Identification methods, systems, and storage media for laboratory safety hazard management

Abstract

This invention discloses a method, system, and storage medium for identifying and managing laboratory safety hazards. The method includes: constructing a three-dimensional digital twin space for the laboratory; acquiring multi-source heterogeneous data such as real-time video streams, environmental parameters, and hazardous chemical storage and retrieval status through visual acquisition devices, environmental sensors, and intelligent cabinet terminals; performing spatiotemporal alignment and feature extraction via edge computing to generate a fused perception data stream containing location information and confidence evaluation, which is then mapped to the twin space. A large language model intelligent agent parses the data, constructs a multi-dimensional risk knowledge graph of personnel, environment, and behavior, and infers behavioral intentions to distinguish between normal operations and violations. Based on the intention results, the D-S evidence theory is used to fuse behavioral and environmental data to calculate the regional risk probability, generate a dynamic risk heatmap, and formulate control strategies. Physical intelligent terminals are driven to execute and record feedback data. This provides reliable technical support for intelligent laboratory safety management.

Description

Technical Field

[0001] This invention belongs to the field of Internet of Things (IoT) sensing technology, and more specifically, relates to an identification method, system, and storage medium for the management and control of laboratory safety hazards. Background Technology

[0002] Laboratories are the core carriers of scientific research, innovation, and teaching practice, encompassing diverse safety scenarios such as the storage and use of hazardous chemicals, the operation of precision instruments, and the control of complex environments. Their safety management level directly affects the personal safety of researchers and the continuous and stable conduct of scientific research activities. Currently, laboratory safety management systems still rely primarily on traditional manual methods, depending on on-site inspections and regular checks by management personnel, and on personal experience to identify potential hazards and assess risks. This approach suffers from numerous shortcomings, including delayed response times, limited coverage, and significant subjective bias.

[0003] The visual acquisition devices, environmental sensors, and intelligent cabinets for hazardous chemicals deployed in the laboratory operate independently, and their data acquisition formats and transmission mechanisms are incompatible. Multi-source information such as real-time video, environmental parameters, and the storage and retrieval status of hazardous chemicals cannot be uniformly aggregated and processed collaboratively, resulting in a large number of data silos that are difficult to support the perception of the overall safety status.

[0004] Laboratory personnel's operational behaviors are complex and varied, and potential hazards such as violations and dangerous actions cannot be identified in real time. Traditional video surveillance can only perform post-event tracing, failing to provide pre-event warnings or timely intervention. The coupling risk between abnormal environmental parameters and personnel behavior lacks scientific quantification methods, and risk assessment results from single data sources are one-sided and lack credibility, making it difficult to form a basis for safety management decisions.

[0005] The existing control system lacks a three-dimensional visualization platform for global control, and key aspects such as hazardous material diffusion simulation and emergency evacuation route planning lack technical support, resulting in low emergency response efficiency. As scientific research scenarios become increasingly complex, the concealment and suddenness of laboratory safety hazards have significantly increased. Traditional control models are no longer suitable for dynamic safety control needs, and missed hazard identification and misjudgment of risks can easily lead to safety accidents.

[0006] The industry urgently needs to build a brand-new intelligent management and control technology system to break through the technical limitations of traditional models, realize the full-domain perception, intelligent identification, dynamic assessment and closed-loop handling of laboratory safety hazards, comprehensively improve the intelligence and automation level of laboratory safety management and control, and provide safety assurance for scientific research activities. Summary of the Invention

[0007] This invention aims to solve the problems of data silos, lagging behavior recognition, one-sided risk assessment, and inefficient emergency response in traditional laboratory safety management. By integrating three-dimensional digital twins, multi-source data perception, large language model intelligent agents, and DS evidence theory, it achieves full-domain perception, intelligent identification, dynamic assessment, and automated closed-loop management of safety hazards, thereby improving the intelligence level of laboratory safety management and emergency response capabilities.

[0008] In view of the above-mentioned deficiencies or improvement needs of the prior art, as a first aspect of the present invention, the present invention provides an identification method for laboratory safety hazard control, comprising: S1. Construct a three-dimensional digital twin space for the laboratory. Multi-source heterogeneous data is acquired in real time through visual acquisition devices, environmental sensors, and smart cabinet terminals deployed in the laboratory. The multi-source heterogeneous data includes real-time video streams of the laboratory, environmental parameter data, and hazardous chemical storage and retrieval status data. Edge computing nodes are used to perform spatiotemporal alignment and feature extraction on the multi-source heterogeneous data to generate a fused perception data stream containing location information and confidence evaluation, which is then mapped into the digital twin space. S2. An intelligent agent driven by a large language model performs semantic parsing and correlation analysis on the fused perception data stream to construct a multi-dimensional risk knowledge graph of personnel, environment, and behavior; the multi-dimensional risk knowledge graph is used to perform intent reasoning on personnel behavior to distinguish between normal experimental operations and violations. S3. Based on the results of the intent reasoning, calculate the risk probability value of each area of ​​the laboratory, generate a dynamic risk heat map and overlay it on the digital twin space, and generate targeted laboratory safety management and control strategies based on the distribution characteristics of the risk heat map. S4. Drive the intelligent terminal of the physical layer to execute the control strategy and record the execution feedback data; Specifically, the calculation of risk probability values ​​for each area of ​​the laboratory includes: acquiring a first basic probability allocation function of video behavior analysis results and a second basic probability allocation function of environmental parameter data; calculating the conflict coefficient between the first and second basic probability allocation functions to characterize the inconsistency between visual behavioral features and environmental physical features; using the Dempster combination rule to fuse the first and second basic probability allocation functions to obtain the trust level of each area under different states, and using the normalized value of the trust level as the risk probability value.

