A laboratory safety inspection method and system based on the Internet of Things and a storage medium

Abstract

The application discloses a kind of laboratory safety inspection method and system based on Internet of Things, method includes: obtaining laboratory digital twin model and dividing logic control unit, real-time monitoring each unit environment state and equipment operation data, calculate dynamic risk weight, generate virtual inspection path accordingly and mapping as physical inspection task is issued to terminal.By corresponding unit, mobile sensing and fixed Internet of Things static environment data are synchronously collected after inspection personnel arrive, and multi-dimensional feature vector of hazardous chemicals and environmental safety is extracted by space-time alignment.The feature vector is input into the hidden danger identification model to output the hidden danger confidence, and the final hidden danger level is determined by combining the laboratory safety knowledge graph verification.The general hidden danger is pushed to the rectification guide, the major hidden danger is linked to the equipment control and the twin model is updated, the rectification review is up to standard, the instruction is removed, the state is synchronized, the physical and virtual space safety inspection closed loop is completed, and reliable technical support is provided for laboratory safety control.

Description

Technical Field

[0001] This invention belongs to the field of Internet of Things (IoT) technology, and more specifically, relates to a laboratory safety inspection method, system, and storage medium based on IoT. Background Technology

[0002] Laboratories are core locations for scientific research and teaching experiments, housing a large number of hazardous chemicals, precision instruments, and special equipment. They encompass multiple safety risk sources, including high temperatures, high pressures, and flammable and explosive materials, highlighting the increasing complexity and importance of safety management. Currently, laboratory safety inspections still rely on traditional manual methods, with pre-set and fixed routes that cannot be flexibly adjusted based on real-time risk conditions. This easily leads to situations where high-risk areas are overlooked and low-risk areas are over-inspected, resulting in a mismatch between the allocation of inspection resources and actual safety needs.

[0003] The existing inspection system relies on a limited data collection method, primarily manual on-site registration. Data collected by various fixed IoT monitoring devices is independent, lacking a unified data fusion and analysis mechanism. Spatiotemporal asynchrony exists between mobile sensing data and static environmental data, making it difficult to eliminate data transmission delays and collection errors, thus failing to fully reconstruct the true safety conditions of the laboratory. Hazard identification excessively depends on the subjective experience of inspection personnel, lacking intelligent analysis and judgment tools, leading to frequent false alarms and missed alarms, and difficulty in assessing risk levels.

[0004] The hazard handling process is significantly lagging, requiring manual reporting and multi-level review after a hazard is discovered, making it impossible to achieve rapid response based on tiered classification. In the face of major safety hazards, it is difficult to promptly coordinate with relevant control equipment for emergency response, compromising the timeliness of safety control. The physical site and virtual control space are disconnected, lacking a globally visualized control platform; inspection execution and hazard rectification cannot be fully documented, hindering the formation of a closed-loop safety management system.

[0005] At present, IoT technology has been initially applied in laboratory safety management, but it has not yet achieved deep integration with technologies such as digital twins, artificial intelligence and knowledge graphs. It cannot support intelligent inspection driven by dynamic risks, collaborative analysis of multi-source data and closed-loop management of the whole process, and is far from meeting the development requirements of modern laboratory safety management for high efficiency and intelligence. Summary of the Invention

[0006] This invention aims to address the pain points of traditional laboratory safety inspections, such as fixed routes, isolated data, inaccurate hazard identification, delayed response, and lack of closed-loop management. It integrates IoT, digital twin, and artificial intelligence technologies to construct a dynamic risk-driven intelligent inspection system, achieving multi-source data fusion, intelligent hazard identification, and hierarchical closed-loop handling, thereby comprehensively improving the timeliness and intelligence level of laboratory safety management.

[0007] To address the aforementioned deficiencies or improvement needs of existing technologies, as a first aspect of this invention, the present invention provides a laboratory safety inspection method based on the Internet of Things, comprising: S1. Obtain the digital twin model of the laboratory and divide the laboratory into several logical control units; monitor the environmental status parameters and equipment operation data of each logical control unit in real time, and calculate the dynamic risk weight of each unit; according to the dynamic risk weight, dynamically generate a virtual inspection path in the digital twin model and map it into a physical space inspection task sequence and send it to the mobile terminal. S2. In response to the arrival of the inspection personnel at the current logical control unit, simultaneously collect the mobile sensing data stream from the perspective of the inspection personnel and the static environmental data stream from the fixed IoT node; perform spatiotemporal alignment on the mobile sensing data stream and the static environmental data stream, and extract multidimensional feature vectors of hazardous chemical status characteristics and environmental safety characteristics; S3. Input the multidimensional feature vector into the pre-trained safety hazard identification model, perform cross-modal association analysis to output the confidence level of potential hazards; use the laboratory safety knowledge graph to perform semantic reasoning verification on the confidence level of potential hazards, eliminate interference from a single data source, and determine the final safety hazard level; S4. Generate a set of control instructions based on the final safety hazard level; if it is a general hazard, output rectification guidance information through the mobile terminal; if it is a major hazard, send a lock or disposal instruction to the associated control device through the IoT gateway, and update the safety status identifier of the digital twin model; receive the verification data after rectification, and when the preset safety threshold is met, release the disposal instruction and update the digital twin model synchronously to complete the closed loop between physical space and virtual space.

