the existing knowledge base, the first vector, and the second vector, are learned and fused using a multi-head attention network to obtain the vectorized representation result of the potential risks of the scene to be monitored, including:

[0020] By introducing GRU to construct a fusion gating system, the potential risks of the monitored scenario, the first vector, and the second vector are dynamically weighted and fused using the fusion gating system to obtain the vectorized representation of the potential risks of the monitored scenario.

[0021] The further technical solution is as follows: The step of using real-time dynamic factors as input to calculate the risk information features after superimposing the vectorized representation result of the potential risk with real-time dynamic factors includes:

[0022] Extract long-term dependency features from the real-time dynamic factors;

[0023] The long-term dependency features and the vectorized representation of potential risks are fused together to obtain risk information features.

[0024] The further technical solution is as follows: extracting long-term dependency features from the real-time dynamic factors includes:

[0025] Long-term dependency features are extracted from the real-time dynamic factors using gated dilated convolution operations.

[0026] The further technical solution is as follows: the fusion of the long-term dependency features and the vectorized representation results of the potential risks to obtain risk information features includes:

[0027] The long-term dependency features and the vectorized representation of potential risks are dynamically weighted and fused to obtain risk information features, wherein the weights used in the dynamic weighting are trained based on historical data.

[0028] The further technical solution is as follows: determining the risk type, specific risk point, and risk value based on the risk information characteristics includes:

[0029] The risk information features are extracted using a shared encoder, and multi-task classification is performed to determine different risk types. An attention mechanism is introduced to weight the output of each risk type.

[0030] Construct a spatiotemporal graph convolutional network to calculate risk values ​​and output risk point information to obtain specific risk points and their magnitudes.

[0031] The further technical solution is as follows: the construction of a spatiotemporal graph convolutional network to calculate risk values ​​and output risk point information, so as to obtain specific risk points and risk value magnitudes, includes:

[0032] A spatiotemporal graph convolutional network is constructed to calculate the risk value of each sensor node through the spatiotemporal correlation between nodes, and output its corresponding location information to obtain the specific risk point and the magnitude of the risk value.

[0033] The present invention also provides an adaptive dynamic risk monitoring and early warning system, comprising:

[0034] The initial acquisition unit is used to acquire static information and dynamic factors in the scene to be monitored in real time to obtain initial data;

[0035] The fusion unit is used to perform entity recognition and risk factor recognition on the initial data, combine the existing knowledge base to obtain the potential risks of the scene to be monitored, and use a multi-head attention network for learning and fusion to obtain the vectorized representation of the potential risks of the scene to be monitored.

[0036] The real-time acquisition unit is used to acquire real-time monitoring data and identify real-time dynamic factors;

[0037] The calculation unit is used to take real-time dynamic factors as input and calculate the risk information features after superimposing the vectorized representation of the potential risk with real-time dynamic factors.

[0038] The first determining unit is used to determine the risk type, specific risk point, and risk value based on the risk information characteristics.

[0039] The second determining unit is used to determine the early warning processing strategy based on the risk type, the specific risk point, and the magnitude of the risk value.

[0040] The output unit is used to output the risk type, the specific risk point, the risk value, and the early warning processing strategy.

[0041] The advantages of this invention compared to existing technologies are as follows: This invention acquires static information and dynamic factors of the monitored scene, identifies entities and risk factors, extracts potential risks using a knowledge base, and employs a multi-head attention network for learning and fusion to obtain a vectorized representation of the potential risks. The system further acquires and processes dynamic factors in real time, fusing them with potential risk features to calculate updated risk information features. Based on these features, the system can accurately identify risk types, risk points, and risk values, thereby formulating appropriate early warning and handling strategies. This method, through adaptive learning and the fusion of real-time data, not only significantly improves risk identification capabilities in complex dynamic environments but also enhances the system's generalization performance, overcoming the limitations of traditional technologies in multiple scenarios and providing more accurate and efficient end-to-end risk monitoring and early warning capabilities.

[0042] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0043] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 A flowchart illustrating an adaptive dynamic risk monitoring and early warning method provided in an embodiment of the present invention;

[0045] Figure 2 This is a schematic diagram of a sub-process of an adaptive dynamic risk monitoring and early warning method provided in an embodiment of the present invention;

[0046] Figure 3 This is a schematic diagram of a sub-process of an adaptive dynamic risk monitoring and early warning method provided in an embodiment of the present invention;

[0047] Figure 4 This is a schematic diagram of a sub-process of an adaptive dynamic risk monitoring and early warning method provided in an embodiment of the present invention;

[0048] Figure 5 A flowchart illustrating an adaptive dynamic risk monitoring and early warning method provided in this embodiment of the invention. Figure 2 ;

[0049] Figure 6 A schematic block diagram of an adaptive dynamic risk monitoring and early warning system provided in an embodiment of the present invention;

[0050] Figure 7 A schematic block diagram of a vectorized unit of an adaptive dynamic risk monitoring and early warning system provided in an embodiment of the present invention;

[0051] Figure 8 A schematic block diagram of a computing unit for an adaptive dynamic risk monitoring and early warning system provided in an embodiment of the present invention;

[0052] Figure 9 A schematic block diagram of the first determining unit of an adaptive dynamic risk monitoring and early warning system provided in an embodiment of the present invention;

[0053] Figure 10 A schematic block diagram of a computer device provided for an embodiment of the present invention. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0056] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0057] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0058] Please see Figure 1 , Figure 1 This is a schematic flowchart illustrating an adaptive dynamic risk monitoring and early warning method provided in an embodiment of the present invention. This method is applied to a server that interacts with various sensors, including a wide-angle camera array, an infrared thermal imager, a pressure sensing network, and a voiceprint sensor. By acquiring static information and dynamic factors of the monitored scene, and combining multi-head attention networks and deep learning techniques, it can identify and quantify potential risks in real time. This method utilizes dynamic weighted fusion technology to combine historical data with real-time dynamic factors, improving the accuracy of risk identification. Simultaneously, it employs a spatiotemporal graph convolutional network and a shared encoder for multi-task classification and risk value calculation, further enhancing the system's generalization ability and adaptability. This method can accurately identify risk types, specific risk points, and risk values ​​in complex dynamic environments, significantly improving the identification and response capabilities of the risk early warning system and overcoming the limitations of traditional technologies in multiple scenarios.

[0059] Figure 2 This is a flowchart illustrating the adaptive dynamic risk monitoring and early warning method provided in an embodiment of the present invention. Figure 2 As shown, the method includes the following steps S110 to S170.

[0060] S110. Acquire static information and dynamic factors in the scene to be monitored in real time to obtain initial data.

[0061] In this embodiment, the initial data refers to the static information and dynamic factors collected in the scene to be monitored. The static information includes the nature, characteristics and load distribution of the area; the dynamic factors are the key dynamic factors that affect the environment and decision-making, including activities, weather conditions, personnel conditions and time nodes.

[0062] In this embodiment, the task of stage S110 is to collect static information and dynamic factors in the monitored scene in real time using various sensors and devices. This information constitutes the initial data, providing a comprehensive understanding of the monitoring environment and potential risks.

[0063] Static information refers to environmental information in a monitoring scenario that does not change over time; it provides fundamental background data for subsequent analysis. This information is typically unaffected by immediate factors, thus playing a foundational role in long-term monitoring and risk assessment. The core components of static information include:

[0064] The nature of a monitoring area refers to its basic characteristics and functional positioning. These areas can be various locations such as shopping malls, stadiums, industrial plants, and public transportation hubs. The nature of each location determines the type of potential risks it faces. For example, stadiums may face the risk of overcrowding and stampedes, while factories may face risks such as equipment failure or chemical leaks. Therefore, understanding the nature of a location helps in determining the design of different risk assessment models and early warning systems.