[0009] Furthermore, in S1, the fused sensing data stream is structurally defined as a time-series data stream containing a state mean vector and a covariance matrix; for each detected target or potential hazard point, each frame of data in this stream is no longer just a simple three-dimensional coordinate. It is not a single entity, but a complete state description package, whose mathematical expression is: ; in, It is the state mean vector after being corrected by extended Kalman filtering, representing the most likely position coordinates and velocity of the object in three-dimensional space; while This is the error covariance matrix after minimizing the trace optimization. It acts as a quality label for the data, and the value of this matrix directly quantifies the credibility of the current fusion result.

[0010] Furthermore, the generation process of the fused sensing data stream in S1 specifically includes: A spatiotemporal registration framework is constructed to adaptively synchronize the multi-source heterogeneous data in the time dimension by introducing a dynamic latency compensation factor. Correcting transmission latency jitter caused by load fluctuations in edge computing nodes; The multi-source heterogeneous data is subjected to spatial dimension probability mapping, targeting the two-dimensional pixel coordinates of the visual acquisition device. We introduce depth uncertainty estimation and map it to a Gaussian distribution in three-dimensional space. Instead of a single coordinate point, it represents the depth blur caused by occlusion or insufficient lighting; The multi-source heterogeneous data is fused and updated using an extended Kalman filter, incorporating the physical measurements from environmental sensors. As the observation input, combined with the Gaussian distribution By minimizing the error covariance matrix The fusion weights of visual and sensor data are dynamically adjusted based on the traces to generate a fused perception data stream that includes location information and confidence evaluation. The probability mapping of the spatial dimension is described by the following nonlinear observation equation: in, The observation vector contains pixel coordinates. ; For observation functions; It is a three-dimensional spatial state vector; For observation noise that follows a Gaussian distribution, the covariance matrix of the observation noise is... It dynamically adjusts according to the ambient light intensity.

[0011] Furthermore, the covariance matrix of the observed noise Dynamically adjusts according to ambient light intensity, specifically including: Will Designed as a reference noise matrix With light sensitivity function The product of, i.e.: in, and In pixel coordinate system direction and The reference measurement variance of the direction is used to characterize the inherent noise level of the sensor under standard conditions; Based on real-time collected ambient light intensity The calculated adaptive weighting coefficients are used to increase the covariance matrix when insufficient lighting or overexposure causes a decrease in the image signal-to-noise ratio. The numerical value is used to reduce the weight of visual observation data in the fusion process, thereby achieving interference-resistant adaptive fusion.

[0012] Furthermore, the construction process of the multidimensional risk knowledge graph in S2 specifically includes: The state mean vector in the fused sensing data stream With error covariance matrix The mapping is a structured semantic description. The intelligent agent driven by the large language model extracts personnel entities, environmental entities and behavioral entities from the pre-defined ontology layer, and identifies the spatial proximity relationship and temporal causal relationship between entities to generate an initial set of triples. An environmental context enhancement mechanism is introduced, in which the physical measurement values ​​of environmental sensors in the fused perception data stream are dynamically attached to the behavioral entities as attribute nodes, and semantic association edges between behavioral nodes and environmental state nodes are constructed to form a local subgraph containing "behavior-environment" coupling features. Based on the graph database, the initial set of triples and local subgraphs are stored and topologically connected. The vectorization capability of the large language model is used to embed entity nodes and calculate the semantic similarity between new input data and historical behavior patterns in the graph in real time. The node attributes and edge weights in the graph are dynamically updated to generate a dynamic risk knowledge graph that can represent the multidimensional coupling relationship between people, environment and behavior.

[0013] Furthermore, in step S2, the intention reasoning of personnel behavior using the multidimensional risk knowledge graph specifically includes: The coordinate sequence of key points of human skeleton is extracted from the real-time video stream using a pose estimation algorithm; A spatiotemporal graph is constructed based on the coordinate sequence, where nodes represent key points of the skeleton and edges represent skeleton connections and relationships between time frames. The spatiotemporal graph is used to extract features from the spatiotemporal graph. Spatial graph convolution aggregates the skeletal structure features of adjacent nodes, and temporal graph convolution captures the dynamic features of actions that change over time. The extracted spatiotemporal feature vector is input into the classifier to calculate the probability value of the current behavior belonging to the preset violation operation category. When the probability value is greater than the set threshold, the intention reasoning result is determined to be a violation.

[0014] Furthermore, in S3, the first basic probability allocation function and the second basic probability allocation function are fused and calculated using the Dempster combination rule, specifically as follows: in, This represents the basic probability allocation value after fusion, indicating that the laboratory area is in a risky state. The final probability; Representation of recognition framework The non-empty subset in the ensemble represents the risk status determination result after fusion; and Let each represent the first basic probability assignment function. Second basic probability assignment function The elements in the focal element are subsets of risk states under their respective identification frameworks; The conflict coefficient, representing the degree of evidentiary conflict between video behavior analysis results and environmental parameter data, is calculated using the following formula: ; This represents the empty set.

[0015] Furthermore, the S3 process generates targeted laboratory safety management strategies, specifically including: When the risk probability value of a specific area in the risk heatmap exceeds a preset leakage alarm threshold, the fluid diffusion simulation model is activated in the three-dimensional digital twin space. The temperature, humidity, and airflow field distribution data from the current environmental parameter data are used as the boundary conditions for the fluid diffusion simulation model. Based on the aforementioned boundary conditions, the convection-diffusion equation is solved using the finite element analysis method to predict the concentration distribution cloud map of hazardous gases within a predetermined time period in the future. Based on the area where the concentration exceeds the safety limit in the concentration distribution cloud map, the system automatically plans evacuation routes to avoid the area and calculates the control parameters of the exhaust equipment used to block gas diffusion.

