A fence area risk early warning method and system based on digital twinning

Abstract

This invention discloses a method and system for risk early warning of fenced areas based on digital twins, belonging to the technical field of risk early warning. It constructs a digital twin model of the fenced area, completing hierarchical digital representation of the model from geometric, behavioral, and rule dimensions; collects multi-source heterogeneous data, and constructs a standardized input set after data normalization; deploys a lightweight deep learning model based on edge computing, identifies unsafe behaviors against behavioral feature benchmarks, classifies risk levels according to standardized risk assessment criteria, and generates corresponding risk assessment results; maps the standardized input set and risk assessment results to the digital twin model of the fenced area in real time, achieving synchronous linkage between virtual and real scenes, and presenting early warning information in a multimodal visualization manner. This invention intuitively presents the safety status of the area through three-dimensional scene display, event recording, and other forms, facilitating managers to have a global grasp of the on-site situation, quickly locate risk locations and hazard types, and assist in on-site dispatch and emergency response.

Description

Technical Field

[0001] This invention belongs to the technical field of risk warning, and in particular relates to a method and system for risk warning of fenced areas based on digital twins. Background Technology

[0002] Safety fences are crucial infrastructure in industrial production sites for isolating automated equipment, delineating hazardous work boundaries, and preventing personnel from accidentally entering high-risk areas. They are widely used in smart manufacturing workshops, equipment maintenance areas, and industrial isolation scenarios. Traditional safety fences only provide physical isolation and cannot automatically sense or monitor personnel activities, equipment operating conditions, or on-site environmental parameters within the area. Daily safety management mainly relies on regular manual inspections, which has practical shortcomings such as limited inspection coverage, delayed hazard detection, high labor costs, and a focus on post-incident handling rather than pre-incident prevention.

[0003] Existing fence monitoring methods mostly rely on single video boundary crossing detection or independent sensor threshold alarms, resulting in fragmented functionality and significant shortcomings in practical applications. First, there is a lack of digital modeling and spatial relational representation of the overall elements of personnel, equipment and environment in the fenced area. The relationship between physical entities and the monitoring platform is limited to simple data collection and cannot form an integrated mapping of the entity's spatial location, operating status and behavioral logic. Second, the risk assessment methods are simple and crude, mostly relying on a single event or a single parameter threshold to trigger alarms, without comprehensively considering the coupled effects between abnormal personnel behavior, equipment status fluctuations, and changes in environmental parameters. There is a lack of systematic quantitative assessment standards and graded handling mechanisms, resulting in strong subjectivity and insufficient accuracy in risk assessment. With the increasing demand for intelligent and visual management and control in industrial sites, technologies such as digital twins, edge acquisition, and multi-parameter fusion assessment are gradually being implemented. However, existing solutions have not yet formed a layered modeling method adapted to fence scenarios, a lightweight detection architecture with multi-feature collaboration, or a dynamic risk assessment system that couples human, machine, and environment. It is difficult to simultaneously take into account scene fidelity, timeliness of behavior recognition, quantification of risk assessment, and synchronization of virtual and real states. Therefore, it is necessary to design a fence area risk early warning method and system based on digital twins to solve the practical defects of existing technologies. Summary of the Invention

[0004] To address the technical problems existing in the background art, the present invention provides a method and system for risk early warning of fenced areas based on digital twins.

[0005] This invention employs the following technical solution: a method for risk early warning of fenced areas based on digital twins, comprising the following steps: A digital twin model of a fenced area is constructed based on digital representation. The digital twin model of the fenced area includes: a personnel model, an equipment and fence model, and an environment model. The model is digitally represented in layers from the geometric, behavioral, and rule dimensions: the geometric dimension maps the relationship between the shape of the entity and its spatial position; the behavioral dimension quantifies and defines the entity's operating state and event response attributes, thereby forming an identifiable behavioral feature benchmark; and the rule dimension configures a multi-parameter fusion risk assessment and early warning judgment logic, thereby establishing a standardized risk judgment basis. Collect heterogeneous data on personnel, equipment, and environment from multiple sources within the area, and construct a standardized input set after data normalization processing; Lightweight deep learning models are deployed based on edge computing, unsafe behaviors are identified by comparing them with behavioral feature benchmarks, risk levels are classified according to standardized risk assessment criteria, and corresponding risk assessment results are generated. The standardized input set and risk assessment results are mapped in real time to the digital twin model of the fenced area, realizing the linkage and synchronization of virtual and real scenes, and presenting the early warning information in a multimodal visualization manner.