[0065] Regional characteristics refer to the physical environment and spatial structure of a region. For example, the size, shape, and functional layout of the region (such as entrances / exits, refuge routes, and floor distribution) are crucial factors influencing the likelihood of risks and emergency response. For instance, a location with multiple exits will be more efficient for evacuation than one with only one exit.

[0066] Load distribution refers to the distribution of weight or pressure across different areas of a scene. Load distribution can be monitored in real time by deploying pressure sensors or other devices. This data helps analyze potential structural risks, such as cracks or tilting caused by excessive ground pressure. For large venues (such as shopping malls or stadiums), load distribution information is crucial for determining whether any areas are overloaded, effectively preventing potential physical structural problems.

[0067] Dynamic factors refer to elements in a scenario that change over time. These factors typically have a significant impact on environmental changes and human behavior, and can help monitor and predict potential risks in real time. Dynamic factors involve the following aspects:

[0068] An event refers to an event or activity that is currently occurring or planned to occur within the monitored area. Examples include sporting events, concerts, exhibitions, and industrial production during weekdays. Each activity may pose different types of risks, and the dynamic changes in events are a key factor in assessing on-site safety risks. Large gatherings of people may lead to overcrowding, increasing the risk of accidents, while certain activities (such as chemical experiments) may pose risks related to chemical spills, fires, etc.

[0069] Weather conditions are a key dynamic factor affecting monitored environments, especially for outdoor environments or areas heavily influenced by weather (such as open spaces and transportation hubs). Weather changes (such as rainfall, wind, and temperature) can directly impact personnel movement, equipment operation, and environmental safety. For example, strong winds may increase the risk of objects being blown away, or large outdoor events may face safety hazards due to sudden downpours. Real-time weather monitoring can help predict potential risks from weather changes and enable timely responses.

[0070] Personnel monitoring primarily involves the behavior and condition of people within the monitored environment. This includes factors such as crowd density, flow direction, and emergency response capabilities. By deploying various sensors (e.g., cameras, pressure sensors, voiceprint sensors), real-time monitoring of personnel behavior is possible. Dynamically changing crowd density can reveal the risk of overcrowding, panic, or stampedes. Behavioral patterns (such as accelerating or changing direction) can sometimes indicate the occurrence of abnormal situations or emergencies.

[0071] Time points refer to factors related to the timing of an event. Specific time points (such as peak weekday hours, holidays, peak pedestrian traffic periods, etc.) may affect the probability of a risk occurring. At certain specific times, the monitored scene may experience higher population density or activity intensity, increasing the likelihood of accidents. For example, during peak hours in some commercial centers or transportation hubs, frequent population movement may increase the risk of traffic accidents or crowd conflicts.

[0072] Specifically, the initial data mentioned above can be acquired through cameras, thermal imagers, voiceprint collectors, and IoT sensors. Specifically, cameras acquire RGB images and video streams; thermal imagers acquire infrared temperature distribution maps; IoT sensors acquire pressure, current, voltage, and vibration; and voiceprint collectors acquire abnormal sounds (such as broken glass).

[0073] Static information and dynamic factors, collected by sensors and other devices, need to be integrated using data fusion and analysis techniques. Data from different sources (such as video surveillance, temperature and humidity sensors, and sound monitoring) are processed to form comprehensive initial data. This initial data not only helps the monitoring system obtain the real-time status of the monitored scene but also provides foundational support for subsequent risk assessment and predictive models. For example, by combining data on pedestrian density, temperature changes, and area load, potential dangerous events such as overcrowding or fires can be predicted, and timely warnings can be issued.

[0074] By acquiring and integrating static information and dynamic factors, the S110 phase provides comprehensive foundational data for subsequent risk analysis and prediction. This data helps the system understand the environmental and personnel conditions of the monitored scenario in real time, effectively identifying potential risks and providing timely information support for decision-makers.

[0075] S120. Entity recognition and risk factor recognition are performed on the initial data. The potential risks of the scene to be monitored are obtained by combining the existing knowledge base. Multi-head attention network is used for learning and fusion to obtain the vectorized representation result of the potential risks of the scene to be monitored.

[0076] In this embodiment, the vectorized representation of potential risks in the monitored scenario refers to a vectorized representation obtained by processing, identifying, modeling, and fusing static information and dynamic factors in the monitored scenario. This vector integrates all relevant risk information. This vector not only reflects the characteristics of various risks in the scenario but also helps the system to perform risk assessment, prediction, and judgment in subsequent processing.

[0077] Specifically, it is a result obtained through the following steps:

[0078] Static information vector (first vector): This includes time-independent, fixed entities and features (such as objects and structures in a scene). This information is extracted using entity recognition techniques (such as CNN, BiLSTM, CRF, etc.) and transformed into a numerical vector representation using vectorization methods (such as word embedding).

[0079] Dynamic factor vector (second vector): refers to factors that change over time (such as pedestrian flow, temperature changes, etc.). These dynamically changing risks are analyzed using multimodal data (such as images, heatmaps, sound, etc.), and the features of different modalities are dynamically modeled using a multi-head attention mechanism. The generated vector is used to represent these dynamically changing risks.

[0080] Fusion Risk Vector: By combining static information vectors and dynamic factor vectors, a weighted fusion using a multi-head attention mechanism and a gated recurrent unit (GRU) is obtained to obtain a comprehensive potential risk vector. This vector integrates the weights of static information and dynamic factors to reflect multiple potential risks in the scenario.

[0081] In summary, this vectorized representation of potential risks is a comprehensive representation of all known risk factors (including time-related and time-independent) in the monitored scenario, providing an accurate and actionable data foundation for subsequent risk calculation, assessment, and intelligent judgment.

[0082] In one embodiment, please refer to Figure 2 The above-mentioned step S120 may include steps S121 to S123.

[0083] S121. Perform entity recognition and vectorization on the static information to obtain a first vector.

[0084] In this embodiment, static information refers to time-independent and relatively fixed factors, such as routine data in sensor reports, the position of objects in the scene, or structural features.

[0085] In this step, the first step is to identify the targets or entities in these static information using entity recognition technology. For example, the CNN+BiLSTM+CRF structure can be used to identify targets in the scene, such as crowds, fire sources, vibration sources, etc.

[0086] For text data (such as sensor reports), CNN is used for feature extraction, BiLSTM captures temporal contextual dependencies, and CRF is used to optimize the global consistency of label prediction.

[0087] The identified entity information is transformed into a first vector through vectorization techniques (such as word embedding or other vectorization methods), which can effectively represent the various entities identified in the static information.

[0088] S122. The dynamic factors are identified and vectorized to obtain a second vector.

[0089] In this embodiment, dynamic factors refer to factors that change over time and affect the evolution of risk, such as temperature change trends, changes in pedestrian flow, and dynamic changes in images.

[0090] Risk factor identification for these dynamic data mainly utilizes multimodal data (such as images, infrared thermograms, acoustic spectrograms, stress distribution maps, etc.) and combines them with an improved multi-head attention mechanism to model these dynamic factors.

[0091] By using a multi-head attention mechanism to dynamically correlate inputs from different modalities, the accuracy of extracting dynamic risk features is improved.

[0092] After risk identification, a second vector will be obtained, which represents the potential risk of dynamic factors.

[0093] Specifically, with the introduction of an improved multi-head attention mechanism, dynamic correlation modeling is performed on input data of different modalities (such as images, infrared heatmaps, voiceprint spectrograms, and stress distribution maps) to improve the accuracy of risk feature extraction. Specifically, the attention mechanism extracts the relationships between different modalities by performing a weighted summation of the query (Q), key (K), and value (V). Where Q represents a query from static information, and K and V come from dynamic data, Q, K, and V are all transformed by a learned projection matrix, specifically represented as: Q = F s W q K = F t W k V = F t W v ;in It is a learnable projection matrix.

