Real-time risk identification and active safety control method and system for industrial operating personnel

Abstract

The invention relates to the field of industrial safety monitoring, provides a real-time risk identification and active safety control method and system for industrial operating personnel in order to improve the accuracy and efficiency of a safety monitoring technology, and realizes real-time acquisition and edge preprocessing of human-machine-environment data by integrating a physiological / environmental / equipment multi-mode sensor. An FPN network is utilized to fuse spatial-temporal characteristics, Temporal-PC causal inference and a GBDT integration algorithm are combined, an interpretable risk propagation model is constructed, risk indexes are calculated, three-level early warning is divided, and hierarchical intervention is executed for different risk levels, including local early warning, environment regulation and control, emergency evacuation and multi-department emergency linkage. Through LSTM trend prediction and reinforcement learning dynamic parameter adjustment, closed-loop optimization of an intervention strategy is realized, the problems of information splitting, response lag and the like of a traditional monitoring system are solved, the risk identification precision and emergency disposal efficiency of a chemical high-risk scene are remarkably improved, and the method is suitable for operation safety control in high-temperature, high-pressure, toxic and harmful environments.

Description

Technical Field

[0001] This invention relates to the field of industrial safety monitoring, specifically a method and system for real-time risk identification and proactive safety control of industrial workers. Background Technology

[0002] With the booming development of the chemical industry, personnel safety monitoring in hazardous work scenarios faces severe challenges. Although traditional safety monitoring methods have ensured operational safety to a certain extent, the limitations of existing technologies are becoming increasingly prominent as production scale expands and process complexity increases, making it difficult to meet the needs of intelligent and precise safety management. This is mainly reflected in the following aspects.

[0003] First, there is a serious problem of information silos. Traditional safety monitoring methods, such as fixed gas detectors, video surveillance, and manual inspections, operate independently and lack linkage with the real-time physiological and environmental conditions of individual workers. This information silo phenomenon makes it difficult to detect sudden and hidden safety risks in a timely manner, such as trace leaks of toxic gases or decreased physical function of workers due to high-temperature work, which can easily lead to missing the best opportunity for response.

[0004] Secondly, risk identification is delayed. Existing monitoring systems mostly rely on single-variable over-limit alarms, using only a single indicator such as excessive gas concentration as the basis for alarms, ignoring the coupling effect between physiological abnormalities and environmental changes. In actual operations, physiological abnormalities such as dehydration and respiratory disorders caused by high-temperature environments often combine with environmental factors to create complex risks. However, existing systems cannot effectively identify risks in their early stages, with response delays often reaching several minutes, greatly increasing the probability of accidents.

[0005] Third, the early warning mechanism is too crude. Traditional alarm methods mainly rely on fixed thresholds and adopt a single "alarm + evacuation" response mode, lacking refined safety strategies such as multi-level classification, gradual intervention, and path repair. This leads to false alarms and missed alarms in practical applications. Frequent false alarms can cause unnecessary production interruptions, while missed alarms may cause potential risks to evolve into serious accidents, severely affecting production safety and efficiency.

[0006] Fourth, there is a lack of intelligent collaborative capabilities. In hazardous chemical operation scenarios, the dangerous interactions between personnel and the environment exhibit complex nonlinear causal relationships. However, existing monitoring solutions struggle to establish interpretable and predictive risk propagation mechanisms, making it impossible to proactively anticipate potential risks. They can only passively respond to risk events that have already occurred, severely hindering the improvement of proactive safety intervention capabilities.

[0007] Therefore, there is an urgent need for a way to improve the accuracy and efficiency of safety monitoring technology. Summary of the Invention

[0008] To improve the accuracy and efficiency of safety monitoring technology, this invention provides a method and system for real-time risk identification and proactive safety control of industrial workers.

[0009] The technical solution adopted by the present invention to solve the above problems is:

[0010] Methods for real-time risk identification and proactive safety control for industrial workers include:

[0011] Step 1: Collect real-time work condition data and preprocess it. The work condition data includes the physiological data of the workers, environmental data and equipment condition data.

[0012] Step 2: Perform multimodal fusion on the preprocessed real-time operation data to obtain multimodal feature vectors, and dynamically assign weights to each operation data item.

[0013] Step 3: Construct and train a causal model to identify causal relationships based on multimodal feature vectors and the trained causal model;

[0014] Step 4: Calculate the hazard index based on the multimodal feature vector and the dynamic weights corresponding to each work condition data item. Train the LSTM model using the multimodal feature vector and the hazard index, and use the trained LSTM model to predict the future hazard index.

