Production safety accident emergency disposal scheme intelligent generation method and system

Abstract

The invention discloses a production safety accident emergency disposal scheme intelligent generation method and system, and the method comprises the steps: obtaining the spatial-temporal characteristics of historical accident cases, calculating the correlation degree, building an accident prediction network based on the correlation degree, simulating uncertain factors, calculating the similarity between a current accident and the historical cases, determining the prior probability, and carrying out the accident development reasoning. The method comprises the steps of tracing accident hazard degrees of different production links, obtaining an cause chain, generating an emergency resource allocation strategy set by utilizing Markov decision, constructing target function optimization, obtaining an emergency disposal scheme, monitoring production process, equipment state and personnel configuration changes in real time, and dynamically adjusting the emergency disposal scheme to adapt to the changes. According to the method, the historical data, the graph neural network and the Markov decision are combined, accident prediction and emergency resource intelligent allocation are realized, the method can adapt to the change of a production environment in real time, the disposal efficiency and safety are improved, and meanwhile, the method has relatively good interpretability.

Description

Technical Field

[0001] The present invention relates to the field of production safety management, and in particular to a method and system for intelligently generating emergency response plans for production safety accidents. Background Art

[0002] With the acceleration of industrialization, various production activities are becoming increasingly complex, and production safety accidents are frequent, posing a serious threat to people's lives and property and social stability. The preparation of traditional production safety accident emergency response plans often relies on manual experience, which has problems such as low efficiency, lack of pertinence, and lack of dynamic adaptability.

[0003] In the process of emergency response, since the situation at the accident site changes rapidly, it is crucial to quickly and accurately generate a response plan that conforms to the actual situation. However, existing technical means are difficult to comprehensively consider many complex factors in a short period of time, such as the type of accident, the on-site environment, the distribution of personnel and materials, etc., resulting in slow emergency response and poor rescue results.

[0004] Faced with diverse production scenarios and ever-changing safety risks, an intelligent method and system for generating emergency response plans for production safety accidents can use artificial intelligence methods to quickly analyze accident situations, automatically generate scientific, reasonable, efficient and practical emergency response plans, improve the level and capabilities of emergency management of production safety accidents, and minimize accident losses. Summary of the invention

[0005] The purpose of the present invention is to provide a method and system for intelligently generating emergency response plans for production safety accidents.

[0006] To achieve the above object, the present invention is implemented according to the following technical solutions: A first aspect of the present invention provides a method for intelligently generating an emergency response plan for a production safety accident, comprising: Obtain the spatiotemporal characteristics of historical accident cases and calculate the correlation between the spatiotemporal characteristics; Based on the correlation between the spatiotemporal features, an accident prediction network is constructed through a graph neural network, and a disturbance term is added to the accident prediction network to simulate the influence of uncertain factors; Calculate the similarity between the current accident and the historical accident cases, determine the prior probability of the current accident based on the similarity, and perform accident development reasoning on the current accident by combining the prior probability with the accident prediction network; Based on the inference results, the accident hazard levels of different production links are traced back to obtain the cause chain; Taking the causal chain as the emergency response path, using Markov decision making to generate an emergency resource allocation strategy set that satisfies the emergency response path, constructing an objective function for optimization, and obtaining an emergency response plan; The emergency response plan will be updated according to changes in production processes, equipment upgrades and personnel changes.

[0007] As a further method, the method of obtaining the spatiotemporal features of historical accident cases and calculating the correlation between the spatiotemporal features includes: Extract time features from historical accident cases, including production shifts, equipment maintenance cycles, and seasonal periods, and count the accident frequency in each period; extract spatial features, including the functional zoning of the factory, process flow layout, and logistics channel direction; Combine the corresponding time features and spatial features in each historical accident case into a feature vector; Calculate the correlation between spatiotemporal features, the expression is: , in, and are the feature vectors of the i-th and j-th spatiotemporal features, r, s and t represent the number of accident types, the number of production links and the number of environmental factors, respectively. For the accident type , production link is m and environmental factor is n, the characteristic vector and The probability of simultaneous occurrence, and The accident types are , production link is m and environmental factor is n, the characteristic vector and the eigenvector Probability of occurrence.