[0008] Furthermore, the calculation process of the dynamic risk weight of each unit in S1 specifically includes: Acquire real-time environmental status parameter sets and equipment operation datasets for each logical control unit; the environmental status parameters include temperature and humidity, harmful gas concentration and smoke index, and the equipment operation data includes high-power equipment current values ​​and fume hood operation status; The real-time environmental state parameter set and the equipment operation dataset are processed using a sliding window method to calculate the mean and standard deviation of each parameter within a preset time window. Based on the Z-score standardization method, the deviation of each parameter from historical normal operating conditions is calculated to obtain an anomaly score. ;in, Indicates the first The identifier index of each logical control unit; Simultaneously, the absolute values ​​of the first-order differences of each parameter within adjacent time windows are calculated to obtain the rate of change. Using weighted fusion formula Calculate the first Dynamic risk weights of each logical control unit ,in This is the preset balance coefficient.

[0009] Further, in step S1, based on the dynamic risk weights, the virtual inspection path is dynamically generated in the digital twin model, including: Dynamic risk weights Logical control units that exceed the preset weight threshold are marked as high-priority inspection points. The three-dimensional coordinates of all high-priority inspection points in the digital twin model are extracted, and the Euclidean distance between any two high-priority inspection points is calculated as the path cost. Ant colony optimization is used to find the path to the high-priority inspection points: Initialize the ant population size. pheromone volatility coefficient Pheromones importance factor Heuristic function importance factor Each ant Select the next node according to the state transition probability formula. The state transition probability formula is: in, Representing a path The concentration of pheromones on the surface Indicates from node To the node Heuristic information, Ants The set of nodes that have not yet been visited; After each iteration, the global pheromone concentration is updated based on the total path length. When the maximum number of iterations is reached, the path with the highest pheromone concentration is output as the virtual inspection path that covers all high-optimal inspection points and has the shortest total path length.

[0010] Furthermore, the spatiotemporal alignment of the mobile sensing data stream and the static environment data stream in S2 includes: The real-time physical coordinates of the current inspection personnel are obtained using the positioning module built into the mobile terminal, and the real-time physical coordinates are mapped to the three-dimensional coordinate system of the digital twin model. Using the current time as the base timestamp, the video surveillance data stream within the range corresponding to the real-time physical coordinates is extracted as the mobile sensing data stream, and the environmental data stream collected by the fixed sensors deployed within the preset radius of the coordinates is retrieved as the static environmental data stream. The time delay difference between the mobile sensing data stream and the static environment data stream is eliminated by using the Kalman filter algorithm, thereby achieving spatiotemporal synchronization of the data.

[0011] Furthermore, in S3, cross-modal correlation analysis is performed to output the confidence level of potential hazards, including: Constructing a feature vector of hazardous chemical states Environmental safety feature vector And concatenate them to form a multidimensional feature vector. ; Will Input a pre-trained security hazard identification model and calculate the cross-modal correlation matrix through a self-attention mechanism. ,set up , , They are respectively The query matrix, key matrix, and value matrix generated by linear transformation; the formula for calculating the cross-modal association matrix is: ;in, For feature dimensions; based on and We perform weighted fusion and output the confidence level of potential hazards through a fully connected layer and a sigmoid function. The calculation formula is: ;in, and The weight matrix and bias terms obtained from model training, This is the Sigmoid activation function.

[0012] Furthermore, in step S3, the confidence level of the potential hazard is verified by semantic reasoning using a laboratory safety knowledge graph, including: The laboratory safety knowledge graph includes entity nodes and relation edges. The entity nodes include personnel, hazardous chemicals, equipment, and environmental parameters. The relation edges include storage relationships, operational relationships, and impact relationships. The target entities corresponding to the confidence levels of the potential hazards are input into the laboratory safety knowledge graph, and multi-hop path queries are performed based on the graph database; If a neighboring entity with a strong correlation risk to the target entity is found, the weight of the confidence level of the potential risk is increased; if a neighboring entity with a mutually exclusive relationship to the target entity is found and the neighboring entity is in normal status, the weight of the confidence level of the potential risk is decreased.

[0013] Furthermore, if the issue in S4 is a significant hidden danger, a locking or handling command is sent to the associated control device via the IoT gateway, including: Identify the source location and type of the major hidden dangers; Simulate the spread path of the aforementioned hazard type in a digital twin model to identify the affected associated control units; Send a locking command to the smart access control within the associated control unit, send a full-speed exhaust command to the ventilation equipment corresponding to the location of the hazard source, and send an evacuation command to the audible and visual alarm within the associated control unit.

[0014] Furthermore, in step S4, the rectified review data is received, and when a preset safety threshold is met, the handling instruction is lifted, including: Receive post-rectification image data and environmental re-measurement data uploaded by inspection personnel via mobile terminals; The differences between the images before and after rectification are compared using image recognition algorithms, and the environmental retest data is combined to determine whether the preset safety threshold is met. If the conditions are met, a release signal is generated and sent to the IoT gateway to unlock the associated control device. The corresponding security status identifier in the digital twin model is updated from alarm status to normal status, and the inspection record is stored on the blockchain node.