[0015] Step 5: Determine the risk level based on the future hazard index and early warning threshold;

[0016] Step 6: Implement graded safety control and intervention measures based on risk level and causal relationship.

[0017] Furthermore, the physiological data of the operators includes: body temperature, heart rate, blood oxygen, respiratory rate, blood pressure and location information; the environmental data includes toxic gas concentration, temperature and humidity, dust, noise, VOC and infrared characteristics of leaked gas; and the equipment operating data includes: valve opening, pipeline pressure, reactor temperature and pump speed.

[0018] Furthermore, the preprocessing includes: employing 3 The criteria for removing outliers are: linear interpolation, spatiotemporal fusion interpolation, or multidimensional interpolation to fill missing values; and the Min-Max method is used for data normalization.

[0019] Furthermore, step 2 specifically includes:

[0020] A spatiotemporal feature fusion network model is constructed, including: a temporal feature extraction module for extracting temporal features, a spatial feature extraction module for extracting spatial features, and a feature fusion module for feature fusion;

[0021] The preprocessed real-time operation data is input into the spatiotemporal feature fusion network model for multimodal fusion.

[0022] The formula for calculating the weight is: ,in, Chemical safety experts assign risk level scores to each parameter. Data items calculated based on a historical accident database The frequency of accidents Adjustments will be made dynamically based on job type. .

[0023] Furthermore, the Temporal-PC algorithm is used to construct a causal model, including: calculating the partial correlation coefficient of parameter pairs and constructing an initial causal graph; filtering strong causal relationships in the initial causal graph based on the partial correlation coefficient; integrating the causal graphs corresponding to all strong causal relationships to generate a weighted causal graph; parameter pairs include combinations of physiological data of workers and environmental data, combinations of environmental data, combinations of equipment operating condition data and environmental data, and combinations of physiological data of workers.

[0024] Furthermore, the causal model is validated using VR scenarios, specifically including: constructing a VR scenario of a chemical accident, inviting safety experts to score the rationality of the causal paths, and retaining causal paths with scores higher than a preset threshold.

[0025] Furthermore, the risk index The calculation method is as follows: , For data items The value after normalization Let be the weight of data item i, and N be the dimension of the multimodal feature vector. and .

[0026] Furthermore, it also includes: calculating intervention effectiveness, and dynamically adjusting the early warning threshold based on the intervention effectiveness using a reinforcement learning algorithm; the intervention effectiveness is calculated as follows: ,in, This represents the actual number of accidents that were successfully avoided. This is the total number of times the warning has been triggered. This represents the pass rate for manual review.

[0027] A real-time risk identification and proactive safety control system for industrial workers, including:

[0028] Data acquisition module: used to collect real-time operational data, including physiological data of workers, environmental data, and equipment operational data;

[0029] Data processing module: preprocesses the collected real-time operation condition data, performs multimodal fusion on the preprocessed real-time operation condition data to obtain multimodal feature vectors, and dynamically assigns weights to each operation condition data item.

[0030] Causal relationship determination module: Constructs and trains a causal model, and identifies causal relationships based on multimodal feature vectors and the trained causal model;

[0031] Hazard index prediction module: The hazard index is calculated based on the multimodal feature vector and the dynamic weights corresponding to each work condition data item. The LSTM model is trained using the multimodal feature vector and the hazard index, and the trained LSTM model is used to predict the future hazard index.

[0032] The risk level determination module determines the risk level based on the future hazard index and early warning threshold.

[0033] Safety control and intervention module: Implements graded safety control and intervention measures based on risk level and causal relationship.

[0034] Furthermore, it also includes an intervention effectiveness calculation module: calculating intervention effectiveness and dynamically adjusting the warning threshold based on the intervention effectiveness using a reinforcement learning algorithm.