[0094] To effectively integrate data from different modalities, a multi-head parallel computing mechanism is used. Each attention head focuses on the feature representation of a different modality subspace, and parallel computation enhances the ability to extract risk features from each modality. The final multi-head attention outputs are merged using the following formula: MultiHead = Concat(head1,...,head2)W o , where each head i The calculation method is as follows:

[0095] Here, multi-head attention is achieved through different projection matrices (W). i Q W i k W i v This provides independent feature extraction for each head, ensuring that risk information from each modality is fully considered and fused, ultimately resulting in an output matrix W. o The results from all heads are merged. In this way, the improved multi-head attention mechanism can more accurately capture and extract risk features from different modalities, thereby enhancing the model's performance and risk prediction capabilities.

[0096] S123. The potential risks of the scene to be monitored are obtained by combining the existing knowledge base, the first vector, and the second vector. A multi-head attention network is used to learn and fuse them to obtain the vectorized representation result of the potential risks of the scene to be monitored.

[0097] In this embodiment, a fusion gating system is constructed by introducing GRU. The fusion gating system is used to dynamically weight and fuse the potential risks of the monitored scenario, the first vector, and the second vector to obtain the vectorized representation of the potential risks of the monitored scenario.

[0098] In this step, the acquisition of potential risks in the monitored scenario needs to be combined with the existing knowledge base (including historical data, scenario-specific data models, expert experience, etc.) to determine which spatial and temporal characteristics are more important for risk prediction, and to conduct analysis based on existing patterns.

[0099] The knowledge base includes:

[0100] Historical data: Historical records can help analyze accidents or anomalies that occurred in similar scenarios in the past. By analyzing historical data, it is possible to reveal which combinations of characteristics often indicate potential risks. For example, the simultaneous occurrence of high temperatures and low humidity in a region may indicate a fire risk.

[0101] Expert rules: The knowledge base may include rules derived from expert experience, such as the risk of stampede when the crowd density exceeds a certain threshold under specific conditions.

[0102] Machine learning models: Models trained on historical data using machine learning algorithms (such as decision trees, neural networks, etc.) can help us identify potential risk factors. For example, a model can be trained to determine the probability of a fire occurring under certain environmental conditions.

[0103] The potential risks of the monitored scenario are not simply a superposition of spatial and temporal features, but rather a weighted fusion using advanced machine learning and deep learning models. The model comprehensively evaluates both spatial and temporal features to obtain a comprehensive risk score or risk level.

[0104] Spatial and temporal features undergo preprocessing operations such as standardization, denoising, and smoothing to ensure data quality and consistency. Algorithms (such as Principal Component Analysis (PCA) and Random Forest) are used to select the most representative features and remove redundant and irrelevant features. Multimodal learning or deep neural network techniques are used to fuse spatial and temporal features. Common fusion methods include weighted averaging, attention mechanisms (such as multi-head attention), recurrent neural networks (RNNs), and gated recurrent units (GRUs). A trained model (such as a neural network or support vector machine) is used for risk prediction, outputting a score or category of the potential risk.

[0105] Risk assessment is not only a static process but also requires dynamic adjustments based on real-time data feedback. For example, during monitoring, if changes in certain characteristics are detected (such as a sudden increase in population density or an abnormal rise in temperature), the system can adjust the parameters of the risk assessment model in real time and update the risk prediction. In this way, spatial and temporal characteristics can be combined with rules in a knowledge base to accurately identify and assess potential safety risks.

[0106] The acquisition of potential risks in the monitored scenarios is achieved through various technical means, including data collection, knowledge base application, feature fusion, and machine learning. Spatial and temporal features are core elements in this process, requiring weighted fusion using advanced models and incorporating historical data and expert experience to ultimately output accurate predictions of potential risks. This process not only focuses on single-dimensional features but also comprehensively considers information from multiple aspects to achieve efficient and accurate risk assessment.

[0107] By leveraging the fusion of multimodal data and using gated recurrent units (GRUs) to dynamically weight and fuse different types of features, a vectorized representation of the potential risks in the monitored scenario is obtained. The following is a detailed explanation of this process step by step:

[0108] Multimodal data refers to data obtained from different sources or types of information. This data usually contains multiple features, such as the potential risks of the scene to be monitored, the spatial features involved in the first vector and the second vector, and the temporal features. Specifically, spatial features, such as population distribution and hot spot areas, reflect the characteristics of the spatial structure in a certain area; temporal features, such as temperature change trends and population flow evolution, reflect the dynamic process that changes over time.

[0109] These characteristics have different levels of importance and impact in monitoring scenarios, and need to be rationally integrated through intelligent algorithms to provide accurate risk assessment.

[0110] Multi-head attention is used to learn data features in parallel across multiple subspaces, thereby achieving more efficient information extraction. In this scenario, the first and second vectors represent spatial and temporal features, respectively, which are processed by a multi-head attention network to obtain a fused potential risk representation.

[0111] In this way, the multi-head attention mechanism can capture the relationship between spatial and temporal features, thereby reflecting risk more comprehensively.

[0112] In traditional fusion methods, spatial and temporal features are often weighted and summed with fixed weights. However, in real-world scenarios, the contributions of different features vary at different time points. Therefore, GRU (Gated Recurrent Unit) is introduced to automatically adjust the fusion weights of these features based on dynamic changes in the scene.

[0113] The gating mechanism learns and dynamically adjusts the weights of different features, enabling the fusion of spatial and temporal features to better meet practical needs at different time steps and in different scenarios. Specifically, GRU generates a fusion gating weight to control the relative contributions of spatial and temporal features.

[0114] The formula for the fusion process is: F fusion =∝·F s +(1-∝)·MultiHead; where ∝ represents the Sigmoid activation function, outputting fusion weights, which are dynamically generated by a gated recurrent unit (GRU). These dynamic weights, generated by the GRU, represent the ratio of fused spatial to temporal features; F s Spatial features (such as population distribution and hot spot areas) are represented by MultiHead, while temporal features (such as temperature change trends and population flow evolution) are represented by MultiHead after processing through a multi-head attention mechanism.

[0115] Specifically, ∝ is controlled by a Sigmoid activation function whose output ranges from [0,1]. The closer its value is to 1, the greater the influence of spatial features on the final fusion result; conversely, the closer it is to 0, the greater the influence of temporal features. Through the gating mechanism of GRU, ∝ is dynamically adjusted according to the current time point and context information to reflect the different importance of features at different times and in different scenarios.

[0116] The final fusion result F fusion This is a vectorized representation of the potential risks of the scenario to be monitored. This representation includes comprehensive information on spatial and temporal characteristics under different times and scenarios, and can accurately reflect the risk status of the scenario.

[0117] Vectorized representation of potential risk F fusionThis will serve as input to the subsequent risk calculation network for further analysis and processing. This includes classifying risk types, predicting risk magnitude, and judging risk evolution trends. The intelligent discrimination module uses these vectorized representations for decision support, automatically determining whether potential risks have reached warning thresholds and predicting future risk change trends.

[0118] By introducing a multi-head attention mechanism and a gated recurrent unit (GRU) for dynamic weighted fusion, this process can efficiently integrate spatial and temporal features, improving the accuracy and flexibility of risk assessment. In particular, the introduction of GRUs makes feature fusion not just a fixed linear combination, but can be dynamically adjusted according to specific scenarios, thus making the entire risk identification process more accurate and adaptive.

[0119] S130. Obtain real-time monitoring data and identify real-time dynamic factors.

[0120] In this embodiment, the real-time dynamic factors are factors that reflect real-time changes in the environment and behavior, including crowd density, motion vectors, carried items, high-temperature areas, and sound change information.

[0121] Real-time dynamic factors refer to factors that reflect real-time changes in the environment and behavior within a scenario. These factors are acquired through real-time monitoring data to promptly identify potential risks or anomalies in dynamic scenarios.

[0122] Crowd density refers to the number of people per unit area within a specific time and space, reflecting the degree of crowding in a location or area. Crowd density is often closely related to the occurrence of safety incidents; for example, overcrowded areas are prone to stampedes.

