A chemical industrial park safety production risk dynamic early warning method

Abstract

The present application relates to a kind of chemical industry park safety production risk dynamic early warning method, belong to the technical field of risk safety monitoring of chemical industry park, solve the problems that early warning method in prior art cannot predict danger in advance, is vulnerable to environmental interference, relies on artificial, data dependency is strong, cannot be applied to complex scene.For each primary index, the comprehensive score corresponding to the index is calculated, the subjective weight is calculated based on the improved AHP method, and the objective weight is calculated based on the entropy weight method of dynamic variational autoencoder;The subjective weight and the objective weight are fused to obtain the comprehensive weight;The risk level of the chemical industry park is calculated based on the comprehensive score of each primary index and the corresponding comprehensive weight.The present application realizes a kind of chemical industry park safety production risk dynamic early warning method, which is timely, accurate and suitable for complex scene.

Description

Technical Field

[0001] This invention relates to the field of risk and safety monitoring technology in chemical industrial parks, and in particular to a dynamic early warning method for safety production risks in chemical industrial parks. Background Technology

[0002] Chemical industrial parks, as concentrated areas for the production, storage, and transportation of hazardous chemicals, face multiple safety risks, including gas leaks, fires, explosions, and equipment malfunctions, making safe production paramount. While various risk early warning methods for chemical industrial parks have been proposed with technological advancements, these methods still have many limitations in practical application and cannot meet the urgent needs of modern chemical industrial parks for "early detection, early warning, and early response" of risks.

[0003] Specifically, existing risk warning technologies can be mainly divided into the following categories: (1) Traditional monitoring-type early warning technology. This type of technology is the basic application scheme for risk warning in existing chemical industrial parks. The core is to achieve early warning by directly sensing the physical or chemical signals of the risk source. By deploying gas sensors, temperature sensors, pressure sensors, etc., data such as the concentration of hazardous chemicals and equipment operating parameters in the park are collected. When the threshold is exceeded, an early warning is triggered. Combined with manual inspection and video monitoring technology, the staff conducts regular on-site inspections and uses fixed or mobile video monitoring equipment to identify and warn of obvious risks (such as open flames and violations). However, this early warning technology has the disadvantages of limited timeliness, insufficient monitoring coverage and stability, and high dependence on manual labor. That is, it can only respond to obvious risks that have occurred and cannot predict hidden equipment failures in advance. Slow leakage and other potential risks make it difficult to achieve "early warning". The deployment cost of sensors is high and they are easily affected by environmental factors such as dust, humidity and extreme temperature, which can lead to data distortion or false alarms. Manual inspection is inefficient and subjective, and there is a risk of human negligence. (2) Model prediction early warning technology. This type of technology achieves risk prediction based on data deduction or logical analysis. It simulates the risk evolution process by constructing mathematical models or logical frameworks. Specifically, it uses numerical methods such as CFD to simulate the diffusion path of hazardous chemical leaks and the impact range of fire and explosion, and assists in risk level assessment. Or, based on historical accident data and equipment operation data, it quantifies the probability of risk occurrence and the degree of impact through regression analysis, Bayesian network, risk matrix and other models. However, this early warning technology has the disadvantages of strong data dependence, poor model adaptability and high difficulty in engineering application.

[0004] Therefore, there is a need to provide a dynamic early warning method for safety production risks in chemical industrial parks that can predict risks in a timely and accurate manner and adapt to complex scenarios. Summary of the Invention

[0005] Based on the above analysis, the embodiments of the present invention aim to provide a dynamic early warning method for safety production risks in chemical industrial parks, in order to solve the problems of existing early warning methods, such as inability to predict dangers in advance, susceptibility to environmental interference, reliance on manual labor, strong data dependence, and inability to be applied to complex scenarios.

[0006] This invention provides a method for dynamic early warning of safety production risks in chemical industrial parks, including: Construct a risk factor table for chemical industrial parks, which includes multiple primary indicators and multiple secondary indicators corresponding to each primary indicator. For each primary indicator, calculate the corresponding comprehensive score, calculate the subjective weight based on the improved AHP method, and calculate the objective weight based on the entropy weight method of the dynamic variational autoencoder. The subjective weights and the objective weights are fused to obtain a comprehensive weight; The risk level of the chemical industrial park is calculated based on the comprehensive score of each primary indicator and its corresponding comprehensive weight. If the risk level is higher than the preset threshold, a risk warning for the park is issued.

[0007] Based on a further improvement to the above method, the calculation of subjective weights based on the improved AHP method includes: S221: Obtain the risk factor table of the chemical industrial park corresponding to the previous N data collection times at the current time. For the risk factor table at each time point, perform the following operations: S2211: Obtain the aggregated value of each primary indicator, wherein the aggregated value is the average of the representative values ​​of all secondary indicators under the primary indicator; S2212: For any two primary indicators, calculate the importance ratio of that primary indicator to the corresponding one; S222: For each primary indicator pair, construct a numerical distribution sequence based on the importance ratio of the indicator pair in the previous N collection times, use the importance ratio corresponding to the 5th percentile in the numerical distribution sequence as the initial lower limit of the indicator pair, and use the importance ratio corresponding to the 95th percentile as the initial upper limit of the indicator pair. S223: Construct an interval fuzzy judgment matrix based on the initial lower limit and initial upper limit of each first-level indicator pair; S224: Perform consistency verification and correction on the interval fuzzy judgment matrix, and determine the target lower limit and target upper limit for each primary indicator based on the corrected interval fuzzy judgment matrix; S225: Calculate the target lower limit and target upper limit for each primary indicator in the interval fuzzy judgment matrix to obtain the weight interval vector corresponding to each primary indicator; S226: Obtain the subjective weight corresponding to each primary indicator based on the weight interval vector corresponding to each primary indicator.

[0008] Based on a further improvement of the above method, the step of calculating the importance ratio of any two primary indicators to their corresponding primary indicators includes: , Where j and k are different primary indicators, Let j and k be the importance ratio between the two primary indicators. The aggregated value of the index corresponding to index j. This represents the aggregated value of the indicator corresponding to indicator k. .