[0035] The advantages of this invention compared to existing technologies are as follows: Real-time operational data is collected and preprocessed, including worker physiological data, environmental parameters, and equipment operating conditions. Then, multimodal fusion and dynamic weight allocation are performed on the preprocessed real-time operational data. A causal model is constructed and pre-trained, and multimodal feature vectors are input into the causal model to identify causal relationships. A hazard index is calculated based on dynamic weights, multimodal feature vectors, and a hazard index prediction model. Finally, the risk level is determined based on the hazard index and a preset warning threshold, and safety control and intervention measures are implemented based on the risk level and causal relationships. This invention integrates physiological / environmental / equipment multimodal sensors to achieve real-time acquisition and edge preprocessing of human-machine-environment data. It utilizes an FPN network to fuse spatiotemporal features, combines Temporal-PC causal inference with the GBDT ensemble algorithm, constructs an interpretable risk propagation model, calculates the hazard index, and classifies warnings into three levels. Tiered interventions are implemented for different risk levels, including local warnings, environmental control, emergency evacuation, and multi-departmental emergency response. By using LSTM trend prediction and reinforcement learning for dynamic parameter tuning, closed-loop optimization of intervention strategies is achieved, solving problems such as information fragmentation and response lag in traditional monitoring systems. This significantly improves the accuracy of risk identification and the efficiency of emergency response in high-risk chemical scenarios, and is suitable for operational safety control in high-temperature, high-pressure, and toxic and hazardous environments. Attached Figure Description

[0036] Figure 1Flowchart for real-time risk identification and proactive safety control methods for industrial workers;

[0037] Figure 2 This is a schematic diagram of the spatiotemporal feature fusion network model structure;

[0038] Figure 3 This is a diagram illustrating the data flow. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0040] like Figure 1 As shown, the method for real-time risk identification and proactive safety control for industrial workers includes:

[0041] Step 1: Collect real-time operational data and preprocess it. The operational data includes the physiological data of the workers, environmental data, and equipment operational data.

[0042] In this embodiment, the collected physiological data includes body temperature, heart rate, blood oxygen saturation, respiratory rate, blood pressure, and location information. These data are collected using wearable devices, including a flexible thermocouple sensor (accuracy ±0.1℃, sampling frequency 10Hz, attached to the wrist artery), a PPG photoplethysmography sensor (wavelength 525nm / 660nm, heart rate accuracy ±1bpm, blood oxygen saturation ±2%, worn on the wrist / earlobe), a strain-type flexible sensor (attached to the chest to detect chest deformation, accuracy ±1 time / min, 10Hz sampling), an oscillatory micro blood pressure module (wrist, accuracy ±3mmHg, measured every 5 minutes), and a positioning unit (GPS, BeiDou, etc.). Environmental data includes toxic gas concentrations (toxic gases such as...). Temperature, humidity / pressure, dust, noise, etc., are collected by fixed sensors, including electrochemical sensors (accuracy ±5%FS, deployed at a height of 1.5m in the work area, 5Hz sampling, RS485 transmission), temperature and humidity transmitters (accuracy ±0.3℃ / ±2%RH, pressure ±0.1kPa, deployed near the equipment, 5Hz sampling), and laser scattering dust meters (0.1-1000mg / m³, accuracy ±10%) and sound level meters (30-130dB, accuracy ±1dB) deployed along the inspection path. It also includes integrated miniature gas sensors (such as PID VOC detection, accuracy ±10%, 10Hz sampling, synchronous transmission with physiological data), and drones equipped with multispectral cameras (detecting infrared characteristics of leaked gas, 5m resolution, patrolling once every 2 minutes, covering the work area). The equipment operating status is read in real time by the chemical production SCADA system (DCS / PLC), which reads valve opening (%), pipeline pressure (kPa), reactor temperature (°C), pump speed (rpm), etc., with a sampling frequency of 1Hz and transmits the data via OPC UA protocol.

[0043] Preprocessing includes: using 3 The criteria for removing outliers are: linear interpolation, spatiotemporal fusion interpolation, or multidimensional interpolation to fill missing values; and the Min-Max method is used for data normalization.

[0044] Specifically, for the collected parameter sequence The statistical calculations are as follows:

[0045] Mean: Standard deviation: ;

[0046] according to The criteria for outlier identification are as follows: That is, satisfy or Data points will be marked as outliers.

[0047] Using time windows ( Dynamic calculation of (sampling points) and , Window size, window sliding step This ensures a balance between real-time performance and stability. For non-stationary sequences (such as CO concentration fluctuations caused by the movement of inspection personnel), an exponentially weighted moving average (EWMA) is used to adjust the window weights. , ,in, It is the mean of time t. It is the standard deviation of time t. It is the observation value at time t. The forgetting factor is set to 0.2 in this embodiment to prioritize recent data trends.