[0123] The number and distribution of people can be calculated and tracked by installing cameras and combining them with computer vision algorithms (such as facial recognition and pedestrian detection). The flow of people in a specific area can be estimated using sensors (such as infrared sensors, pressure sensors, and RFID devices). For example, infrared sensors can detect the number of people crossing a specific area. People density changes over time, so dynamic analysis using real-time data streams (video, sensors, etc.) is necessary to identify trends in density. Excessively high density may indicate potential safety risks.

[0124] In large public venues (such as train stations, shopping malls, and stadiums), monitor pedestrian density to promptly identify overcrowded areas and implement crowd control measures or issue warnings. In traffic management, monitor traffic flow in different areas to identify high-density traffic conditions.

[0125] Motion vectors refer to the direction and velocity of motion of an individual or object, and are typically used to describe the movement of an object in space. In risk assessment, motion vectors can help identify abnormal behavior (such as fast-moving people or objects) and predict potential safety incidents.

[0126] By capturing the movement trajectories of people or objects using surveillance cameras and employing optical flow and motion detection algorithms, the direction, speed, and acceleration of the motion can be calculated. Combined with sensors such as accelerometers and gyroscopes, the movement trajectories of objects or people can be tracked and their speed and direction analyzed. Based on motion vectors and combined with behavior recognition technology, abnormal movement patterns can be identified, such as sudden acceleration or escape behavior, which may be an indication of an emergency.

[0127] In location surveillance, identifying fast-moving targets (such as running or escaping) can provide early warnings of potential emergencies, such as stampedes, conflicts, or violence. In road surveillance, monitoring vehicle motion vectors can identify abnormal behaviors such as traffic accidents, driving against traffic, or speeding.

[0128] Carried items refer to items carried by individuals in a scenario, especially items that may affect safety, such as suspicious packages, hazardous chemicals, and sources of ignition. Monitoring carried items helps identify potential threats.

[0129] Image recognition technologies (such as deep learning algorithms) are used to identify the types of items carried by individuals. Intelligent video surveillance systems can detect items such as packages and backpacks, thereby determining whether there are potential dangers. In certain locations (such as airports and train stations), X-ray equipment is used to scan carried items to identify whether there are prohibited or potentially dangerous items. Technologies such as RFID and NFC are used to identify the carrier and the attributes of the items themselves.

[0130] In security check scenarios, it automatically identifies and analyzes carried items, and issues real-time alerts for suspicious items. In public safety monitoring, it analyzes whether there are any unusual items carried by people in the crowd and issues an alarm.

[0131] High-temperature areas refer to regions with abnormally high temperatures, which are usually a sign of potential fires or equipment malfunctions. High-temperature monitoring can effectively prevent dangerous events such as fires.

[0132] Infrared cameras monitor temperature distribution in real time, quickly identifying areas of localized high temperatures. Infrared imaging technology can detect temperature changes in environments where direct contact is not possible. Temperature sensors are installed in key locations to monitor ambient temperature in real time. These sensors can determine the presence of high-temperature risks based on set thresholds. By comprehensively using temperature and humidity sensors and gas monitoring equipment, abnormally high temperatures in the environment can be detected, enabling timely detection of fire hazards.

[0133] In high-risk locations (such as chemical plants and power plants), high-temperature monitoring can promptly detect equipment malfunctions or fire hazards. In building monitoring, it can analyze indoor and outdoor temperature changes to detect the presence of fires or overheated equipment.

[0134] Sound change information refers to changes in sound in the environment. Abnormal sound patterns (such as sudden screams, explosions, and rapid footsteps) can often indicate potential safety risks.

[0135] By installing microphones and sound sensors, ambient sounds are captured and analyzed in real time. Sound features can be extracted using acoustic analysis algorithms (such as FFT and audio pattern recognition). Machine learning algorithms are used to identify unusual sounds in the environment, such as screams, explosions, and gunshots, to quickly determine if a security threat exists. Changes in ambient noise are analyzed in real time to determine if illegal activities or abnormal events (such as fighting sounds or the sound of weapons colliding) are taking place.

[0136] Sound monitoring systems are used in public places (such as shopping malls, train stations, and airports) to identify abnormal sounds and trigger alarms. In security management, by analyzing changes in sound inside and outside buildings, emergencies (such as fires, fights, and cries for help) can be detected in a timely manner.

[0137] The acquisition and identification of real-time dynamic factors are achieved through various sensors and data analysis technologies, including pedestrian density, movement vectors, carried items, high-temperature areas, and sound change information. These factors can be acquired in real time through video surveillance, sensors, and behavior recognition technologies, and then analyzed to identify potential safety risks. The comprehensive analysis and monitoring of real-time dynamic factors provides crucial data support for public safety, risk warning, and emergency response.

[0138] S140. Using real-time dynamic factors as input, calculate the risk information features after superimposing the vectorized representation of the potential risk with real-time dynamic factors.

[0139] Risk information features refer to numerical information or vectors that reflect potential risks in a system. These features are typically generated by combining historical data, sensor inputs, and external environmental factors, after a certain process of modeling, calculation, and analysis. These features can reflect the system's risk status, abnormal situations, or hazard levels. For example, in fields such as industrial monitoring, health monitoring, and intelligent transportation, risk information features may include the probability of equipment failure, the risk of health anomalies, and the probability of traffic accidents.

[0140] In this embodiment, the risk information features are an output obtained by modeling and analyzing historical risk data of the system and weighted fusion of dynamic data input from the current environment. These features help predict issues related to the system's security and stability.

[0141] The approach employs dynamic weighted fusion, trained using long-term dependent features and historical data, to ultimately extract potential risk information features. These steps reflect a comprehensive method for extracting and fusing multiple dynamic factors, including sensor data and video features, and deriving risk features through convolution and weighting strategies.

[0142] In one embodiment, please refer to Figure 3 The above-mentioned step S140 may include steps S141 to S143.

[0143] S141. Extract long-term dependency features from the real-time dynamic factors.

[0144] In this embodiment, long-term dependency features refer to features extracted from time series data that exhibit dependencies over a long period. These features are typically obtained by analyzing long-term series data to capture dynamic changes, trends, and periodicities that span a considerable timeframe. These features reflect long-term trends or patterns in system behavior, rather than just short-term fluctuations.

[0145] For example, when monitoring equipment data, long-term dependent features may include the performance trends of the equipment over a long period of time, the long-term impact of environmental changes, etc. These features are often extracted and calculated using models that can process time series data (such as RNN, LSTM, gated dilated convolution, etc.).

[0146] In this embodiment, long-term dependency features are extracted using methods such as gated dilated convolution, aiming to capture dynamic dependencies with long time spans in order to more accurately analyze real-time data and predict potential risks.

[0147] Specifically, gated dilated convolution operations are used to extract long-term dependent features from the real-time dynamic factors.

[0148] In this embodiment, gated dilated convolution is used to extract long-term dependent features from real-time dynamic factors. Gated dilated convolution combines the scalability of convolution operations with the flexibility of gating mechanisms, which is helpful for processing data with long-term sequence properties, such as temperature sequences, pressure values, or video frame features.

[0149] Dilated convolutions expand the receptive field by increasing the "holes" in the convolutional kernel, enabling the network to capture data dependencies over longer distances without increasing the number of convolutional layers. This is particularly effective when processing long-term data series.

[0150] Gating mechanisms introduce a gating coefficient g. kThis mechanism controls the contribution of input information at different time steps to the convolution computation, allowing the model to more flexibly focus on important moments and suppress irrelevant fluctuations or noise. Specifically, the formula for the gating mechanism is: Where *d represents a convolution operation with dilation rate d, g k X is the gating coefficient. sensor It is the sensor data at the current moment (temperature sequence, video frame features, pressure value, etc.), F t This represents the vector representation of the extracted real-time dynamic factors.

[0151] S142. The long-term dependency features and the vectorized representation results of the potential risks are fused to obtain risk information features.