[0009] Based on a further improvement to the above method, the step of performing consistency verification and correction on the interval fuzzy judgment matrix, and determining the corresponding target lower limit and target upper limit for each primary indicator based on the corrected interval fuzzy judgment matrix, includes: S2251: Constructing a minimum adjustment optimization model: , For any pair of first-level indicators j and k, the following constraints must be satisfied: Reciprocity constraint: , , Boundary constraints: , Transitivity constraints: , , in, This is the corrected interval fuzzy judgment matrix. This refers to the lower limit of the target for j and k in the first-level index of the corrected interval fuzzy judgment matrix. Let i be the target upper limit of the first-level indicator for j and k in the corrected interval fuzzy judgment matrix, and i be a first-level indicator that is different from j and k. S2252: The minimum adjustment optimization model is solved by using an iterative adjustment algorithm. In each iteration, the constraint conditions are checked. If they are not met, the lower limit and upper limit of the index pair that do not meet the constraint conditions are corrected. The iteration is repeated until the total deviation is less than the preset threshold. The lower limit and upper limit of each index pair after the iteration are taken as the target lower limit and target upper limit of the index pair.

[0010] Based on a further improvement of the above method, the step of calculating the weight interval vector corresponding to each primary indicator based on the lower and upper limits of the target for each primary indicator in the interval fuzzy judgment matrix includes: For any primary metric, perform the following operations: S2251: Initialize the weight interval vector; S2252: For the t-th iteration, the update formula for the weight interval vector is: , in, , Let be the lower and upper bounds of the weight of index j in the (t-1)th iteration, respectively. Let be the weight interval vector corresponding to the t-th iteration index j; S2253: Normalize the weight interval vector after each iteration; S2254: Determine whether convergence has occurred based on the convergence condition. If convergence has not occurred, return to step S2252; if convergence has occurred, proceed to step S2255. The convergence condition is: , in, Let j be the weight interval vector after normalization for the index j in the t-th iteration. Let j be the normalized weight interval vector of index j in the (t-1)th iteration. To take the mean function, where, , This is the convergence threshold; S2255: After convergence The weight interval vector of index j .

[0011] Based on a further improvement of the above method, the calculation of objective weights using the entropy weight method based on a dynamic variational autoencoder includes: S227: For any first-level indicator, perform the following operation: S2271: Obtain the risk factor table of the chemical industrial park at the current moment, and use the secondary indicator values ​​under the primary indicator as input to the dynamic variational autoencoder to obtain the latent variables corresponding to each secondary indicator. , The mean, Let p = 1, 2, 3, ..., N, representing the variance. k N k This represents the number of secondary indicators under this primary indicator; S2272: Based on the latent variables corresponding to each secondary indicator, the dynamic weight corresponding to that secondary indicator is calculated using the entropy weight method: ; S2273: For any secondary indicator, its dynamic objective weight is obtained through the following formula: , h=1,2,3,…,N k N kThis represents the number of secondary indicators under this primary indicator; S2274: The objective weight of a primary indicator is obtained based on the dynamic objective weights of all secondary indicators under the primary indicator. ; S228: Repeat step S227 until all primary indicators have obtained their corresponding objective weights.

[0012] A further improvement to the above method, the fusion of the subjective weights and the objective weights to obtain a comprehensive weight, includes: , in, The comprehensive weight of the first-level indicator j, The subjective weight of the first-level indicator j, The objective weight of the primary indicator j, This is the fusion coefficient.

[0013] Based on further improvements to the above method, λ=0.7 is set in scenarios emphasizing expert experience; and λ=0.3 is set in scenarios with high data quality and high automation. Based on further improvements to the above method, the risk level of the chemical industrial park is calculated based on the comprehensive score and comprehensive weight of each primary indicator, including: S41: Calculate the risk value of the chemical industrial park at the current moment: , in, This is the comprehensive score corresponding to the primary indicator j; S42: If the risk value A value greater than 90 indicates an urgent risk; if the risk value is... A value greater than 70 indicates high risk; if the risk value is... If the value is greater than 50, it is considered ordinary risk; if the risk value is... A value less than 50 indicates low risk.

[0014] Based on further improvements to the above methods, the primary indicators include macro-environmental indicators, inherent risk indicators, safety management indicators, online monitoring indicators, and emergency response capability indicators; among them, The secondary indicators corresponding to the macro-environmental indicators include the park's geographical layout map, surrounding population density, real-time meteorological parameters, and terrain slope data; The secondary indicators corresponding to the inherent risk indicators include equipment corrosion index, chemical storage volume, historical accident heat map, and equipment operating pressure level; The secondary indicators corresponding to the safety management indicators include the completeness of emergency plan documents, safety training records, inspection task completion rate, and response time for hazard rectification. The secondary indicators corresponding to the online monitoring indicators include the average temperature sensor value, gas concentration residual, real-time abnormal events, and number of pressure sudden changes. The secondary indicators corresponding to the emergency response capability indicators include the amount of fire-fighting materials in reserve, emergency drill records, real-time personnel location density, fire truck arrival time, and the unobstructed status of emergency access routes.

[0015] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects: This invention provides a dynamic early warning method for safety production risks in chemical industrial parks. By categorizing available parameters of chemical industrial parks into five types and conducting risk assessments based on these five types of indicators, a standardized risk early warning method applicable to various types of chemical industrial parks is provided. Furthermore, in implementing risk early warning, all parameters are standardized, and the weights of each indicator are calculated using an improved AHP (Advanced Personality Hierarchy Process) and entropy weight method. This not only reduces data dependence and eliminates the influence of human experience but also improves the accuracy of risk prediction.

[0016] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0017] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. Figure 1 This is an example diagram illustrating a dynamic early warning method for safety production risks in chemical industrial parks, provided as an embodiment of the present invention. Detailed Implementation

[0018] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0019] A specific embodiment of the present invention discloses a method for dynamic early warning of safety production risks in chemical industrial parks, such as... Figure 1 As shown, the method includes: S1: Construct a risk factor table for the chemical industrial park. The risk factor table includes multiple primary indicators and multiple secondary indicators corresponding to each primary indicator.