[0048] Interpolation methods include linear interpolation, spatiotemporal fusion interpolation, and multidimensional interpolation. Assuming that in a data sequence... and It refers to two adjacent known data points, with one missing value between them. If the corresponding time position is t, then the formula for calculating the interpolation coefficient k is:

[0049] ,

[0050] The linear interpolation formula for missing values ​​is:

[0051] ,

[0052] For mobile monitoring devices (such as inspection robots), weighted interpolation is performed by combining spatial location information (x, y):

[0053] ,

[0054] in, and These represent the spatial distances from the missing point to the known points before and after it (calculated via GPS), ensuring that the interpolation results conform to physical spatial continuity.

[0055] When multiple parameters are missing simultaneously (such as heart rate, body temperature, and CO concentration), Kriging interpolation is used, taking advantage of the correlation between the parameters: ,in, It is an estimate of the unknown point. These are the values ​​of known data points, and their weighting coefficients. By semi-variogram Calculations reflect the spatial dependence between parameters (such as the strong correlation between heart rate and body temperature under high temperature conditions).

[0056] Then, the Min-Max method was used to process the original data. Mapping to the [0,1] interval, the formula is as follows:

[0057] ,

[0058] in, and These are the minimum and maximum values ​​for historical data (updated monthly).

[0059] Extreme value handling: When new data exceeds the historical range, saturation mapping is used.

[0060] ,

[0061] At the same time, for more than Extreme values ​​are marked as high-risk points and their weight is increased by 20%.

[0062] The coefficient of variation for each parameter is calculated automatically each week. ,right For parameters (such as large fluctuations in dust concentration), an adaptive window is used. Calculate local and This improves the accuracy of normalization.

[0063] Step 2: Perform multimodal fusion on the preprocessed real-time operation data to obtain multimodal feature vectors, and dynamically assign weights to each operation data item.

[0064] In this embodiment, a spatiotemporal feature fusion network model is used for multimodal fusion. This model includes a temporal feature extraction module for extracting temporal features, a spatial feature extraction module for extracting spatial features, and a feature fusion module for feature fusion. Specifically, as shown... Figure 2 As shown, the temporal feature extraction module includes a 1D-CNN unit and an LSTM unit, while the spatial feature extraction module is specifically a 2D-CNN module, using the concat layer of FPN as the feature fusion module.

[0065] For time-series data such as heart rate (HR) and pipeline pressure (P), a 1D-CNN module is used to extract local features, such as the mean, variance, slope, maximum, and minimum values ​​of HR within a 3-second sliding window, generating 5-dimensional features. Then, LSTM is used to capture long-term dependencies, such as the trend changes of HR within 10 minutes, generating 5-dimensional time-series features, for a total of 10 dimensions. For spatial data such as gas concentration (C) and personnel location (X, Y), a 2D-CNN is used to extract spatial distribution features. For example, within the work area, a convolutional network is used to compress the entire CO concentration map into 8 thermodynamic features to describe the overall shape of the gas distribution. This is then combined with the location neighborhood risk of the personnel, such as the maximum CO concentration within 5 m around the personnel (1st dimension); and the distance between the personnel's current location and the center of the region with the highest CO concentration (2nd dimension). These 8 overall features plus 2 location neighborhood risk values ​​constitute 10-dimensional spatial features. Subsequently, the temporal features (10-dimensional) and spatial features (10-dimensional) are concatenated through the concat layer of the FPN to output a 20-dimensional multimodal feature vector. For example, the temporal features of CO concentration (such as the average value over the past 3 minutes) and spatial features (such as the concentration value at the current location) are fused to solve the problems of noise (such as instantaneous data fluctuations caused by equipment vibration) and semantic missingness (such as the impact of personnel location on risk not being considered separately).

[0066] The model's pre-training utilizes a large amount of historical chemical engineering data from various operating conditions, labeled with spatiotemporal features. The temporal branch extracts features using a 1D-CNN (3-second window) and a 128-unit LSTM, while the spatial branch processes a 1m×1m heatmap using a 2D-CNN (3×3 convolutional kernel). End-to-end training minimizes the reconstruction loss, ensuring that the cosine distance between the safety / risk feature vectors is ≥0.6. The preprocessed real-time operating condition data is then input into the spatiotemporal feature fusion network model, which outputs a multimodal feature vector.

