Industrial safety production risk perception method and system based on LLM intelligent agent

Abstract

This invention discloses an industrial safety production risk perception method and system based on an LLM (Limited Least Mechanism) agent, relating to the field of industrial data processing technology. The method includes: during tank leak repair, predicting the set of hazardous gas concentrations in area A within a repair time window based on historical hazardous gas concentration monitoring records; acquiring images of protective equipment worn by personnel going to area A for repair within the preset repair time window, performing wear feature recognition to obtain the fit, and obtaining the filter canister model and effective pressure difference accumulation duration; using an abnormal risk perception engine trained on an LLM agent, deriving the predicted probability of abnormal protection risk based on the predicted hazardous gas concentration set, fit, filter canister model, and effective pressure difference accumulation duration; and issuing a protective risk warning to personnel if the risk exceeds a preset threshold. This invention effectively improves the accuracy and reliability of industrial safety production risk perception.

Description

Technical Field

[0001] This invention relates to the field of industrial data processing technology, specifically to an industrial safety production risk perception method and system based on LLM agents. Background Technology

[0002] With the increasing demands for industrial safety in industries such as petrochemicals, the pressure to prevent and control safety risks during tank leak repair operations is growing. Safety accidents caused by hazardous gas leaks seriously threaten the personal safety of maintenance personnel, highlighting the growing importance of industrial safety risk perception technology. Existing technologies include personnel safety risk perception solutions based on visual recognition and sensor monitoring, used for preliminary control of basic safety matters such as the wearing of protective equipment by maintenance personnel.

[0003] However, traditional vision-based personnel risk perception methods can only identify whether workers are wearing respiratory protective equipment, but fail to accurately identify the actual protective effect of respiratory protective equipment in combination with the current work scenario. This results in insufficient accuracy and reliability of risk perception, making it difficult to meet the refined and precise safety protection and management needs in high-risk scenarios such as tank leak repair. Summary of the Invention

[0004] This invention provides a method and system for industrial safety production risk perception based on LLM agents, aiming to solve the technical problem of insufficient accuracy and reliability of industrial safety production risk perception in the prior art.

[0005] In view of the above problems, the present invention provides an industrial safety production risk perception method and system based on LLM intelligent agents.

[0006] In a first aspect, the present invention provides an industrial safety production risk perception method based on LLM agents, including: When carrying out tank leak repairs, the predicted hazardous gas concentration set of area A within the preset repair time window is obtained based on historical hazardous gas concentration monitoring records. The protective equipment images of the personnel going to the A area for maintenance within the preset maintenance time window are captured by on-site monitoring video. Wearing features are identified in the protective equipment images to obtain the fit. The filter canister model and effective differential pressure accumulation time are also obtained by monitoring the filter canister through sensors. Using an abnormal risk perception engine trained on an LLM agent, the probability of predicted abnormal protection risk is inferred based on the predicted harmful gas concentration set, wearing fit, filter canister model and effective pressure difference accumulation time. If the predicted probability of abnormal protection risk exceeds the preset risk probability threshold, a protection risk warning will be issued to the staff.

[0007] Secondly, this invention provides an industrial safety production risk perception system based on LLM agents, comprising: The environmental risk prediction module is used to predict the set of hazardous gas concentrations in area A within a preset maintenance time window based on historical hazardous gas concentration monitoring records during tank leak repairs. The protective status monitoring module is used to collect images of protective equipment worn by personnel going to the A area for maintenance within the preset maintenance time window through on-site monitoring video, and to perform wearing feature recognition on the protective equipment wearing images to obtain the wearing tightness, and to obtain the filter canister model and effective differential pressure accumulation time through sensor monitoring. The intelligent risk perception module is used to infer the probability of abnormal protection risk based on the predicted harmful gas concentration set, wearing fit, filter canister model and effective pressure difference accumulation time by using the abnormal risk perception engine trained based on LLM agent. The dynamic early warning response module is used to issue a protective risk warning to the staff if the predicted probability of abnormal protection risk exceeds a preset risk probability threshold.

[0008] One or more technical solutions provided in this invention have at least the following technical effects or advantages: This invention provides an industrial safety production risk perception method and system based on LLM agents. In the scenario of tank leakage repair, the system first predicts the gas concentration in area A within a preset repair time window based on historical hazardous gas concentration monitoring records, thus gaining an early understanding of the gas environment risk in the work area and providing accurate data support for subsequent protective risk assessment. Next, it collects and identifies the characteristics of protective equipment worn by workers through on-site monitoring video to obtain the fit, and simultaneously monitors the filter canister model and effective differential pressure accumulation time using sensors to comprehensively capture the actual protective status of personnel. Then, using an abnormal risk perception engine trained on an LLM agent, it derives the probability of abnormal protective risks that fits the actual situation on-site, achieving a scientific assessment of protective risks. Finally, it judges based on preset risk probability thresholds, and promptly issues protective risk warnings to maintenance workers when the risk probability exceeds the threshold. This effectively improves the accuracy and reliability of industrial safety production risk perception, enabling timely and effective control of protective risks in high-risk operation scenarios, and comprehensively ensuring the safety of personnel and operations during tank leakage repair. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 A flowchart illustrating the industrial safety production risk perception method based on LLM agents provided in this embodiment of the invention; Figure 2 A schematic diagram of the structure of an industrial safety production risk perception system based on LLM agents provided in an embodiment of the present invention; The components represented by each number in the attached diagram are explained below: Environmental risk prediction module 11, protection status monitoring module 12, intelligent risk perception module 13, dynamic early warning response module 14. Detailed Implementation

[0011] This invention provides an industrial safety production risk perception method and system based on LLM agents, which is used to address the technical problem of insufficient accuracy and reliability of industrial safety production risk perception in the prior art.

[0012] Example 1, as Figure 1 As shown, this invention provides an industrial safety production risk perception method based on LLM agents, the method comprising: S100: When carrying out tank leak repairs, predict the set of hazardous gas concentrations in area A within a preset repair time window based on historical hazardous gas concentration monitoring records.

[0013] In this embodiment of the invention, during tank leak repair, the predicted hazardous gas concentration set for area A within a preset repair time window is obtained based on historical hazardous gas concentration monitoring records. During tank leak repair, hazardous gases exhibit spatial diffusion and dynamic concentration changes. The distribution of hazardous gas concentrations in the work area cannot be fully reflected by data from a single monitoring point, and the gas concentration continuously changes with the leak situation and environmental factors during the repair process. Relying solely on single-point historical monitoring records for direct prediction fails to reflect the overall distribution characteristics of hazardous gases within the plant area. Furthermore, traditional prediction methods do not dynamically adapt to the gas's hazard characteristics and spread patterns, easily leading to significant deviations between predicted and actual concentrations. Subsequent risk inference for protection requires accurate and comprehensive hazardous gas concentration data for the repair area. Therefore, it is necessary to achieve accurate prediction of hazardous gas concentrations within the preset repair time window for area A through multi-monitoring information integration, spatial correlation interpolation, construction of a multi-branch predictor using a long short-term memory network, and dynamic branch adaptation, providing reliable basic data support for subsequent risk assessment.

[0014] Step S100 in the method provided in this embodiment of the invention includes: The system monitors and obtains several hazardous gas concentration sequence sets from several monitoring points in a chemical plant within a historical time zone. Based on the location coordinates of these monitoring points, it integrates information and performs interpolation estimation based on spatial correlation to construct a hazardous gas concentration distribution sequence set. A harmful gas spread predictor was constructed based on a long short-term memory network. Using the harmful gas spread predictor, the predicted harmful gas concentration set of area A within a preset maintenance time window is obtained based on the harmful gas concentration distribution sequence set.

[0015] First, several hazardous gas concentration sequences from multiple monitoring points within a historical time zone are acquired. Based on the location coordinates of these monitoring points, information is integrated and spatial correlation-based interpolation estimation is performed to construct a hazardous gas concentration distribution sequence set. This hazardous gas concentration sequence set refers to a dataset formed by arranging the concentration data of various hazardous gases monitored by a single monitoring point at different time points within a historical time zone in chronological order, including gas type, monitoring time, and concentration value. Spatial correlation interpolation estimation utilizes the basic assumption that spatially adjacent or nearby points typically have similar concentration values. Using the measured concentrations at monitoring points as known input, and combining the spatial relationships of each point, algorithms such as Kriging interpolation and inverse distance weighting are used to calculate the virtual location concentration values ​​for key areas within the plant where no monitoring points are located, achieving coverage of hazardous gas concentrations from point to area. The hazardous gas concentration distribution sequence set is a dataset formed by integrating the concentration data of various hazardous gases from all monitoring points and the interpolated virtual locations within the plant at different time points within a historical time zone, arranged by spatial location and time order. This dataset comprehensively reflects the overall hazardous gas concentration distribution of the plant at different historical moments.

[0016] Specifically, continuous monitoring is conducted at several pre-set monitoring points in and around the chemical plant's tank area to obtain hazardous gas concentration data for each monitoring point within a historical time zone. This data is then organized chronologically into several hazardous gas concentration sequence sets, with each monitoring point corresponding to one hazardous gas concentration sequence set. The actual geographic coordinates of all monitoring points are collected, and the spatial location information of the aforementioned hazardous gas concentration sequence sets is integrated to form a concentration dataset with location labels. For each specific moment within the historical time zone, a spatial correlation-based interpolation estimation algorithm, such as Kriging interpolation, is used as input to calculate the estimated concentration values ​​of all virtual locations within the key area of ​​the tank area, using the measured concentrations of all monitoring points at that moment as input. The measured concentrations of the monitoring points at each moment and the estimated concentrations of the virtual locations are then integrated according to spatial location and arranged chronologically to construct a hazardous gas concentration distribution sequence set.