[0006] In a further embodiment, the environment model includes: a static environment model and a dynamic environment model; the static environment model is defined as a set of fixed scene elements, and the dynamic environment model is a set of real-time environment data streams, employing a layered decoupled modeling approach between the set of fixed scene elements and the set of real-time environment data streams. When constructing a static environment model, the overall spatial position of fixed facilities in the fenced area and the three-dimensional coordinates of the installation of various sensors are solidified using a set of fixed elements of the scene. When constructing a dynamic environment model, the corresponding real-time environmental parameters are retrieved based on the real-time environmental data stream set using the installation three-dimensional coordinates. The real-time data labels and parameter time-series polyline curves are then fixedly attached to the corresponding sensor three-dimensional coordinate positions, thereby achieving a fixed static scene structure and independent fixed-point refresh of dynamic monitoring data.

[0007] In a further embodiment, the behavioral dimension is quantified in the following way: Construct a bimodal parameterized model of entity behavior and define a set of behavior parameters. ,in, These are the entity's normal dynamic parameters, used to reflect the baseline distribution of the entity's operating state; For entity event response parameters, use entity normal dynamic parameters. Using the baseline distribution as a reference, the deviation of the current state from the baseline distribution is calculated to form an abnormal feature vector of the event response; Based on the baseline-deviation correspondence of dual-modal parameters, an identifiable behavioral feature baseline is formed through encoding.

[0008] In a further embodiment, the rule dimension is defined as follows: Construct a multi-parameter fusion risk assessment model and define the risk assessment parameter set. ;in: These are static risk weight parameters used to reflect the baseline distribution of the inherent risk level of the fenced area; These are dynamic risk correction parameters, which are correction terms constructed based on behavioral deviation and environmental conditions. ; Based on formula By integrating static benchmarks and dynamic adjustments, the comprehensive risk value is calculated. ; By setting up multiple risk threshold ranges, different levels of early warning responses can be triggered, forming a standardized basis for risk assessment.

[0009] In a further embodiment, the output flow of the lightweight deep learning model is as follows: Define the standardized input set as ,in, For video frame features, For equipment status characteristics, Characteristics of environmental operating conditions; Set a lightweight deep learning model as , These are the network weight parameters of the model; using video frame features Main feature input, coupled with device state features and environmental operating conditions Construct multi-source correlation constraints and adaptively simplify convolutional channels through a feature sensitivity screening mechanism to retain effective representation pathways for unsafe behaviors; Define the standardized input set Input Model Perform forward inference to locate the human target, extract behavioral features, and classify behavioral types, and output a normalized behavioral feature vector. .

[0010] In a further embodiment, the real-time mapping method for mapping the standardized input set and risk assessment results to the digital twin model of the fenced area is as follows: The standardized input set is mapped to the personnel model, equipment and fence model, and environment model of the digital twin model of the fenced area according to the hierarchical decoupling modeling method; The risk assessment results are visualized and overlaid onto the corresponding physical locations in the digital twin model of the fenced area, achieving a one-to-one correspondence between risk levels and physical locations. At the same time, a two-way data link is established between the digital twin model and the physical fence system, and the early warning instructions in the model are back-synchronized to the physical monitoring terminal to trigger response actions.

[0011] In a further embodiment, the multi-source association constraint is constructed as follows: Video frame features Equipment status characteristics and environmental operating conditions Mapped to behavioral deep features respectively Equipment characteristics and environmental characteristics Construct a multi-source correlation constraint loss function: ;in, Identify mission losses by identifying unsafe acts by personnel. For device-environment joint mapping function, This is the coefficient for balancing the associated constraints.

[0012] In a further embodiment, the feature sensitivity screening mechanism is as follows: During model training, backpropagation is used to calculate the weights of each convolutional channel relative to the total loss value. gradient sensitivity : , It is the first The weight parameters of each convolutional channel; Based on gradient sensitivity The channels are dynamically filtered. For high-sensitivity channels with gradient sensitivity higher than a preset threshold, their weight update paths are retained and the weights are continuously strengthened through gradient backpropagation, so that their ability to represent the behavior-device-environment correlation features is gradually enhanced. For low-sensitivity channels with gradient sensitivity below a preset threshold, the weights of redundant channels are made close to zero.

[0013] In a further embodiment, the risk levels include: low risk, medium risk, high risk, and emergency risk.

[0014] A digital twin-based fenced area risk warning system, used to implement the fenced area risk warning method described above, includes: The multi-source data acquisition module is configured on-site in the fenced area to collect video surveillance data, equipment operation status data, and environmental monitoring data, thereby constructing a multi-source heterogeneous data acquisition link for personnel, equipment, and environment within the area. The edge computing and behavior recognition module deploys a lightweight deep learning model to process the collected data, construct a standardized input set, identify unsafe behaviors of people within the fenced area, and output a normalized behavior feature vector. The risk assessment and early warning determination module is configured with a multi-parameter fusion risk assessment model, which is used to calculate the comprehensive risk value based on behavioral deviation, equipment status and environmental conditions, and to divide the risk level based on preset multi-level risk thresholds and generate graded early warning instructions. The digital twin modeling and visualization module constructs a digital twin scene that includes personnel models, equipment and fence models, and environmental models. It enables real-time mapping and synchronization of multi-source data and risk assessment results, and presents risk status and early warning information in a multimodal visualization manner. The linkage control and emergency response module establishes a two-way data link with the physical fence system to receive early warning commands from the digital twin model, triggering response actions such as audible and visual alarms, equipment linkage control, and emergency information push, thereby achieving closed-loop early warning management.