[0016] As a second aspect of the present invention, an identification system for laboratory safety hazard management is also provided, for implementing the aforementioned identification method for laboratory safety hazard management, comprising: A multi-source data fusion sensing unit is used to construct a three-dimensional digital twin space for the laboratory. It acquires multi-source heterogeneous data in real time through visual acquisition devices, environmental sensors, and smart cabinet terminals deployed in the laboratory. The multi-source heterogeneous data includes real-time video streams of the laboratory, environmental parameter data, and hazardous chemical storage and retrieval status data. Edge computing nodes are used to perform spatiotemporal alignment and feature extraction on the multi-source heterogeneous data to generate a fusion sensing data stream containing location information and confidence evaluation, which is then mapped into the digital twin space. The behavioral intent reasoning unit is used to perform semantic parsing and correlation analysis on the fused perception data stream based on a large language model-driven intelligent agent, and construct a multi-dimensional risk knowledge graph of personnel, environment, and behavior; and use the multi-dimensional risk knowledge graph to perform intent reasoning on personnel behavior, and distinguish between normal experimental operations and violations. The strategy formulation unit is used to calculate the risk probability value of each area of ​​the laboratory based on the result of the intent reasoning, generate a dynamic risk heat map and overlay it on the digital twin space, and generate a targeted laboratory safety management strategy based on the distribution characteristics of the risk heat map. The control and execution unit is used to drive the intelligent terminal of the physical layer to execute the control and execution strategy and record the execution feedback data.

[0017] As a third aspect of the invention, a computer-readable storage medium is also provided, having a computer program stored thereon, the computer program being executed by a processor as described in any one of the inventions, an identification method for laboratory safety hazard control.

[0018] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: 1. The identification method for laboratory safety hazard management of the present invention constructs a three-dimensional digital twin space of the laboratory, utilizes visual acquisition devices, environmental sensors, and intelligent cabinet terminals to acquire real-time video streams, environmental parameter data, and hazardous chemical storage and retrieval status data, leverages edge computing nodes to perform spatiotemporal alignment and feature extraction on multi-source heterogeneous data, introduces a dynamic delay compensation factor to correct transmission delay jitter, maps two-dimensional pixel coordinates to a three-dimensional Gaussian distribution to represent deep ambiguity, and completes data fusion and updating through extended Kalman filtering to generate a fused sensing data stream containing location information and confidence evaluation, which is then mapped to the digital twin space. This processing method can solve the problems of asynchronous multi-source data transmission, insufficient positioning accuracy, and environmental interference, achieve the fusion of heterogeneous data, output standardized data with confidence evaluation, provide data support for subsequent behavior analysis and risk assessment, and ensure the reliability of the sensing process.

[0019] 2. The identification method for laboratory safety hazard management of the present invention uses a large language model-driven intelligent agent to perform semantic parsing and correlation analysis on the fused perception data stream, extracts personnel, environment, and behavioral entities, and constructs spatial and temporal associations between entities. It then builds a multi-dimensional risk knowledge graph of personnel, environment, and behavior by combining an environmental context enhancement mechanism. A pose estimation algorithm extracts the coordinate sequence of key points of the human skeleton, and a spatiotemporal graph convolutional network is used to extract spatiotemporal features of actions and complete behavior classification, enabling the inference of personnel's intentions and distinguishing between normal experimental operations and violations. This technique can transform fused data into structured risk association information, establish a coupled relationship between behavior and environment, improve the fine-grainedness and accuracy of behavior recognition, and solve the problems of traditional behavior recognition lacking contextual support and having a single judgment result, providing behavioral basis for risk quantification calculation.

[0020] 3. The identification method for laboratory safety hazard management of the present invention obtains the basic probability allocation function corresponding to video behavior analysis and environmental parameter data based on behavioral intent reasoning results. It calculates the conflict coefficient of the two types of functions to characterize the inconsistency of data features, uses Dempster's combination rule to complete data fusion calculation, normalizes the trust level as the regional risk probability value, generates a dynamic risk heatmap and overlays it onto the digital twin space, formulates management strategies based on risk distribution, drives the physical layer intelligent terminal to execute the strategies and records execution feedback data. This process can resolve feature conflicts between multi-source data, quantify the risks in various areas of the laboratory, generate suitable management solutions based on risk distribution, complete the closed-loop process from risk identification to management execution, improve the targeting and execution efficiency of laboratory safety hazard management, and ensure the implementation of management measures. Attached Figure Description

[0021] Figure 1 This is a flowchart of an identification method for laboratory safety hazard management according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the system units in an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0023] Example 1 Please refer to Figure 1 This embodiment 1 provides an identification method for laboratory safety hazard management, including: S1. Construct a three-dimensional digital twin space for the laboratory. Multi-source heterogeneous data is acquired in real time through visual acquisition devices, environmental sensors, and smart cabinet terminals deployed in the laboratory. The multi-source heterogeneous data includes real-time video streams of the laboratory, environmental parameter data, and hazardous chemical storage and retrieval status data. Edge computing nodes are used to perform spatiotemporal alignment and feature extraction on the multi-source heterogeneous data to generate a fused perception data stream containing location information and confidence evaluation, which is then mapped into the digital twin space. S2. An intelligent agent driven by a large language model performs semantic parsing and correlation analysis on the fused perception data stream to construct a multi-dimensional risk knowledge graph of personnel, environment, and behavior; the multi-dimensional risk knowledge graph is used to perform intent reasoning on personnel behavior to distinguish between normal experimental operations and violations. S3. Based on the results of the intent reasoning, calculate the risk probability value of each area of ​​the laboratory, generate a dynamic risk heat map and overlay it on the digital twin space, and generate targeted laboratory safety management and control strategies based on the distribution characteristics of the risk heat map. S4. Drive the intelligent terminal of the physical layer to execute the control strategy and record the execution feedback data; Specifically, the calculation of risk probability values ​​for each area of ​​the laboratory includes: acquiring a first basic probability allocation function of video behavior analysis results and a second basic probability allocation function of environmental parameter data; calculating the conflict coefficient between the first and second basic probability allocation functions to characterize the inconsistency between visual behavioral features and environmental physical features; using the Dempster combination rule to fuse the first and second basic probability allocation functions to obtain the trust level of each area under different states, and using the normalized value of the trust level as the risk probability value.