[0017] For example, 10 monitoring points, numbered M1 to M10, were set up around the No. 1 storage tank area of ​​a petrochemical plant. The historical time zone was taken from 0:00 to 24:00 every day for the past 6 months, and the monitored gas types were H2S, CO, and VOCs. First, the H2S concentration at point M1 at 8:00 every day for the past 6 months was obtained: 1.2 mg / m³. 3 1.1 mg / m 3 1.3 mg / m 3 ...CO concentration: 0.8 mg / m³ 3 0.7 mg / m 3 ...VOCs concentration: 3.5 mg / m³ 3 3.4 mg / m 3 ...and organize them into a hazardous gas concentration sequence set for M1; similarly, obtain the sequence sets for each point from M2 to M10. Collect the geographic coordinates of M1 to M10 and add location labels to each sequence set. For 8:00 AM on June 1st, the measured H2S concentrations of M1 to M10 were 1.0-1.5 mg / m³. 3 Using the Kriging interpolation algorithm, the estimated H2S concentrations at 20 virtual locations, such as between M1 and M2, and between M3 and area A, were calculated. Area A is the specific area within Tank Area No. 1 designated for planned maintenance. Similarly, the estimated CO and VOCs concentrations at all virtual locations at that time were calculated. The measured concentrations at all monitoring points and the estimated concentrations at virtual locations at 8:00 AM on June 1st were integrated by coordinates. Then, the concentrations at all times over the past 6 months were integrated sequentially. After arranging them in chronological order, the hazardous gas concentration distribution sequence set of Tank Area No. 1 of the petrochemical plant was obtained.

[0018] Secondly, a predictor of the spread of harmful gases is constructed based on a long short-term memory network.

[0019] Among them, the harmful gas spread predictor built based on long short-term memory network includes: Based on the historical tank leakage and maintenance records of the chemical plant, several sample hazardous gas concentration distribution sequence sets were collected, and the historical maximum hazardous gas concentration distribution of different sample hazardous gas concentration distribution sequence sets within the historical maintenance time window was obtained as the sample predicted hazardous gas concentration distribution, thus obtaining several sample predicted hazardous gas concentration distributions. The aforementioned sample harmful gas concentration distribution sequence set and the sample predicted harmful gas concentration distribution are used as training data, and K-fold cross-partitioning is performed to obtain K training sets, where K is greater than or equal to 10. The long short-term memory network is trained to convergence using the K training sets, generating K branches for predicting the spread of harmful gases. These branches are then combined to obtain a predictor for the spread of harmful gases.

[0020] First, based on historical tank leak maintenance records of the chemical plant, several sample hazardous gas concentration distribution sequence sets were collected. The historical maximum hazardous gas concentration distribution within the historical maintenance time window was then obtained for each sample hazardous gas concentration distribution sequence set as the sample predicted hazardous gas concentration distribution, resulting in several sample predicted hazardous gas concentration distributions. The sample hazardous gas concentration distribution sequence set refers to the hazardous gas concentration distribution sequence set extracted from the historical tank leak maintenance records of the chemical plant, providing basic samples for predictor training. The sample predicted hazardous gas concentration distribution refers to the dataset formed by integrating the maximum hazardous gas concentration values ​​at various spatial locations within the historical maintenance time window from a single sample hazardous gas concentration distribution sequence set, reflecting the peak gas concentration distribution within the region during the maintenance process under that sample scenario.

[0021] Specifically, historical tank leakage maintenance records of the chemical plant are retrieved, and records similar to the current maintenance scenario in Area A are selected, such as those with the same tank type, medium, and leakage type. Several sample hazardous gas concentration distribution sequence sets are extracted from these records. For each sample hazardous gas concentration distribution sequence set, its corresponding historical maintenance time window is determined, and the historical maximum concentration values ​​of various hazardous gases at each spatial location within that window are extracted. The historical maximum concentration values ​​at each spatial location are integrated according to location to obtain the sample predicted hazardous gas concentration distribution for each sample, ultimately forming several sample predicted hazardous gas concentration distributions.

[0022] For example, from the nearly 10 years of tank leakage maintenance records of the aforementioned petrochemical plant, 2000 maintenance records of similar media leakage in Tank No. 1 were selected. The corresponding hazardous gas concentration distribution sequence set for each record was extracted, resulting in 2000 sample hazardous gas concentration distribution sequence sets. The historical maintenance time window for each record was 4 hours. For the first sample, the maximum H2S concentration of 2.5 mg / m³ at point M1 within the 4-hour maintenance window was extracted. 3 The maximum H2S concentration at the virtual point in region A is 3.0 mg / m³. 3 The maximum concentration values ​​of CO and VOCs at each location are integrated according to location to obtain the predicted hazardous gas concentration distribution of the first sample; similarly, the processing of 1000 samples is completed to obtain the predicted hazardous gas concentration distribution of 2000 samples.

[0023] Secondly, the aforementioned sample hazardous gas concentration distribution sequence sets and the corresponding sample predicted hazardous gas concentration distributions are used as training data, and K-fold cross-partitioning is performed to obtain K training sets, where K is greater than or equal to 10. K-fold cross-partitioning refers to randomly dividing all training data into K equal parts, randomly selecting one part with replacement K times, each time selecting one part as the validation set and the rest as the training set. After iterating K times, K independent training sets are constructed, where K ≥ 10. The training sets are datasets formed by pairing the sample hazardous gas concentration distribution sequence sets with the corresponding sample predicted hazardous gas concentration distributions one-to-one, used to train the Long Short-Term Memory network, where the hazardous gas concentration distribution sequence sets are the input features, and the corresponding sample predicted hazardous gas concentration distributions are the output labels.

[0024] Specifically, several sample harmful gas concentration distribution sequence sets are paired one-to-one with the corresponding sample predicted harmful gas concentration distributions to form complete training data; a K value is set (K≥10), and the above training data is randomly divided into K-fold cross-partitions with replacement, dividing the training data into K parts; each part of the data is selected as the validation set in turn, and the remaining K-1 parts are used as the training set. After K iterations, K training sets are obtained.

[0025] For example, the above 2000 sample harmful gas concentration distribution sequence sets are paired one-to-one with the 2000 sample predicted harmful gas concentration distribution sets to form 2000 sets of training data; K=10 is set, and the 2000 sets of data are randomly divided into 10 equal parts, with 200 sets in each part; the first part is selected as the validation set, and the second to tenth parts are selected as the training set to obtain the first training set; the second part is selected as the validation set, and the first and third to tenth parts are selected as the training set to obtain the second training set; and so on, after 10 iterations, 10 training sets are obtained, each training set containing 1800 sets of training data and 200 sets of validation data.

[0026] Based on this, the Long Short-Term Memory (LSTM) networks are trained to convergence using the K training sets, generating K branches for predicting the spread of hazardous gases. These branches are then combined to obtain a hazardous gas spread predictor. LSTM is a special type of recurrent neural network that can effectively capture long-term dependencies in time-series data. Each hazardous gas spread prediction branch refers to an LSTM network trained to convergence on a single training set, capable of independently predicting the concentration distribution of hazardous gases. Different branches have different prediction focuses due to variations in training data. The hazardous gas spread predictor is a multi-branch prediction model formed by combining the K hazardous gas spread prediction branches. The accuracy and stability of the results can be improved through multi-branch collaborative prediction.

[0027] Specifically, a basic Long Short-Term Memory (LSTM) network model is constructed. The input layer dimension is adapted to the feature set of harmful gas concentration distribution sequences. The hidden layer has two layers with 128 neurons each. The LSTM layer uses the tanh activation function. The output layer dimension is adapted to the sample predicted harmful gas concentration distribution and uses a linear activation function. Hyperparameters such as the number of iterations and learning rate are set. The mean squared error (MSE) is selected as the loss function. The convergence condition is that the MSE of the model validation set ≤ 1 × 10⁻⁶. -4 Furthermore, the prediction error tends to stabilize after 30 consecutive iterations without significant decrease; K training sets are input into the constructed LSTM model, and each model is trained sequentially until the convergence condition is met, generating K independent branches for predicting the spread of harmful gases; the K branches for predicting the spread of harmful gases are combined to construct a predictor for the spread of harmful gases that can realize simultaneous calling of multiple branches and integration of results.

[0028] For example, an LSTM model with two hidden layers and 128 neurons in each layer is constructed. The input layer dimension is adapted to the gas concentration sequence features of storage tank area No. 1, and the LSTM layer uses the tanh activation function. The output layer dimension is adapted to the concentration distribution prediction of region A and uses a linear activation function. The number of iterations is set to 500, the learning rate is 0.001, the loss function is the mean squared error (MSE), and the convergence condition is that the validation set MSE ≤ 1 × 10⁻⁶. -4 Furthermore, the prediction error tends to stabilize after 30 consecutive rounds. The above 10 training sets are sequentially input into the LSTM model and trained until convergence, generating 10 branches for predicting the spread of hazardous gases, K=10, numbered F1~F10. Combining F1~F10 yields the hazardous gas spread predictor for the No. 1 storage tank area of ​​the petrochemical plant.