[0015] The beneficial effects of this invention are as follows: This invention completes the hierarchical digital representation of various entities in the fenced area from three levels: geometry, behavior, and rules. At the same time, it decouples the environmental model into static fixed elements and dynamic real-time data streams for modeling. This ensures that the virtual scene and the physical site are highly consistent in terms of shape, size, and spatial layout. It also enables the static scene structure to be fixed and the dynamic monitoring data to be refreshed independently at fixed points. This provides a high-fidelity digital foundation for the synchronization of fenced area status, situation display, and risk assessment.

[0016] This invention constructs a standardized input set based on multi-source data from the field, and forms an association constraint relationship with equipment status and environmental conditions. By using channel sensitivity filtering and sparse constraint methods, it simplifies redundant parameters of the model, achieving a lightweight model structure while ensuring the behavior recognition effect remains unchanged. It can be deployed on the field edge computing terminal without relying on centralized computing on the backend server, effectively reducing data transmission pressure, shortening the latency of unsafe behavior recognition and status feedback, and adapting to the application needs of real-time monitoring and early warning in industrial fields.

[0017] This invention establishes a benchmark for entity behavior and a quantitative method for state deviation. It introduces static risk benchmark parameters and dynamic correction parameters to perform multi-parameter fusion calculations to obtain the comprehensive risk value of the fenced area. It also divides the risk threshold intervals into multiple levels to match the corresponding early warning levels and handling strategies. This changes the traditional extensive mode of relying on experience-based qualitative judgment and single threshold alarms, and realizes the quantification, standardization and grading of risk assessment. The basis for risk judgment is more sufficient and the classification of early warning levels is more reasonable.

[0018] This invention maps multi-source monitoring data, personnel behavior recognition results, equipment operating status, and comprehensive risk assessment results to a unified digital twin virtual scene, enabling real-time linkage and synchronization between the physical fence and the virtual scene in terms of spatial location, physical status, and risk situation. It intuitively presents the regional safety status through various forms such as 3D scene display, data dashboard, status indicators, and event records, facilitating managers to have a comprehensive grasp of the on-site situation, quickly locate risk locations and hazard types, and assist in on-site dispatch and emergency response. Attached Figure Description

[0019] Figure 1 This is a flowchart of the risk warning method for fenced areas in Example 1.

[0020] Figure 2 This is a physical layout diagram of the fenced area in Example 1.

[0021] Figure 3 This is the basic layout diagram of Example 2. Detailed Implementation

[0022] The present invention will now be further described with reference to the accompanying drawings and embodiments.

[0023] Example 1 This embodiment is based on a security fence entity: its basic components include a steel fence, a camera for target detection, a power distribution box, and an audible and visual alarm. Each side of the fence has a security door and a security light curtain. The right side of the fence is equipped with an all-in-one machine with an edge computing device, which can also be used to display environmental data.

[0024] Based on this, such as Figure 1 This embodiment discloses a method for risk warning of fenced areas based on digital twins, including the following steps: A digital twin model of a fenced area is constructed based on digital representation. The digital twin model of the fenced area includes: a personnel model, an equipment and fence model, and an environment model. The model is digitally represented in layers from the geometric, behavioral, and rule dimensions: the geometric dimension maps the relationship between the shape of the entity and its spatial position; the behavioral dimension quantifies and defines the entity's operating state and event response attributes, thereby forming an identifiable behavioral feature benchmark; and the rule dimension configures a multi-parameter fusion risk assessment and early warning judgment logic, thereby establishing a standardized risk judgment basis. Collect heterogeneous data on personnel, equipment, and environment from multiple sources within the area, and construct a standardized input set after data normalization processing; Lightweight deep learning models are deployed based on edge computing, unsafe behaviors are identified by comparing them with behavioral feature benchmarks, risk levels are classified according to standardized risk assessment criteria, and corresponding risk assessment results are generated. The standardized input set and risk assessment results are mapped in real time to the digital twin model of the fenced area, realizing the linkage and synchronization of virtual and real scenes, and presenting the early warning information in a multimodal visualization manner.

[0025] The environment model includes: a static environment model and a dynamic environment model; the static environment model is defined as a set of fixed elements of the scene, such as... Figure 2 As shown, 3D modeling and engine development tools such as Unity3D, Solidworks, and Blender are generally used for construction.