[0015] As a second aspect of the present invention, an Internet of Things-based laboratory safety inspection system is also provided, comprising: The digital twin path planning unit is used to acquire the digital twin model of the laboratory, divide the laboratory into several logical control units, monitor the environmental status parameters and equipment operation data of each logical control unit in real time, calculate the dynamic risk weight of each unit, and dynamically generate virtual inspection paths in the digital twin model according to the dynamic risk weights, and map them into a sequence of inspection tasks in the physical space and send them to the mobile terminal. The multi-source data feature extraction unit is used to respond to the arrival of inspection personnel at the current logical control unit, and simultaneously collect the mobile sensing data stream from the perspective of the inspection personnel and the static environmental data stream from the fixed Internet of Things node; perform spatiotemporal alignment on the mobile sensing data stream and the static environmental data stream, and extract multi-dimensional feature vectors of hazardous chemical status features and environmental safety features; The hazard identification and verification unit is used to input the multi-dimensional feature vector into a pre-trained safety hazard identification model, perform cross-modal correlation analysis to output the confidence level of potential hazards, and use a laboratory safety knowledge graph to perform semantic reasoning verification on the confidence level of potential hazards, eliminate interference from a single data source, and determine the final safety hazard level. The graded handling and closed-loop unit is used to generate a set of control instructions based on the final safety hazard level. If it is a general hazard, it outputs rectification guidance information through a mobile terminal. If it is a major hazard, it sends a locking or handling instruction to the associated control equipment through the Internet of Things gateway and updates the safety status identifier of the digital twin model. It receives the review data after rectification, and when the preset safety threshold is met, it releases the handling instruction and updates the digital twin model synchronously to complete the closed loop between the physical space and the virtual space.

[0016] As a third aspect of the invention, a computer-readable storage medium is also provided, on which a computer program is stored, which is executed by a processor as described in any one of the inventions: a laboratory safety inspection method based on the Internet of Things.

[0017] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: 1. The IoT-based laboratory safety inspection method of the present invention acquires a digital twin model of the laboratory and divides it into several logical control units. It monitors the environmental status parameters and equipment operation data of each unit in real time, calculates the data anomaly score and change rate, and obtains dynamic risk weights by weighting. Based on the weights, high-priority inspection points are marked. The optimal virtual inspection path is generated in the digital twin model using the ant colony algorithm. The path is mapped into a physical space inspection task sequence and sent to the mobile terminal. This method can get rid of the limitations of traditional fixed inspection routes, dynamically allocate inspection resources according to the real-time risk status, prioritize the coverage of high-risk areas, shorten the inspection path length, improve the pertinence and efficiency of inspection work, and keep the inspection layout consistent with the real-time safety status of the laboratory.

[0018] 2. The IoT-based laboratory safety inspection method of the present invention, in response to the arrival of inspection personnel at a designated logical control unit, simultaneously collects mobile sensing data streams from the personnel's perspective and static environmental data streams from fixed IoT nodes. Using coordinate mapping and Kalman filtering, the two types of data are aligned in time and space. Multi-dimensional feature vectors of hazardous chemical status and environmental safety are extracted. These feature vectors are then input into a pre-trained safety hazard identification model to perform cross-modal correlation analysis. After outputting the confidence level of potential hazards, semantic reasoning verification is performed using a laboratory safety knowledge graph. This method can eliminate the time and space bias of multi-source data, integrate heterogeneous data information, remove interference from a single data source, improve the accuracy of hazard judgment, and objectively determine the level of safety hazards.

[0019] 3. The IoT-based laboratory safety inspection method of the present invention generates a corresponding control instruction set based on the final safety hazard level. For general hazards, rectification guidance information is pushed to mobile terminals. For major hazards, locking or disposal instructions are sent to associated devices via IoT gateway and the safety status identifier of the digital twin model is updated. Rectification review data is received and the disposal instruction is released when the safety threshold is met, and the status of the digital twin model is updated synchronously. This enables the rapid and graded handling of hazards, timely control of the spread of major risks, and synchronization of the status of physical and virtual spaces. It constructs a closed loop for the entire process from inspection, identification, disposal to review, standardizes the laboratory safety management process, and ensures the integrity and traceability of management work. Attached Figure Description

[0020] Figure 1 This is a flowchart of a laboratory safety inspection method based on the Internet of Things 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

[0021] 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.

[0022] Example 1 Please refer to Figure 1 This embodiment 1 provides a laboratory safety inspection method based on the Internet of Things, including: S1. Obtain the digital twin model of the laboratory and divide the laboratory into several logical control units; monitor the environmental status parameters and equipment operation data of each logical control unit in real time, and calculate the dynamic risk weight of each unit; according to the dynamic risk weight, dynamically generate a virtual inspection path in the digital twin model and map it into a physical space inspection task sequence and send it to the mobile terminal. S2. In response to the arrival of the inspection personnel at the current logical control unit, simultaneously collect the mobile sensing data stream from the perspective of the inspection personnel and the static environmental data stream from the fixed IoT node; perform spatiotemporal alignment on the mobile sensing data stream and the static environmental data stream, and extract multidimensional feature vectors of hazardous chemical status characteristics and environmental safety characteristics; S3. Input the multidimensional feature vector into the pre-trained safety hazard identification model, perform cross-modal association analysis to output the confidence level of potential hazards; use the laboratory safety knowledge graph to perform semantic reasoning verification on the confidence level of potential hazards, eliminate interference from a single data source, and determine the final safety hazard level; S4. Generate a set of control instructions based on the final safety hazard level; if it is a general hazard, output rectification guidance information through the mobile terminal; if it is a major hazard, send a lock or disposal instruction to the associated control device through the IoT gateway, and update the safety status identifier of the digital twin model; receive the verification data after rectification, and when the preset safety threshold is met, release the disposal instruction and update the digital twin model synchronously to complete the closed loop between physical space and virtual space.