[0029] Furthermore, using the harmful gas spread predictor, the predicted harmful gas concentration set of region A within a preset maintenance time window is predicted based on the harmful gas concentration distribution sequence set.

[0030] Specifically, the hazardous gas spread predictor uses the hazardous gas concentration distribution sequence set to predict the set of hazardous gas concentrations in area A within a preset maintenance time window, including: Extract all hazardous gas types from the hazardous gas concentration distribution sequence set, and perform a weighted evaluation according to a preset gas hazard level to determine the hazardous gas hazard coefficient; Calculate the concentration spread and growth rate of all hazardous gas types in the hazardous gas concentration distribution sequence set, and determine the hazardous gas spread and growth coefficient by weighting according to the preset gas hazard level; The ratio of the hazard coefficient of the harmful gas to the preset hazard coefficient of the gas is used as the first prediction complexity; The ratio of the harmful gas spread growth coefficient to the preset gas spread growth coefficient is used as the second prediction complexity. The combined prediction complexity is obtained by weighted summation of the first prediction complexity and the second prediction complexity. The product of the comprehensive prediction complexity and the initial prediction branch selection number P is rounded down to obtain the adaptive prediction branch selection number J, where P is 4. If the calculated J is less than 2, then J is equal to 2; if the calculated J is greater than K, then J is equal to K. J prediction branches are randomly selected from the K branches for predicting the spread of harmful gases. Each branch is used to make predictions based on the set of harmful gas concentration distribution sequences. The J prediction results are then averaged to obtain the set of predicted harmful gas concentrations in region A within a preset maintenance time window.

[0031] First, all hazardous gas types are extracted from the hazardous gas concentration distribution sequence set, and a weighted evaluation is performed according to a preset gas hazard level to determine the hazardous gas hazard coefficient. The hazardous gas hazard coefficient is a quantitative coefficient obtained by weighted calculation, comprehensively considering the types and hazard levels of various gases in the hazardous gas concentration distribution sequence set. It reflects the overall hazard level of hazardous gases in the area; the higher the hazardous gas hazard coefficient value, the higher the hazard level. The preset gas hazard level refers to a pre-set weight value based on the toxicity, explosion limits, irritant properties, and other characteristics of various hazardous gases. Gases with higher hazard levels have higher weight values. From the hazardous gas concentration distribution sequence set, all hazardous gas types are extracted, and duplicate types are removed to form a gas type set. The preset gas hazard level is retrieved, and a corresponding weight value is assigned to each gas type. The average concentration of each gas in the concentration distribution sequence set is calculated, multiplied by the corresponding preset gas hazard level, and then the calculation results for all gases are summed to obtain the hazardous gas hazard coefficient.

[0032] For example, from the set of hazardous gas concentration distribution sequences, the gas types H2S, CO, and VOCs are extracted, with preset gas hazard levels: H2S hazard level = 0.4, CO hazard level = 0.3, and VOCs hazard level = 0.3; the calculated average H2S concentration in this area over the past 6 months is 1.2 mg / m³. 3 The average CO concentration was 0.8 mg / m³. 3 The average concentration of VOCs was 3.5 mg / m³. 3 The hazard coefficient of harmful gases is 1.2×0.4+0.8×0.3+3.5×0.3=1.77.

[0033] Secondly, the concentration spread and growth rate of all hazardous gas types in the hazardous gas concentration distribution sequence is calculated, and a weighted assessment based on a preset gas hazard level is performed to determine the hazardous gas spread and growth coefficient. The hazardous gas spread and growth rate refers to the average concentration increase of a certain type of hazardous gas per unit time within the concentration distribution sequence, reflecting the speed of gas diffusion and spread. The hazardous gas spread and growth coefficient is a quantitative coefficient obtained through weighted calculation, comprehensively considering the spread and growth rate and hazard level of various hazardous gases. It reflects the overall spread trend of hazardous gases within the area; the larger the hazardous gas spread and growth coefficient value, the faster the spread and the higher the risk.

[0034] For example, the concentration spread rate of each type of hazardous gas in the concentration distribution sequence set is calculated, which is the average of the concentration differences between adjacent time points. The spread rate of each gas is multiplied by the corresponding preset gas hazard level, and then the calculated results for all gases are summed to obtain the hazardous gas spread rate coefficient. For example, the calculated H2S spread rate in the No. 1 storage tank area of ​​the aforementioned petrochemical plant is 0.15 mg / m³. 3 h、CO0.10mg / m 3 •h, VOCs 0.20mg / m³ 3 •h; The coefficient for the spread and growth of harmful gases = 0.15 × 0.4 + 0.10 × 0.3 + 0.20 × 0.3 = 0.15.

[0035] Furthermore, the ratio of the hazardous gas hazard coefficient to the preset gas hazard coefficient is used as the first prediction complexity. The first prediction complexity, equal to the ratio of the hazardous gas hazard coefficient to the preset gas hazard coefficient, is a quantitative indicator reflecting the impact of the hazardous gas hazard level on the prediction difficulty. The larger the ratio, the higher the degree of gas hazard, and the more precise the prediction needs to be. The preset gas hazard coefficient refers to a fixed constant, a baseline value for the hazardous gas hazard coefficient set in advance based on historical maintenance data of the chemical plant and industry standards. The calculation formula is: First prediction complexity = Hazardous gas hazard coefficient / Preset gas hazard coefficient. For example, if the preset gas hazard coefficient of this petrochemical plant is 1.5, the first prediction complexity = 1.77 / 1.5 = 1.18.

[0036] Furthermore, the ratio of the harmful gas spread growth coefficient to the preset gas spread growth coefficient is used as the second prediction complexity. The second prediction complexity, equal to the ratio of the harmful gas spread growth coefficient to the preset gas spread growth coefficient, is a quantitative indicator reflecting the impact of the harmful gas spread trend on prediction difficulty. A larger ratio indicates a faster gas spread rate and higher prediction difficulty. The preset gas spread growth coefficient is a fixed constant, a benchmark value for the harmful gas spread growth coefficient set in advance based on historical maintenance data of the chemical plant and industry standards. The calculation formula is: Second prediction complexity = Harmful gas spread growth coefficient / Preset gas spread growth coefficient. For example, if the preset gas spread growth coefficient for this petrochemical plant is 0.12, the second prediction complexity = 0.15 / 0.12 = 1.25.

[0037] Based on this, the first prediction complexity and the second prediction complexity are weighted and summed to obtain the comprehensive prediction complexity. The comprehensive prediction complexity is a quantitative index obtained by weighting and summing the first and second prediction complexities according to preset weights. It comprehensively reflects the overall impact of the degree of hazard and the spread trend of harmful gases on the prediction difficulty. The higher the comprehensive prediction complexity value, the higher the prediction difficulty, and the more prediction branches need to be invoked. Preset weights are set for the first and second prediction complexities, with a preset weight sum of 1. The two prediction complexities are multiplied by their respective weights, and the results are summed to obtain the comprehensive prediction complexity. For example, if the weight of the first prediction complexity is set to 0.5 and the weight of the second prediction complexity is also set to 0.5, the comprehensive prediction complexity = 1.18 × 0.5 + 1.25 × 0.5 = 1.215.

[0038] Subsequently, the product of the comprehensive prediction complexity and the initial number of prediction branches selected, P, is rounded down to obtain the appropriate number of prediction branches selected, J. Here, P is 4. If the calculated J is less than 2, J is set to 2; if the calculated J is greater than K, J is set to K. The initial number of prediction branches selected, P, is a pre-set base number of prediction branches, a fixed constant P=4, serving as the benchmark for calculating the appropriate number of branches. The appropriate number of prediction branches selected, J, is a dynamically adjusted number of prediction branches based on the comprehensive prediction complexity, achieving a match between prediction branches and prediction difficulty. The value of J ranges from 2 ≤ J ≤ K. The comprehensive prediction complexity is multiplied by the initial number of prediction branches selected, P, and the result is rounded down to obtain a preliminary value of J. A range check is performed on the preliminary J value: if J < 2, J = 2; if J > K, J = K; if 2 ≤ J ≤ K, the preliminary J value remains unchanged. For example, the overall prediction complexity is 1.215, P=4, K=10; the initial J value is 1.215×4=4.86, which is rounded down to 4; verification: 2≤4≤10, therefore the number of adapted prediction branches selected is J=4.

[0039] Finally, J prediction branches are randomly selected from the K hazardous gas spread prediction branches, and predictions are made based on the hazardous gas concentration distribution sequence set. The J prediction results are then arithmetically averaged to obtain the predicted hazardous gas concentration set for region A within the preset maintenance time window. The predicted hazardous gas concentration set is a dataset formed by arranging the predicted concentration values ​​of various hazardous gases at different time points in region A within the preset maintenance time window in chronological order. From the K hazardous gas spread prediction branches, J prediction branches are randomly selected; the hazardous gas concentration distribution sequence set is input into the selected J prediction branches, and predictions are made respectively to obtain J predicted hazardous gas concentration results for region A within the preset maintenance time window; the J prediction results are then arithmetically averaged to eliminate prediction bias from individual branches, ultimately obtaining the predicted hazardous gas concentration set for region A within the preset maintenance time window.