[0026] The dynamic environment model is a collection of real-time environment data streams, employing a layered decoupled modeling approach that combines a set of fixed scene elements with a set of real-time environment data streams. When constructing a static environment model, the overall spatial position of fixed facilities in the fenced area and the three-dimensional coordinates of the installation of various sensors are solidified using a set of fixed elements of the scene. When constructing a dynamic environment model, the corresponding real-time environment parameters are retrieved based on the real-time environment data stream set using the installed three-dimensional coordinates. The real-time data labels and parameter time-series polyline curves are then fixedly attached to the corresponding sensor three-dimensional coordinate positions, thereby achieving a fixed static scene structure and independent fixed-point refresh of dynamic monitoring data, reducing the overall rendering computation load of the three-dimensional scene.

[0027] This application adopts a coordinate index matching mechanism, using the sensor's three-dimensional coordinates pre-stored in the static model as a unique index. In the massive real-time environmental data stream, it filters and retrieves the exclusive environmental data collected by the sensor at that coordinate point, thus preventing data corruption across points.

[0028] Taking fenced area monitoring equipment as an example: The fixed facilities in the fenced area mainly include permanent fixed structures such as the factory ground, walls, ceiling, basic lighting equipment, fence posts, protective nets, and safety door frames. These facilities together constitute the complete physical space base of the fenced area. The overall spatial position is replicated and solidified in the digital twin static model, and the structure, position, and shape remain fixed throughout the process, providing a stable spatial reference system for dynamic data loading.

[0029] On this fixed spatial base, all kinds of environmental sensing devices deployed in the scene are preset with independent and unique three-dimensional spatial coordinates, which serve as points for data matching and visualization, such as temperature and humidity sensors, noise sensors, infrared sensing devices, etc.

[0030] Meanwhile, the retrieved environmental parameters will be presented in a visual form with real-time data tags and parameter time-series curves attached to the corresponding sensor's three-dimensional coordinate position. The real-time data tags will display the current instantaneous monitoring values, intuitively presenting the current environmental status. The parameter time-series curves will continuously record and display the time-series change trends of environmental parameters, clearly reflecting the fluctuation patterns of environmental elements such as temperature, humidity, and noise, achieving a modeling effect of fixed static scenes and precise dynamic data linkage visualization within the fenced area.

[0031] It should be noted that the mounting described in this embodiment uses the sensor's three-dimensional coordinates as the positioning reference, binds and fixes the retrieved numerical labels and time-series change curves to the corresponding spatial points, and displays them synchronously with the scene's perspective. Only the data content is dynamically updated, while the position and shape of the scene's base remain unchanged.

[0032] The method for quantifying the behavioral dimension described in this embodiment is as follows: Construct a bimodal parameterized model of entity behavior and define a set of behavior parameters. ,in, The entities described in this embodiment, which are considered as normal dynamic parameters, include at least personnel, protective equipment, and sensing devices. Therefore, the normal dynamic parameters of the entities... The temporal-spatial coupled projection algorithm is used for quantization, which can be further understood as mapping the normal operating conditions of equipment and the routine trajectories of personnel into normalized steady-state feature vectors to establish a baseline distribution of the entity's operating state. The normal operating conditions of equipment are reflected by the equipment sampling frequency, operating current fluctuation range, and average response delay, while the routine trajectories of personnel are reflected by the path coordinate sequence, movement speed range, and dwell time distribution.

[0033] For entity event response parameters, use entity normal dynamic parameters. Using the baseline distribution as a reference, the deviation of the current state from the baseline distribution is calculated to form an anomaly feature vector of the event response. The calculation formula in this embodiment is as follows: In the formula, This is the state feature vector of the entity at the previous moment (such as the current position / speed of a person, or the current current / delay of a device). For entity's normal dynamic parameters In the mean vector of the normal state characteristics of the entity (i.e., the center of the baseline distribution). For entity's normal dynamic parameters In this context, the standard deviation vector of the entity's normal state characteristics (i.e., the fluctuation range of the baseline distribution). The Euclidean distance between the current state and the normal baseline (the direct difference in feature space).

[0034] Among them, standardized deviation The larger the value, the further the current state deviates from the normal baseline, and the higher the probability of an anomaly; Standardized Deviation The smaller the value, the closer the current state is to the normal baseline, which is considered normal behavior / operating condition.

[0035] Furthermore, entity event response parameters can also be used to identify unsafe behaviors such as personnel not wearing safety helmets. Head images of personnel wearing safety helmets in compliant work scenarios are pre-collected, and typical features of the helmet's color, outline, size, and positional proportion are extracted to construct a baseline vector set of head features for helmet-wearing personnel, which serves as part of the normal dynamic parameters. On the edge computing side, features of the personnel's head region in real-time video frames are extracted, and the real-time head feature vectors are matched with the baseline feature vector set for similarity. The feature deviation is calculated; if the similarity is below a preset threshold, it is determined to be an abnormal state of not wearing a safety helmet, and corresponding event response parameters are generated. These parameters can be input along with multiple parameters such as personnel location, equipment status, and environmental conditions into a multi-parameter fusion risk assessment model to calculate a comprehensive risk value and trigger a corresponding level of early warning response.