[0023] This embodiment 1 further elaborates on the above steps.

[0024] (1) Digital Twin Path Planning This embodiment details how to use the "environmental status parameters" and "equipment operation data" of each logical control unit to calculate dynamic risk weights and generate virtual inspection paths accordingly. This process completely abandons the weighted logic that relies on historical static data, and instead focuses on the deviation and trend of real-time data to ensure that inspection tasks can respond immediately to the current safety situation of the laboratory.

[0025] First, real-time data collection is performed on each logical control unit within the laboratory. Assume the laboratory is divided into... A logical control unit, index Representing the One logical control unit ( For each unit The IoT sensor network uploads environmental status parameter sets (including temperature, humidity, harmful gas concentration, smoke index, etc.) and equipment operation datasets (including current, voltage, and fume hood airflow rate of high-power equipment, etc.) in real time. To quantify the current safety status of the unit, a sliding window mechanism is used, within a preset time window. Data is continuously collected within a short period of time (e.g., the last 10 minutes).

[0026] Next, in order to calculate the first Anomaly scoring of each logical control unit First, calculate the real-time average value of each monitoring parameter within the window. and standard deviation Subsequently, the historical benchmark average established by the unit under long-term stable operating conditions is invoked. and the benchmark standard deviation The Z-score normalization method is used to calculate the deviation of the current real-time data from historical normal operating conditions. For the ... The first unit in the Each monitoring parameter has a single-point anomaly score. The calculation formula is as follows: in, For the current moment Unit 1 The monitored values ​​of each parameter. For all monitored parameters... A weighted average or maximum value is applied to obtain the overall anomaly score for that unit. . The higher the value, the further the current environment or equipment status deviates from the normal range, and the higher the possibility of potential faults or leaks.

[0027] While calculating the anomaly score, it is also necessary to capture the evolution trend of the risk, i.e., the rate of change. This is achieved by calculating the absolute value of the first-order difference of data within adjacent time windows. Let the current time window be... The previous time window was Then the first The parameter change rate of each logic control unit It can be represented as: in, This represents the total number of monitored parameters within the unit. Indicates the first Unit 1 The parameters are in The mean within the time window. This reflects the severity of data fluctuations. Even if the current value is within the acceptable range, severe fluctuations often indicate that the equipment is about to malfunction or the environment is about to get out of control.

[0028] Obtain an anomaly score and rate of change Then, the first step is calculated by linear weighted fusion. The final dynamic risk weight of each logical control unit To balance the impact of the "current degree of abnormality" and the "future deterioration trend," a balance coefficient is introduced. (The value range is usually set between 0.6 and 0.8, focusing on the current anomaly), the calculation formula is: Calculate the values ​​of all units Afterwards, Logical control units exceeding a preset risk threshold are selected and marked as "high-priority inspection points." The three-dimensional coordinates of these high-priority inspection points are then extracted from the digital twin model. And calculate any two points and Euclidean distance between As path cost: Finally, the ant colony algorithm is used to plan the shortest path covering all high-priority inspection points. Initialization Only ants, setting pheromone importance factors Heuristic function importance factor and pheromone volatility coefficient .Ant From node Transfer to node probability Based on the pheromone concentration along the current path and heuristic information (Usually take the reciprocal of the distance) Jointly decided: By iteratively updating the pheromone matrix multiple times until the maximum number of iterations is reached, the optimal path output by the algorithm after convergence becomes the final virtual inspection path issued to the inspection personnel. This path is mapped to a physical inspection task sequence and sent to the mobile terminal held by the inspection personnel. This process ensures that inspection resources are preferentially allocated to risk weights. The highest level of inspection has enabled truly dynamic risk-driven inspections.

[0029] (2) Feature extraction from multi-source data When the inspection personnel arrive at the corresponding logical control unit, the synchronous data acquisition process is immediately initiated. First, the real-time physical coordinates of the inspection personnel are obtained using the mobile terminal's built-in positioning module, and these coordinates are mapped to the three-dimensional coordinate system of the digital twin model. Specifically, the mobile terminal uses its integrated positioning module to obtain the real-time coordinates of the current inspection personnel in the laboratory's physical space, denoted as... To accurately map this physical location to the virtual space of the digital twin model, a coordinate transformation matrix between the physical and virtual spaces was pre-established. Through matrix operations, real-time physical coordinates are... Converted into 3D virtual coordinates in a digital twin model The calculation formula is as follows: in, This step involves converting physical coordinates into homogeneous coordinates. This process enables real-time location tracking of inspection personnel within the digital twin model, providing a spatial reference for subsequent data association.

[0030] After determining the virtual coordinates, use the current time Using the base timestamp, multi-source data is extracted and retrieved. This is done using the mapped virtual coordinates. Centered on the coordinates, the video surveillance data stream within the corresponding range is extracted as the motion sensing data stream. This data stream is typically generated by the camera on the inspector's handheld terminal or by a wearable data collection device. Simultaneously, it is retrieved from the digital twin model... Center and preset radius All fixed sensor nodes within the range are used, and the environmental data streams collected by these nodes at the same time are retrieved as static environmental data streams. At this time, due to differences in network transmission latency and hardware sampling frequency, the mobile sensing data stream... data stream in static environment There are often deviations on the timeline, making direct fusion analysis impossible.