[0020] Among them, the chemical industrial park risk factor table is summarized based on the acquisition difficulty and importance degree of each available parameter index in the existing chemical industrial park. Specifically, the risk factor table consists of five primary indicators, including macro-environment indicators, inherent risk indicators, safety management indicators, online monitoring indicators, and emergency response capacity indicators. Each primary indicator includes multiple secondary indicators. Specifically, the macro-environment indicators include secondary indicators such as the park geographical layout map, surrounding population density, real-time meteorological parameters, and terrain slope data. The inherent risk indicators include secondary indicators such as equipment corrosion index, chemical storage volume, historical accident heat map, and equipment operating pressure level. The safety management includes secondary indicators such as the completeness of the emergency plan document, safety training records, inspection task completion rate, and hidden danger rectification response time. The online monitoring indicators include secondary indicators such as the average value of temperature sensors, gas concentration residuals, real-time abnormal events, and pressure mutation times. The emergency response capacity indicators include secondary indicators such as the reserve quantity of fire-fighting materials, emergency drill records, real-time positioning density of personnel, arrival time of fire trucks, and the unobstructed state of the emergency passage.

[0021] Among them, the present invention further classifies the data types of the above secondary indicators into static data and dynamic data. This classification method is to adopt different data preprocessing methods for different data types, thereby improving the efficiency of risk early warning.

[0022] Exemplarily, for the text data in the static data, including unstructured text (such as accident reports, process documents, regulatory provisions (such as HAZOP analysis reports, etc.)), semi-structured data (such as equipment inventory tables, SDS safety specifications, etc.), and multi-source heterogeneous data (such as sensor logs, emergency drill records, etc.), all need to be preprocessed to obtain standardized inputs.

[0023] To avoid the interference of invalid words in model training, cleaning and filtering are first carried out: (1) Format standardization processing (such as PDF to Markdown, table structure extraction); (2) Remove special characters (such as #%&) and stop words (such as "of", "is") through regular expressions to ensure the retention of core information; (3) Split continuous text into lexical units, using Jieba / HanLP for Chinese and Spacy for English to provide structured units for entity recognition and semantic analysis; (4) Eliminate the impact of morphological changes on the model and unify words into their original forms (such as "operating→operate"). Then, a standardized word sequence (such as ["reaction kettle", "T-101", "temperature", ">200℃"]) is output through preprocessing. Entity recognition (NER) is carried out according to the word sequence, which is a technology to automatically identify and classify key elements from the text (such as equipment names, process parameters), and convert unstructured text into computable structured data, providing an information cornerstone for risk assessment.

[0024] For complex entity recognition in long texts, such as the combination of device ID and parameters, the BiLSTM-CRF model is first used to solve the local context dependency problem. The context dependency is captured by bidirectional LSTM and the label transfer is constrained by CRF (e.g., BIO label: B-device, I-device). Then, BERT is fine-tuned to understand professional semantics and terminology associations: domain data (e.g., chemical equipment entity library) is injected based on pre-trained models (e.g., BERT).

[0025] For hazardous chemical industry entity types: equipment ("Reaction vessel T-101", "Storage tank V-200"), process parameters (such as "Temperature > 200℃", "Pressure rating PN16"), and risk words (such as "Seal failure", "Corrosion rate ≥ 0.5mm / year"), the output is: a list of triples (entity, type, position), for example: ("T-101", "Equipment", [12, 17]) (meaning that the entity "T-101" is located at the 12th-17th character position in the text).

[0026] After identifying isolated entities (e.g., "storage tank" and "temperature"), relationships are extracted to construct dynamic associations between equipment and risks, revealing the risk transmission logic (e.g., "storage tank leakage → gas diffusion → explosion"). First, dependency parsing extracts subject-verb-object structures (e.g., "reactor T-101 → existence → overtemperature risk"). Then, agent / patient relationships are identified (e.g., "[nitration reaction] leads to [explosion]"). The generalization ability of large models is used to handle complex sentence structures, and structured relationships are generated through LLM (Prompt example: "Extract entity relationships, output JSON"), thus obtaining key hazardous chemical relationships, including equipment-risk (e.g., "storage tank → leakage → toxicity diffusion") and measures-effects (e.g., "temperature chain → inhibition → overpressure").

[0027] After obtaining the causal and dependency relationships between entities, fragmented information is integrated to construct a global dynamic knowledge network that supports risk reasoning and real-time updates. First, entities are disambiguated and aligned to resolve the issue of multiple names for the same entity (e.g., "R-201 = nitration reactor"), avoiding knowledge redundancy. Then, a graph database is used for storage, employing Neo4j to store nodes (equipment / risk) and edges (relationships), supporting efficient queries (e.g., using Cypher to retrieve equipment-related risks). Finally, a dynamic update mechanism is set up to periodically update maintenance records and associate them with data from different timestamps (e.g., a 2025 inspection report), ensuring that the knowledge graph reflects the latest status of equipment and risks in real time.

[0028] To quantify its compliance with security standards, BERT vectors are bound to national standard documents, achieving an end-to-end mapping of "regulatory provisions - semantic vectors - risk indicators," avoiding manual line-by-line verification and enabling automated compliance review.

[0029] For the preprocessed text entities, a domain-fine-tuned BERT model (injected with professional corpora such as chemical equipment libraries and accident reports) is used to obtain initial vectors: , Then, by using the entity's position information in the original text (such as [12,17] in the NER output), a local context window (20 words before and after) is extracted, and the contextual semantics are fused: , For national standard clauses (such as the sealing structure requirements in GB / T 34525-2017), the clauses are first broken down into "subject-condition-constraint" triples, such as "pressure vessel", "wall thickness inspection", "≥1 time every 3 years"), and then the same BERT model (sharing weights with Edoc) is used to generate the base vectors: , Add a rule weight matrix Wrule (based on the clause type preset): .