[0152] In this embodiment, the long-term dependency features and the vectorized representation of potential risks are dynamically weighted and fused to obtain risk information features, wherein the weights used in the dynamic weighting are trained based on historical data.

[0153] Once long-term dependent features have been extracted from real-time dynamic factors, the next goal is to integrate these features with existing vectorized representations of potential risks. Potential risk vectors typically represent known risk information derived from historical data within the system, while real-time dynamic factors represent current external conditions that may influence the risk.

[0154] Potential risk vectorization representation result F fusion It was derived through the analysis and modeling of historical data, reflecting the potential risks of the system.

[0155] Real-time dynamic factor vector F t It is a long-term dependent feature extracted from sensor data.

[0156] The fusion of these two vectors is achieved through dynamic weighting, using the formula: F final =γF fusion +(1-γ)F t γ is a dynamic weight obtained through training on historical data. γ∈[0,1], and its specific value is obtained through training on historical data. Through this weighted fusion, the model can dynamically adjust the calculation of risk characteristics based on the balance between changes in real-time dynamic factors and historical risk information.

[0157] When γ is large, the impact of potential risks is greater, indicating that the influence of historical information is dominant.

[0158] When γ is small, the influence of real-time dynamic factors is greater, and the model relies more on current sensor data.

[0159] After weighted fusion using S142, the resulting risk information feature F final By integrating historical risk data with real-time dynamic factors, this feature vector can more comprehensively reflect the potential risks of the current system. This feature vector can then be further used for applications such as risk prediction and decision support.

[0160] Long-term dependency features of real-time dynamic factors are extracted by gated dilated convolution, and then dynamically weighted and fused with the vectorized representation of potential risks to obtain the final risk information features. This process helps to more accurately capture potential risks in a system, especially in complex environments that depend on multiple dynamic factors.

[0161] Conventional methods typically involve simply overlaying dynamic and static information or performing basic fusion analysis. For example, real-time dynamic data is directly combined with static information such as equipment parameters and site layout, and risk assessment is conducted using simple mathematical models or rule thresholds. This approach often deals with dynamic factors superficially and fails to effectively uncover the deeper features and patterns within the dynamic data.

[0162] This embodiment innovatively extracts long-term dependency features from real-time dynamic factors and dynamically weights and fuses them with the vectorized representation of potential risks. The extraction of long-term dependency features is the core element. Through advanced time series processing techniques such as gated dilated convolution, it focuses on grasping key information such as trends and periodicity across a longer time span in dynamic data, thereby more accurately understanding the potential risk patterns of the system evolving over time.

[0163] The method in this embodiment, by extracting long-term dependent features, can capture subtle long-term change patterns in dynamic factors, such as the slow decline trend of equipment performance and the periodic impact of environmental factors. This information is easily overlooked in conventional simple fusion methods. Fusing these long-term dependent features with vectorized representations of potential risks allows the risk assessment model to more comprehensively consider the interactions of various risk factors, thereby significantly improving the accuracy of risk assessment.

[0164] By focusing on the long-term trends and dependencies of dynamic factors, the method in this embodiment can better predict potential future risks. For example, by analyzing the long-term trends of equipment performance, the potential time of equipment failure can be predicted in advance; understanding the long-term impact of environmental changes can also more accurately predict when environmental factors will pose a potential threat to the system, providing more time to take preventative measures in advance.

[0165] A dynamic weighted fusion mechanism is introduced, enabling the model to automatically adjust the calculation of risk characteristics based on the balance between changes in real-time dynamic factors and historical risk information. When the system is operating relatively stably and historical risk information is more representative, the model can appropriately increase the weight of the potential risk vector; while when sudden and significant changes in dynamic factors occur, the model can promptly increase the weight of real-time dynamic factors to quickly respond to the current risk situation and better adapt to the complex and ever-changing system environment.

[0166] S150. Determine the risk type, specific risk point, and risk value based on the risk information characteristics.

[0167] In one embodiment, please refer to Figure 4 The above-mentioned step S150 may include steps S151 to S152.

[0168] S151. Use a shared encoder to extract high-order features of the risk information features, perform multi-task classification to determine different risk types, and introduce an attention mechanism to weight the output of each risk type.

[0169] In this embodiment, the fusion risk feature vector F generated in the previous step is used as the basis for... final These features are processed using a shared encoder. The role of the shared encoder is to process F... final High-order, representative, and discriminative features are extracted. These high-order features can capture subtle differences between different risk types and provide rich information for subsequent classification tasks.

[0170] High-order features extracted by a shared encoder are integrated into multiple classification branches to achieve multi-task classification. Each classification branch corresponds to a risk type. For example, the following risk types can be determined:

[0171] Unusual crowd gatherings; signs of fire; abnormal heat spread; abnormal sound signatures (such as broken glass).

[0172] Each branch will generate a score output related to its corresponding risk type.

[0173] To improve the classification model's ability to identify key features, an attention mechanism is introduced. This mechanism automatically assigns weights to different risk types, allowing the model to focus more on important features and thus optimize classification results.

[0174] Specifically, the attention mechanism will adjust the input risk information features F final The output score for each risk type is generated through a weighted average: S k =Attention(F final )·W k ; where: Sk This represents the score output for the k-th type of risk. Attention(F) final ) represents the risk information feature F final The attention weights are calculated. W k It is the weight matrix for the k-th type of risk.

[0175] S152. Construct a spatiotemporal graph convolutional network to calculate risk values ​​and output risk point information to obtain specific risk points and their magnitudes.

[0176] In this embodiment, a spatiotemporal graph convolutional network is constructed to calculate the risk value of each sensor node through the spatiotemporal correlation between nodes and output its corresponding location information to obtain the specific risk point and the magnitude of the risk value.

[0177] Specifically, a Spatio-temporal Graph Convolutional Network (ST-GCN) is constructed to capture the spatio-temporal correlations between nodes. This network uses a graph structure to simulate the spatial layout and dynamic evolution of sensor nodes.

[0178] In this diagram, each node represents a sensor unit, such as a camera, infrared sensor, pressure sensor, or acoustic sensor. The edge weights between nodes represent their spatiotemporal correlation, that is, the mutual influence of the risk information of these nodes in time and space.

[0179] The risk value r of each node i It is calculated using a spatiotemporal graph convolutional network, and this value represents the risk state of each sensor node. The calculation formula is as follows: r i =σ[W g ·(∑A ij · j )];A ij σ is the adjacency weight between node i and node j, usually determined by historical co-occurrence data or physical proximity; j is the risk feature vector of the j-th node, representing the current risk information of that node; σ is the normalized output of the activation function (such as Sigmoid); r i ∈[0,1] represents the risk value at the i-th position. The information of each node can be output as the location information of the risk point.

[0180] Finally, the spatiotemporal graph convolutional network will output the location information of the risk points corresponding to each node based on the risk value of each node. These locations can be the specific locations or regions of sensor nodes, thereby determining the specific risk points in the system and the magnitude of the risk value of each risk point.

[0181] As can be seen, S151 extracts high-order features through a shared encoder, determines different risk types through a multi-task classifier, and introduces an attention mechanism to weight the outputs of different risk types to enhance the identification ability of key features.

[0182] S152 constructs a spatiotemporal graph convolutional network, which captures the spatiotemporal relationships between sensor nodes, calculates the risk value of each node, and outputs the location information of each risk point.

[0183] These two steps together complete the process from extracting risk information features to calculating risk values ​​and locating risk points, providing the system with accurate risk warning and location functions.

[0184] S160. Determine the early warning handling strategy based on the risk type, the specific risk point, and the magnitude of the risk value.

[0185] In this embodiment, the system determines the early warning handling strategy based on dynamically changing thresholds, combined with the risk type, specific risk point, and risk value. The specific steps are as follows:

[0186] The system triggers different response mechanisms based on different risk levels. The early warning mechanism is divided into the following levels:

[0187] Yellow Alert: Low risk. This alert reminds relevant personnel to strengthen monitoring and take preventative measures. Typically, a yellow alert corresponds to a low risk level, possibly indicating a slow-developing event with only minor anomalies. The system recommends increased observation and implementation of preventative measures.