[0040] For example, among the 10 prediction branches F1-F10, four branches F2, F5, F7, and F9 are randomly selected; the hazardous gas concentration distribution sequence set is input into the four branches, and a maintenance time window is preset for area A (October 1, 2025, 9:00-13:00, 4 hours), resulting in four sets of prediction results: such as F2 predicting an H2S concentration of 2.8 mg / m³ at 9:00. 3 F5 predicts the H2S concentration at 9:00 AM to be 2.9 mg / m³. 3 F7 predicts the H2S concentration at 9:00 AM to be 2.7 mg / m³. 3 F9 predicts the H2S concentration at 9:00 AM to be 2.8 mg / m³. 3 The predicted H2S concentration at 9:00 is (2.8 + 2.9 + 2.7 + 2.8) / 4 = 2.8 mg / m³. Similarly, the average of the prediction results for all time points and all gas types within the 4-hour window is completed to finally obtain the predicted hazardous gas concentration set for region A from 9:00 to 13:00 on October 1, 2025.

[0041] In this embodiment of the invention, by integrating information from multiple monitoring points and using spatial correlation interpolation, the limitations of traditional single-point monitoring are overcome, realizing the construction of the distribution of hazardous gas concentration from point to area, and fully reflecting the spatial characteristics of gas concentration in the region. The multi-branch hazardous gas spread predictor based on LSTM fully captures the time series change pattern of hazardous gas concentration, and the K-fold cross-partitioning training method improves the model's generalization ability and prediction stability. The prediction branches are dynamically adapted through the hazardous gas hazard coefficient and spread growth coefficient, so that the prediction strategy matches the actual prediction difficulty, taking into account both prediction efficiency and accuracy. The final predicted hazardous gas concentration set for region A accurately reflects the dynamic change characteristics of hazardous gas concentration within the preset maintenance time window, providing comprehensive, reliable, and on-site-appropriate basic concentration data for subsequent protection anomaly risk inference, and solving the problems of one-sided data and insufficient accuracy of traditional prediction methods.

[0042] S200: The protective equipment images of the personnel going to the A area for maintenance within the preset maintenance time window are collected by on-site monitoring video, and the wearing characteristics of the protective equipment images are identified to obtain the wearing tightness, and the filter canister model and effective differential pressure accumulation time are obtained by sensor monitoring.

[0043] In this embodiment of the invention, images of protective equipment worn by personnel heading to Area A for maintenance within the preset maintenance time window are captured via on-site monitoring video. Wearing feature recognition is performed on these images to obtain the fit, and sensor monitoring is used to obtain the filter canister model and effective differential pressure accumulation duration. In high-risk scenarios involving tank leak maintenance, respiratory protective equipment is the core line of defense for personnel safety, and its actual protective effect directly determines personnel's ability to resist harmful gases. Traditional vision-based monitoring methods can only determine whether equipment is worn, but cannot accurately assess the fit and effective protection duration of the filter canister. Furthermore, they do not dynamically adjust the assessment weights based on the work scenario, resulting in significant blind spots in risk perception. Therefore, this step achieves accurate quantification of the actual protective effect of the protective equipment through multimodal data fusion, intelligent feature recognition, and dynamic weighted evaluation.

[0044] Step S200 in the method provided in this embodiment of the invention includes: First, images of personnel wearing protective equipment (PEE) are captured via on-site monitoring video when they enter Area A for maintenance within the preset maintenance time window. These PPE images, captured by on-site monitoring video, show the personnel wearing respiratory protective equipment upon entering Area A within the preset maintenance time window, including complete visual information of their face and the equipment. Based on the preset maintenance time window, the high-definition monitoring cameras deployed on-site are triggered to capture images of the PPE worn by maintenance personnel heading to Area A, storing them as an image sequence with timestamps. Simultaneously, the model and specifications of the PPE, as well as the employee ID of the maintenance personnel, are obtained through equipment sensors, providing basic information for subsequent fit assessment.

[0045] For example, a weld leak inspection was conducted in Area A of Tank Area No. 1 at a petrochemical plant, with the scheduled inspection window from 9:00 AM to 1:00 PM on January 28, 2026. At 9:05 AM, the on-site monitoring camera captured an image of maintenance personnel E001 entering Area A, clearly showing their full-face respirator and the fit to their face. Simultaneously, the employee ID E001 and the protective equipment model (R-6800 full-face respirator, medium size) were obtained.

[0046] Secondly, wear feature recognition is performed on the images of the protective equipment being worn to obtain the degree of fit. The process of performing wear feature recognition on the protective equipment wearing image to obtain the wearing fit includes: A set of images of protective equipment being worn and a set of wearable feature labels are collected as training data. A convolutional neural network is trained until convergence to generate a wearable feature recognizer. The wearable features include the horizontal offset deviation of the mask, the similarity of the left and right contours of the mask, the proportion of the exposed area of ​​facial skin, and the proportion of abnormal pixel area in the bridge of the nose. Using the wear feature recognizer, the wear feature of the protective equipment is recognized to obtain the current wear feature, and the dynamic wearing fit is evaluated and determined based on the current wear feature; The wearing fit is determined by a weighted evaluation of the static wearing fit and the dynamic wearing fit.

[0047] First, a set of images of protective equipment being worn and a set of wearable feature labels were collected as training data. A convolutional neural network was trained until convergence to generate a wearable feature recognizer. The wearable features include mask horizontal offset deviation, similarity of the mask's left and right contour shapes, percentage of exposed facial skin area, and percentage of abnormal pixel area in the bridge of the nose region. Specifically, mask horizontal offset deviation refers to the horizontal distance between the center of the mask and the center of the face; similarity of the mask's left and right contour shapes refers to the degree of matching between the left and right edges of the mask and the facial contours; percentage of exposed facial skin area refers to the proportion of facial skin not covered by the mask; and percentage of abnormal pixel area in the bridge of the nose region refers to the proportion of pixels within the sealed band area of ​​the bridge of the nose and nostrils where the brightness gradient direction is perpendicular to the local edge and the brightness value deviates from the standard range.

[0048] First, a training dataset was collected. The data sources were real-world monitoring images from chemical plant storage tank maintenance scenarios, as well as controlled images from the laboratory simulating different abnormal wearing conditions. These images covered workers of different body types, mainstream respiratory protective equipment models, and different working light / angle conditions, forming a sample set of protective equipment wearing images that included typical conditions such as good fit, misalignment, gaps, and nose bridge seal failure. Simultaneously, based on industrial visual annotation standards, four types of wearing feature labels were manually labeled for each image: horizontal mask offset deviation, similarity of the left and right contour shapes of the mask, percentage of exposed facial skin area, and percentage of abnormal pixel area in the nose bridge region, forming a sample wearing feature label set.

[0049] For example, 2,000 images of full-face respirators were collected as a sample set. Among them, 1,500 images were real images captured from the maintenance site monitoring video of the No. 1-3 storage tank area of ​​a petrochemical plant over the past 3 years, and 500 images were controlled images from the laboratory simulating abnormal states such as nose bridge seal failure and mask displacement. Each image was manually labeled with four types of wearing feature values: horizontal mask offset deviation, similarity of the left and right contour shapes of the mask, proportion of exposed facial skin area, and proportion of abnormal pixel area in the nose bridge region, forming a corresponding sample wearing feature label set.

[0050] Secondly, a convolutional neural network (CNN) was built and trained. The input layer dimension was adapted to the image resolution of 224×224×3, corresponding to an RGB three-channel image; the hidden layer contained three convolutional pooling modules and two fully connected layers; the output layer had four neurons, using a linear activation function, and outputting quantized values ​​of four types of wearable features. The sample data was preprocessed, and batch training was used. The optimizer was Adam, and the loss function was the mean squared error (MSE) to suit the feature value regression task; the convergence condition was that the validation set MSE ≤ 1×10⁻⁶. -3 Furthermore, the prediction error fluctuation range for 20 consecutive iterations is less than 5×10. -4Simultaneously, if the feature recognition accuracy on the validation set is ≥95%, training stops and a wearable feature recognizer is generated if any of these conditions are met. For example, the model is trained using a batch size of 32, the Adam optimizer, and the loss function is the mean squared error (MSE). Training continues until the MSE on the validation set decreases to 8 × 10⁻⁶. -4 Furthermore, it remained stable for 20 consecutive rounds, achieving a validation set feature recognition accuracy of 96%. Once the convergence condition was met, training was stopped, generating a wear feature recognizer that can automatically identify four types of wear features based on images of protective equipment being worn.

[0051] Secondly, using the wearable feature recognizer, the wearable feature of the protective equipment image is recognized to obtain the current wearable features, and the dynamic fit is evaluated and determined based on the current wearable features. The dynamic fit is a quantized value obtained by normalizing and weighting the wearable features based on real-time wearable image recognition, with a value range of 0-1, reflecting the mask sealing effect under the current dynamic operation. The protective equipment image is input into the wearable feature recognizer, which outputs the quantized values ​​of the current four types of wearable features; the feature values ​​are normalized: horizontal offset score S1 = 1 - (actual horizontal offset distance / preset maximum allowable offset distance), if S1 < 0, it is taken as 0, and if S1 > 1, it is taken as 1; contour similarity score S2 directly uses the contour shape similarity value output by the recognizer; skin exposure score S3 = 1 - (actual skin exposure percentage / preset maximum allowable exposure percentage), the result is limited to 0-1; nose bridge abnormality score S4 = 1 - (actual nose bridge abnormal pixel percentage / preset maximum allowable abnormal percentage), the result is limited to 0-1. The dynamic fit is obtained by weighting the scores of the four categories based on the experience of industrial experts.