[0036] Based on the baseline-deviation correspondence of dual-modal parameters, an identifiable behavioral feature baseline is formed through encoding.

[0037] To further illustrate, for a person entity, the distribution of conventional trajectory coordinates, movement speed range, and dwell time is used to construct... The steady-state eigenvector, when the standardized deviation is calculated Exceeding the preset threshold (e.g.) When this condition occurs, the state is marked as an abnormal behavior of out of bounds / delay / climbing, with the corresponding codes being E01 / E02 / E03.

[0038] For the sensing device entity, the sampling frequency, operating current fluctuation, and average response delay constitute... The steady-state eigenvector, when the standardized deviation is calculated Exceeding a preset threshold (e.g.) When the status is marked as an abnormal behavior of sampling abnormality / equipment failure / communication interruption, the corresponding codes are D01 / D02 / D03.

[0039] For the physical protective equipment, the feedback of fence status and the alarm triggering delay constitute the basis. The steady-state eigenvector, when the standardized deviation is calculated Exceeding the preset threshold (e.g.) When this condition occurs, the status is marked as an abnormal behavior of fence failure / alarm delay, with the corresponding codes F01 / F02; By establishing a one-to-one mapping relationship between different entities, different deviation ranges, anomaly types, and unique codes, a standardized behavioral feature benchmark library can be formed that can be used for subsequent behavior identification and risk assessment.

[0040] Correspondingly, the definition of the rule dimension is as follows: Construct a multi-parameter fusion risk assessment model and define the risk assessment parameter set. ;in: These are static risk weight parameters, used to reflect the baseline distribution of the inherent risk level of the fenced area, determined by the area level weights. and entity attribute weights Weighted composition, that is: , .

[0041] in, This represents the inherent risk value of the region. Both are inherent risk values ​​of the entity. Normalized risk value within the interval.

[0042] For ease of understanding, the following examples illustrate the possible values ​​in this embodiment: For high-protection-level fenced areas, the following settings are provided. , This indicates that the area itself has a high risk level; For operations and maintenance personnel entities, set up , This indicates that the entity attribute has a low risk; therefore, the static risk weight parameter... .

[0043] These are dynamic risk correction parameters, which are correction terms constructed based on behavioral deviation and environmental conditions. : In the formula, and These are all correction factors, with a value range of [value range missing]. , For environmental operating conditions, The normalized environmental risk value within the interval can be obtained by mapping environmental indicators such as visibility, wind speed, nighttime conditions, and rainfall intensity. These environmental condition variables... The use of normalization design aims to maintain a consistent quantitative system with the previously mentioned risk values ​​and deviations, while also avoiding integration issues arising from differences in the dimensions of indicators across different environments. Based on the above example, , , ,but .

[0044] Based on formula By integrating static benchmarks and dynamic adjustments, the comprehensive risk value is calculated. ; Spatiotemporal alignment, noise filtering, missing data completion, and format normalization are performed on multi-source heterogeneous sensing data to construct a standardized input set. Specifically, a timestamp synchronization algorithm is used to achieve temporal alignment of data from different devices, Kalman filtering is used to remove environmental noise and device jitter data, and linear interpolation / trend prediction is used to complete missing data in sensing blind spots. The raw data of different formats are uniformly mapped into a structured data format compatible with the digital twin model, providing a reliable data foundation for subsequent behavioral dimension quantification and rule dimension evaluation.

[0045] Define the standardized input set as ,in, For video frame features, For equipment status characteristics, Characteristics of environmental operating conditions; Set a lightweight deep learning model as , These are the network weight parameters of the model; using video frame features Main feature input, coupled with device state features and environmental operating conditions Construct multi-source correlation constraints and adaptively simplify convolutional channels through a feature sensitivity screening mechanism to retain effective representation pathways for unsafe behaviors; Define the standardized input set Input Model Perform forward reasoning to locate human targets, extract behavioral features, and classify behavior types. Identify two types of violations: not wearing reflective work clothes and climbing over fences. Output a normalized behavior feature vector. ; behavioral feature vector Behavioral dimensions are quantified, and unsafe behaviors are identified by comparing them with behavioral characteristic benchmarks. This involves generating corresponding multi-level risk threshold ranges based on the rules outlined above for identification.

[0046] In this embodiment, the multi-source association constraint is constructed as follows: Video frame features Equipment status characteristics and environmental operating conditions Mapped to behavioral deep features respectively Equipment characteristics and environmental characteristics Construct a multi-source correlation constraint loss function: ;in, To identify unsafe behaviors of personnel and reduce task losses, this method is used to drive the model to learn the characteristic patterns of target behaviors such as personnel not wearing reflective work clothes and climbing over fences in fenced areas. It is a device-environment joint mapping function used to characterize the coupling constraint relationship between device status and environmental conditions on personnel behavior characteristics. It is generally a trainable multimodal fusion module. This is the correlation constraint balance coefficient, used to adjust the weight of multi-source correlation constraint terms in the total loss. It is generally set to a value of 0 to 1.