[0024] This embodiment 1 further elaborates on the above steps.

[0025] (1) Multi-source data fusion sensing Traditional laboratory safety sensing suffers from problems such as independent data transmission from multiple devices, difficulty in time and space synchronization, susceptibility to obstruction and lighting interference in positioning, and inability to quantify data reliability, failing to provide a stable and reliable foundation for identifying safety hazards. To address these practical issues, a three-dimensional digital twin space for the laboratory is first constructed. Visual acquisition devices, environmental sensors, and intelligent cabinet terminals are then strategically deployed within the laboratory to continuously collect real-time video streams, environmental parameter data, and data on the storage and retrieval status of hazardous chemicals.

[0026] Subsequently, to achieve the generation of fused sensing data streams under complex laboratory electromagnetic environments and network fluctuations, the edge computing nodes first perform a time synchronization step based on a dynamic latency compensation factor. Since visual acquisition devices and various environmental sensors in the laboratory often experience non-deterministic network transmission jitter, this embodiment does not use fixed timestamp alignment, but instead introduces a dynamic latency compensation factor. This allows for real-time correction of transmission latency. In practice, edge computing nodes maintain a sliding time window to monitor data packet arrival jitter in real time and dynamically adjust based on the current network congestion status. The value of . When network load fluctuations are detected, which cause an increase in packet delay, the current transmission lag is predicted using historical delay sequences and superimposed on the original timestamp as a compensation value. This aligns the logical time of different source data at the software level, ensuring that subsequent spatial registration is based on the state at the same physical moment.

[0027] After completing time alignment, the next step is the spatial dimension probabilistic mapping stage. This step aims to solve the problem of inaccurate positioning caused by missing depth information or occlusion in traditional coordinate transformations. This embodiment abandons the traditional single-pixel coordinate mapping method and instead maps the two-dimensional pixel coordinates of the visual acquisition device... This is considered an observation with uncertainty. First, based on the current ambient light intensity and image sharpness, a depth uncertainty estimate is calculated. Then, the two-dimensional pixels are back-projected into three-dimensional space to generate a Gaussian distribution in three-dimensional space. This Gaussian distribution not only contains the most probable three-dimensional spatial coordinates of the object. Furthermore, the reliability of the location is quantified through the covariance matrix. For example, when a person is in an area with edge occlusion, the variance of the Gaussian distribution automatically increases, thus mathematically representing a state of "depth blur" rather than incorrectly locking a fixed coordinate.

[0028] Next, to generate the final fused sensing data stream, this embodiment utilizes the extended Kalman filter algorithm to perform deep fusion of multi-source heterogeneous data. The output fused sensing data stream is defined in terms of data structure as a time series data stream containing a state mean vector and a covariance matrix. For each detected target or potential hazard point, each frame of data in this stream is no longer just a simple three-dimensional coordinate. It is not a single entity, but a complete state description package, whose mathematical expression is: ; in, Represents a point in time or a time index. It is the state mean vector after being corrected by extended Kalman filtering, representing the most likely position coordinates and velocity of the object in three-dimensional space; while This is the error covariance matrix after trace minimization optimization. It acts as a quality label for the data, and the magnitude of this matrix directly quantifies the reliability of the current fusion result—if If the diagonal elements have small values, it indicates that the fused positional information is highly accurate, allowing for direct triggering of precise control commands (such as automatic fire suppression); conversely, if... A large value indicates severe environmental interference (such as dense smoke obscuring visual weights) and data ambiguity. In such cases, the response priority for this data should be reduced or the data should be switched to manual review mode.

[0029] During this process, the physical measurements of environmental sensors (such as infrared and smoke sensors) This is used as the observation input and interacts with the probability distribution of the aforementioned visual generation. The posterior estimation error covariance matrix is ​​minimized. The filter dynamically calculates the optimal weights for fusing visual and environmental sensor data. If the visual data has high uncertainty (i.e., a large Gaussian distribution variance), the filter automatically reduces the weight of the visual data, relying more on the physical readings of the environmental sensors, and vice versa. This fusion process strictly follows the following nonlinear observation equation for state updates: in, This is the observation vector at the current moment, containing pixel coordinates. ; A nonlinear observation function that maps three-dimensional world coordinates to a two-dimensional image plane; It is a three-dimensional spatial state vector; The observed noise follows a Gaussian distribution with a mean of zero; and This refers to the principal point coordinates in the camera's intrinsic parameter matrix, also known as the optical center offset. Specifically, to adapt to the dynamically changing environment of the laboratory, the covariance matrix of the observation noise... It is not a fixed constant, but rather a function designed to dynamically adjust with the ambient light intensity. The specific process of dynamic adjustment is as follows: To achieve dynamic response to ambient light intensity, the average gray value or luminance component within the current field of view is first calculated using real-time image frames acquired by the vision acquisition device. This is used as a quantitative indicator of ambient light intensity. Subsequently, a light sensitivity function is defined. The function is designed to characterize the weight of the impact of current lighting conditions on observation accuracy. Its specific expression is as follows: in, This is a preset noise gain coefficient used to control the degree of noise amplification under low light conditions; The optimal center value for imaging brightness calibrated for the equipment; This is the brightness tolerance parameter. The physical meaning of this function is: when the ambient light... Deviation from optimal imaging brightness When the environment is too dark or too bright, the exponential term approaches 1, making... An increase in ambient light indicates a decrease in the reliability of the observation; when ambient light... near When the exponential term approaches 0, A value close to 1 indicates that the observation is in its optimal state.