[0031] To eliminate the aforementioned time delay difference and achieve spatiotemporal synchronization of data, this embodiment employs a Kalman filter algorithm to align the two sets of data streams. The timestamps of the mobile sensing data stream are used as observations, and the timestamps of the static environment data stream are used as reference ground truths to construct a state-space model. State vectors are defined. For the first Time deviation and drift rate, observation vector This is a measurement of the time difference between the two data streams.

[0032] First, a time update (prediction) process is performed, using the best estimate from the previous time step to predict the current state: in, For prior state estimation, Here is the state transition matrix. To estimate the covariance a priori, Let be the process noise covariance matrix.

[0033] A measurement update (correction) process is then performed, incorporating current observations to revise the prediction results and calculate the Kalman gain. And update the state estimate: in, For the observation matrix, To measure the noise covariance matrix, the time delay of the mobile sensing data stream relative to the static environment data stream can be estimated in real time through the above iterative calculation. Based on this, the mobile sensing data stream is shifted and compensated on the time axis, ultimately achieving strict synchronization of the two types of data streams under a unified spatiotemporal benchmark, providing a high-quality data foundation for subsequent multimodal data fusion analysis.

[0034] After completing spatiotemporal alignment, multi-dimensional feature vectors of hazardous chemical state features and environmental safety features are extracted from the aligned data. First, hazardous chemical state features are extracted from the aligned motion sensing data stream. The aligned motion sensing data stream is then input into a pre-trained lightweight object detection network. This network is specifically trained for laboratory scenarios and can identify key objects such as reagent bottles, reaction vessels, and pipeline valves. For the detected... For each hazardous chemical object, the network outputs its bounding box coordinates. and visual state feature vectors This feature vector contains visual attributes of the object, such as liquid level, liquid color changes, whether the container is damaged, whether the label is clear, and visual texture features indicating smoke or liquid leakage. Simultaneously, environmental safety features are extracted from the aligned static environmental data stream. For the aforementioned identified... For each hazardous chemical object, extract the set of environmental parameters associated with its spatial location. To quantify the safety status of the environment, an environmental safety feature vector is calculated. This vector contains not only the original temperature, humidity, and gas concentration values, but also risk indicators calculated based on these values. For example, it can be used to calculate the current gas concentration. With respect to the lower explosive limit of the substance The ratio, and the rate of temperature change Environmental safety feature vector It can be represented as: in, This indicates the operating status of the ventilation equipment (0 or 1).

[0035] Finally, visual features are combined with environmental features to construct a multidimensional feature vector for this hazardous chemical object. Visual state feature vector Perform dimensionality reduction on the fully connected layer to make its dimension similar to that of the environmental safety feature vector. The features are matched, and then the final multidimensional feature vector is generated through feature concatenation. This multidimensional feature vector It encompasses both the physical state visible to the human eye and the microscopic state perceived by the environment. Through this spatiotemporal alignment and feature fusion, it is possible to detect hidden dangers that cannot be detected by a single modality, thereby providing data support for subsequent hidden danger identification.

[0036] (3) Hazard identification and verification Complete the multidimensional feature vector based on the state characteristics and environmental safety characteristics of hazardous chemicals. After the construction is completed, hazard identification is required. First, the constructed multidimensional feature vector... The input is fed into a pre-trained safety hazard identification model, which internally employs a cross-modal correlation analysis algorithm based on an attention mechanism. Within the model, it first uses three trainable linear transformation matrices... , , input vector Mapped to query matrix Key matrix Sum matrix To uncover the implicit coupling between the state of hazardous chemicals (e.g., reagent volatilization) and environmental factors (e.g., temperature rise), the model calculates the cross-modal correlation matrix. This matrix measures the correlation between different feature dimensions by calculating the dot product of the query vector and the key vector, and introduces a scaling factor. The stable gradient is calculated using the following formula: in, Key matrix Feature dimensions, This represents the transpose of the key matrix. The function is used to normalize association scores into a probability distribution. Matrix Each element in the table represents the spatial or semantic association weight between hazardous chemical characteristics and environmental characteristics.

[0037] Based on the calculated cross-modal correlation matrix The model further performs weighted fusion to generate a fused feature representation. This is achieved using the correlation matrix. Log-value matrix Weighted summation is performed to aggregate cross-modal contextual information, resulting in a fused feature vector. Finally, The input is fed into a fully connected layer and mapped to the [0,1] interval using a Sigmoid activation function. The final confidence score of the potential hazard is then output. This confidence level represents the probability of a safety incident occurring in the current scenario, and is calculated using the following formula: in, and These are the weight matrix and bias term obtained from training the output layer, respectively. This represents the Sigmoid function.

[0038] To further improve the accuracy of judgments and avoid interference from single data sources or model calculations, a laboratory safety knowledge graph is introduced to perform semantic reasoning verification of hazard confidence. First, a laboratory safety knowledge graph containing rich semantic information is constructed. The laboratory safety knowledge graph includes entity nodes and relational edges. Entity nodes include personnel, hazardous chemicals, equipment, and environmental parameters, while relational edges include storage relationships, operational relationships, and impact relationships. Each entity node in the graph maintains real-time state attributes, which are synchronized with sensor data streams in the physical world through an IoT interface.