[0052] For example, taking the image recognition results of the maintenance personnel E001 as an example: the actual horizontal offset distance was 2mm, the preset maximum allowable offset distance was 5mm, and the horizontal offset score was S1=1−(2 / 5)=0.6; the similarity of the mask's left and right contour shapes was 0.85, and the contour similarity score was S2=0.85; the facial skin exposure ratio was 5%, the preset maximum allowable exposure ratio was 10%, and the skin exposure score was S3=1−(5 / 10)=0.5; the abnormal pixel ratio of the bridge of the nose was 3%, the preset maximum allowable abnormal ratio was 8%, and the abnormal score of the bridge of the nose was S4=1−(3 / 8)=0.625. The weights for the following scores are set: 0.40 for abnormal nose bridge score, 0.25 for horizontal deviation score, 0.20 for contour similarity score, and 0.15 for skin exposure score. The dynamic fit score is calculated as follows: 0.625×0.4+0.6×0.25+0.85×0.2+0.5×0.15=0.645.

[0053] In addition, static fit is obtained. Static fit is a quantitative value obtained through static matching analysis based on the model and specifications of the protective equipment and the facial contour data of the personnel. The value ranges from 0 to 1 and reflects the theoretical sealing effect between the equipment and the face.

[0054] The steps for obtaining static fit include: The device model and specifications of the protective equipment worn are obtained through sensor monitoring, as well as the employee number of the staff member; The employee's facial contour data is retrieved based on the employee ID. Static fit analysis was performed based on the device model, device specifications, and facial contour data to assess and determine the static fit degree.

[0055] First, the model and specifications of the protective equipment being worn are obtained through sensor monitoring, as well as the employee's ID number. For example, the protective equipment model is R-6800; the specifications are: a sealing width suitable for a 13-15cm face, and a nose bridge pad height suitable for a 2.8-3.2cm face; the employee's ID number is E001.

[0056] Secondly, the facial contour data of the employee is retrieved based on the employee number. This facial contour data includes facial width, nasal bridge height, jawline curvature, and cheekbone distance. For example, the facial contour data of employee E001 is: facial width 14cm, nasal bridge height 3cm.

[0057] Then, a static fit analysis is performed based on the equipment model, specifications, and facial contour data to assess and determine the static fit degree. First, key sealing parameters are extracted from the protective equipment model and specifications, and the corresponding dimensions of the person's facial contour are simultaneously retrieved to form matching dimension pairs. Second, an adaptation score is calculated for each dimension, using the rule of subtracting the deviation between the equipment and the person's dimension from 1 and dividing by the preset tolerance for that dimension, with a score range of 0-1. Finally, the static fit degree is obtained by weighting and summing the scores according to the influence of each dimension on the sealing effect. For example, if the equipment is calibrated to fit a facial width of 13-15cm and a nose bridge height of 2.8-3.2cm, while the person's actual facial width is 14cm and nose bridge height is 3cm, the calculated facial width adaptation score is 0.5, and the nose bridge height score is 1.0. Combined with the scores for jaw curvature and cheekbone distance, and weighted and summed, the final static fit degree is 0.82.

[0058] Furthermore, the wearing fit is determined by a weighted evaluation based on the static wearing fit and the dynamic wearing fit.

[0059] The wearing fit is determined by a weighted evaluation based on the static wearing fit and the dynamic wearing fit, including: The intensity of the workers' dynamic work activities is assessed according to the preset maintenance plan and used as the mask seal failure factor; The ratio of the harmful gas spread growth coefficient to the preset gas spread growth coefficient is used as the first compensation coefficient. The ratio of the mask sealing failure coefficient to the preset standard mask sealing failure coefficient is used as the second compensation coefficient; The dynamic weighted compensation coefficient is obtained by weighted summation of the first compensation coefficient and the second compensation coefficient. The product of the dynamic weight compensation coefficient and the preset dynamic weight is used as the adaptive dynamic weight, wherein the preset dynamic weight is 0.65. If the calculated adaptive dynamic weight is less than 0.5, it is taken as 0.5; if the calculated adaptive dynamic weight is greater than 0.8, it is taken as 0.8. The adaptation static weight is obtained by subtracting the adaptation dynamic weight from 1; Based on the adaptive dynamic weights and adaptive static weights, the static wearing fit and dynamic wearing fit are weighted and summed to obtain the wearing fit.

[0060] First, the intensity of the workers' dynamic work activities is assessed according to the preset maintenance plan, and this is used as the mask seal failure coefficient. The dynamic work activity intensity is a quantitative indicator that comprehensively considers the amplitude, frequency, duration, and environmental interaction of the work steps, ranging from 0 to 1. It measures the potential interference of the work on the stability of the mask's sealing interface. The mask seal failure coefficient is a predicted value between 0 and 1, representing the probability and extent to which the mask's seal will measurably deteriorate from its initial state due to personnel movements under the current planned work. 0 indicates no impact, and 1 indicates a very high probability of severe seal failure. Based on the preset maintenance plan, the amplitude, frequency, duration, and environmental interaction intensity of the workers' movements are comprehensively assessed, and the assessment results are quantified as the mask seal failure coefficient.

[0061] For example, the pre-set maintenance plan for maintenance worker E001 in Area A of Tank Area 1 of a petrochemical plant includes steps such as bending over to operate valves and standing inspections. Bending over to operate the valves lasts for 15 minutes, with an operation frequency of 2 times per minute, and the environmental risk level is medium. After comprehensive assessment, the dynamic work activity intensity is 0.6, therefore the mask sealing failure factor is 0.6.

[0062] Secondly, the ratio of the harmful gas spread growth coefficient to the preset gas spread growth coefficient is used as the first compensation coefficient. The first compensation coefficient reflects the degree of compensation for the weight adjustment caused by the harmful gas spread rate, and is calculated by the ratio of the harmful gas spread growth coefficient to the preset gas spread growth coefficient. The preset gas spread growth coefficient refers to a benchmark value for the gas spread growth coefficient set in advance based on historical maintenance data of the chemical plant and industry standards. The previously calculated harmful gas spread growth coefficient and the preset gas spread growth coefficient are retrieved, and their ratio is calculated; the result is the first compensation coefficient. For example, if the harmful gas spread growth coefficient for the E001 work area is known to be 0.15, and the preset gas spread growth coefficient for this petrochemical plant is 0.12, then the first compensation coefficient = 0.15 / 0.12 = 1.25.

[0063] Furthermore, the ratio of the mask sealing failure coefficient to the preset standard mask sealing failure coefficient is used as the second compensation coefficient. The second compensation coefficient reflects the degree of compensation for the weight adjustment caused by the operational action, and is calculated by the ratio of the mask sealing failure coefficient to the preset standard mask sealing failure coefficient. The preset standard mask sealing failure coefficient refers to a benchmark value for the mask sealing failure coefficient set in advance based on protective equipment performance standards and industry experience. The previously calculated mask sealing failure coefficient and the preset standard mask sealing failure coefficient are retrieved, and their ratio is calculated; the result is the second compensation coefficient. For example, if the mask sealing failure coefficient of E001 is known to be 0.6, and the preset standard mask sealing failure coefficient of this petrochemical plant is 0.5, then the second compensation coefficient = 0.6 / 0.5 = 1.2.

[0064] Then, the dynamic weight compensation coefficient is obtained by weighted summation of the first compensation coefficient and the second compensation coefficient. The dynamic weight compensation coefficient is a quantified coefficient obtained by weighted summation, taking into account the combined effects of the first and second compensation coefficients, and is used to dynamically adjust and adapt the dynamic weights. Preset weights are assigned to the first and second compensation coefficients, with a weight sum of 1. The two coefficients are multiplied by their corresponding weights and then summed to obtain the dynamic weight compensation coefficient. For example, if the weight of the first compensation coefficient is set to 0.5 and the weight of the second compensation coefficient is also set to 0.5, the dynamic weight compensation coefficient = 1.25 × 0.5 + 1.2 × 0.5 = 1.225.

[0065] Furthermore, the product of the dynamic weight compensation coefficient and the preset dynamic weight is used as the adaptive dynamic weight, where the preset dynamic weight is 0.65. If the calculated adaptive dynamic weight is less than 0.5, it is set to 0.5; if the calculated adaptive dynamic weight is greater than 0.8, it is set to 0.8. The adaptive dynamic weight is the dynamic fit weight adjusted based on the dynamic weight compensation coefficient, with a fixed value range of 0.5-0.8, used to balance the influence ratio of static and dynamic fit. The preset dynamic weight refers to the pre-set basic dynamic fit weight, which is a fixed constant of 0.65. Multiplying the dynamic weight compensation coefficient by the preset dynamic weight yields the preliminary adaptive dynamic weight; the preliminary result is then checked for range: if < 0.5, it is set to 0.5; if > 0.8, it is set to 0.8. The final result is the adaptive dynamic weight. For example, the dynamic weight compensation coefficient is 1.225, the preset dynamic weight is 0.65, and the initial adapted dynamic weight is 1.225 × 0.65 ≈ 0.796; after verification, 0.796 is within the range of 0.5-0.8, and the final adapted dynamic weight is 0.796.