[0030] For example, calculating the compliance score (CS) between the document to be evaluated and the standard document is a composite function of semantic similarity and rule weights:

[0031] The compliance matching degree is a composite function of semantic similarity and rule weight. SimBERT is the semantic similarity output by the BERT model (range [0,1]), which is obtained by calculating the cosine similarity between the document entity vector Edoc and the national standard clause vector EGB. wi is the weight of the i-th rule (determined by the vulnerability type), for example, for incident reporting, Increase by 0.3 to highlight the risk of historical vulnerabilities recurring; for equipment ledgers, remove... (Items using only similarity to avoid overweighting static parameters; for emergency plans, add a timeliness factor.) This is used to penalize expired files (t = number of years). Ii is an indicator function (1 if document matching rule i is used, 0 otherwise). For example, for spatial data, the park is divided into an N×N grid (e.g., 100m×100m), and the number of hazardous sources within each grid is counted. Equipment density is calculated to quantify the concentration of risk in the area.

[0032] For example, the equipment type coefficient can be set in advance (e.g., tank = 0.9, pipeline = 0.7).

[0033] For example, the weighting of historical incidents:

[0034] For dynamic data, the process involves transforming real-time sensor data (e.g., temperature, pressure, gas concentration) and video surveillance alarm signals within the chemical industrial park into standardized, quantifiable dynamic indicators, providing input for subsequent risk weight calculations and spatiotemporal modeling. The process can be divided into four main stages: data classification, time-series statistics, anomaly detection, and indicator standardization, as detailed below: (1) Classification and source of dynamic data. Dynamic data is divided into two categories, and their sources and characteristics are shown in Table 1 below. Sensor data must be accompanied by timestamps and spatial location information. Sensor data is bound to device / area IDs according to the nearest neighbor principle. Event-type data are discrete abnormal events that do not participate in sliding window statistics. They are directly mapped to preset risk impact chain nodes and the transmission path is updated in real time.

[0035] Table 1. Sources, processing methods, and characteristic descriptions of dynamic data

[0036] (2) Time-series statistics. For the sensor time-series data, the sliding window method is used to calculate statistical indicators, eliminate noise, and extract key features. The specific steps are as follows: Sliding window division: Window length: dynamically adjusted according to data frequency (e.g., if the temperature sensor sampling frequency is 5 seconds, the window length is set to 5 minutes, covering 60 data points). Window sliding step size: usually 1 / 3 of the window length (e.g., sliding once every 1 minute), balancing computational efficiency and real-time performance.

[0037] Statistical indicator calculation: For the data within each window, calculate the following indicators: Mean: Reflects the average level of data within the window. , in, The data value at the current moment: the i-th data point in the time series.

[0038] Variance: Measures the dispersion of data and reflects its volatility. .

[0039] Mutation count: Counts the number of times the difference between adjacent data points exceeds a threshold, identifying mutation events.

[0040] in, For a preset threshold (such as a temperature change threshold) =5 ), This is an indicator function: if the condition within the parentheses is true, the value is 1; otherwise, it is 0.

[0041] Residual: For each time point t within the window, calculate the absolute difference between the actual concentration and the reference concentration.

[0042] in, Determined based on process parameters or historical normal operating data.

[0043] (3) Anomaly Detection. Anomaly points or patterns are identified through time series decomposition and statistical testing to avoid noise interference in risk calculations. STL decomposition (Seasonal-Trend Decomposition) and the residual threshold method are used, as detailed below: STL Decomposition: When data exhibits periodic or trend-like changes (such as seasonal temperature fluctuations), directly using a fixed baseline can lead to false alarms. Time series decomposition methods, such as STL decomposition, can break down the original time series data into a trend component, a seasonal component, and a residual component.

[0044] Among them, the trend term is used to represent long-term change trends (such as a slow rise in temperature caused by equipment aging), the seasonal term is used to represent periodic fluctuations (such as periodic temperature changes caused by the temperature difference between day and night), and the residual term is used to represent the remaining part after removing the trend and seasonality. By decomposing and stripping away predictable patterns, the residual abnormally reflects the real sudden events (such as sudden equipment failure or network attack) more purely, reflecting abnormal fluctuations.

[0045] Residual threshold alarm. The standard deviation of the residuals is calculated. Based on the normal distribution characteristics, 99.7% of the data points fall within this range, and the probability of exceeding this range is extremely low. If the absolute value of the residual at any point exceeds this range, it is marked as abnormal.

[0046] (4) Standardization of dynamic indicators. The time-series statistics and anomaly detection results are converted into standardized indicator values ​​to eliminate dimensional differences and unify the numerical range ([0,1]). The specific method is as follows: Normalization is performed using linear normalization for each dynamic indicator (such as mean temperature, residual, and number of abrupt changes):

[0047] in, and These are the minimum and maximum values ​​of historical data (such as the lowest and highest average temperatures over the past 30 days).

[0048] Abnormal event quantification involves setting dynamic thresholds based on sensor type (e.g., temperature, pressure, gas concentration) and process requirements. If a certain window meets a preset threshold (e.g., gas concentration residual > 0.8), a binary index (0 / 1) is generated.

[0049] in, Anomaly detection threshold (e.g., number of stress mutations) =3 times / window).

[0050] Examples of dynamic data output are shown in Table 2.

[0051] Table 2. Examples of Dynamic Data Output

[0052] In this study, temperature is measured using the average value because temperature in chemical production is typically a slowly changing continuous variable (e.g., reactor temperature, ambient temperature). Single fluctuations may be caused by measurement noise or brief disturbances, while the average temperature over a long time window better reflects the true thermal state of the equipment. Gas concentration is measured using the residual value because the safety of gas concentrations (e.g., toxic gases, flammable gases) needs to be considered in conjunction with their baseline values ​​(expected concentrations under normal operating conditions), and the residual value directly reflects the degree of abnormal deviation. Pressure is measured using the number of abrupt changes because pressure in chemical systems often experiences instantaneous changes (e.g., sudden pressure increases or decreases) due to valve opening / closing, pump start / stop operations, etc. A single abrupt change may lead to equipment fatigue or seal failure, requiring real-time monitoring.