[0039] When receiving a confidence level of potential hazards for a specific target entity First, the target entity is input into the laboratory safety knowledge graph as the query entry node. Then, using the graph traversal algorithm of the graph database, a multi-hop path query is performed outward from this entry node. Specifically, the query depth is set to... The purpose is to discover other entity nodes that are physically adjacent to the target entity or logically related to it.

[0040] After obtaining the set of related entities, the confidence levels of potential risks are weighted and adjusted according to a predefined set of inference rules. The inference rules are divided into risk enhancement rules and risk mitigation rules.

[0041] For risk enhancement rules, if a neighboring entity found in the query has a strong correlation risk with the target entity, and the current state of that neighboring entity is abnormal, the confidence level is increased. For example, the graph shows that "ethanol" and "open flame" have a "prohibited contact" relationship. If the query finds that the "alcohol lamp" entity near the target entity "ethanol" is in a "lit" state, then risk superposition is determined. In this case, the original confidence level is corrected using a sigmoid function or linear weighting, and the corrected confidence level is... The calculation formula can be expressed as: in, To enhance the weighting coefficient (with values ​​between 0 and 1), representing the influence of associated risks.

[0042] For risk mitigation rules, if a neighboring entity found in the query has a mutually exclusive or causal relationship with the target entity, and the neighboring entity is in a normal state or provides a reasonable explanation, the confidence level is reduced. For example, an increase in "smoke concentration" is detected, initially indicating a fire hazard. However, in the knowledge graph, the "smoke sensor" and the "ventilation system" have a "controlled" relationship. If the query finds that the "ventilation system" is "off" and the "temperature and humidity" node shows "high humidity," this may mean that the sensor readings are drifting due to environmental condensation or airflow stagnation, rather than a real fire. In this case, it is determined to be interference from a single data source, and the confidence level is reduced. in, To suppress the weighting coefficient.

[0043] Finally, the corrected final confidence level The final safety hazard level is determined by comparing it with a preset threshold level, which then determines whether to trigger an alarm or generate a rectification work order. The closer the value is to 1, the higher the likelihood of a safety hazard in the area. This process effectively eliminates false alarms caused by single sensor malfunctions or environmental background noise, while enhancing the ability to identify complex coupled risks and ensuring safe laboratory management.

[0044] (4) Tiered treatment and closed loop After determining the final safety hazard level, corresponding control instruction sets are generated based on the level differences. Differentiated handling methods are adopted for different risk levels to ensure that control measures match the hazard risk. For cases determined to be general hazards, rectification guidance information is directly output to inspection personnel via mobile terminals, clarifying the rectification content and operation standards, guiding on-site personnel to quickly complete the hazard rectification, and improving the efficiency of handling routine hazards.

[0045] For situations identified as major hazards, the source location and type of the hazard are first identified. Then, the environmental simulation engine in the digital twin model is invoked to simulate the diffusion path of this hazard type under the current laboratory environmental parameters based on computational fluid dynamics algorithms. Based on the simulation results, affected associated control units are delineated. Subsequently, based on the identified associated control units, specific control command sets are generated and issued. For all smart access control devices within the associated control units, a locking command is sent to prohibit personnel from entering the hazardous area and prevent secondary injuries. For ventilation equipment corresponding to the hazard source location, a full-speed exhaust command is sent to accelerate the discharge of harmful gases at maximum power. For audible and visual alarms within the associated control units, high-frequency flashing and voice evacuation commands are sent to prompt personnel in the area to evacuate quickly. All commands are protocol-converted through an IoT gateway to ensure compatibility with hardware devices of different brands and protocols. Simultaneously, in the digital twin model, the safety status indicator of the associated control unit is updated in real time from "normal" (green) to "critical alarm" (red), and the diffusion simulation path is highlighted on the 3D visualization interface, providing intuitive decision-making basis for the remote command center.

[0046] After issuing the emergency response order, the rectification and verification phase begins. Once inspection or maintenance personnel have completed on-site handling, they upload rectified image data and environmental re-measurement data via mobile terminals. Using image recognition algorithms, the rectified image is compared pixel-by-pixel with the original image of the potential hazard, and the difference is calculated to verify whether the physical rectification is adequate. Simultaneously, real-time values ​​from environmental sensors in the area are read to determine if they have fallen below a preset safety threshold. Only when the image difference meets the requirements and the environmental data is deemed satisfactory is the rectification deemed complete.

[0047] After rectification is deemed satisfactory, a release signal is generated and sent to the IoT gateway, automatically unlocking the previously locked smart access control and restoring the ventilation equipment to normal operating speed. Simultaneously, in the digital twin model, the corresponding safety status indicator is updated from "alarm status" back to "normal status," achieving synchronization between the physical and virtual spaces. Finally, to ensure the inspection records are tamper-proof and traceable, the entire inspection process data (including hazard discovery time, hazard type, handling instructions, before-and-after images, and verification data) is packaged into a data block and stored on the blockchain node via a smart contract. This block contains the hash value of the previous block, forming a chain structure that ensures the integrity and reliability of laboratory safety data, completing a closed-loop management process from hazard discovery and handling to verification.