[0008] As a further method, the method of constructing an accident prediction network through a graph neural network based on the correlation between the temporal and spatial features includes: The spatiotemporal features are used as nodes in the graph, and the correlation degree is used as the connection weight to obtain the accident feature graph; The convolutional layer of the graph neural network is used to extract features and aggregate information from the accident feature graph. Through convolution operations, the feature representation of accident development under different combinations of spatiotemporal features is learned. In the fully connected layer of the graph neural network, a prediction model for the development of an accident is constructed based on the learned feature representation. The initial state information of the accident is used as input, and the development state and trend of the accident at different moments in the future are output, including the severity and scope of the accident. The graph neural network is trained using historical accident case data. By adjusting the network parameters, the error between the prediction results and the accident development situation is minimized to obtain an accident prediction model.

[0009] As a further method, the method of adding a disturbance term to the accident prediction network to simulate the influence of uncertain factors includes: Through the probabilistic reasoning model of Bayesian network, the probability distribution of uncertain factors on the development of accidents is quantified; A disturbance term generation model is constructed through generative adversarial networks, which takes historical accident data and the probability distribution of uncertain factors on accident development as input to generate disturbance terms. The generated disturbance term is integrated into the accident prediction network through the attention mechanism. Specifically, the similarity score between the disturbance term and the features of each layer of the accident prediction network is used as the attention weight, the features are weighted summed up, and the disturbance term is integrated.

[0010] As a further method, the method of calculating the similarity between the current accident and the historical accident cases and determining the prior probability of the current accident based on the similarity includes: Calculate the similarity between the current accident and the historical accident case. The expression is: , in, and are the feature vectors of the current accident and historical accident case i respectively, q is the total number of feature dimensions, is the weight coefficient of the jth feature dimension, is the balance coefficient of the Sigmoid function, e is a natural constant, is the slope of the Sigmoid function of the jth feature dimension, is the jth eigenvalue in the eigenvector of the current accident, is the jth eigenvalue in the eigenvector of historical accident case i, is the upper limit of the value range of the jth feature dimension; According to the calculated similarity, the probability distribution fitting method based on kernel density estimation is used to determine the prior probability of the current accident, which is expressed as: , in, is the conditional probability of an accident developing to state A given the similarity S, p is the number of historical accident cases used to calculate the similarity, h is the smoothing parameter of the probability density function, is the bandwidth parameter of the Gaussian kernel function, S is the value of the given similarity, is the similarity between the current accident and the i-th historical accident case.

[0011] As a further method, the method of performing accident development reasoning on the current accident by combining the prior probability with the accident prediction network includes: Assign the prior probability of the current accident as the initial weight to the input layer nodes of the accident prediction network; Using the information fusion module based on the attention mechanism, the prior probability is dynamically fused with the features extracted from each hidden layer in the accident prediction network, so that the network can focus on the feature information with high correlation with the prior probability; The output of the accident prediction network is used as the accident development reasoning result of the current accident.

[0012] As a further method, the method for tracing the degree of accident hazards in different production links based on the reasoning results includes: Calculate the accident hazard degree of different production links in the reasoning results, the expression is: , in, is the degree of accident hazard in the i-th production link, is the number of risk factors included in the i-th production link, is the importance of the jth risk factor in the i-th production link, is the probability of occurrence of the jth risk factor in the i-th production link, is the degree of loss caused by the occurrence of the jth risk factor in the i-th production link, is the spatiotemporal adjustment coefficient, and are the time and location information of the occurrence of the jth risk factor in the i-th production link, and They are the time with the highest accident rate and the spatial location of the equipment with the most serious damage in historical accidents; The production links where the degree of accident harm is greater than the average value are regarded as the key links. Starting from the key links, the fault tree analysis method is used, with the accident harm result as the top event, and the factors leading to the result are gradually decomposed downwards; Use Bayesian network to conduct probability analysis on the decomposed factors, determine the causal relationship strength and conditional probability between factors, and construct a complete causal network; The causal path with the largest total accident hazard degree is selected from the causal network as the final causal chain.