[0047] By minimizing the total loss value Iterative optimization of model parameters This allows the model to automatically strengthen the weights of convolutional channels strongly correlated with fence-related unsafe behaviors during training, while suppressing interference from irrelevant background and weakly correlated features, thus intrinsically achieving model lightweighting. Further, this means that during backpropagation... It will affect the weight updates of each convolutional channel in the model: If the output characteristics of a certain channel can effectively drive video behavior characteristics Approximating the combined constraints of the device and the environment If the channel weight contributes positively to the gradient of the constraint term, the model will continuously amplify and strengthen the channel weight through gradient descent, making it an effective feature channel strongly correlated with unsafe fence behavior. If the output features of a certain channel are unrelated to the device state and environmental constraints (such as background noise, irrelevant textures, etc.), its gradient contribution to the constraint terms is minimal or even zero. The model will weaken its weight update through a sparse regularization mechanism, eventually pushing the weights of redundant channels to zero, thereby achieving automatic elimination of irrelevant feature channels.

[0048] Through the above mechanism, the model intrinsically strengthens key channels and simplifies redundant channels during training, achieving lightweighting without the need for additional independent pruning modules, while improving the robustness of unsafe behavior detection under complex working conditions.

[0049] In a further embodiment, the real-time mapping method for mapping the standardized input set and risk assessment results to the digital twin model of the fenced area is as follows: The standardized input set is mapped to the personnel model, equipment and fence model, and environment model of the digital twin model of the fenced area using a hierarchical decoupled modeling approach. Specifically, video frame features, after human target localization and behavioral feature extraction, are synchronously updated to update the dynamic posture and behavioral state of the personnel model. For example, if a person is not wearing reflective work clothes, this behavior is identified and marked in real time, and the corresponding behavioral feature vector is synchronously updated to the digital twin model of the fenced area, triggering subsequent risk assessment processes. Similarly, if a person climbs over or crosses the fence, this violation is detected and marked in real time, and the location and behavioral feature data are synchronously mapped to the personnel model in the virtual scene, achieving real-time synchronization of the person's location and behavioral state between the virtual and real worlds.

[0050] Equipment status characteristics and environmental operating condition characteristics are attached to the corresponding nodes of the equipment, fence model, and environment model according to preset spatial coordinates, so as to realize the real-time alignment of virtual and real status of multi-source heterogeneous data.

[0051] The risk assessment results are visualized (such as risk heat maps, highlighted marks, and warning pop-ups) and overlaid onto the corresponding physical locations in the digital twin model of the fenced area to achieve a one-to-one correspondence between risk levels and physical locations. At the same time, a two-way data link is established between the digital twin model and the physical fence system to synchronize the warning instructions in the model back to the physical monitoring terminal, triggering response actions such as audible and visual alarms and equipment linkage.

[0052] Based on the virtual-real linkage and synchronization mechanism, the system realizes full-element visualization of personnel status, equipment operation, environmental parameters, and risk level in the physical fence area, providing real-time and intuitive digital twin support for on-site control and emergency response.

[0053] To achieve a quantitative classification of risk levels, three risk thresholds are set. , , Normalized composite risk value It is divided into four zones, with corresponding risk levels: low risk, medium risk, high risk, and emergency risk.

[0054] Specifically, the low-risk range is: The corresponding personnel's normal activities, equipment operation, and environmental parameters are all within the safe baseline range; The medium-risk range is: This could be due to personnel exhibiting actions that slightly deviate from behavioral benchmarks, or minor fluctuations in equipment or environmental parameters. The high-risk range is: This is due to unsafe acts by relevant personnel, or obvious abnormalities in equipment or environmental parameters. The range of emergency risk is: This indicates that personnel have climbed over the fence, or that an emergency has occurred, such as equipment malfunction or severely excessive environmental parameters.

[0055] when When the value falls into the high-risk range, the system determines it to be a high-risk state, triggers a high-risk warning, and outputs a linkage warning signal between personnel crossing the boundary and harsh environmental conditions.

[0056] In a further embodiment, corresponding early warning response strategies are configured respectively: Low risk: The overall risk value is below the first threshold. This corresponds to situations where personnel are engaged in normal activities within the fenced area, equipment operation, and environmental parameters are all within the safe baseline range; the system only records monitoring data and does not trigger proactive warnings. Medium risk: The overall risk value is at the first threshold. With the second threshold In between, when a person makes a slight deviation from the behavioral baseline (such as approaching a fence but not climbing over it) or when there is a small fluctuation in equipment / environmental parameters, an audible and visual alert is triggered and a reminder message is pushed to the monitoring terminal. High risk: The overall risk value is at the second threshold. With the third threshold In the meantime, if personnel engage in unsafe behavior (such as climbing or scaling fences) or if equipment / environment parameters show obvious abnormalities, the system will determine it as a high-risk situation, trigger a high-risk warning, sound and light alarms, lock the corresponding monitoring screen, and push the warning information to the on-site management personnel. Emergency Risk: When the comprehensive risk value exceeds the third threshold, it corresponds to an emergency event such as personnel climbing over the fence, equipment failure, or severely excessive environmental parameters. This triggers linkage control commands, such as closing the fence passage, activating emergency lighting, and pushing the highest-level warning information to the management platform and emergency response terminal.