[0053] As shown in Table 3, it provides the data types, data sources, and preprocessing methods for each primary indicator and its secondary indicators.

[0054] Table 3. Data types, data sources, and preprocessing methods for various indicators within the chemical industrial park.

[0055] S2: For each primary indicator, calculate the comprehensive score corresponding to that indicator, calculate the subjective weight based on the improved AHP method, and calculate the objective weight based on the entropy weight method of the dynamic variational autoencoder.

[0056] For each primary indicator, its comprehensive score is the standardized mean of all secondary indicators under that primary indicator. Of course, for certain special indicators, adjustments can be made to further improve the accuracy of the prediction. For example, for a certain static equipment, the static inherent risk value can be calculated by combining the equipment type coefficient, corrosion correction factor (based on the annual corrosion rate), and management score (such as the completeness of the emergency plan). , Corrosion correction: (The carbon steel material is doubled). Furthermore, whether to adjust the specifications of a certain piece of equipment is set by the technicians according to actual needs.

[0057] The method proposed in this invention calculates two types of weights for each primary indicator: subjective weights and objective weights. This approach overcomes the inherent limitations of using either subjective or objective weighting methods alone, resulting in a more scientific, comprehensive, and reliable risk assessment outcome that better meets the actual needs of risk management in chemical industrial parks. Specifically, the subjective weights (i.e., the improved Analytic Hierarchy Process, AHP) incorporate domain knowledge to ensure the "logical rationality" of the assessment and reflect management orientation and decision-making intent, such as prioritizing the control of a certain type of risk within a specific timeframe. The objective weights (i.e., the improved entropy weighting method) are entirely data-driven, reflecting the distribution characteristics and interrelationships of the indicator data themselves, revealing information variability, and identifying which indicators show the greatest difference and strongest discriminative power in the assessment sample. Combining subjective and objective weights enhances the predictive power of the results, significantly improving their stability and reliability.

[0058] The calculation of subjective weights based on the improved AHP method includes: S221: Obtain the risk factor table of the chemical industrial park corresponding to the previous N data collection times at the current time. For the risk factor table at each time point, perform the following operations: S2211: Obtain the aggregated value of each primary indicator, where the aggregated value is the average of the representative values ​​of all secondary indicators under that primary indicator.

[0059] The representative value of a secondary indicator refers to the average value of the secondary indicator for the same type of equipment at the same data collection time. Understandably, within the same chemical industrial park, there may be multiple pieces of equipment of the same type; the average value of the indicator parameters of that same type of equipment is used as the representative value of that secondary indicator.

[0060] S2212: For any two primary indicators, calculate the importance ratio of that primary indicator to the corresponding primary indicator, including: , Where j and k are different primary indicators, Let j and k be the importance ratio between the two primary indicators. The aggregated value of the index corresponding to index j. This represents the aggregated value of the indicator corresponding to indicator k. .

[0061] As can be seen, step S221 obtains the importance ratio of each primary indicator at different collection times through the historical chemical industrial park risk factor table. This importance ratio is based on the aggregated value of the corresponding primary indicator.

[0062] For example, N can be 10, meaning that based on the current data collection time, the risk factor table of the chemical industrial park corresponding to the previous 10 data collection times is obtained. The value of N can be determined by technicians according to actual needs.

[0063] S222: For each primary indicator pair, construct a numerical distribution sequence based on the importance ratio of the indicator pair in the previous N collection times. Use the importance ratio corresponding to the 5th percentile in the numerical distribution sequence as the initial lower limit of the indicator pair, and use the importance ratio corresponding to the 95th percentile as the initial upper limit of the indicator pair.

[0064] This step requires sorting the importance ratios obtained in step S221. This sorting is based on the importance ratios of the same level of indicators. For example, if the current level of indicators are macroeconomic environment indicators and inherent risk indicators, then the importance ratios of the macroeconomic environment indicators and inherent risk indicators in step S221 are sorted. Of course, the order of the indicators must also be considered to ensure that the importance ratios of the macroeconomic environment indicators and inherent risk indicators satisfy a reciprocal constraint with those of the inherent risk indicators and macroeconomic environment indicators. This invention uses the fluctuation range of statistical indicators, with the importance ratio corresponding to the 5th percentile as the initial lower limit of the indicator pair and the importance ratio corresponding to the 95th percentile as the initial upper limit. If no corresponding importance ratio exists at a given percentage, the average of the two values ​​before and after that percentile is taken as the importance ratio corresponding to that percentile.

[0065] S223: Construct an interval fuzzy judgment matrix based on the initial lower limit and initial upper limit of each first-level indicator pair.

[0066] S224: Perform consistency verification and correction on the interval fuzzy judgment matrix, and determine the corresponding target lower limit and target upper limit for each primary indicator based on the corrected interval fuzzy judgment matrix, including: S2251: Constructing a minimum adjustment optimization model: , For any pair of first-level indicators j and k, the following constraints must be satisfied: Reciprocity constraint: , , Boundary constraints: , Transitivity constraints: , , in, This is the corrected interval fuzzy judgment matrix. This refers to the lower limit of the target for j and k in the first-level index of the corrected interval fuzzy judgment matrix. Let i be the target upper limit of the first-level indicator for j and k in the corrected interval fuzzy judgment matrix, and i be a first-level indicator that is different from j and k. S2252: The minimum adjustment optimization model is solved by using an iterative adjustment algorithm. In each iteration, the constraint conditions are checked. If they are not met, the lower limit and upper limit of the index pair that do not meet the constraint conditions are corrected. The iteration is repeated until the total deviation is less than the preset threshold. The lower limit and upper limit of each index pair after the iteration are taken as the target lower limit and target upper limit of the index pair.

[0067] The constraint condition is that the change in all first-level indicators between the two consecutive times is less than 0.001.