[0013] As a further method, the method of using Markov decision to generate an emergency resource allocation strategy set that satisfies the emergency response path and constructing an objective function for optimization includes: Determine the emergency resource requirements for each link based on the degree of harm caused by the accident; According to the sequence of emergency response paths, the Markov decision process is used to simulate the state transition in the emergency resource allocation process, and a set of emergency resource allocation strategies that meet the emergency resource needs of each link is generated; Taking the minimum loss as the first objective function, the expression is: , in, , and They are the number of facility categories, the number of production links, and the number of environmental impact factors. is the weight coefficient of the damage cost of the i-th type of facility, is the damage cost suffered by the i-th type of facility in time t, is the weight coefficient of production loss in the jth production link, is the production loss caused by the accident in the jth production link within time t, is the weight coefficient of the damage cost of the kth environmental factor, is the environmental damage cost caused by the accident of the kth environmental factor within time t; Taking the highest disposal efficiency as the second objective function, the expression is: , in, is the number of processing tasks, is the weight coefficient of the success degree of the rth disposal task, is the success rate of the rth task in time t, v is the number of response links, is the weight coefficient of the u-th response link time, is the time spent on the uth response link within time t, x is the number of categories of resource investment, is the weight coefficient of the w-th type of disposal resource cost, is the input cost of the wth type of disposal resource in time t; The same initial weight coefficient is assigned to the first objective function and the second objective function, and the optimal weight coefficient is found through the multi-objective optimization algorithm. Finally, the optimal emergency resource allocation strategy is selected from the emergency resource scheduling strategy set as the emergency response plan.

[0014] As a further method, the method for updating the emergency response plan according to changes in production processes, equipment updates and personnel changes includes: monitoring changes in production processes, equipment status and personnel configurations, modifying the emergency resource allocation strategy set, the first objective function and the second objective function according to the changes, and dynamically adjusting the emergency response plan.

[0015] A second aspect of the present invention provides an intelligent generation system for emergency response plans for production safety accidents, comprising: The spatiotemporal feature extraction module is used to obtain the spatiotemporal features of historical accident cases and calculate the correlation between spatiotemporal features; An accident prediction network module is used to construct an accident prediction network through a graph neural network based on the correlation between the temporal and spatial features, and to add a disturbance term to the accident prediction network to simulate the influence of uncertain factors; The accident development reasoning module is used to calculate the similarity between the current accident and the historical accident cases, determine the prior probability of the current accident based on the similarity, and perform accident development reasoning on the current accident through the prior probability combined with the accident prediction network The cause analysis module is used to trace the degree of accident hazards in different production links based on the reasoning results and obtain the cause chain; A disposal plan generation module is used to use the cause chain as an emergency response path, use Markov decision making to generate an emergency resource allocation strategy set that satisfies the emergency response path, construct an objective function for optimization, and obtain an emergency disposal plan; The plan update module is used to update the emergency response plan according to changes in production processes, equipment updates and personnel changes.