[0057] In summary, this embodiment achieves a digital representation of all elements of the fenced area—humans, machines, and environment—by constructing a multi-dimensional, hierarchical digital twin model encompassing geometry, behavior, and rules. It also achieves real-time detection of unsafe behaviors while maintaining recognition accuracy through edge deployment of a lightweight deep learning model. Furthermore, it quantifies and calculates the comprehensive risk value and classifies it into four warning levels using a multi-parameter fusion risk assessment model. Finally, it provides an intuitive presentation of the risk situation and tiered response through real-time virtual-real mapping and multimodal visualization. This embodiment fully verifies the feasibility and effectiveness of the method and system described in this invention, effectively improving the automation and standardization of safety management in industrial fenced areas and providing reliable support for on-site safety management and emergency response.

[0058] Example 2 This embodiment discloses a fenced area risk warning system based on digital twins, used to implement the fenced area risk warning method described in the embodiment, including: The multi-source data acquisition module is configured on-site in the fenced area to collect video surveillance data, equipment operation status data, and environmental monitoring data, thereby constructing a multi-source heterogeneous data acquisition link for personnel, equipment, and environment within the area. The edge computing and behavior recognition module deploys a lightweight deep learning model to process the collected data, construct a standardized input set, identify unsafe behaviors of people within the fenced area, and output a normalized behavior feature vector. The risk assessment and early warning determination module is configured with a multi-parameter fusion risk assessment model, which is used to calculate the comprehensive risk value based on behavioral deviation, equipment status and environmental conditions, and to divide the risk level based on preset multi-level risk thresholds and generate graded early warning instructions. The digital twin modeling and visualization module constructs a digital twin scene that includes personnel models, equipment and fence models, and environmental models. It enables real-time mapping and synchronization of multi-source data and risk assessment results, and presents risk status and early warning information in a multimodal visualization manner. The linkage control and emergency response module establishes a two-way data link with the physical fence system to receive early warning commands from the digital twin model, triggering response actions such as audible and visual alarms, equipment linkage control, and emergency information push, thereby achieving closed-loop early warning management.

Claims

1. A method for risk early warning of fenced areas based on digital twins, characterized in that, Includes the following steps: A digital twin model of a fenced area is constructed based on digital representation. The digital twin model of the fenced area includes: a personnel model, an equipment and fence model, and an environment model. The model is digitally represented in layers from the geometric, behavioral, and rule dimensions: the geometric dimension maps the relationship between the shape of the entity and its spatial position; the behavioral dimension quantifies and defines the entity's operating state and event response attributes, thereby forming an identifiable behavioral feature benchmark; and the rule dimension configures a multi-parameter fusion risk assessment and early warning judgment logic, thereby establishing a standardized risk judgment basis. Collect heterogeneous data on personnel, equipment, and environment from multiple sources within the area, and construct a standardized input set after data normalization processing; Lightweight deep learning models are deployed based on edge computing, unsafe behaviors are identified by comparing them with behavioral feature benchmarks, risk levels are classified according to standardized risk assessment criteria, and corresponding risk assessment results are generated. The standardized input set and risk assessment results are mapped in real time to the digital twin model of the fenced area, realizing the linkage and synchronization of virtual and real scenarios, and presenting the early warning information in a multimodal visualization manner.

2. The method for risk early warning of fenced areas based on digital twins according to claim 1, characterized in that, The environment model includes a static environment model and a dynamic environment model. The static environment model is defined as a set of fixed scene elements, and the dynamic environment model is defined as a set of real-time environment data streams. A layered, decoupled modeling approach is adopted, combining the set of fixed scene elements and the set of real-time environment data streams. When constructing a static environment model, the overall spatial position of fixed facilities in the fenced area and the three-dimensional coordinates of the installation of various sensors are solidified using a set of fixed elements of the scene. When constructing a dynamic environment model, the corresponding real-time environmental parameters are retrieved based on the real-time environmental data stream set using the installation three-dimensional coordinates, and the real-time data labels and parameter time series line curves are fixedly attached to the corresponding sensor three-dimensional coordinate positions.

3. The method for risk early warning of fenced areas based on digital twins according to claim 1, characterized in that, The method for quantifying the behavioral dimensions is as follows: Construct a bimodal parameterized model of entity behavior and define a set of behavior parameters. ,in, These are the entity's normal dynamic parameters, used to reflect the baseline distribution of the entity's operating state; For entity event response parameters, use entity normal dynamic parameters. Using the baseline distribution as a reference, the deviation of the current state from the baseline distribution is calculated to form an abnormal feature vector of the event response; Based on the baseline-deviation correspondence of dual-modal parameters, an identifiable behavioral feature baseline is formed through encoding.