[0188] Red Alert: High risk, requiring immediate emergency response measures. This may include emergency measures such as evacuation, area closure, and resource mobilization. Red alerts are typically accompanied by high risk values, indicating a sharp increase in the potential threat of the incident, which may escalate rapidly and necessitates immediate action.

[0189] The system adjusts its warning levels based on constantly changing risk information and dynamic thresholds. These thresholds are based on historical data, environmental changes, and real-time monitoring results, and can intelligently adapt to different scenarios and real-time risk assessments.

[0190] S170, Output the risk type, the specific risk point, the risk value, and the early warning processing strategy.

[0191] The system will output the following information to the terminal display according to the early warning processing strategy in S160:

[0192] Risk type: The specific type of risk identified (e.g., abnormal crowd gathering, potential fire hazards, etc.).

[0193] Specific risk points: the location or area where the risk originates.

[0194] Risk value: The specific risk value calculated for each risk point, used to indicate the severity of the risk.

[0195] Early warning handling strategy: Based on the risk type, risk value and its threshold, output the corresponding early warning strategy (such as yellow warning, red warning, etc.).

[0196] Through this step, the system can assess risks in real time and intelligently, and output corresponding handling strategies according to different warning levels, thereby ensuring a rapid response and effective countermeasures when facing potential threats.

[0197] Please see Figure 5 The method in this embodiment acquires static information and dynamic factors in the monitored scene in real time to perform entity recognition and risk factor identification. It combines a knowledge base and a multi-head attention network to vectorize the data, thereby extracting risk feature information. Based on this information, the system calculates the risk type, risk probability, and evolution trend, and formulates early warning strategies accordingly.

[0198] Specifically, the method incorporates static information (such as regional load distribution) and dynamic factors (such as pedestrian density and motion vectors), employing spatial and temporal feature encoders to extract spatial and temporal features of different risk factors. Risk points and risk values ​​are then assessed using a multi-task classification head, a spatiotemporal graph convolutional network, and reinforcement learning strategies, thereby achieving intelligent early warning processing.

[0199] The knowledge base is built using a fundamental large-scale model and risk knowledge distillation. This risk knowledge base is continuously improved by combining professional knowledge and data obtained from academic papers and textbooks, enabling it to identify potential risks. The knowledge base includes a risk list and a risk magnitude assessment database, supporting agents in calculating risks and providing accurate early warning information.

[0200] The aforementioned adaptive dynamic risk monitoring and early warning method acquires static information and dynamic factors of the monitored scenario, identifies entities and risk factors, extracts potential risks using a knowledge base, and employs a multi-head attention network for learning and fusion to obtain a vectorized representation of the potential risks. The system further acquires and processes dynamic factors in real time, fusing them with potential risk features to calculate updated risk information features. Based on these features, the system can accurately identify risk types, risk points, and risk values, thereby formulating appropriate early warning and handling strategies. This method, through the fusion of adaptive learning and real-time data, not only significantly improves risk identification capabilities in complex dynamic environments but also enhances the system's generalization performance, overcoming the limitations of traditional technologies in multiple scenarios and providing more accurate and efficient end-to-end risk monitoring and early warning capabilities.

[0201] Figure 6This is a schematic block diagram of an adaptive dynamic risk monitoring and early warning system 300 provided in an embodiment of the present invention. Figure 6 As shown, corresponding to the above adaptive dynamic risk monitoring and early warning method, the present invention also provides an adaptive dynamic risk monitoring and early warning system 300. This adaptive dynamic risk monitoring and early warning system 300 includes a unit for executing the above-described adaptive dynamic risk monitoring and early warning method, and the system can be configured in a server. Specifically, please refer to... Figure 6 The adaptive dynamic risk monitoring and early warning system 300 includes an initial acquisition unit 301, a fusion unit 302, a real-time acquisition unit 303, a calculation unit 304, a first determination unit 305, a second determination unit 306, and an output unit 307.

[0202] An initial acquisition unit 301 is used to acquire static information and dynamic factors in the scene to be monitored in real time to obtain initial data; a fusion unit 302 is used to perform entity recognition and risk factor recognition on the initial data, combine it with an existing knowledge base to obtain the potential risks of the scene to be monitored, and use a multi-head attention network for learning and fusion to obtain the vectorized representation result of the potential risks of the scene to be monitored; a real-time acquisition unit 303 is used to acquire real-time monitoring data and identify real-time dynamic factors; a calculation unit 304 is used to take the real-time dynamic factors as input and calculate the risk information features after superimposing the vectorized representation result of the potential risks with the real-time dynamic factors; a first determination unit 305 is used to determine the risk type, specific risk point, and risk value based on the risk information features; a second determination unit 306 is used to determine the early warning processing strategy based on the risk type, the specific risk point, and the risk value; and an output unit 307 is used to output the risk type, the specific risk point, the risk value, and the early warning processing strategy.

[0203] In one embodiment, such as Figure 7 As shown, the fusion unit 302 includes a first vectorization subunit 3021, a second vectorization subunit 3022, and a fusion subunit 3023.

[0204] The first vectorization subunit 3021 is used to perform entity recognition and vectorization on the static information to obtain a first vector; the second vectorization subunit 3022 is used to perform risk dynamic factor recognition and vectorization on the dynamic factors to obtain a second vector; the fusion subunit 3023 is used to combine the potential risks of the scene to be monitored with the existing knowledge base, the first vector, and the second vector and learn and fuse them using a multi-head attention network to obtain the vectorized representation result of the potential risks of the scene to be monitored.

[0205] In one embodiment, the fusion subunit 3023 is used to construct a fusion gating by introducing a GRU, and to dynamically weight and fuse the potential risks of the monitored scenario, the first vector, and the second vector using the fusion gating to obtain a vectorized representation of the potential risks of the monitored scenario.

[0206] In one embodiment, such as Figure 8 As shown, the calculation unit 304 includes an extraction subunit 3041 and a weighted fusion subunit 3042.

[0207] Extraction subunit 3041 is used to extract long-term dependency features from the real-time dynamic factors; weighted fusion subunit 3042 is used to fuse the long-term dependency features and the vectorized representation results of the potential risks to obtain risk information features.

[0208] In one embodiment, the extraction subunit 3041 is used to extract long-term dependent features from the real-time dynamic factors using a gated dilated convolution operation.

[0209] In one embodiment, the weighted fusion subunit 3042 is used to dynamically weight and fuse the long-term dependency features and the vectorized representation results of the potential risks to obtain risk information features, wherein the weights used for dynamic weighting are trained based on historical data.

[0210] In one embodiment, such as Figure 9 As shown, the first determining unit 305 includes a type determining subunit 3051 and an information determining subunit 3052.

[0211] The type determination subunit 3051 is used to extract high-order features of the risk information features using a shared encoder, perform multi-task classification to determine different risk types, and introduce an attention mechanism to weight the output of each risk type.

[0212] The information determination subunit 3052 is used to construct a spatiotemporal graph convolutional network to calculate risk values ​​and output risk point information to obtain specific risk points and risk value magnitudes.

[0213] In one embodiment, the information determination subunit 3052 is used to construct a spatiotemporal graph convolutional network, calculate the risk value of each sensor node through the spatiotemporal correlation between nodes, and output its corresponding location information to obtain the specific risk point and the magnitude of the risk value.

[0214] It should be noted that those skilled in the art can clearly understand that the specific implementation process of the above-mentioned adaptive dynamic risk monitoring and early warning system 300 and its various units can be referred to the corresponding descriptions in the aforementioned method embodiments. For the sake of convenience and brevity, these details will not be repeated here.