[0016] Compared with the prior art, the embodiments of the present invention have at least the following advantages or beneficial effects: (1) The present invention obtains the spatiotemporal characteristics of historical accident cases and calculates their correlation, and combines graph neural networks to build an accident prediction network, which fully considers the complex relationship between accidents in time and space dimensions, and can more accurately capture the potential laws of accident development, providing a more reliable basis for emergency response; (2) The present invention reversely traces the degree of accident hazards in different production links in the reasoning results to obtain the cause chain, which can quickly find the key factors and links of the accident, so as to formulate emergency response strategies more targetedly; (3) The present invention updates the emergency response plan according to changes in production processes, equipment relationships and personnel changes, so that the emergency response plan can always adapt to the actual production situation, maintain its effectiveness and practicality, and ensure that the most appropriate response plan can be provided for accident emergencies in different production stages and environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a flowchart of a method for intelligently generating an emergency response plan for a production safety accident in an embodiment of the present invention; Figure 2 This is a logical schematic diagram of an intelligent generation system for emergency response plans for production safety accidents in an embodiment of the present invention. DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0019] Reference Figure 1 As shown, the present invention provides a method for intelligently generating an emergency response plan for a production safety accident, comprising: S100 obtains the spatiotemporal features of historical accident cases and calculates the correlation between the spatiotemporal features; In the actual assessment, a large chemical company mainly produces various chemical products. Its production process includes raw material storage and pretreatment, chemical reaction synthesis, product separation and purification, and finished product packaging. It is distributed in different production areas and has many large reactors, storage tanks, pipelines and other equipment. The production process involves high temperature, high pressure, flammable, explosive, and toxic and hazardous chemicals; In the actual assessment, the company had a total of 60 production safety accidents of various types in the past 10 years, involving 12 major production links and 6 environmental factors (temperature, humidity, ventilation conditions, static electricity intensity, pressure stability and material concentration). Time feature extraction was performed. In spring every year (March-May), due to the rise in temperature and relatively dry air, the incidence of accidents caused by static electricity is relatively high, with an average of 1.5 accidents per month; during the equipment maintenance period from 2 a.m. to 6 a.m., due to personnel fatigue and negligent operation, the frequency of mechanical failures and material leakage accidents is 1 per week; on the first day of work after holidays, due to equipment restart and personnel status adjustment problems, the probability of various accidents is higher than usual. 30%, about 2 cases per month; spatial feature extraction, where the chemical reaction synthesis area, due to the presence of complex chemical reactions and a large number of hazardous chemicals, once an accident occurs, it is very easy to trigger a chain reaction, affecting the surrounding pretreatment area and separation and purification area, causing serious casualties and property losses, and may cause long-term pollution to the surrounding environment; the raw material storage area is close to the main traffic arteries in the factory. If a fire or explosion occurs, it may cause damage to the transport vehicles, hinder the rescue channel, and expand the scope of the accident; combine them into a feature vector in the format of [time period, chemical material spatial position]; calculate the spatiotemporal feature correlation, where for the feature vectors [spring, chemical reaction synthesis area] and [equipment maintenance period, raw material storage area], according to statistics, under the conditions of explosion type, synthesis in the production link and dry environmental factors (corresponding to spring), the probability of the two occurring at the same time is 0.05.

[0020] S200 constructs an accident prediction network through a graph neural network based on the correlation between the temporal and spatial features, and adds a disturbance term to the accident prediction network to simulate the influence of uncertain factors; It should be explained that building an accident prediction network based on the correlation of spatiotemporal features is to use the powerful feature extraction and relationship modeling capabilities of graph neural networks to explore the potential patterns of accidents in the spatiotemporal dimensions. Adding disturbance terms takes into account uncertain factors in the actual environment, such as sudden weather changes and unexpected equipment failures. By simulating the impact of these uncertain factors on the development of accidents, the prediction network can cover more possible situations and enhance the robustness and practicality of the prediction results. In the actual evaluation, the spatiotemporal features are used as nodes in the graph, and the correlation is used as the connection weight to construct an accident feature graph. The convolutional layer of the graph neural network is used to extract features and aggregate information on the accident feature graph. By studying the data of 60 historical accident cases, the characteristic representation of accident development under different combinations of spatiotemporal features is gradually mastered. An accident development prediction model is constructed in the fully connected layer. The initial information of an accident (initial temperature, pressure, material flow, equipment operating status) is used as input, and the development status of the accident at different times in the future (whether the fire will expand, whether an explosion will occur, and the scope of toxic gas leakage) and trend (minor, moderate, severe, catastrophic) are output.