[0068] Specifically, the iterative process is as follows: (1) Initialization Set the initial value of the corrected matrix to the original interval fuzzy judgment matrix, that is... , .

[0069] (2) Calculate the initial total deviation Calculate the initial total adjustment based on the objective function. Initially, .

[0070] (3) Constraint checking and correction Verify whether the iteration process satisfies the reciprocity constraint, boundary constraint, and transitivity constraint.

[0071] If the reciprocity constraint is not satisfied, then let , And perform step (4).

[0072] If the boundary constraints are not satisfied, then let , And perform step (4).

[0073] If the transitivity constraint is not satisfied, then let , And perform step (4).

[0074] Where t is the number of iterations.

[0075] (4) Iteration termination judgment Calculate the corrected total deviation, if If the condition is met, stop the iteration; otherwise, return to step (3) to continue the correction until the condition is met.

[0076] (5) Result output After the iteration ends, , This refers to the lower and upper limits of the target for j and k in the primary indicator.

[0077] S225: Based on the lower and upper limits of the target corresponding to each primary indicator in the interval fuzzy judgment matrix, calculate the weight interval vector corresponding to each primary indicator, including: For any primary metric, perform the following operations: S2251: Initialize the weight interval vector, i.e. .

[0078] S2252: For the t-th iteration, the update formula for the weight interval vector is: , in, , Let be the lower and upper bounds of the weight of index j in the (t-1)th iteration, respectively. Let be the weight interval vector corresponding to the t-th iteration index j.

[0079] S2253: After each iteration, the weight interval vector is normalized, including: , in, Let be the normalized weight interval vector of the first-level indicator j at the t-th iteration, and m be the number of first-level indicator pairs. The lower limit weight of the primary indicator j in the t-th iteration. Let be the upper limit weight of the first-level indicator j in the t-th iteration.

[0080] S2254: Determine whether convergence has occurred based on the convergence condition. If convergence has not occurred, return to step S2252; if convergence has occurred, proceed to step S2255. The convergence condition is: , in, Let j be the weight interval vector after normalization for the index j in the t-th iteration. Let j be the normalized weight interval vector of index j in the (t-1)th iteration. To take the mean function, where, , Let the normalized lower bound weight of index j be the weight of index j in the t-th iteration. Let j be the normalized upper limit weight of index j in the t-th iteration. This is the convergence threshold.

[0081] For example, .

[0082] S2255: After convergence The weight interval vector of index j .

[0083] S226: Based on the weight interval vector corresponding to each primary indicator, the subjective weight corresponding to that primary indicator is obtained, including: , This step involves assigning a weight interval vector to each primary indicator. The interval is removed, transforming it into a single, defined point value. This process uses the geometric mean method to calculate the lower limit of the interval. and the upper limit of the interval The geometric mean not only retains the information of the upper and lower limits, but also reflects the central tendency within the range of uncertainty.

[0084] Traditional entropy weighting methods rely on data from a single point in time, failing to capture dynamic changes (such as sudden anomalies in "equipment vibration frequency"). Therefore, a Dynamic Variational Autoencoder (DVAE) is used to analyze the spatiotemporal evolution of the data, extract latent variables (similar to "hidden risk patterns"), and then calculate weights. Since the primary indicator data are aggregated values, they lose the original data's distribution characteristics, while the entropy weighting method relies on the degree of variation of the indicators (i.e., the dispersion of the data distribution) to determine the weights. If the entropy weighting method is applied to primary indicators, the input is the aggregated value of the primary indicator within each time window (e.g., the "safety management" score). These aggregated values ​​have already smoothed the fluctuations of the underlying secondary indicators, causing the entropy weighting method to fail to capture fine-grained data changes. All secondary indicators are converted into an input matrix Z for the DVAE. This invention assumes that all secondary indicators are continuous time-series data after preprocessing, suitable for differential entropy calculation based on the Gaussian assumption. By capturing the spatiotemporal evolution characteristics of the indicator data through the DVAE, reconstructing the data distribution, and calculating entropy weights, the problem of insufficient sensitivity to data distribution in traditional entropy weighting methods is solved.

[0085] Among them, the entropy weight method based on dynamic variational autoencoder is used to calculate objective weights, including: S227: For any first-level indicator, perform the following operation: S2271: Obtain the risk factor table of the chemical industrial park at the current moment, and use the secondary indicator values ​​under the primary indicator as input to the dynamic variational autoencoder to obtain the latent variables corresponding to each secondary indicator. , The mean, Let p = 1, 2, 3, ..., N, representing the variance. k N k This represents the number of secondary indicators under this primary indicator.

[0086] The standardized secondary indicator values ​​are input as a window of data into the encoder to obtain the latent variable distribution parameters (mean). and variance ),Right now ,in, The hidden state from the previous time step is used for temporal dependency modeling. Then, the latent variables are... The reconstructed data is obtained through the decoder. By comparing the original data With reconstructing data The differences (reconstruction error) ensure that latent variables truly capture the essential characteristics of the data, providing reliable input for subsequent entropy weight calculations. Furthermore, when data in a certain window is missing, latent variables are used to reconstruct the data. Replace the original data: The loss function of the dynamic variational autoencoder is as follows: Reconstruction loss: KL divergence loss: The total loss is: ,in, This is a balancing coefficient used to adjust the trade-off between reconstruction accuracy and the normality of latent variable distribution.

[0087] For example, Debugging can be performed within the range of [0.01, 0.5]. An initial debugging setting of 0.1 is recommended. S2272: Based on the latent variables corresponding to each secondary indicator, the dynamic weight corresponding to that secondary indicator is calculated using the entropy weight method: , The greater the entropy, the more unpredictable the changes in the indicator, and the lower the weight.

[0088] S2273: For any secondary indicator, its dynamic objective weight is obtained through the following formula: , h=1,2,3,…,N k N k This represents the number of secondary indicators under this primary indicator.

[0089] S2274: The objective weight of a primary indicator is obtained based on the dynamic objective weights of all secondary indicators under the primary indicator. .