[0066] Subsequently, the adaptive static weight is obtained by subtracting the adaptive dynamic weight from 1. The adaptive static weight is a static fit weight that complements the adaptive dynamic weight; the sum of the two is 1, used to balance the influence ratio of static fit. Adaptive static weight = 1 - adaptive dynamic weight. For example, if the adaptive dynamic weight = 0.796, the adaptive static weight = 1 - 0.796 = 0.204.

[0067] Finally, based on the adaptive dynamic weight and the adaptive static weight, the static fit and dynamic fit are weighted and summed to obtain the fit score. The fit score is a quantitative value obtained by combining the static fit score and the dynamic fit score through weighted summation, taking a value of 0-1, reflecting the actual sealing effect of the protective equipment. Fit score = Adaptive dynamic weight × Dynamic fit score + Adaptive static weight × Static fit score. For example, given that the static fit score of E001 is 0.82, the dynamic fit score is 0.645, the adaptive dynamic weight is 0.8, the adaptive static weight is 0.2, and the fit score is 0.82 × 0.2 + 0.645 × 0.8 = 0.68.

[0068] In addition, the filter canister model and effective differential pressure accumulation time are obtained through sensor monitoring. The effective differential pressure accumulation time refers to the cumulative usage time during filter canister use, from the initial state to reaching the manufacturer's specified limit due to the increase in pressure difference across its two ends caused by contaminant adsorption. This time is used to determine the effectiveness of the filter canister. Differential pressure sensors installed at the inlet and outlet of the filter canister monitor the pressure difference changes in real time, and combined with the built-in timing module, the effective differential pressure duration is recorded cumulatively. For example, the respiratory protective equipment used by maintenance personnel E001 is model A-300; the maintenance starts at 9:00 AM. The data acquisition terminal automatically reads the model as A-300 and starts timing; the differential pressure sensor monitors in real time: initial differential pressure 10 Pa at 9:00 AM, differential pressure 35 Pa at 9:30 AM, differential pressure reaches 50 Pa at 10:15 AM; the terminal determines that the differential pressure meets the standard and stops accumulating time, with an effective differential pressure accumulation time of 75 minutes.

[0069] In this embodiment of the invention, multi-source data fusion and intelligent recognition overcome the limitations of traditional visual monitoring, which only determines whether protective equipment is worn, and achieve precise quantification of the actual protective effect of protective equipment. On the one hand, convolutional neural networks are used to identify multi-dimensional wearing features, combined with static and dynamic weighted evaluation, to accurately calculate the wearing fit, solving the problem that traditional methods cannot assess the sealing effect. On the other hand, differential pressure sensors and dual algorithms monitor the status of the filter canister, ensuring real-time perception of the filter canister's effectiveness. The final output of wearing fit, filter canister model, and effective differential pressure accumulation time provides comprehensive and accurate data for subsequent protective anomaly risk inference, improving the accuracy and reliability of personnel protective risk perception in high-risk work scenarios.

[0070] S300: Utilizing an abnormal risk perception engine trained on an LLM agent, the system infers the probability of predicted abnormal protection risk based on the predicted harmful gas concentration set, wearing fit, filter canister model, and effective pressure difference accumulation time.

[0071] In this embodiment of the invention, an anomaly risk perception engine trained based on an LLM agent is used to infer the probability of predicted protective anomaly risks based on the predicted hazardous gas concentration set, wearing fit, filter canister model, and effective pressure difference accumulation time. In tank leakage maintenance scenarios, protective anomaly risks are coupled with multiple factors such as gas concentration, wearing fit, and filter canister status, making it difficult for traditional rule engines to handle the dynamic reasoning requirements of complex scenarios. This step constructs an anomaly risk perception engine based on a Large Language Model (LLM), combining semantic encoding of multi-source data with instruction fine-tuning to achieve accurate reasoning of protective anomaly risks, solving the problems of weak generalization ability and insufficient reasoning accuracy of traditional methods.

[0072] Step S300 in the method provided in this embodiment of the invention includes: Among them, the anomaly risk perception engine, trained based on LLM agents, includes: Based on the historical tank leakage and maintenance records of the chemical plant, several sets of hazardous gas concentrations, several sets of wearing tightness, several sets of filter canister models, and several sets of effective differential pressure accumulation duration were collected as sample training data, and several sets of protection abnormality risk probabilities were collected as supervision labels. Using the sample training data and supervision labels, the LLM agent is trained and optimized through feature semantic encoding and instruction fine-tuning until the preset convergence condition is reached, thereby generating an abnormal risk perception engine.

[0073] The probability of abnormal risk in sample protection refers to the proportion of historical abnormal risk events under different scenarios, such as different concentrations of harmful gases in samples, sample fit, sample filter canister model, and cumulative duration of effective pressure difference.

[0074] First, based on the historical tank leak maintenance records of the chemical plant, several sets of hazardous gas concentrations, several sets of clothing fit, several sets of filter canister models, and several sets of effective differential pressure accumulation durations were collected as training data. Several sets of protection anomaly risk probabilities were also collected as supervisory labels. The training data refers to the multi-source feature data extracted from the historical tank leak maintenance records of the chemical plant, including the hazardous gas concentration sets, clothing fit, filter canister models, and effective differential pressure accumulation durations. The protection anomaly risk probability refers to the proportion of historical protection anomaly risk events occurring in the corresponding sample scenario, ranging from 0 to 1. From the chemical plant's historical maintenance database, records similar to the current maintenance scenario in Area A were selected, and multi-source feature data were extracted as training data. The number of historical protection anomaly events and the total number of operations corresponding to each sample were statistically analyzed, and the event proportion was calculated as a supervisory label. The structured feature data and vector features were uniformly stored in a PostgreSQL (pgvector) database to support efficient retrieval and access to subsequent training data.

[0075] For example, from the tank leakage maintenance records of a petrochemical plant over the past 10 years, 2,000 samples of the same scenario were selected. The concentration sets of H2S, CO, and VOCs, the fit of the filter, the model of the filter canister, and the cumulative duration of the effective pressure difference were extracted for each sample. Statistical analysis revealed that 120 of the samples had experienced protective abnormality events, with a corresponding supervision label of 120 / 2000=0.06, which means that the probability of the sample protective abnormality risk is 6%. All data was stored in a PostgreSQL (pgvector) database.

[0076] Secondly, using the sample training data and supervision labels, the LLM agent is trained and optimized through feature semantic encoding and instruction fine-tuning until a preset convergence condition is met, generating an anomaly risk perception engine. Feature semantic encoding refers to converting structured numerical features into natural language descriptions and vector representations that the LLM can understand, combining multimodal models such as Qwen-VL-Max / ERNIE-Vision to enhance feature expression capabilities. Instruction fine-tuning refers to constructing instruction-based training samples that fit industrial scenarios to perform targeted fine-tuning of the LLM, improving the model's adaptability to risk protection reasoning tasks. The anomaly risk perception engine is the converged LLM agent after instruction fine-tuning, possessing multi-source feature semantic understanding and anomaly risk reasoning capabilities, supporting Docker+K8s containerized deployment, and achieving highly available and scalable service operation. Convergence condition: Risk probability prediction error MSE on the validation set ≤ 5 × 10⁻⁶. -3 Furthermore, the error did not decrease significantly after 20 consecutive iterations.

[0077] Specifically, the sample training data is semantically encoded: numerical features are converted into natural language descriptions, and corresponding vector representations are generated through a multimodal model; instruction-based training samples are constructed: using "inferring the probability of anomaly risk protection based on multi-source features of the input" as the instruction, training pairs are formed by combining the encoded features with supervision labels; DeepSeek / GLM-4 is selected as the base LLM, and the constructed training samples are used for instruction fine-tuning to optimize the model's risk inference capability. The sample training data is divided into training and validation sets in an 8:2 ratio, and the LLM is fine-tuned using the training set; after each iteration, the model's prediction error is evaluated using the validation set, and training stops when the convergence condition is met; the converged model is packaged into a Docker image and deployed in a containerized manner through a K8s cluster, generating an anomaly risk perception engine that can provide inference services externally.

[0078] For example, semantically encode the sample data of maintenance personnel E001: Maintenance scenario in area A, H2S concentration 2.8 mg / m³ 3 CO concentration 0.9 mg / m³ 3 The fit was 0.672, the filter canister was type A-300, and the effective differential pressure accumulation time was 75 minutes. A command-based training sample was constructed: Input: [encoded features], Output: Probability of abnormal protection risk 0.06. This sample was used to fine-tune GLM-4, enabling it to learn risk association rules in industrial scenarios. The 2000 samples were divided into a 8:2 ratio: 1600 training samples and 400 validation samples. The training set was used to fine-tune GLM-4; after 120 iterations, the validation set MSE decreased to 3.8 × 10⁻⁶. -3Furthermore, it remained stable for 20 consecutive rounds, meeting the convergence conditions; the model was packaged into a Docker image and deployed to a K8s cluster to generate an anomaly risk perception engine.

[0079] Finally, using the anomaly risk perception engine trained on an LLM agent, the predicted probability of anomaly risk is inferred based on the predicted hazardous gas concentration set, wearing fit, filter canister model, and effective pressure difference accumulation time. The predicted hazardous gas concentration set generated by S100, the wearing fit, filter canister model, and effective pressure difference accumulation time generated by S200 are retrieved from the PostgreSQL (pgvector) database. The real-time input data is semantically encoded, converted into natural language descriptions and vector representations understandable by the LLM, and input into the anomaly risk perception engine. The engine performs semantic understanding and feature association on the input semantic features, combining them with risk patterns from historical samples. Through the contextual reasoning capabilities of the LLM, the probability value of anomaly risk in the current scenario is calculated.