[0021] S300 calculates the similarity between the current accident and the historical accident cases, determines the prior probability of the current accident based on the similarity, and performs accident development reasoning on the current accident by combining the prior probability with the accident prediction network; It needs to be explained that the purpose of calculating the similarity between the current accident and the historical cases is to find situations with similar characteristics to the current accident from the historical data, determine the prior probability based on the similarity, give quantitative values ​​to the prior knowledge, and reflect the reference value of historical experience to the current accident. Combining it with the accident prediction network can enable the network to fully consider historical laws and actual conditions when predicting, optimize the reasoning process, and improve the accuracy and reliability of the prediction of the current accident development; In the actual evaluation, a simulated accident was generated. The accident occurred in mid-April (spring) in the chemical reaction synthesis area. The reactor pressure increased abnormally and a small amount of toxic gas leaked. The similarity with historical accident cases was calculated, and the similarities with 6 historical accidents were 0.75, 0.68, 0.8, 0.7, 0.72, and 0.65, respectively. The prior probability of the current accident was determined according to the probability distribution fitting method based on kernel density estimation. The prior probability of the current accident was assigned to the input layer nodes of the accident prediction network as the initial weight, and the input feature nodes related to spring and the chemical reaction synthesis area were given a higher initial weight. The information fusion module based on the attention mechanism was used to dynamically fuse the prior probability with the features extracted from each hidden layer in the hidden layer of the accident prediction network, including the pressure increase rate and the amount of toxic gas leakage. Finally, the accident prediction network outputs that the probability that the current accident may develop into a minor accident in the next hour is 0.4, the probability of developing into a moderate accident is 0.45, and the probability of developing into a serious accident is 0.15.

[0022] S400 traces the accident hazard levels of different production links based on the reasoning results and obtains the cause chain; In the actual assessment, the accident hazard degree of different production links in the reasoning results is calculated. Among them, the chemical reaction synthesis area has 5 main risk factors: reaction out of control, equipment sealing failure, cooling system failure, operator misoperation, and excessive ambient temperature. The highest accident rate in historical accidents occurred in late April. The spatial location of the equipment with the most serious damage is the core reaction area of ​​the chemical reaction synthesis area. The accident hazard degree of the chemical reaction synthesis area is calculated to be 0.769. After comparison, the hazard degree of the chemical reaction synthesis area is greater than the average value. It is taken as a key link. Starting from the chemical reaction synthesis area, the fault tree analysis method is used, with the accident hazard result as the top event, and the factors leading to the result are gradually decomposed downward. For example, reaction out of control may be due to temperature control failure, material ratio imbalance or catalyst abnormality, and equipment sealing failure may be due to aging of seals, improper installation or pressure shock, etc. In the actual assessment, the Bayesian network is used to conduct probability analysis on the decomposed factors to determine the causal relationship strength and conditional probability between the factors. After constructing a complete causal network, the causal path with the largest total accident hazard level is selected from the causal network, specifically temperature control failure-rapid increase in reaction pressure due to uncontrolled reaction-equipment rupture-large-scale explosion, as the final causal chain.