[0067] The dynamic weights of real-time operational data items are calculated based on expert ratings, accident frequency caused by parameters, and job type. The calculation formula is as follows: ,in, Each parameter is scored by a chemical safety expert (1-10 points, such as CO concentration score of 9 and heart rate score of 7). : Calculate parameters based on historical accident database The frequency of accidents (normalized to [0,1], e.g., CO concentration accident frequency 0.8, heart rate accident frequency 0.3); : Adjust dynamically according to job type (e.g., gas concentration in confined space operations). Heart rate During routine inspections, gas concentration Heart rate ), Updated hourly using LSTM model (For example, when the work environment is detected to have entered a confined space,) The weight of gas concentration was dynamically adjusted from 0.3 to 0.5 to increase its weighting.

[0068] Step 3: Construct and train a causal model to identify causal relationships based on multimodal feature vectors and the trained causal model.

[0069] In this embodiment, the Temporal-PC algorithm is used to construct a causal model and perform pre-training. Multimodal feature vectors are input into the model to obtain a causal relationship graph and identify causal relationships. The causal model includes: calculating the partial correlation coefficients between parameters such as physiological, environmental, and equipment operating conditions to construct an initial causal graph; filtering strong causal relationships in the initial causal graph based on the partial correlation coefficients; and integrating the causal graphs corresponding to all strong causal relationships to generate a weighted causal graph.

[0070] Specifically, for each parameter pair (a "parameter pair" refers to various interrelated combinations of monitoring parameters in an industrial operation scenario, specifically including the following categories: 1. combinations of physiological and environmental parameters; 2. combinations of environmental parameters; 3. combinations of equipment and environmental parameters; 4. combinations of physiological parameters, such as CO concentration and heart rate), the partial correlation coefficients (PCs) at lags of 1, 2, and 3 are calculated to construct an initial causal graph. For example, the change in CO concentration at time t and the change in heart rate at time t+2 have a strong PC (PC = 0.75), forming a causal edge. Next, Fisher-Z transform was applied to the PC values ​​to test conditional independence (p < 0.01) and screen for strong causal relationships (mutual information MI > 0.6). For example, the MI of CO and HR was 0.85, far exceeding the threshold, confirming a strong causal relationship. Then, causal graphs for different time windows (10 seconds, 30 seconds, 60 seconds) and spatial scales (5 meters, 10 meters, 20 meters) were integrated using GBDT to generate a weighted DAG. ,in, It refers to multiple initial causal graphs constructed for different time windows (10 seconds, 30 seconds, 60 seconds) and spatial scales (5 meters, 10 meters, 20 meters). Each graph presents the parametric causal relationships at a specific spatiotemporal scale in the form of a DAG. Refers to each initial DAG ( The weights corresponding to the graphs are proportional to the F1-score (model evaluation metric, measuring the accuracy of causal relationship identification) of the graph on the validation set, and require F1 ≥ 0.92 (to ensure the reliability of the base graph). For example, a graph with F1 = 0.95 has a higher weight than a graph with F1 = 0.92, meaning that the former's causal relationship is more credible and it will account for a higher proportion in the final fusion. This refers to integrating multiple initial DAGs through weighted summation to form a final weighted DAG, which comprehensively reflects the strong causal relationships between parameters at different spatiotemporal scales, providing interpretable causal evidence for subsequent risk assessment and intervention (such as identifying the critical pathway of "increased CO concentration → abnormal heart rate").

[0071] For the pre-training of the causal model, the partial correlation coefficient (PC) of parameters with lags of 1-3 steps is first calculated to construct an initial causal graph. Strong causal relationships are confirmed using the Fisher-Z test (p<0.01) and mutual information screening (MI>0.6). Then, causal graphs for different time windows (10 / 30 / 60 seconds) and spatial scales (5 / 10 / 20 meters) are fused using a method where the GBDT ensemble weights are proportional to the validation set F1-score (≥0.92). Finally, experts are invited to score the causal paths (≥4 points) through a VR scene, and reasonable paths are stored in the knowledge base to ensure the model accurately captures the causal relationships between chemical parameters, serving risk warning purposes. VR simulation verification specifically involves constructing a VR scene (such as...). (For typical chemical accident scenarios such as leaks and high-temperature dehydration), safety experts are invited to wear VR devices and score the reasonableness of causal paths (≥4 points, out of 5 points). For example, experts verify the reasonableness of the causal path "high temperature + high humidity → dehydration risk" and store it in the risk communication knowledge base for subsequent risk assessment.