[0090] S228: Repeat step S227 until all primary indicators have obtained their corresponding objective weights.

[0091] S3: The subjective weights and the objective weights are fused to obtain a comprehensive weight, including: , in, The comprehensive weight of the first-level indicator j, The subjective weight of the first-level indicator j, The objective weight of the primary indicator j, This is the fusion coefficient.

[0092] For example, in scenarios that emphasize expert experience, λ is set to 0.7; in scenarios with high data quality and high degree of automation, λ is set to 0.3.

[0093] Preferably, it can be set .

[0094] S4: The risk level of the chemical industrial park is calculated based on the comprehensive score of each primary indicator and its corresponding comprehensive weight. If the risk level is higher than a preset threshold, a risk warning for the park is issued, including: S41: Calculate the risk value of the chemical industrial park at the current moment: , in, This is the comprehensive score corresponding to the primary indicator j.

[0095] S42: If the risk value A value greater than 90 indicates an urgent risk; if the risk value is... A value greater than 70 indicates high risk; if the risk value is... If the value is greater than 50, it is considered ordinary risk; if the risk value is... A value less than 50 indicates low risk.

[0096] The preset threshold is 50. That is, when the risk value exceeds 50, an early warning notification will be issued to the technical staff.

[0097] Furthermore, step S4 can also achieve automatic identification and extraction of key indicators, that is, automatically identify the dominant risk factors, sort them by contribution, and improve the efficiency of positioning. For example, if the "equipment corrosion index" has the highest contribution in the "inherent risk" dimension, then the key indicator output is the "equipment corrosion index", and its calculation method is as follows:

[0098] Where p is a secondary indicator. 0.4 is the real-time value of indicator p, and 0.4 is the contribution threshold coefficient (an empirical parameter that is adjusted according to the actual scenario and used to screen for significant risk factors).

[0099] Furthermore, this invention also develops a visualization system for a dynamic early warning method for safety production risks in chemical industrial parks. This visualization system can display information visually according to different risk levels, as shown in Table 4. Table 4 Risk Level - Color Dynamic Mapping Table

[0100] The visualization system includes: (1) Global risk heat map Technology stack: React + MapboxGL + Deck.gl Features: Dynamic heatmap (maps risk values ​​to color intensity and updates in real time); diffusion impact range simulation (combines a lightweight CFD proxy model to display the direction of gas diffusion with particle animation); click-to-interact (view details of each device or area, such as risk score, key indicators, emergency resources, etc.).

[0101] (2) Augmented Reality View Technology stack: Unity + ARFoundation Functions: Risk area marking (overlaying high-risk areas (red semi-transparent layer) on camera footage); navigation guidance (green arrows marking the nearest safe exit); voice alarms (real-time announcements, such as "50 meters northeast, storage tank leak, please evacuate!"). (3) Health monitoring of 3D equipment Technology stack: Three.js + Blender model Functions: Equipment coloring (dynamically rendering colors based on corrosion index, such as a green to red gradient); anomaly location (highlighting abnormal equipment in red and displaying real-time sensor data curves); profile analysis (cutting the equipment model to view internal risk points, such as weld corrosion).

[0102] (4) Interpretability analysis panel Technology stack: D3.js + SHAP library Features: Contribution Waterfall Chart (displays the top 5 metrics by SHAP value and their contribution); Weight Conflict Warning (compares the weight differences between AHP and entropy weight method, highlighting metrics with a difference >30%); Comprehensive Report (displays all management and rating metrics).

[0103] Compared with existing technologies, this embodiment provides a dynamic early warning method for safety production risks in chemical industrial parks. By categorizing available parameters of chemical industrial parks into five types and conducting risk assessments based on these five types of indicators, it provides a standardized risk early warning method applicable to various types of chemical industrial parks. Furthermore, in implementing risk early warning, all parameters are standardized, and the weights of each indicator are calculated using an improved AHP and entropy weight method. This not only reduces data dependence and eliminates the influence of human experience but also improves the accuracy of risk prediction.

[0104] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0105] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for dynamic early warning of safety production risks in chemical industrial parks, characterized in that, include: Construct a risk factor table for chemical industrial parks, which includes multiple primary indicators and multiple secondary indicators corresponding to each primary indicator. For each primary indicator, calculate the corresponding comprehensive score, calculate the subjective weight based on the improved AHP method, and calculate the objective weight based on the entropy weight method of the dynamic variational autoencoder. The subjective weights and the objective weights are fused to obtain a comprehensive weight; The risk level of the chemical industrial park is calculated based on the comprehensive score of each primary indicator and its corresponding comprehensive weight. If the risk level is higher than the preset threshold, a risk warning for the park is issued.

2. The method for dynamic early warning of safety production risks in chemical industrial parks according to claim 1, characterized in that, The calculation of subjective weights based on the improved AHP method includes: S221: Obtain the risk factor table of the chemical industrial park corresponding to the previous N data collection times at the current time. For the risk factor table at each time point, perform the following operations: S2211: Obtain the aggregated value of each primary indicator, wherein the aggregated value is the average of the representative values ​​of all secondary indicators under the primary indicator; S2212: For any two primary indicators, calculate the importance ratio of that primary indicator to the corresponding one; S222: For each primary indicator pair, construct a numerical distribution sequence based on the importance ratio of the indicator pair in the previous N collection times, use the importance ratio corresponding to the 5th percentile in the numerical distribution sequence as the initial lower limit of the indicator pair, and use the importance ratio corresponding to the 95th percentile as the initial upper limit of the indicator pair. S223: Construct an interval fuzzy judgment matrix based on the initial lower limit and initial upper limit of each first-level indicator pair; S224: Perform consistency verification and correction on the interval fuzzy judgment matrix, and determine the target lower limit and target upper limit for each primary indicator based on the corrected interval fuzzy judgment matrix; S225: Calculate the target lower limit and target upper limit for each primary indicator in the interval fuzzy judgment matrix to obtain the weight interval vector corresponding to each primary indicator; S226: Obtain the subjective weight corresponding to each primary indicator based on the weight interval vector corresponding to each primary indicator.