[0023] S500 uses the cause chain as an emergency response path, uses Markov decision making to generate an emergency resource allocation strategy set that satisfies the emergency response path, constructs an objective function for optimization, and obtains an emergency response plan; It needs to be explained that the causal chain is used as the emergency response path because the causal chain clearly presents the logical context of the occurrence and development of the accident, and targeted measures can be taken along this path. The use of Markov decision to generate a resource allocation strategy set is based on its effective simulation of the state transition of the random process, taking into account the uncertainty of the accident development, and reasonably arranging the allocation of resources at different stages; In the actual assessment, for the chemical reaction synthesis area, due to the risk of runaway reaction and toxic gas leakage, it is necessary to equip the emergency resources of fire foam (10 cubic meters), chemical neutralizer (5 tons), explosion-proof tools (3 sets), respiratory protection equipment (20 sets), leakage plugging equipment (2 sets), emergency lighting equipment (5), environmental monitoring instruments (2 sets), and spare coolant (3 tons). According to the sequence of the emergency response path, the Markov decision process is used to simulate the state transition in the emergency resource allocation process, including: Plan 1: Within the first 4 minutes after the accident, deploy 5 cubic meters of firefighting foam, 10 sets of respiratory protection equipment and 1 set of environmental monitoring equipment to the chemical reaction synthesis area, and arrange 2 rescue personnel to carry explosion-proof tools to the product purification area, and then deploy 3 tons of chemical neutralizers and 1 set of leakage plugging equipment to the material storage area; deploy 2 tons of spare coolant to the finished product packaging area, and organize the evacuation of surrounding personnel to a safe area; Plan 2: Within the first 4 minutes after the accident, deploy 6 cubic meters of fire-fighting foam, 8 sets of respiratory protection equipment and 2 sets of emergency lighting equipment to the chemical reaction synthesis area, and arrange 3 rescue personnel to carry some explosion-proof tools and plugging equipment to the purification area, and deploy 4 tons of chemical neutralizers and 1 set of environmental monitoring instruments to the material storage area; deploy 3 tons of spare coolant to the finished product packaging area, and organize the evacuation of surrounding personnel to a safe area; In the actual evaluation, the minimum loss is taken as the first objective function, and the highest disposal efficiency is taken as the second objective function. The same initial weight coefficient is assigned to the first and second objective functions. Through the multi-objective optimization algorithm and multiple iterative calculations, the optimal weight coefficients are found, which are 0.36 and 0.74 respectively. Finally, the optimal emergency resource allocation strategy, namely Plan 2, is selected from the emergency resource scheduling strategy set as the emergency disposal plan.

[0024] S600 updates the emergency response plan according to production process changes, equipment updates and personnel changes.

[0025] In the actual evaluation, the changes in production processes, equipment status and personnel configuration were monitored, and the emergency resource allocation strategy set, the first objective function and the second objective function were modified according to the changes, and the emergency disposal plan was dynamically adjusted. Specifically, special emergency inhibitors (2 tons) and special emergency maintenance tools (4 sets) for new equipment were equipped for the new chemical reaction process; the storage location and retrieval routes of emergency resources were re-planned according to the distribution of new equipment, and key emergency materials were stored in a special emergency material warehouse close to the new equipment to ensure that they can be quickly obtained in the event of an accident.

[0026] In this embodiment, the method of obtaining the spatiotemporal features of historical accident cases and calculating the correlation between the spatiotemporal features includes: Extract time features from historical accident cases, including production shifts, equipment maintenance cycles, and seasonal periods, and count the accident frequency in each period; extract spatial features, including the functional zoning of the factory, process flow layout, and logistics channel direction; Combine the corresponding time features and spatial features in each historical accident case into a feature vector; Calculate the correlation between spatiotemporal features, the expression is: , in, and are the feature vectors of the i-th and j-th spatiotemporal features, r, s and t represent the number of accident types, the number of production links and the number of environmental factors, respectively. For the accident type , production link is m and environmental factor is n, the characteristic vector and The probability of simultaneous occurrence, and The accident types are , production link is m and environmental factor is n, the characteristic vector and the eigenvector Probability of occurrence.