[0072] Step 4: Calculate the hazard index based on the multimodal feature vector and the dynamic weights corresponding to each work condition data item. Train the LSTM model using the multimodal feature vector and the hazard index, and use the trained LSTM model to predict the future hazard index.

[0073] First, the risk index is calculated based on dynamic weights. , , The value of parameter i after normalization. and , The value range is [0,1].

[0074] Using the multimodal feature vectors from the past 10 minutes of historical chemical engineering data as input, and the hazard index for the next 5 minutes as the prediction target, the calculated hazard index serves as the supervision target. A composite loss function combining mean squared error (MSE) and attention mechanism weights is employed. A bidirectional LSTM network is used to capture temporal dependencies, and the Adam optimizer is used to iteratively update the model parameters until the prediction accuracy on the test set reaches over 85%. The trained LSTM model is then used to predict future hazard indices.

[0075] Step 5: Determine the risk level based on the future hazard index and early warning threshold. In this embodiment, three threshold levels are set, as shown in the table below:

[0076] Table 1 Risk Level Classification Table

[0077]

[0078] Step 6: Implement graded safety control and intervention measures based on risk level and causal relationship.

[0079] This embodiment adopts tiered intervention measures to address different risk levels, specifically as follows:

[0080] Level I Intervention (Local Early Warning):

[0081] Wearable devices:

[0082] Wristband: 50Hz vibration (0.5g, 3 seconds) + yellow light flashing (500cd / m²), voice prompt "Slight environmental risk, check equipment" (60dB).

[0083] Data synchronization: Risk parameters (including GPS location, error ≤ 5m) are sent to the central control every 2 seconds.

[0084] Central control room:

[0085] Large screen: Marks personnel location and displays risk parameter trends (such as CO concentration 25ppm, heart rate 110bpm, equipment pressure 110%).

[0086] Generate a report and push it to safety management personnel (mobile APP / PC, including similar historical cases, such as "Inspector A suffered mild poisoning in a similar environment on June 15, 2023").

[0087] Optimization: Job-specific adaptation. Based on the real-time job type of the operator, the system adjusts the weight of various monitoring parameters (such as equipment pressure, gas concentration, etc.) in risk assessment, making risk identification more aligned with the actual safety requirements of the current job. For example, when an operator is in a "maintenance" position, the system will increase the weight of the "equipment pressure" parameter (increase by 0.2) while decreasing the weight of the "gas concentration" parameter (decrease by 0.2). This is because in maintenance work, the operating status of equipment (such as abnormal pressure) has a more direct impact on safety (e.g., equipment failure may lead to mechanical injury, leakage, etc.), therefore, priority should be given to the risk of equipment failure, rather than focusing on environmental parameters such as gas concentration.

[0088] Level II intervention (environmental control + equipment management):

[0089] Environmental control:

[0090] Ventilation: Start local fans (wind speed 3m / s, coverage 5m, CO concentration ≤20ppm). ≥19.5%, response time ≤10 seconds).

[0091] Cooling: Evaporative cooler (temperature ≤30℃, humidity ≤70%RH, response time ≤15 seconds).

[0092] Equipment Management and Control:

[0093] Lock-up: Valves ≤50%, pump speed ≤80% (via SCADA system, delay ≤5 seconds), sampling frequency of critical equipment (such as reactors) increased to 10Hz.

[0094] Personal protection:

[0095] AR safety helmet: Displays "Wearing gas mask (model: XX), moving towards the safe zone (distance 30m, coordinates XX.XXX, YY.YYY), estimated time 15 seconds".

[0096] Multi-platform synchronization: simultaneous warnings from central control, AR, mobile APP, and on-site sound and light (80dB, 5 seconds), with a reach rate of ≥95% (MEC edge computing, latency ≤50ms).

[0097] Level III Intervention (Emergency Evacuation + System Coordination):

[0098] Emergency evacuation:

[0099] Wearable device: 100Hz vibration (1.0g, 10 seconds) + 100dB alarm (2000Hz, penetrates 80dB noise), voice "Evacuate immediately!".

[0100] AR: AI plans the shortest path (avoiding toxic gas, high temperature, and leaks, with a length of ≤50m, an update frequency of 1 second, and based on real-time environmental data).

[0101] System linkage:

[0102] Equipment: Emergency cooling of reactor (≤5℃ / min, target ≤ set value +10℃), depressurization of storage tank (≥50% opening, ≤10kPa / min), shutdown of non-essential power supply (delay ≤3 seconds).