3. The method for dynamic early warning of safety production risks in chemical industrial parks according to claim 2, characterized in that, For any two primary indicators, the importance ratio of that primary indicator to the corresponding primary indicator is calculated, including: ， Where j and k are different primary indicators, Let j and k be the importance ratio between the two primary indicators. The aggregated value of the index corresponding to index j. This represents the aggregated value of the indicator corresponding to indicator k. .

4. The method for dynamic early warning of safety production risks in chemical industrial parks according to claim 3, characterized in that, The process of performing consistency verification and correction on the interval fuzzy judgment matrix, and determining the corresponding target lower limit and target upper limit for each primary indicator based on the corrected interval fuzzy judgment matrix, includes: S2251: Constructing a minimum adjustment optimization model: ， For any pair of first-level indicators j and k, the following constraints must be satisfied: Reciprocity constraint: , , Boundary constraints: , Transitivity constraints: , , in, This is the corrected interval fuzzy judgment matrix. This refers to the lower limit of the target for j and k in the first-level index of the corrected interval fuzzy judgment matrix. Let i be the target upper limit of the first-level indicator for j and k in the corrected interval fuzzy judgment matrix, and i be a first-level indicator that is different from j and k. S2252: The minimum adjustment optimization model is solved by using an iterative adjustment algorithm. In each iteration, the constraint conditions are checked. If they are not met, the lower limit and upper limit of the index pair that do not meet the constraint conditions are corrected. The iteration is repeated until the total deviation is less than the preset threshold. The lower limit and upper limit of each index pair after the iteration are taken as the target lower limit and target upper limit of the index pair.

5. The method for dynamic early warning of safety production risks in chemical industrial parks according to claim 4, characterized in that, The calculation of the weight interval vector corresponding to each primary indicator based on the lower and upper limits of the target in the interval fuzzy judgment matrix includes: For any primary metric, perform the following operations: S2251: Initialize the weight interval vector; S2252: For the t-th iteration, the update formula for the weight interval vector is: ， in, , Let be the lower and upper bounds of the weight of index j in the (t-1)th iteration, respectively. Let be the weight interval vector corresponding to the t-th iteration index j; S2253: Normalize the weight interval vector after each iteration; S2254: Determine whether convergence has occurred based on the convergence condition. If convergence has not occurred, return to step S2252; if convergence has occurred, proceed to step S2255. The convergence condition is: ， in, Let j be the weight interval vector after normalization for the index j in the t-th iteration. Let j be the normalized weight interval vector of index j in the (t-1)th iteration. To take the mean function, where, , This is the convergence threshold; S2255: After convergence The weight interval vector of index j .

6. The method for dynamic early warning of safety production risks in chemical industrial parks according to claim 5, characterized in that, The entropy weight method based on dynamic variational autoencoder for calculating objective weights includes: S227: For any first-level indicator, perform the following operation: S2271: Obtain the risk factor table of the chemical industrial park at the current moment, and use the secondary indicator values ​​under the primary indicator as input to the dynamic variational autoencoder to obtain the latent variables corresponding to each secondary indicator. , The mean, Let p = 1, 2, 3, ..., N, representing the variance. k N k This represents the number of secondary indicators under this primary indicator; S2272: Based on the latent variables corresponding to each secondary indicator, the dynamic weight corresponding to that secondary indicator is calculated using the entropy weight method: ； S2273: For any secondary indicator, its dynamic objective weight is obtained through the following formula: ， h=1,2,3,…,N k N k This represents the number of secondary indicators under this primary indicator; S2274: The objective weight of a primary indicator is obtained based on the dynamic objective weights of all secondary indicators under the primary indicator. ； S228: Repeat step S227 until all primary indicators have obtained their corresponding objective weights.

7. The method for dynamic early warning of safety production risks in chemical industrial parks according to claim 6, characterized in that, The process of fusing the subjective weights and the objective weights to obtain a comprehensive weight includes: ， in, The comprehensive weight of the first-level indicator j, The subjective weight of the first-level indicator j, The objective weight of the primary indicator j. This is the fusion coefficient.

8. The method for dynamic early warning of safety production risks in chemical industrial parks according to claim 7, characterized in that, In scenarios where expert experience is emphasized, λ is set to 0.7; in scenarios with high data quality and high degree of automation, λ is set to 0.

3.

9. The method for dynamic early warning of safety production risks in chemical industrial parks according to claim 8, characterized in that, The risk level of the chemical industrial park is calculated based on the comprehensive score and comprehensive weight of each primary indicator, including: S41: Calculate the risk value of the chemical industrial park at the current moment: ， in, This is the comprehensive score corresponding to the primary indicator j; S42: If the risk value A value greater than 90 indicates an urgent risk; if the risk value is... A value greater than 70 indicates high risk; if the risk value is... A value greater than 50 indicates ordinary risk; if the risk value is... A value less than 50 indicates low risk.

10. The method for dynamic early warning of safety production risks in chemical industrial parks according to claim 1, characterized in that, The primary indicators include macro-environmental indicators, inherent risk indicators, safety management indicators, online monitoring indicators, and emergency response capability indicators; among them, The secondary indicators corresponding to the macro-environmental indicators include the park's geographical layout map, surrounding population density, real-time meteorological parameters, and terrain slope data; The secondary indicators corresponding to the inherent risk indicators include equipment corrosion index, chemical storage volume, historical accident heat map, and equipment operating pressure level; The secondary indicators corresponding to the safety management indicators include the completeness of emergency plan documents, safety training records, inspection task completion rate, and response time for hazard rectification. The secondary indicators corresponding to the online monitoring indicators include the average temperature sensor value, gas concentration residual, real-time abnormal events, and number of pressure sudden changes. The secondary indicators corresponding to the emergency response capability indicators include the amount of fire-fighting materials in reserve, emergency drill records, real-time personnel location density, fire truck arrival time, and the unobstructed status of emergency access routes.

Images