[0215] The aforementioned adaptive dynamic risk monitoring and early warning system 300 can be implemented as a computer program, which can, for example... Figure 10 It runs on the computer device shown.

[0216] Please see Figure 10 , Figure 10 This is a schematic block diagram of a computer device provided in an embodiment of this application. The computer device 500 can be a server, wherein the server can be a standalone server or a server cluster composed of multiple servers.

[0217] See Figure 10 The computer device 500 includes a processor 502, a memory, and a network interface 505 connected via a system bus 501. The memory may include a non-volatile storage medium 503 and internal memory 504.

[0218] The non-volatile storage medium 503 may store an operating system 5031 and a computer program 5032. The computer program 5032 includes program instructions that, when executed, cause the processor 502 to perform an adaptive dynamic risk monitoring and early warning method.

[0219] The processor 502 provides computing and control capabilities to support the operation of the entire computer device 500.

[0220] The internal memory 504 provides an environment for the operation of the computer program 5032 in the non-volatile storage medium 503. When the computer program 5032 is executed by the processor 502, the processor 502 can execute an adaptive dynamic risk monitoring and early warning method.

[0221] This network interface 505 is used for network communication with other devices. Those skilled in the art will understand that... Figure 10 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device 500 to which the present application is applied. The specific computer device 500 may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0222] The processor 502 is used to run a computer program 5032 stored in the memory to perform the following steps:

[0223] The process involves: acquiring static information and dynamic factors in the scene to be monitored to obtain initial data; performing entity recognition and risk factor identification on the initial data; combining the initial data with an existing knowledge base to obtain the potential risks of the scene to be monitored; employing a multi-head attention network for learning and fusion to obtain a vectorized representation of the potential risks of the scene to be monitored; acquiring real-time monitoring data and identifying real-time dynamic factors; using the real-time dynamic factors as input to calculate the risk information features after superimposing the vectorized representation of the potential risks with the real-time dynamic factors; determining the risk type, specific risk point, and risk value based on the risk information features; determining an early warning processing strategy based on the risk type, the specific risk point, and the risk value; and outputting the risk type, the specific risk point, the risk value, and the early warning processing strategy.

[0224] The static information includes the nature, characteristics, and load distribution of the area; the dynamic factors are key dynamic factors affecting the environment and decision-making, including activities, weather conditions, personnel conditions, and time nodes; the real-time dynamic factors are factors reflecting real-time changes in the environment and behavior, including crowd density, movement vectors, carried items, high-temperature areas, and sound change information.

[0225] In one embodiment, when the processor 502 performs entity recognition and risk factor identification on the initial data, obtains the potential risks of the scene to be monitored by combining the existing knowledge base, and performs learning and fusion using a multi-head attention network to obtain the vectorized representation result of the potential risks of the scene to be monitored, the processor 502 specifically implements the following steps:

[0226] The static information is subjected to entity recognition and vectorization to obtain a first vector; the dynamic factors are subjected to risk dynamic factor recognition and vectorization to obtain a second vector; the potential risks of the scene to be monitored are obtained by combining the existing knowledge base, the first vector, and the second vector, and then a multi-head attention network is used for learning and fusion to obtain the vectorized representation result of the potential risks of the scene to be monitored.

[0227] In one embodiment, when implementing the step of combining the existing knowledge base to obtain the potential risks of the scene to be monitored, the first vector, and the second vector to learn and fuse them using a multi-head attention network to obtain the vectorized representation result of the potential risks of the scene to be monitored, the processor 502 specifically implements the following steps:

[0228] By introducing GRU to construct a fusion gating system, the potential risks of the monitored scenario, the first vector, and the second vector are dynamically weighted and fused using the fusion gating system to obtain the vectorized representation of the potential risks of the monitored scenario.

[0229] In one embodiment, when the processor 502 implements the step of taking real-time dynamic factors as input and calculating the risk information feature after superimposing the vectorized representation result of the potential risk with real-time dynamic factors, the specific steps are as follows:

[0230] Long-term dependency features are extracted from the real-time dynamic factors; the long-term dependency features and the vectorized representation results of the potential risks are fused to obtain risk information features.

[0231] In one embodiment, when implementing the step of extracting long-term dependent features from the real-time dynamic factors, the processor 502 specifically implements the following steps:

[0232] Long-term dependency features are extracted from the real-time dynamic factors using gated dilated convolution operations.

[0233] In one embodiment, when the processor 502 fuses the long-term dependent features and the vectorized representation results of the potential risks to obtain risk information features, it specifically implements the following steps:

[0234] The long-term dependency features and the vectorized representation of potential risks are dynamically weighted and fused to obtain risk information features, wherein the weights used in the dynamic weighting are trained based on historical data.

[0235] In one embodiment, when implementing the step of determining the risk type, specific risk point, and risk value based on the risk information characteristics, the processor 502 specifically implements the following steps:

[0236] The risk information features are extracted using a shared encoder to perform multi-task classification to determine different risk types, and an attention mechanism is introduced to weight the output of each risk type. A spatiotemporal graph convolutional network is constructed to calculate the risk value and output the risk point information to obtain the specific risk point and the magnitude of the risk value.

[0237] In one embodiment, when the processor 502 implements the step of constructing the spatiotemporal graph convolutional network to calculate risk values ​​and output risk point information to obtain specific risk points and risk value magnitudes, the following steps are specifically implemented:

[0238] A spatiotemporal graph convolutional network is constructed to calculate the risk value of each sensor node through the spatiotemporal correlation between nodes, and output its corresponding location information to obtain the specific risk point and the magnitude of the risk value.

[0239] It should be understood that in the embodiments of this application, the processor 502 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0240] It will be understood by those skilled in the art that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program includes program instructions and can be stored in a storage medium, which is a computer-readable storage medium. The program instructions are executed by at least one processor in the computer system to implement the process steps of the embodiments of the above methods.

[0241] Therefore, the present invention also provides a storage medium. This storage medium can be a computer-readable storage medium. The storage medium stores a computer program, wherein when executed by a processor, the computer program causes the processor to perform the following steps:

[0242] The process involves: acquiring static information and dynamic factors from the scene to be monitored to obtain initial data; performing entity recognition and risk factor identification on the initial data; combining the initial data with an existing knowledge base to obtain the potential risks of the scene to be monitored; employing a multi-head attention network for learning and fusion to obtain a vectorized representation of the potential risks of the scene to be monitored; acquiring real-time monitoring data and identifying real-time dynamic factors; using the real-time dynamic factors as input to calculate the risk information features after superimposing the vectorized representation of the potential risks with the real-time dynamic factors; determining the risk type, specific risk point, and risk value based on the risk information features; determining an early warning processing strategy based on the risk type, the specific risk point, and the risk value; and outputting the risk type, the specific risk point, the risk value, and the early warning processing strategy.

[0243] The static information includes the nature, characteristics, and load distribution of the area; the dynamic factors are key dynamic factors affecting the environment and decision-making, including activities, weather conditions, personnel conditions, and time nodes; the real-time dynamic factors are factors reflecting real-time changes in the environment and behavior, including crowd density, movement vectors, carried items, high-temperature areas, and sound change information.

[0244] In one embodiment, when the processor executes the computer program to perform entity recognition and risk factor identification on the initial data, obtain the potential risks of the scene to be monitored by combining the existing knowledge base, and perform learning and fusion using a multi-head attention network to obtain the vectorized representation result of the potential risks of the scene to be monitored, the processor specifically implements the following steps:

[0245] The static information is subjected to entity recognition and vectorization to obtain a first vector; the dynamic factors are subjected to risk dynamic factor recognition and vectorization to obtain a second vector; the potential risks of the scene to be monitored are obtained by combining the existing knowledge base, the first vector, and the second vector, and then a multi-head attention network is used for learning and fusion to obtain the vectorized representation result of the potential risks of the scene to be monitored.