[0103] Firefighting: Sprinkler system (0.5MPa, covering 100m², 10 minutes, response ≤5 seconds) to dilute toxic gas and cool down.

[0104] Emergency: Send a distress signal via 5G (including location, risk type, number of people, response time ≤ 3 seconds, via URLLC slicing).

[0105] Optimization: Intelligent shutdown (based on causal models, such as "reactor overheating → CO increase" strong causal relationship, prioritize shutting down the reactor to reduce secondary disasters).

[0106] First, an intelligent shutdown protocol is constructed, relying on causal models to identify strong causal relationships (such as "reactor overheating → CO increase"), prioritizing the shutdown of high-risk equipment (such as reactors rather than pumps). Using a pre-trained decision tree model, the shutdown decision delay is ≤2 seconds, with an accuracy rate ≥95%, effectively reducing secondary disasters such as explosions and leaks. Second, multi-departmental collaboration is achieved: the fire department uses the park's GIS to plan the optimal route, arriving at the scene within 5 minutes to avoid congestion; the medical department automatically assesses injuries based on physiological parameters such as heart rate, blood oxygen, and body temperature (such as poisoning level and heatstroke severity), preparing emergency equipment such as CPR devices, oxygen masks, and antidotes in advance; the environmental protection department monitors leak diffusion through multispectral imaging by drones (updated every 2 minutes, accuracy ≤10m), initiating environmental purification measures such as activated carbon adsorption and spray absorption to prevent pollution spread. These measures optimize emergency response from multiple dimensions, including equipment management, rescue response, and pollution control, improving the timeliness, accuracy, and safety of chemical accident handling, reducing disaster losses, and providing comprehensive protection for chemical operation safety. Data flow is as follows: Figure 3 As shown.

[0107] Furthermore, it also includes: calculating the intervention effectiveness, and dynamically adjusting the preset warning threshold based on the intervention effectiveness using a reinforcement learning algorithm.

[0108] In one embodiment of this application, the intervention efficacy is calculated after the intervention, and the calculation formula is as follows: ,in, This represents the actual number of accidents that were successfully avoided. This is the total number of times the warning has been triggered. This represents the pass rate for manual review.

[0109] Set intervention goals as intervention effectiveness At the same time, the false alarm rate should be controlled within Within this range. To enable the system to adapt to various changes in chemical operating conditions, such as seasonal changes and equipment aging, the DQN algorithm or PPO algorithm from reinforcement learning is used to dynamically adjust the early warning threshold. , , Whenever the intervention effectiveness E increases by 1%, the system will give a +1 reward. In this way, the threshold is continuously optimized to ensure that the system can efficiently carry out risk warning and intervention under different working conditions, thereby realizing closed-loop feedback optimization of risk prediction and intervention, and further improving the overall safety performance of chemical operations.

[0110] Correspondingly, this application also provides a real-time risk identification and proactive safety control system for industrial workers, including:

[0111] Data acquisition module: used to collect real-time operational data, including physiological data of workers, environmental data, and equipment operational data;

[0112] Data processing module: preprocesses the collected real-time operation condition data, performs multimodal fusion on the preprocessed real-time operation condition data to obtain multimodal feature vectors, and dynamically assigns weights to each operation condition data item.

[0113] Causal relationship determination module: Constructs and trains a causal model, and identifies causal relationships based on multimodal feature vectors and the trained causal model;

[0114] Hazard index prediction module: The hazard index is calculated based on the multimodal feature vector and the dynamic weights corresponding to each work condition data item. The LSTM model is trained using the multimodal feature vector and the hazard index, and the trained LSTM model is used to predict the future hazard index.

[0115] The risk level determination module determines the risk level based on the future hazard index and early warning threshold.

[0116] Safety control and intervention module: Implements graded safety control and intervention measures based on risk level and causal relationship.

[0117] Furthermore, it also includes an intervention effectiveness calculation module: calculating intervention effectiveness and dynamically adjusting the warning threshold based on the intervention effectiveness using a reinforcement learning algorithm.