4. The method for risk early warning of fenced areas based on digital twins according to claim 1, characterized in that, The rule dimensions are defined as follows: Construct a multi-parameter fusion risk assessment model and define the risk assessment parameter set. ;in: These are static risk weight parameters used to reflect the baseline distribution of the inherent risk level of the fenced area; These are dynamic risk correction parameters, which are correction terms constructed based on behavioral deviation and environmental conditions. ; Based on formula By integrating static benchmarks and dynamic adjustments, the comprehensive risk value is calculated. ; By setting up multiple risk threshold ranges, different levels of early warning responses can be triggered, forming a standardized basis for risk assessment.

5. The method for risk early warning of fenced areas based on digital twins according to claim 1, characterized in that, The output process of the lightweight deep learning model is as follows: Define the standardized input set as ,in, For video frame features, For equipment status characteristics, Characteristics of environmental operating conditions; Set a lightweight deep learning model as , These are the network weight parameters of the model; using video frame features Main feature input, coupled with device state features and environmental operating conditions Construct multi-source correlation constraints and adaptively simplify convolutional channels through a feature sensitivity screening mechanism to retain effective representation pathways for unsafe behaviors; Define the standardized input set Input Model Perform forward inference to locate the human target, extract behavioral features, and classify behavioral types, and output a normalized behavioral feature vector. .

6. The method for risk early warning of fenced areas based on digital twins according to claim 1, characterized in that, The real-time mapping method for mapping the standardized input set and risk assessment results to the digital twin model of the fenced area is as follows: The standardized input set is mapped to the personnel model, equipment and fence model, and environment model of the digital twin model of the fenced area according to the hierarchical decoupling modeling method; The risk assessment results are visualized and overlaid onto the corresponding physical locations in the digital twin model of the fenced area, achieving a one-to-one correspondence between risk levels and physical locations. At the same time, a two-way data link is established between the digital twin model and the physical fence system, and the early warning instructions in the model are back-synchronized to the physical monitoring terminal to trigger response actions.

7. A method for risk early warning of fenced areas based on digital twins according to claim 5, characterized in that, The multi-source association constraints are constructed as follows: Video frame features Equipment status characteristics and environmental operating conditions Mapped to behavioral deep features respectively Equipment characteristics and environmental characteristics Construct a multi-source correlation constraint loss function: ;in, Identifying task losses through unsafe acts by personnel. For device-environment joint mapping function, This is the coefficient for balancing the associated constraints.

8. A method for risk early warning of fenced areas based on digital twins according to claim 5, characterized in that, The feature sensitivity screening mechanism is as follows: During model training, backpropagation is used to calculate the weights of each convolutional channel relative to the total loss value. gradient sensitivity : , It is the first The weight parameters of each convolutional channel; Based on gradient sensitivity The channels are dynamically filtered. For high-sensitivity channels with gradient sensitivity higher than a preset threshold, their weight update paths are retained and the weights are continuously strengthened through gradient backpropagation, so that their ability to represent the behavior-device-environment correlation features is gradually enhanced. For low-sensitivity channels with gradient sensitivity below a preset threshold, the weights of redundant channels are made close to zero.

9. A method for risk early warning of fenced areas based on digital twins according to claim 1, characterized in that, The risk levels are categorized as: low risk, medium risk, high risk, and emergency risk.

10. A fenced area risk early warning system based on digital twins, used to implement the fenced area risk early warning method as described in any one of claims 1 to 9, characterized in that, include: The multi-source data acquisition module is configured on-site in the fenced area to collect video surveillance data, equipment operation status data, and environmental monitoring data, thereby constructing a multi-source heterogeneous data acquisition link for personnel, equipment, and environment within the area. The edge computing and behavior recognition module deploys a lightweight deep learning model to process the collected data, construct a standardized input set, identify unsafe behaviors of people within the fenced area, and output a normalized behavior feature vector. The risk assessment and early warning determination module is configured with a multi-parameter fusion risk assessment model, which is used to calculate the comprehensive risk value based on behavioral deviation, equipment status and environmental conditions, and to divide the risk level based on preset multi-level risk thresholds and generate graded early warning instructions. The digital twin modeling and visualization module constructs a digital twin scene that includes personnel models, equipment and fence models, and environmental models. It enables real-time mapping and synchronization of multi-source data and risk assessment results, and presents risk status and early warning information in a multimodal visualization manner. The linkage control and emergency response module establishes a two-way data link with the physical fence system to receive early warning commands from the digital twin model, triggering response actions such as audible and visual alarms, equipment linkage control, and emergency information push, thereby achieving closed-loop early warning management.

Images