[0048] Example 2 Please refer to Figure 2 This embodiment 2 provides an Internet of Things-based laboratory safety inspection system, including: The digital twin path planning unit is used to acquire the digital twin model of the laboratory, divide the laboratory into several logical control units, monitor the environmental status parameters and equipment operation data of each logical control unit in real time, calculate the dynamic risk weight of each unit, and dynamically generate virtual inspection paths in the digital twin model according to the dynamic risk weights, and map them into a sequence of inspection tasks in the physical space and send them to the mobile terminal. The multi-source data feature extraction unit is used to respond to the arrival of inspection personnel at the current logical control unit, and simultaneously collect the mobile sensing data stream from the perspective of the inspection personnel and the static environmental data stream from the fixed Internet of Things node; perform spatiotemporal alignment on the mobile sensing data stream and the static environmental data stream, and extract multi-dimensional feature vectors of hazardous chemical status features and environmental safety features; The hazard identification and verification unit is used to input the multi-dimensional feature vector into a pre-trained safety hazard identification model, perform cross-modal correlation analysis to output the confidence level of potential hazards, and use a laboratory safety knowledge graph to perform semantic reasoning verification on the confidence level of potential hazards, eliminate interference from a single data source, and determine the final safety hazard level. The graded handling and closed-loop unit is used to generate a set of control instructions based on the final safety hazard level. If it is a general hazard, it outputs rectification guidance information through a mobile terminal. If it is a major hazard, it sends a locking or handling instruction to the associated control equipment through the Internet of Things gateway and updates the safety status identifier of the digital twin model. It receives the review data after rectification, and when the preset safety threshold is met, it releases the handling instruction and updates the digital twin model synchronously to complete the closed loop between the physical space and the virtual space.

[0049] 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 Internet of Things-based laboratory safety inspection method.

[0050] 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.

[0051] 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.

[0052] 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 laboratory safety inspection method based on the Internet of Things, characterized in that, include: S1. Obtain a digital twin model of the laboratory and divide the laboratory into several logical control units; Real-time monitoring of environmental status parameters and equipment operation data of each logical control unit; calculation of dynamic risk weights for each unit; based on the dynamic risk weights, dynamic generation of virtual inspection paths in the digital twin model; mapping of virtual inspection paths into physical space inspection task sequences and sending them to mobile terminals. S2. In response to the arrival of the inspection personnel at the current logical control unit, simultaneously collect the mobile sensing data stream from the perspective of the inspection personnel and the static environmental data stream from the fixed IoT node; Spatiotemporal alignment is performed on the mobile sensing data stream and the static environment data stream to extract multidimensional feature vectors of hazardous chemical status characteristics and environmental safety characteristics. S3. Input the multidimensional feature vector into the pre-trained safety hazard identification model, perform cross-modal association analysis to output the confidence level of potential hazards; use the laboratory safety knowledge graph to perform semantic reasoning verification on the confidence level of potential hazards, eliminate interference from a single data source, and determine the final safety hazard level; S4. Generate a set of control instructions based on the final safety hazard level; if it is a general hazard, output rectification guidance information through the mobile terminal; if it is a major hazard, send a lock or disposal instruction to the associated control device through the IoT gateway, and update the safety status identifier of the digital twin model; receive the verification data after rectification, and when the preset safety threshold is met, release the disposal instruction and update the digital twin model synchronously to complete the closed loop between the physical space and the virtual space. In S3, cross-modal correlation analysis is performed to output the confidence level of potential hazards, including: Constructing a feature vector of hazardous chemical states Environmental safety feature vector And concatenate them to form a multidimensional feature vector. ; Will Input a pre-trained security hazard identification model and calculate the cross-modal correlation matrix through a self-attention mechanism. ,set up , , They are respectively The query matrix, key matrix, and value matrix generated by linear transformation; the formula for calculating the cross-modal association matrix is: ;in, For feature dimensions; based on and We perform weighted fusion and output the confidence level of potential hazards through a fully connected layer and a sigmoid function. The calculation formula is: ;in, and The weight matrix and bias terms obtained from model training, Use the Sigmoid activation function; The S3 step utilizes a laboratory safety knowledge graph to perform semantic reasoning verification on the confidence level of the potential hazard, including: The laboratory safety knowledge graph includes entity nodes and relation edges. The entity nodes include personnel, hazardous chemicals, equipment, and environmental parameters. The relation edges include storage relationships, operational relationships, and impact relationships. The target entities corresponding to the confidence levels of the potential hazards are input into the laboratory safety knowledge graph, and multi-hop path queries are performed based on the graph database; If a neighboring entity with a strong correlation risk to the target entity is found, the weight of the confidence level of the potential risk is increased; if a neighboring entity with a mutually exclusive relationship to the target entity is found and the neighboring entity is in normal status, the weight of the confidence level of the potential risk is decreased.

2. The laboratory safety inspection method based on the Internet of Things according to claim 1, characterized in that, The calculation process of the dynamic risk weight of each unit in S1 specifically includes: Acquire real-time environmental status parameter sets and equipment operation datasets for each logical control unit; the environmental status parameters include temperature and humidity, harmful gas concentration and smoke index, and the equipment operation data includes high-power equipment current values ​​and fume hood operation status; The real-time environmental state parameter set and the equipment operation dataset are processed using a sliding window method to calculate the mean and standard deviation of each parameter within a preset time window. Based on the Z-score standardization method, the deviation of each parameter from historical normal operating conditions is calculated to obtain an anomaly score. ;in, Indicates the first The identifier index of each logical control unit; Simultaneously, the absolute values ​​of the first-order differences of each parameter within adjacent time windows are calculated to obtain the rate of change. Using weighted fusion formula Calculate the first Dynamic risk weights of each logical control unit ,in This is the preset balance coefficient.