Claims

1. A method for real-time risk identification and active safety control of industrial workers, characterized in that, Comprise: Step 1, collecting real-time operation condition data and preprocessing, the operation condition data includes physiological data of operation personnel, environmental data and equipment condition data; Step 2, multi-modal fusion is carried out on the preprocessed real-time operation condition data to obtain a multi-modal feature vector, and dynamic weight distribution is carried out on each operation condition data item; Step 3, build causal model and train, identify causal relationship based on multi-modal feature vector and trained causal model; Step 4, calculate the risk index based on the multi-modal feature vector and the dynamic weight corresponding to each operation condition data item, train the LSTM model using the multi-modal feature vector and the risk index, and predict the future risk index using the trained LSTM model; Step 5, determine the risk level based on the future risk index and the warning threshold; Step 6, take hierarchical safety control intervention measures based on the risk level and the causal relationship.

2. The industrial worker real-time risk identification and active safety control method according to claim 1, characterized in that, The physiological data of the operation personnel includes body temperature, heart rate, blood oxygen, respiratory rate, blood pressure and position information, the environmental data includes toxic gas concentration, temperature and humidity, dust, noise, VOC and leaked gas infrared characteristics, and the equipment condition data includes valve opening, pipeline pressure, reactor temperature and pump speed.

3. The industrial worker real-time risk identification and active safety control method according to claim 1, characterized in that, The preprocessing includes: using 3 criteria to remove outliers; using linear interpolation, spatio-temporal fusion interpolation or multi-dimensional interpolation to fill in missing values; using Min-Max method for data normalization.

4. The industrial worker real-time risk identification and active safety control method according to claim 1, characterized in that, Step 2 specifically comprises: Building a spatio-temporal feature fusion network model, including a time series feature extraction module for extracting time series features, a spatial feature extraction module for extracting spatial features, and a feature fusion module for feature fusion; Input the preprocessed real-time operation condition data into the spatio-temporal feature fusion network model for multi-modal fusion; The weight calculation formula is: wherein, is the risk level score of each parameter by a chemical safety expert, is a data item calculated based on a historical accident database the frequency of causing accidents, is dynamically adjusted according to the post type, .

5. The industrial worker real-time risk identification and active safety control method according to claim 1, characterized in that, Temporal-PC algorithm is used to build a causal model, including calculating the partial correlation coefficient of the parameter pair, building an initial causal graph, screening strong causal relationships in the initial causal graph based on the partial correlation coefficient, integrating all strong causal relationships corresponding to the causal graph to generate a weighted causal graph, and the parameter pair includes the combination of operation personnel physiological data and environmental data, the combination of environmental data, the combination of equipment condition data and environmental data, and the combination of operation personnel physiological data.

6. The industrial worker real-time risk identification and active safety control method according to claim 5, characterized in that, The causal model is verified by VR scene, specifically including: building a chemical accident VR scene, inviting safety experts to score the rationality of the causal path, and retaining the causal path with a score higher than a preset threshold.

7. The industrial worker real-time risk identification and active safety control method according to claim 1, characterized in that, Risk index The calculation is: , is the data item the normalized value, is the weight of the data item i, and N is the dimension of the multi-modal feature vector, and .

8. The industrial worker real-time risk identification and active safety control method according to any one of claims 1-7, characterized in that, Further comprising: Calculate the intervention effectiveness, and dynamically adjust the warning threshold based on the intervention effectiveness using reinforcement learning algorithm; The intervention effectiveness calculation method is: Wherein, represents the number of actual successful avoidance accidents, is the total number of early warning triggers, is the pass rate of manual review.

9. An industrial worker real-time risk identification and active safety control system, characterized in that, Comprise: Data acquisition module: used for collecting real-time operation condition data, the operation condition data includes physiological data of operation personnel, environmental data and equipment condition data; Data processing module: pre-process the collected real-time operation condition data, multi-modal fusion is carried out on the preprocessed real-time operation condition data to obtain a multi-modal feature vector, and dynamic weight distribution is carried out on each operation condition data item; Causal relationship determination module: build causal model and train, identify causal relationship based on multi-modal feature vector and trained causal model; Risk index prediction module: calculate the risk index based on the multi-modal feature vector and the dynamic weight corresponding to each operation condition data item, train the LSTM model using the multi-modal feature vector and the risk index, and predict the future risk index using the trained LSTM model; The risk level determination module determines the risk level based on the future danger index and the early warning threshold; The safety control intervention module takes hierarchical safety control intervention measures based on the risk level and the causal relationship.

10. The industrial worker real-time risk identification and active safety control system, according to claim 9, wherein, The intervention effectiveness calculation module calculates the intervention effectiveness, and dynamically adjusts the early warning threshold based on the intervention effectiveness using a reinforcement learning algorithm.

Images