[0030] After calculating the light sensitivity function Then, it was applied to the observation noise covariance matrix. In real-time construction. Specifically, will Designed as a reference noise matrix With light sensitivity function The product of, i.e.: in, and In pixel coordinate system direction and The variance of the reference measurement in the direction. Using the above formula, when insufficient ambient light leads to... When decreasing, The increase in numerical value directly leads to an increase in the covariance matrix. The value increases. In the update step of the extended Kalman filter, the increase... This will lead to Kalman gain. The filter automatically reduces the weight of the current frame's visual observations when fusing data, relying more on predicted values ​​or data from other sensors (such as infrared or lidar).

[0031] (2) Reasoning about behavioral intentions After completing the multi-source data fusion perception, it is necessary to rely on the intelligent agent driven by the large language model to carry out comprehensive semantic parsing and correlation analysis on the fused perception data stream, construct a multi-dimensional risk knowledge graph, and complete the intention reasoning of personnel behavior based on the graph to determine the nature of the behavior.

[0032] First, during the construction phase of the multidimensional risk knowledge graph, the processing unit performs deep semantic parsing on the fused sensing data stream. Specifically, the edge computing nodes convert the state mean vector output by Kalman filtering or other state estimation algorithms... With error covariance matrix Transform it into a structured semantic description in natural language form. For example, describe vector data as "in coordinates..." There exists a confidence level of The large language model-driven agent automatically extracts three core entities—"personnel," "environmental facilities," and "experimental behavior"—from these descriptions based on the preset ontology layer definition. It also identifies spatial proximity relationships (such as "located next to...") and temporal causal relationships (such as "after operating...") between entities, thereby generating an initial set of triples.

[0033] To enhance the graph's ability to represent risks, this embodiment introduces an environmental context enhancement mechanism. Physical measurements from environmental sensors in the fused data stream (such as temperature and smoke concentration) are used as attribute nodes and attached to the corresponding behavioral entities. Semantic association edges between behavioral nodes and environmental state nodes are constructed to form a local subgraph containing strong coupling features between behavior and environment.

[0034] Subsequently, the processing unit enters the dynamic updating and evolution stage of the graph to generate a dynamic risk knowledge graph that can reflect the laboratory's safety status in real time. The specific process is as follows: Based on the graph database, the initial set of triples and local subgraphs are stored and topologically connected, and the following incremental update logic is executed: First, the vectorization capability of the large language model is used to embed entity nodes, generating node feature vectors at the current moment. The processing unit calculates the semantic similarity between the feature vectors of new input data and historical behavior pattern nodes in the graph in real time. If the similarity is higher than a preset threshold, it is determined to be a continuation of the same entity or behavior, triggering the attribute fusion mechanism to update the attributes of the corresponding node in the graph (such as updating personnel location, refreshing ambient temperature) and increasing the weight of the associated edges. If the similarity is lower than the threshold, it is determined to be a new entity or new behavior, and a new node is created in the graph. At the same time, in order to maintain the real-time performance and accuracy of the graph, the processing unit introduces a time decay mechanism, which automatically reduces the weight of inactive nodes and edges over time, and prunes nodes with weights lower than the cleanup threshold. Through this incremental update based on semantic similarity and the aging process of old data based on time decay, a dynamic risk knowledge graph is finally generated.

[0035] Secondly, in the stage of reasoning about human behavioral intentions based on a multidimensional risk knowledge graph, the processing unit uses computer vision technology to perform fine-grained analysis of the real-time video stream. The specific process is as follows: First, a sequence of coordinates of key points on the human skeleton is extracted from the real-time video stream using a pose estimation algorithm. Based on this coordinate sequence, a spatiotemporal graph is constructed. , where the set of nodes Representing the key skeletal points of various parts of the human body, edge set This includes spatial edges representing the physical connections of the human skeleton and temporal edges representing the association of the same keypoint between adjacent time frames. Subsequently, a spatiotemporal graph convolutional network is used to extract features from this spatiotemporal graph: in the spatial dimension, the skeletal structure features of adjacent nodes are aggregated through spatial graph convolution to capture the human posture information; in the temporal dimension, the dynamic features of the action changing over time are captured through temporal graph convolution, thus forming a feature representation containing spatiotemporal context information.

[0036] Finally, the processing unit inputs the extracted spatiotemporal feature vector into a fully connected classifier to calculate the probability value that the current behavior belongs to a preset violation category (such as running, falling, or improper operation of instruments). This probability value quantifies the degree of matching between the current behavior and a known hazard pattern. When the calculated probability value is greater than a set judgment threshold, the agent determines that the intent reasoning result is a violation, and combines the environmental context information in the aforementioned constructed multidimensional risk knowledge graph (such as whether the current environment is high-temperature or flammable) to generate the final safety hazard control instruction. Through this dual mechanism combining graph semantic reasoning and skeletal motion recognition, the differentiation and real-time early warning of the behavioral intentions of laboratory personnel are achieved.