3. The laboratory safety inspection method based on the Internet of Things according to claim 2, characterized in that, In step S1, based on the dynamic risk weights, a virtual inspection path is dynamically generated in the digital twin model, including: Dynamic risk weights Logical control units that exceed the preset weight threshold are marked as high-priority inspection points. The three-dimensional coordinates of all high-priority inspection points in the digital twin model are extracted, and the Euclidean distance between any two high-priority inspection points is calculated as the path cost. Ant colony optimization is used to find the path to the high-priority inspection points: Initialize the ant population size. pheromone volatility coefficient Pheromones importance factor Heuristic function importance factor Each ant Select the next node according to the state transition probability formula. The state transition probability formula is: in, Representing a path The concentration of pheromones on the surface Indicates from node To the node Heuristic information, Ants The set of nodes that have not yet been visited; After each iteration, the global pheromone concentration is updated based on the total path length. When the maximum number of iterations is reached, the path with the highest pheromone concentration is output as the virtual inspection path that covers all high-optimal inspection points and has the shortest total path length.

4. The laboratory safety inspection method based on the Internet of Things according to claim 1, characterized in that, The S2 step of performing spatiotemporal alignment between the mobile sensing data stream and the static environment data stream includes: The real-time physical coordinates of the current inspection personnel are obtained using the positioning module built into the mobile terminal, and the real-time physical coordinates are mapped to the three-dimensional coordinate system of the digital twin model. Using the current time as the base timestamp, the video surveillance data stream within the range corresponding to the real-time physical coordinates is extracted as the mobile sensing data stream, and the environmental data stream collected by the fixed sensors deployed within the preset radius of the coordinates is retrieved as the static environmental data stream. The time delay difference between the mobile sensing data stream and the static environment data stream is eliminated by using the Kalman filter algorithm, thereby achieving spatiotemporal synchronization of the data.

5. The laboratory safety inspection method based on the Internet of Things according to claim 1, characterized in that, If the S4 step involves a major hidden danger, a locking or handling command will be sent to the associated control device via the IoT gateway, including: Identify the source location and type of the major hidden dangers; Simulate the spread path of the aforementioned hazard type in a digital twin model to identify the affected associated control units; Send a locking command to the smart access control within the associated control unit, send a full-speed exhaust command to the ventilation equipment corresponding to the location of the hazard source, and send an evacuation command to the audible and visual alarm within the associated control unit.

6. The laboratory safety inspection method based on the Internet of Things according to claim 1, characterized in that, In step S4, the rectified review data is received. When a preset safety threshold is met, the handling instruction is lifted, including: Receive post-rectification image data and environmental re-measurement data uploaded by inspection personnel via mobile terminals; The differences between the images before and after rectification are compared using image recognition algorithms, and the environmental retest data is combined to determine whether the preset safety threshold is met. If the conditions are met, a release signal is generated and sent to the IoT gateway to unlock the associated control device. The corresponding security status identifier in the digital twin model is updated from alarm status to normal status, and the inspection record is stored on the blockchain node.

7. An Internet of Things (IoT)-based laboratory safety inspection system, used to implement the IoT-based laboratory safety inspection method as described in claim 1, characterized in that, include: The digital twin path planning unit is used to obtain the digital twin model of the laboratory and divide the laboratory into several logical control units; Real-time monitoring of environmental status parameters and equipment operation data of each logical control unit; calculation of dynamic risk weights for each unit; based on the dynamic risk weights, dynamic generation of virtual inspection paths in the digital twin model; mapping of virtual inspection paths into physical space inspection task sequences and sending them to mobile terminals. The multi-source data feature extraction unit is used to respond to the arrival of inspection personnel at the current logical control unit by simultaneously collecting mobile sensing data streams from the perspective of inspection personnel and static environmental data streams from fixed IoT nodes. Spatiotemporal alignment is performed on the mobile sensing data stream and the static environment data stream to extract multidimensional feature vectors of hazardous chemical status characteristics and environmental safety characteristics. The hazard identification and verification unit is used to input the multi-dimensional feature vector into the pre-trained safety hazard identification model, perform cross-modal correlation analysis, and output the confidence level of potential hazards. The confidence level of the potential hazards was verified by semantic reasoning using a laboratory safety knowledge graph, eliminating interference from a single data source, and determining the final safety hazard level. The graded handling and closed-loop unit is used to generate a set of control instructions based on the final safety hazard level. If it is a general hazard, it outputs rectification guidance information through a mobile terminal. If it is a major hazard, it sends a locking or handling instruction to the associated control equipment through the Internet of Things gateway and updates the safety status identifier of the digital twin model. It receives the review data after rectification, and when the preset safety threshold is met, it releases the handling instruction and updates the digital twin model synchronously to complete the closed loop between the physical space and the virtual space.

8. 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 1-6: a laboratory safety inspection method based on the Internet of Things.

Images