[0080] For example, input the real-time characteristic data of maintenance personnel E001: the predicted H2S concentration in the preset maintenance window of area A is 2.8 mg / m³. 3 CO predicted concentration: 0.9 mg / m³ 3 The wearer's fit was 0.68, the filter canister was Type A-300, and the effective differential pressure accumulation time was 75 minutes. This data was coded and input into the abnormal risk perception engine. After inputting the feature data of E001, the engine inferred that the probability of abnormal protection risk in the current scenario was 0.06.

[0081] This step overcomes the generalization bottleneck of traditional rule engines by constructing an anomaly risk perception engine based on LLM, achieving accurate risk inference under multi-source coupled features. Combining multimodal semantic encoding and instruction fine-tuning improves the model's adaptability to industrial scenarios. Docker+Kubernetes containerized deployment ensures the engine's high availability and scalability. The final output, predicting the probability of anomaly risks, provides a quantitative basis for on-site safety management, reducing protection risks in high-risk operational scenarios.

[0082] S400: If the predicted probability of abnormal protection risk exceeds the preset risk probability threshold, then a protection risk warning will be issued to the staff.

[0083] In this embodiment of the invention, if the predicted probability of abnormal protection risk exceeds a preset risk probability threshold, a protection risk warning is issued to the personnel. Based on historical accident data of the chemical plant and industry safety standards, the preset protection risk probability threshold is typically set to 0.1, or 10%. The predicted protection risk probability output by S300 is compared with the preset risk probability threshold in real time. If the predicted value is greater than or equal to the preset risk probability threshold, a pop-up warning is pushed to the smart terminal worn by the personnel, and a warning message containing the risk scenario, triggering cause, and handling suggestions is sent to the on-site safety control center to ensure that the warning message reaches relevant personnel and supports rapid response.

[0084] For example, if the maintenance personnel predict that the probability of abnormal protection risk is 0.12, which exceeds the preset threshold of 0.1, an early warning will be triggered immediately: the maintenance personnel's smart terminal will display a pop-up window showing "The current H2S concentration in area A is too high and the fit of the mask is not tight enough. It is recommended to check the mask seal or replace the filter canister immediately", and the safety control center will receive the warning details at the same time.

[0085] In this embodiment of the invention, proactive perception and rapid response to protection risks are achieved through real-time threshold comparison and multi-channel early warning, effectively shortening the response time for risk handling, reducing the probability of abnormal events in high-risk operation scenarios, and providing closed-loop protection for on-site safety management.

[0086] Through the specific implementation methods described above, the embodiments of the present invention achieve the following technical effects: This invention provides an industrial safety production risk perception method and system based on LLM intelligent agents. First, by using historical data-driven concentration prediction, the dynamic distribution characteristics of hazardous gases in area A are obtained in advance, solving the lag problem of traditional real-time monitoring. Second, through multimodal visual recognition and sensor monitoring, it overcomes the limitation of traditional protective equipment monitoring that only determines whether it is worn, achieving a quantitative assessment of the fit and effectiveness of the filter canister. Third, through an LLM intelligent agent-based abnormal risk perception engine, semantic reasoning is performed on multi-source coupled features, improving the accuracy and scenario generalization ability of risk assessment and solving the problem of insufficient adaptability of traditional rule engines to complex scenarios. Finally, through threshold-triggered multi-channel early warning, proactive risk perception and rapid response are achieved, effectively shortening risk handling time, reducing the probability of abnormal protective events in high-risk operation scenarios, and overall improving the intelligence level and reliability of industrial safety production risk prevention and control.

[0087] Example 2, as Figure 2 As shown, this invention provides an industrial safety production risk perception system based on LLM agents, the system comprising: Environmental risk prediction module 11 is used to predict the set of hazardous gas concentrations in area A within a preset maintenance time window based on historical hazardous gas concentration monitoring records during tank leak repair. The protective status monitoring module 12 is used to collect images of protective equipment worn by personnel going to the A area for maintenance within the preset maintenance time window through on-site monitoring video, and to perform wearing feature recognition on the protective equipment wearing images to obtain the wearing tightness, and to obtain the filter canister model and effective differential pressure accumulation time through sensor monitoring. The intelligent risk perception module 13 is used to use an abnormal risk perception engine trained based on LLM agent to infer the probability of predicted protective abnormal risk based on the predicted harmful gas concentration set, wearing tightness, filter canister model and effective pressure difference accumulation time. The dynamic early warning response module 14 is used to issue a protection risk warning to the staff if the predicted probability of protection anomalies exceeds a preset risk probability threshold.

[0088] In one embodiment, the environmental risk prediction module 11 is further configured to: The system monitors and obtains several hazardous gas concentration sequence sets from several monitoring points in a chemical plant within a historical time zone. Based on the location coordinates of these monitoring points, it integrates information and performs interpolation estimation based on spatial correlation to construct a hazardous gas concentration distribution sequence set. A harmful gas spread predictor was constructed based on a long short-term memory network. Using the harmful gas spread predictor, the predicted harmful gas concentration set of area A within a preset maintenance time window is obtained based on the harmful gas concentration distribution sequence set.

[0089] Among them, the harmful gas spread predictor built based on long short-term memory network includes: Based on the historical tank leakage and maintenance records of the chemical plant, several sample hazardous gas concentration distribution sequence sets were collected, and the historical maximum hazardous gas concentration distribution of different sample hazardous gas concentration distribution sequence sets within the historical maintenance time window was obtained as the sample predicted hazardous gas concentration distribution, thus obtaining several sample predicted hazardous gas concentration distributions. The aforementioned sample harmful gas concentration distribution sequence set and the sample predicted harmful gas concentration distribution are used as training data, and K-fold cross-partitioning is performed to obtain K training sets, where K is greater than or equal to 10. The long short-term memory network is trained to convergence using the K training sets, generating K branches for predicting the spread of harmful gases. These branches are then combined to obtain a predictor for the spread of harmful gases.

[0090] Specifically, the hazardous gas spread predictor uses the hazardous gas concentration distribution sequence set to predict the set of hazardous gas concentrations in area A within a preset maintenance time window, including: Extract all hazardous gas types from the hazardous gas concentration distribution sequence set, and perform a weighted evaluation according to a preset gas hazard level to determine the hazardous gas hazard coefficient; Calculate the concentration spread and growth rate of all hazardous gas types in the hazardous gas concentration distribution sequence set, and determine the hazardous gas spread and growth coefficient by weighting according to the preset gas hazard level; The ratio of the hazard coefficient of the harmful gas to the preset hazard coefficient of the gas is used as the first prediction complexity; The ratio of the harmful gas spread growth coefficient to the preset gas spread growth coefficient is used as the second prediction complexity. The combined prediction complexity is obtained by weighted summation of the first prediction complexity and the second prediction complexity. The product of the comprehensive prediction complexity and the initial prediction branch selection number P is rounded down to obtain the adaptive prediction branch selection number J, where P is 4. If the calculated J is less than 2, then J is equal to 2; if the calculated J is greater than K, then J is equal to K. J prediction branches are randomly selected from the K branches for predicting the spread of harmful gases. Each branch is used to make predictions based on the set of harmful gas concentration distribution sequences. The J prediction results are then averaged to obtain the set of predicted harmful gas concentrations in region A within a preset maintenance time window.

[0091] In one embodiment, the protection status monitoring module 12 is further configured to: The process of performing wear feature recognition on the protective equipment wearing image to obtain the wearing fit includes: A set of images of protective equipment being worn and a set of wearable feature labels are collected as training data. A convolutional neural network is trained until convergence to generate a wearable feature recognizer. The wearable features include the horizontal offset deviation of the mask, the similarity of the left and right contours of the mask, the proportion of the exposed area of ​​facial skin, and the proportion of abnormal pixel area in the bridge of the nose. Using the wear feature recognizer, the wear feature of the protective equipment is recognized to obtain the current wear feature, and the dynamic wearing fit is evaluated and determined based on the current wear feature; The wearing fit is determined by a weighted evaluation of the static wearing fit and the dynamic wearing fit.

[0092] The steps for obtaining static fit include: The device model and specifications of the protective equipment worn are obtained through sensor monitoring, as well as the employee number of the staff member; The employee's facial contour data is retrieved based on the employee ID. Static fit analysis was performed based on the device model, device specifications, and facial contour data to assess and determine the static fit degree.

[0093] The wearing fit is determined by a weighted evaluation based on the static wearing fit and the dynamic wearing fit, including: The intensity of the workers' dynamic work activities is assessed according to the preset maintenance plan and used as the mask seal failure factor; The ratio of the harmful gas spread growth coefficient to the preset gas spread growth coefficient is used as the first compensation coefficient. The ratio of the mask sealing failure coefficient to the preset standard mask sealing failure coefficient is used as the second compensation coefficient; The dynamic weighted compensation coefficient is obtained by weighted summation of the first compensation coefficient and the second compensation coefficient. The product of the dynamic weight compensation coefficient and the preset dynamic weight is used as the adaptive dynamic weight, wherein the preset dynamic weight is 0.65. If the calculated adaptive dynamic weight is less than 0.5, it is taken as 0.5; if the calculated adaptive dynamic weight is greater than 0.8, it is taken as 0.8. The adaptation static weight is obtained by subtracting the adaptation dynamic weight from 1; Based on the adaptive dynamic weights and adaptive static weights, the static wearing fit and dynamic wearing fit are weighted and summed to obtain the wearing fit.