[0037] (3) Strategy Formulation Based on the results of behavioral intention reasoning, the risk probability values ​​for each area of ​​the laboratory are calculated. The processing unit first needs to clearly define the identification frame and focal elements, and then construct the first basic probability assignment function. Identification Frame It is defined as a set containing all mutually exclusive risk states, i.e. These represent "safe," "mild risk," and "severe risk," respectively. Within this framework, the focal element refers to the subset of the basic probability assignment function assigned a non-zero probability value. Based on the video behavior analysis results, the original confidence scores for each type of action are first obtained using a pose estimation algorithm. These scores are then normalized to obtain the confidence scores for each risk state. Subsequently, the focal element is defined as a single-element subset. ( (and the complete series) (Representing uncertainty). Confidence adjustment coefficient dynamically calibrated based on behavioral temporal stability and historical data. Construct the first basic probability assignment function The focal element allocation logic is as follows: the weighted confidence level is assigned to the single-element focal element, i.e. Assign the remaining probability mass to the focal element of the entire set, i.e. This step ensures and This completes the mathematical mapping from visual behavioral characteristics to risk status.

[0038] Furthermore, the processing unit is based on the same recognition framework. A second basic probability assignment function is constructed to quantify the support of environmental parameter data for risk states. Parameter vectors such as temperature, humidity, and gas concentration are collected through a sensor network. And utilize predefined environmental risk membership functions Physical parameters are mapped to membership degrees for each risk state. Similar to behavioral analysis, the focal element is defined as a single-element subset. and the complete series Introducing environmental data reliability weights. Construct the second basic probability assignment function Its focal element allocation logic is as follows: , This step completes the mathematical mapping from environmental physical characteristics to risk status, and together with the first basic probability assignment function, it forms the basis of multi-source evidence.

[0039] Subsequently, to measure the inconsistency between visual behavioral characteristics and environmental physical characteristics, the processing unit calculates... and Conflict coefficient between According to Dempster's theory, the conflict coefficient... It characterizes the degree of mutual exclusion between focal elements from different evidence sources, and the calculation formula is as follows: ,in and They are respectively and The focal point. If If the value approaches 1, it indicates that the two sources of evidence are highly conflicting and a verification mechanism needs to be triggered; conversely, it indicates that the consensus of the evidence is high and they can be directly merged.

[0040] Finally, the Dempster combination rule was used to fuse the two sources of evidence, through the formula... ( Calculate the comprehensive posterior probability allocation In the fusion results, the trust level of each risk state After normalization, this becomes the final risk probability value. This process, through evidence theory, achieves uncertainty modeling and conflict resolution of multi-source heterogeneous data, providing quantifiable and implementable decision-making basis for laboratory risk early warning.

[0041] Based on the calculated risk probability values, a dynamic risk heatmap is generated in a three-dimensional digital twin space. First, the laboratory space is divided into several grid cells, and the risk probability value of each cell is mapped to a color gradient (e.g., green represents safe, yellow represents mild risk, and red represents severe risk), which is then overlaid on the corresponding location in the digital twin model. When the risk probability value of a specific area in the heatmap exceeds a preset leak alarm threshold, the fluid diffusion simulation model is activated in the digital twin space. The processing unit acquires temperature, humidity, and airflow field distribution data from the current environmental parameters as boundary conditions. Based on these conditions, the convection-diffusion equations are solved using the finite element analysis method. ,in For gas concentration, The diffusion coefficient is... For airflow velocity field, The source term is used. Through numerical solutions, the concentration distribution cloud map of hazardous gases is predicted over a predetermined future time period.

[0042] Based on the predicted concentration distribution cloud map, a targeted laboratory safety management strategy is generated. The specific process is as follows: First, areas where the concentration exceeds safety limits in the cloud map are identified and marked as "restricted areas" in the digital twin model. Based on this, a path planning algorithm automatically plans evacuation routes to avoid these areas, and pushes the optimal escape route to personnel via audible and visual alarms and mobile terminals within the laboratory. Simultaneously, control parameters for ventilation equipment used to block gas diffusion are calculated, such as adjusting the fan speed of fume hoods and activating negative pressure ventilation systems in specific areas to accelerate the removal of harmful gases. This process achieves closed-loop management from risk identification and simulation prediction to proactive intervention, ensuring the elimination of laboratory safety hazards.

[0043] (4) Control and Implementation Upon receiving the control instruction set generated by the strategy formulation unit, the control execution unit first sends the converted control instructions to the corresponding intelligent terminal devices at the physical layer via the laboratory's industrial Ethernet or wireless communication network. For security equipment, it drives the generation of audible and visual alarm signals, plays pre-recorded voice evacuation prompts through speakers deployed in the laboratory, and activates flashing alarm lights to attract personnel's attention. For access control systems, it automatically unlocks the access locks of escape routes and locks entrances to high-risk areas to prevent unauthorized personnel from entering. For environmental control equipment, based on calculated exhaust equipment control parameters, it adjusts the speed of the variable frequency fan to operate in emergency exhaust mode, or controls the opening of electric dampers to alter airflow organization and block the diffusion of harmful gases to safe areas.

[0044] Secondly, during execution, the control and execution unit monitors the response status of the physical terminals in real time. Upon receiving an instruction, the intelligent terminal immediately executes the corresponding action (such as relay activation, motor start-up, or valve rotation) and collects feedback data during the execution process. This feedback data includes the actual operating status of the equipment (such as the actual speed of the fan, the actual opening degree of the valve, and the operating current of the alarm), execution result confirmation signals (such as "action successful" or "equipment failure"), and execution timestamps. The feedback data is transmitted back to the control and execution unit in real time to confirm whether the instruction has been executed correctly.

[0045] Subsequently, the control and execution unit analyzes and records the received feedback data. The feedback data is compared with the original command to verify the accuracy of the execution result. For example, if the command requires the fan speed to increase to 1500 rpm, the control and execution unit will check whether the actual fan speed in the feedback data is within the allowable error range to reach that value. If an excessive deviation is detected or a fault signal is detected from the equipment (such as motor overload or communication timeout), an abnormal handling mechanism will be immediately triggered to attempt to resend the command or switch to backup equipment, and an equipment maintenance alarm will be generated to notify maintenance personnel for repair.