[0246] In one embodiment, when the processor executes the computer program to implement the step of combining the existing knowledge base to obtain the potential risks of the scene to be monitored, and learning and fusing the first vector and the second vector using a multi-head attention network to obtain the vectorized representation result of the potential risks of the scene to be monitored, the specific implementation is as follows:

[0247] By introducing GRU to construct a fusion gating system, the potential risks of the monitored scenario, the first vector, and the second vector are dynamically weighted and fused using the fusion gating system to obtain the vectorized representation of the potential risks of the monitored scenario.

[0248] In one embodiment, when the processor executes the computer program to implement the step of calculating the risk information features after superimposing the vectorized representation result of the potential risk with real-time dynamic factors as input, the specific implementation is as follows:

[0249] Long-term dependency features are extracted from the real-time dynamic factors; the long-term dependency features and the vectorized representation results of the potential risks are fused to obtain risk information features.

[0250] In one embodiment, when the processor executes the computer program to implement the step of extracting long-term dependent features from the real-time dynamic factors, it specifically implements the following steps:

[0251] Long-term dependency features are extracted from the real-time dynamic factors using gated dilated convolution operations.

[0252] In one embodiment, when the processor executes the computer program to fuse the long-term dependency features and the vectorized representation results of the potential risks to obtain risk information features, it specifically implements the following steps:

[0253] The long-term dependency features and the vectorized representation of potential risks are dynamically weighted and fused to obtain risk information features, wherein the weights used in the dynamic weighting are trained based on historical data.

[0254] In one embodiment, when the processor executes the computer program to implement the step of determining the risk type, specific risk point, and risk value based on the risk information characteristics, it specifically implements the following steps:

[0255] The risk information features are extracted using a shared encoder to perform multi-task classification to determine different risk types, and an attention mechanism is introduced to weight the output of each risk type. A spatiotemporal graph convolutional network is constructed to calculate the risk value and output the risk point information to obtain the specific risk point and the magnitude of the risk value.

[0256] In one embodiment, when the processor executes the computer program to implement the step of constructing a spatiotemporal graph convolutional network to calculate risk values ​​and output risk point information to obtain specific risk points and risk value magnitudes, the processor specifically implements the following steps:

[0257] A spatiotemporal graph convolutional network is constructed to calculate the risk value of each sensor node through the spatiotemporal correlation between nodes, and output its corresponding location information to obtain the specific risk point and the magnitude of the risk value.

[0258] The storage medium can be any computer-readable storage medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0259] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0260] In the embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For example, the division of each unit is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0261] The steps in the method of this invention can be adjusted, merged, or reduced in order according to actual needs. The units in the system of this invention can be merged, divided, or reduced according to actual needs. Furthermore, the functional units in the various embodiments of this invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0262] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a terminal, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0263] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An adaptive dynamic risk monitoring and early warning method, characterized in that, include: Acquire static information and dynamic factors in the scene to be monitored to obtain initial data; The initial data is used for entity recognition and risk factor identification. The potential risks of the scene to be monitored are obtained by combining the existing knowledge base. A multi-head attention network is used for learning and fusion to obtain the vectorized representation of the potential risks of the scene to be monitored. Acquire real-time monitoring data and identify real-time dynamic factors; Using real-time dynamic factors as input, the risk information features are calculated by superimposing the vectorized representation of the potential risk with real-time dynamic factors. The risk type, specific risk point, and risk value are determined based on the risk information characteristics. The early warning handling strategy is determined based on the risk type, the specific risk point, and the magnitude of the risk value. Output the risk type, the specific risk point, the risk value, and the early warning handling strategy.

2. The adaptive dynamic risk monitoring and early warning method according to claim 1, characterized in that, The static information includes the nature, characteristics, and load distribution of the area; the dynamic factors are key dynamic factors affecting the environment and decision-making, including activities, weather conditions, personnel conditions, and time nodes; the real-time dynamic factors are factors reflecting real-time changes in the environment and behavior, including crowd density, movement vectors, carried items, high-temperature areas, and sound change information.

3. The adaptive dynamic risk monitoring and early warning method according to claim 2, characterized in that, The process involves performing entity recognition and risk factor identification on the initial data, combining this with an existing knowledge base to obtain the potential risks of the scene to be monitored, and then using a multi-head attention network for learning and fusion to obtain a vectorized representation of the potential risks of the scene to be monitored. This includes: The static information is subjected to entity recognition and vectorization to obtain a first vector; The dynamic factors are identified and vectorized to obtain a second vector; The potential risks of the scene to be monitored are obtained by combining the existing knowledge base, the first vector, and the second vector. A multi-head attention network is then used to learn and fuse them to obtain the vectorized representation of the potential risks of the scene to be monitored.

4. The adaptive dynamic risk monitoring and early warning method according to claim 3, characterized in that, The step of combining the potential risks of the scene to be monitored obtained from the existing knowledge base, the first vector, and the second vector, and then learning and fusing them using a multi-head attention network to obtain a vectorized representation of the potential risks of the scene to be monitored, includes: By introducing GRU to construct a fusion gating system, the potential risks of the monitored scenario, the first vector, and the second vector are dynamically weighted and fused using the fusion gating system to obtain the vectorized representation of the potential risks of the monitored scenario.

5. The adaptive dynamic risk monitoring and early warning method according to claim 1, characterized in that, The step of taking real-time dynamic factors as input and calculating the risk information features after superimposing the vectorized representation of the potential risk with real-time dynamic factors includes: Extract long-term dependency features from the real-time dynamic factors; The long-term dependency features and the vectorized representation of potential risks are fused together to obtain risk information features.

6. The adaptive dynamic risk monitoring and early warning method according to claim 5, characterized in that, The extraction of long-term dependency features from the real-time dynamic factors includes: Long-term dependency features are extracted from the real-time dynamic factors using gated dilated convolution operations.

7. The adaptive dynamic risk monitoring and early warning method according to claim 5, characterized in that, The process of fusing the long-term dependency features and the vectorized representation of potential risks to obtain risk information features includes: The long-term dependency features and the vectorized representation of potential risks are dynamically weighted and fused to obtain risk information features, wherein the weights used in the dynamic weighting are trained based on historical data.

8. The adaptive dynamic risk monitoring and early warning method according to claim 1, characterized in that, The step of determining the risk type, specific risk point, and risk value based on the risk information characteristics includes: The risk information features are extracted using a shared encoder, and multi-task classification is performed to determine different risk types. An attention mechanism is introduced to weight the output of each risk type. Construct a spatiotemporal graph convolutional network to calculate risk values ​​and output risk point information to obtain specific risk points and their magnitudes.

9. The adaptive dynamic risk monitoring and early warning method according to claim 1, characterized in that, The construction of the spatiotemporal graph convolutional network to calculate risk values ​​and output risk point information, thereby obtaining specific risk points and their magnitudes, includes: A spatiotemporal graph convolutional network is constructed to calculate the risk value of each sensor node through the spatiotemporal correlation between nodes, and output its corresponding location information to obtain the specific risk point and the magnitude of the risk value.

10. An adaptive dynamic risk monitoring and early warning system, characterized in that, include: The initial acquisition unit is used to acquire static information and dynamic factors in the scene to be monitored in real time to obtain initial data; The fusion unit is used to perform entity recognition and risk factor recognition on the initial data, combine the existing knowledge base to obtain the potential risks of the scene to be monitored, and use a multi-head attention network for learning and fusion to obtain the vectorized representation of the potential risks of the scene to be monitored. The real-time acquisition unit is used to acquire real-time monitoring data and identify real-time dynamic factors; The calculation unit is used to take real-time dynamic factors as input and calculate the risk information features after superimposing the vectorized representation of the potential risk with real-time dynamic factors. The first determining unit is used to determine the risk type, specific risk point, and risk value based on the risk information characteristics. The second determining unit is used to determine the early warning processing strategy based on the risk type, the specific risk point, and the magnitude of the risk value. The output unit is used to output the risk type, the specific risk point, the risk value, and the early warning processing strategy.

Images