[0094] In one embodiment, the intelligent risk perception module 13 is further configured to: Among them, the anomaly risk perception engine, trained based on LLM agents, includes: Based on the historical tank leakage and maintenance records of the chemical plant, several sets of hazardous gas concentrations, several sets of wearing tightness, several sets of filter canister models, and several sets of effective differential pressure accumulation duration were collected as sample training data, and several sets of protection abnormality risk probabilities were collected as supervision labels. Using the sample training data and supervision labels, the LLM agent is trained and optimized through feature semantic encoding and instruction fine-tuning until the preset convergence condition is reached, thereby generating an abnormal risk perception engine.

[0095] The probability of abnormal risk in sample protection refers to the proportion of historical abnormal risk events under different scenarios, such as different concentrations of harmful gases in samples, sample fit, sample filter canister model, and cumulative duration of effective pressure difference.

[0096] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for perceiving industrial safety production risks based on LLM agents, characterized in that the method... include: When carrying out tank leak repairs, the predicted hazardous gas concentration set of area A within the preset repair time window is obtained based on historical hazardous gas concentration monitoring records. The protective equipment images of the personnel going to the A area for maintenance within the preset maintenance time window are captured by on-site monitoring video. Wearing features are identified in the protective equipment images to obtain the fit. The filter canister model and effective differential pressure accumulation time are also obtained by monitoring the filter canister through sensors. Using an abnormal risk perception engine trained on an LLM agent, the probability of predicted abnormal protection risk is inferred based on the predicted harmful gas concentration set, wearing fit, filter canister model and effective pressure difference accumulation time. If the predicted probability of abnormal protection risk exceeds the preset risk probability threshold, a protection risk warning will be issued to the staff.

2. The industrial safety production risk perception method based on LLM intelligent agents according to claim 1, characterized in that, Based on historical hazardous gas concentration monitoring records, the predicted hazardous gas concentration set for area A within a preset maintenance time window is obtained, including: The system monitors and obtains several hazardous gas concentration sequence sets from several monitoring points in a chemical plant within a historical time zone. Based on the location coordinates of these monitoring points, it integrates information and performs interpolation estimation based on spatial correlation to construct a hazardous gas concentration distribution sequence set. A harmful gas spread predictor was constructed based on a long short-term memory network. Using the harmful gas spread predictor, the predicted harmful gas concentration set of area A within a preset maintenance time window is obtained based on the harmful gas concentration distribution sequence set.

3. The industrial safety production risk perception method based on LLM intelligent agents according to claim 2, characterized in that, A hazardous gas spread predictor is constructed based on a long short-term memory network, including: Based on the historical tank leakage and maintenance records of the chemical plant, several sample hazardous gas concentration distribution sequence sets were collected, and the historical maximum hazardous gas concentration distribution of different sample hazardous gas concentration distribution sequence sets within the historical maintenance time window was obtained as the sample predicted hazardous gas concentration distribution, thus obtaining several sample predicted hazardous gas concentration distributions. The aforementioned sample harmful gas concentration distribution sequence set and the sample predicted harmful gas concentration distribution are used as training data, and K-fold cross-partitioning is performed to obtain K training sets, where K is greater than or equal to 10. The long short-term memory network is trained to convergence using the K training sets, generating K branches for predicting the spread of harmful gases. These branches are then combined to obtain a predictor for the spread of harmful gases.

4. The industrial safety production risk perception method based on LLM intelligent agents according to claim 3, characterized in that, Using the aforementioned hazardous gas spread predictor, a predicted hazardous gas concentration set for region A within a preset maintenance time window is obtained based on the hazardous gas concentration distribution sequence set, including: Extract all hazardous gas types from the hazardous gas concentration distribution sequence set, and perform a weighted evaluation according to a preset gas hazard level to determine the hazardous gas hazard coefficient; Calculate the concentration spread and growth rate of all hazardous gas types in the hazardous gas concentration distribution sequence set, and determine the hazardous gas spread and growth coefficient by weighting according to the preset gas hazard level; The ratio of the hazard coefficient of the harmful gas to the preset hazard coefficient of the gas is used as the first prediction complexity; The ratio of the harmful gas spread growth coefficient to the preset gas spread growth coefficient is used as the second prediction complexity. The combined prediction complexity is obtained by weighted summation of the first prediction complexity and the second prediction complexity. The product of the comprehensive prediction complexity and the initial prediction branch selection number P is rounded down to obtain the adaptive prediction branch selection number J, where P is 4. If the calculated J is less than 2, then J is equal to 2; if the calculated J is greater than K, then J is equal to K. J prediction branches are randomly selected from the K branches for predicting the spread of harmful gases. Each branch is used to make predictions based on the set of harmful gas concentration distribution sequences. The J prediction results are then averaged to obtain the set of predicted harmful gas concentrations in region A within a preset maintenance time window.

5. The industrial safety production risk perception method based on LLM intelligent agents according to claim 1, characterized in that, Perform wear feature recognition on the image of the protective equipment being worn to obtain the fit, including: A set of images of protective equipment being worn and a set of wearable feature labels are collected as training data. A convolutional neural network is trained until convergence to generate a wearable feature recognizer. The wearable features include the horizontal offset deviation of the mask, the similarity of the left and right contours of the mask, the proportion of the exposed area of ​​facial skin, and the proportion of abnormal pixel area in the bridge of the nose. Using the wear feature recognizer, the wear feature of the protective equipment is recognized to obtain the current wear feature, and the dynamic wearing fit is evaluated and determined based on the current wear feature; The wearing fit is determined by a weighted evaluation of the static wearing fit and the dynamic wearing fit.

6. The industrial safety production risk perception method based on LLM intelligent agents according to claim 5, characterized in that, The steps to obtain static fit include: The device model and specifications of the protective equipment worn are obtained through sensor monitoring, as well as the employee number of the staff member; The employee's facial contour data is retrieved based on the employee ID. Static fit analysis was performed based on the device model, device specifications, and facial contour data to assess and determine the static fit degree.

7. The industrial safety production risk perception method based on LLM intelligent agents according to claim 5, characterized in that, The wearing fit is determined by a weighted evaluation of the static wearing fit and the dynamic wearing fit, including: The intensity of the workers' dynamic work activities is assessed according to the preset maintenance plan and used as the mask seal failure factor. The ratio of the harmful gas spread growth coefficient to the preset gas spread growth coefficient is used as the first compensation coefficient. The ratio of the mask sealing failure coefficient to the preset standard mask sealing failure coefficient is used as the second compensation coefficient; The dynamic weighted compensation coefficient is obtained by weighted summation of the first compensation coefficient and the second compensation coefficient. The product of the dynamic weight compensation coefficient and the preset dynamic weight is used as the adaptive dynamic weight, wherein the preset dynamic weight is 0.

65. If the calculated adaptive dynamic weight is less than 0.5, it is taken as 0.5; if the calculated adaptive dynamic weight is greater than 0.8, it is taken as 0.

8. The adaptation static weight is obtained by subtracting the adaptation dynamic weight from 1; Based on the adaptive dynamic weights and adaptive static weights, the static wearing fit and dynamic wearing fit are weighted and summed to obtain the wearing fit.

8. The industrial safety production risk perception method based on LLM intelligent agents according to claim 1, characterized in that, An anomaly risk perception engine, trained based on an LLM agent, is obtained, including: Based on the historical tank leakage and maintenance records of the chemical plant, several sets of hazardous gas concentrations, several sets of wearing tightness, several sets of filter canister models, and several sets of effective differential pressure accumulation duration were collected as sample training data, and several sets of protection abnormality risk probabilities were collected as supervision labels. Using the sample training data and supervision labels, the LLM agent is trained and optimized through feature semantic encoding and instruction fine-tuning until the preset convergence condition is reached, thereby generating an abnormal risk perception engine.

9. The industrial safety production risk perception method based on LLM intelligent agents according to claim 8, characterized in that, The probability of abnormal risk in sample protection refers to the proportion of historical abnormal risk events under different scenarios, such as different concentrations of harmful gases in samples, fit of sample wearing, model of sample filter canister, and cumulative duration of effective pressure difference in samples.

10. An industrial safety production risk perception system based on LLM intelligent agents, characterized in that, The system for implementing the LLM-based industrial safety production risk perception method according to any one of claims 1-9 comprises: The environmental risk prediction module is used to predict the set of hazardous gas concentrations in area A within a preset maintenance time window based on historical hazardous gas concentration monitoring records during tank leak repairs. The protective status monitoring module is used to collect images of protective equipment worn by personnel going to the A area for maintenance within the preset maintenance time window through on-site monitoring video, and to perform wearing feature recognition on the protective equipment wearing images to obtain the wearing tightness, and to obtain the filter canister model and effective differential pressure accumulation time through sensor monitoring. The intelligent risk perception module is used to infer the probability of abnormal protection risk based on the predicted harmful gas concentration set, wearing fit, filter canister model and effective pressure difference accumulation time by using the abnormal risk perception engine trained based on LLM agent. The dynamic early warning response module is used to issue a protective risk warning to the staff if the predicted probability of abnormal protection risks exceeds a preset risk probability threshold.

Images