[0046] Finally, all instruction issuance records, device response data, execution result verification information, and anomaly handling logs during the execution process are uniformly recorded and stored in the execution feedback database. These records not only provide a complete chain of evidence for subsequent security incident tracing but also serve as important training data for optimizing the model parameters of the strategy formulation unit.

[0047] Example 2 Please refer to Figure 2 This embodiment 2 provides an identification system for laboratory safety hazard management, used to implement the aforementioned identification method for laboratory safety hazard management, including: A multi-source data fusion sensing unit is used to construct a three-dimensional digital twin space for the laboratory. It acquires multi-source heterogeneous data in real time through visual acquisition devices, environmental sensors, and smart cabinet terminals deployed in the laboratory. The multi-source heterogeneous data includes real-time video streams of the laboratory, environmental parameter data, and hazardous chemical storage and retrieval status data. Edge computing nodes are used to perform spatiotemporal alignment and feature extraction on the multi-source heterogeneous data to generate a fusion sensing data stream containing location information and confidence evaluation, which is then mapped into the digital twin space. The behavioral intent reasoning unit is used to perform semantic parsing and correlation analysis on the fused perception data stream based on a large language model-driven intelligent agent, and construct a multi-dimensional risk knowledge graph of personnel, environment, and behavior; and use the multi-dimensional risk knowledge graph to perform intent reasoning on personnel behavior, and distinguish between normal experimental operations and violations. The strategy formulation unit is used to calculate the risk probability value of each area of ​​the laboratory based on the result of the intent reasoning, generate a dynamic risk heat map and overlay it on the digital twin space, and generate a targeted laboratory safety management strategy based on the distribution characteristics of the risk heat map. The control and execution unit is used to drive the intelligent terminal of the physical layer to execute the control and execution strategy and record the execution feedback data.

[0048] Example 3 This embodiment 3 also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, can implement any step of an identification method for laboratory safety hazard control.

[0049] The computer-readable storage medium may include various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0050] For a description of the computer-readable storage medium provided in this application, please refer to the above method embodiments; further details will not be repeated here.

[0051] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying potential safety hazards in laboratories, characterized in that, include: S1. Construct a three-dimensional digital twin space for the laboratory. Multi-source heterogeneous data is acquired in real time through visual acquisition devices, environmental sensors, and smart cabinet terminals deployed in the laboratory. The multi-source heterogeneous data includes real-time video streams of the laboratory, environmental parameter data, and hazardous chemical storage and retrieval status data. Edge computing nodes are used to perform spatiotemporal alignment and feature extraction on the multi-source heterogeneous data to generate a fused perception data stream containing location information and confidence evaluation, which is then mapped into the digital twin space. S2. An intelligent agent driven by a large language model performs semantic parsing and correlation analysis on the fused perception data stream to construct a multi-dimensional risk knowledge graph of personnel, environment, and behavior; The multidimensional risk knowledge graph is used to infer the intent of personnel behavior and distinguish between normal experimental operations and violations. S3. Based on the results of the intent reasoning, calculate the risk probability value of each area of ​​the laboratory, generate a dynamic risk heat map and overlay it on the digital twin space, and generate targeted laboratory safety management and control strategies based on the distribution characteristics of the risk heat map. S4. Drive the intelligent terminal of the physical layer to execute the control strategy and record the execution feedback data; Specifically, the calculation of risk probability values ​​for each area of ​​the laboratory includes: acquiring a first basic probability allocation function of video behavior analysis results and a second basic probability allocation function of environmental parameter data; calculating the conflict coefficient between the first and second basic probability allocation functions to characterize the inconsistency between visual behavioral features and environmental physical features; using the Dempster combination rule to fuse the first and second basic probability allocation functions to obtain the trust level of each area under different states, and using the normalized value of the trust level as the risk probability value.

2. The identification method for laboratory safety hazard control according to claim 1, characterized in that, The fusion perception data stream in the S1 is defined as a time series data stream containing state mean vector and covariance matrix in data structure; for each detected target or hidden point, each frame of data in the data stream is no longer a simple three-dimensional coordinate but a complete state description package, the mathematical expression form of which is ​ in, Represents a point in time or a time index. It is the state mean vector after being corrected by extended Kalman filtering, representing the most likely position coordinates and velocity of the object in three-dimensional space; while This is the error covariance matrix after minimizing the trace optimization. It acts as a quality label for the data, and the value of this matrix directly quantifies the credibility of the current fusion result.

3. The identification method for laboratory safety hazard control according to claim 2, characterized in that, The process of generating the fused sensing data stream in S1 specifically includes: A spatiotemporal registration framework is constructed to adaptively synchronize the multi-source heterogeneous data in the time dimension by introducing a dynamic latency compensation factor. Correcting transmission latency jitter caused by load fluctuations in edge computing nodes; The multi-source heterogeneous data is subjected to spatial dimension probability mapping, targeting the two-dimensional pixel coordinates of the visual acquisition device. We introduce depth uncertainty estimation and map it to a Gaussian distribution in three-dimensional space. Instead of a single coordinate point, it represents the depth blur caused by occlusion or insufficient lighting; The multi-source heterogeneous data is fused and updated using an extended Kalman filter, incorporating the physical measurements from environmental sensors. As the observation input, combined with the Gaussian distribution By minimizing the error covariance matrix The fusion weights of visual and sensor data are dynamically adjusted based on the traces to generate a fused perception data stream that includes location information and confidence evaluation. The probability mapping of the spatial dimension is described by the following nonlinear observation equation: in, The observation vector contains pixel coordinates. ; For observation functions; It is a three-dimensional spatial state vector; For observation noise that follows a Gaussian distribution, the covariance matrix of the observation noise is... Dynamically adjusts according to ambient light intensity; and These are the coordinates of the principal point in the camera intrinsic parameter matrix.

4. The identification method for laboratory safety hazard control according to claim 3, characterized in that, The covariance matrix of the observation noise Dynamically adjusts according to ambient light intensity, specifically including: Will Designed as a reference noise matrix With light sensitivity function The product of, i.e.: in, and In pixel coordinate system direction and The reference measurement variance of the direction is used to characterize the inherent noise level of the sensor under standard conditions; Based on real-time collected ambient light intensity The calculated adaptive weighting coefficients are used to increase the covariance matrix when insufficient lighting or overexposure causes a decrease in the image signal-to-noise ratio. The numerical value is used to reduce the weight of visual observation data in the fusion process, thereby achieving interference-resistant adaptive fusion.

5. The identification method for laboratory safety hazard control according to claim 1, characterized in that, The construction process of the multidimensional risk knowledge graph in S2 specifically includes: The state mean vector in the fused sensing data stream With error covariance matrix The mapping is a structured semantic description. The intelligent agent driven by the large language model extracts personnel entities, environmental entities and behavioral entities from the pre-defined ontology layer, and identifies the spatial proximity relationship and temporal causal relationship between entities to generate an initial set of triples. An environmental context enhancement mechanism is introduced, in which the physical measurement values ​​of environmental sensors in the fused perception data stream are attached to the behavioral entities as attribute nodes, and semantic association edges between behavioral nodes and environmental state nodes are constructed to form a local subgraph containing behavioral-environment coupling features. Based on the graph database, the initial set of triples and local subgraphs are stored and topologically connected. The vectorization capability of the large language model is used to embed entity nodes and calculate the semantic similarity between new input data and historical behavior patterns in the graph in real time. The node attributes and edge weights in the graph are dynamically updated to generate a dynamic risk knowledge graph that can represent the multidimensional coupling relationship between people, environment and behavior.

6. The identification method for laboratory safety hazard control according to claim 1, characterized in that, The S2 step, which utilizes the multidimensional risk knowledge graph to infer the intent of personnel behavior, specifically includes: The coordinate sequence of key points of human skeleton is extracted from the real-time video stream using a pose estimation algorithm; A spatiotemporal graph is constructed based on the coordinate sequence, where nodes represent key points of the skeleton and edges represent skeleton connections and relationships between time frames. The spatiotemporal graph is used to extract features from the spatiotemporal graph. Spatial graph convolution aggregates the skeletal structure features of adjacent nodes, and temporal graph convolution captures the dynamic features of actions that change over time. The extracted spatiotemporal feature vector is input into the classifier to calculate the probability value of the current behavior belonging to the preset violation operation category. When the probability value is greater than the set threshold, the intention reasoning result is determined to be a violation.

7. The identification method for laboratory safety hazard control according to claim 1, characterized in that, In step S3, the Dempster combination rule is used to fuse the first basic probability allocation function and the second basic probability allocation function for calculation. Specifically: in, This represents the basic probability allocation value after fusion, indicating that the laboratory area is in a risky state. The final probability; Representation of recognition framework The non-empty subset in the ensemble represents the risk status determination result after fusion; and Let each represent the first basic probability assignment function. Second basic probability assignment function The elements in the focal element are subsets of risk states under their respective identification frameworks; The conflict coefficient, representing the degree of evidentiary conflict between video behavior analysis results and environmental parameter data, is calculated using the following formula: ; This represents the empty set.

8. The identification method for laboratory safety hazard control according to claim 1, characterized in that, The S3 process generates targeted laboratory safety management strategies, specifically including: When the risk probability value of a specific area in the risk heatmap exceeds a preset leakage alarm threshold, the fluid diffusion simulation model is activated in the three-dimensional digital twin space. The temperature, humidity, and airflow field distribution data from the current environmental parameter data are used as the boundary conditions for the fluid diffusion simulation model. Based on the aforementioned boundary conditions, the convection-diffusion equation is solved using the finite element analysis method to predict the concentration distribution cloud map of hazardous gases within a predetermined time period in the future. Based on the area where the concentration exceeds the safety limit in the concentration distribution cloud map, the system automatically plans evacuation routes to avoid the area and calculates the control parameters of the exhaust equipment used to block gas diffusion.

9. An identification system for laboratory safety hazard control, used to implement the identification method for laboratory safety hazard control as described in claim 1, characterized in that, include: A multi-source data fusion sensing unit is used to construct a three-dimensional digital twin space for the laboratory. It acquires multi-source heterogeneous data in real time through visual acquisition devices, environmental sensors, and smart cabinet terminals deployed in the laboratory. The multi-source heterogeneous data includes real-time video streams of the laboratory, environmental parameter data, and hazardous chemical storage and retrieval status data. Edge computing nodes are used to perform spatiotemporal alignment and feature extraction on the multi-source heterogeneous data to generate a fusion sensing data stream containing location information and confidence evaluation, which is then mapped into the digital twin space. The behavioral intent reasoning unit is used to perform semantic parsing and correlation analysis on the fused perception data stream based on a large language model-driven intelligent agent, and to construct a multi-dimensional risk knowledge graph of personnel, environment and behavior. The multidimensional risk knowledge graph is used to infer the intent of personnel behavior and distinguish between normal experimental operations and violations. The strategy formulation unit is used to calculate the risk probability value of each area of ​​the laboratory based on the result of the intent reasoning, generate a dynamic risk heat map and overlay it on the digital twin space, and generate a targeted laboratory safety management strategy based on the distribution characteristics of the risk heat map. The control and execution unit is used to drive the intelligent terminal of the physical layer to execute the control and execution strategy and record the execution feedback data.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program is executed by a processor as described in any one of claims 1-8: an identification method for laboratory safety hazard control.

Images