[0027] In this embodiment, the method of constructing an accident prediction network through a graph neural network based on the correlation between the temporal and spatial features includes: The spatiotemporal features are used as nodes in the graph, and the correlation degree is used as the connection weight to obtain the accident feature graph; The convolutional layer of the graph neural network is used to extract features and aggregate information from the accident feature graph. Through convolution operations, the feature representation of accident development under different combinations of spatiotemporal features is learned. In the fully connected layer of the graph neural network, a prediction model for the development of an accident is constructed based on the learned feature representation. The initial state information of the accident is used as input, and the development state and trend of the accident at different moments in the future are output, including the severity and scope of the accident. The graph neural network is trained using historical accident case data. By adjusting the network parameters, the error between the prediction results and the accident development situation is minimized to obtain an accident prediction model.

[0028] In this embodiment, the method of adding a disturbance term to the accident prediction network to simulate the influence of uncertain factors includes: Through the probabilistic reasoning model of Bayesian network, the probability distribution of uncertain factors on the development of accidents is quantified; A disturbance term generation model is constructed through generative adversarial networks, which takes historical accident data and the probability distribution of uncertain factors on accident development as input to generate disturbance terms. The generated disturbance term is integrated into the accident prediction network through the attention mechanism. Specifically, the similarity score between the disturbance term and the features of each layer of the accident prediction network is used as the attention weight, the features are weighted summed up, and the disturbance term is integrated.

[0029] In this embodiment, the method of calculating the similarity between the current accident and the historical accident cases and determining the prior probability of the current accident based on the similarity includes: Calculate the similarity between the current accident and the historical accident case. The expression is: , in, and are the feature vectors of the current accident and historical accident case i respectively, q is the total number of feature dimensions, is the weight coefficient of the jth feature dimension, is the balance coefficient of the Sigmoid function, e is a natural constant, is the slope of the Sigmoid function of the jth feature dimension, is the jth eigenvalue in the eigenvector of the current accident, is the jth eigenvalue in the eigenvector of historical accident case i, is the upper limit of the value range of the jth feature dimension; According to the calculated similarity, the probability distribution fitting method based on kernel density estimation is used to determine the prior probability of the current accident, which is expressed as: , in, is the conditional probability of an accident developing to state A given the similarity S, p is the number of historical accident cases used to calculate the similarity, h is the smoothing parameter of the probability density function, is the bandwidth parameter of the Gaussian kernel function, S is the value of the given similarity, is the similarity between the current accident and the i-th historical accident case.

[0030] In this embodiment, the method of performing accident development reasoning on the current accident by combining the prior probability with the accident prediction network includes: Assign the prior probability of the current accident as the initial weight to the input layer nodes of the accident prediction network; Using the information fusion module based on the attention mechanism, the prior probability is dynamically fused with the features extracted from each hidden layer in the accident prediction network, so that the network can focus on the feature information with high correlation with the prior probability; The output of the accident prediction network is used as the accident development reasoning result of the current accident.

[0031] In this embodiment, the method for tracing the degree of accident hazards in different production links based on the inference results includes: Calculate the accident hazard degree of different production links in the reasoning results, the expression is: , in, is the degree of accident hazard in the i-th production link, is the number of risk factors included in the i-th production link, is the importance of the jth risk factor in the i-th production link, is the probability of occurrence of the jth risk factor in the i-th production link, is the degree of loss caused by the occurrence of the jth risk factor in the i-th production link, is the spatiotemporal adjustment coefficient, and are the time and location information of the occurrence of the jth risk factor in the i-th production link, and They are the time with the highest accident rate and the spatial location of the equipment with the most serious damage in historical accidents; The production links where the degree of accident harm is greater than the average value are regarded as the key links. Starting from the key links, the fault tree analysis method is used, with the accident harm result as the top event, and the factors leading to the result are gradually decomposed downwards; Use Bayesian network to conduct probability analysis on the decomposed factors, determine the causal relationship strength and conditional probability between factors, and construct a complete causal network; The causal path with the largest total accident hazard degree is selected from the causal network as the final causal chain.

[0032] In this embodiment, the method of using Markov decision to generate an emergency resource allocation strategy set that satisfies the emergency response path and constructing an objective function for optimization includes: Determine the emergency resource requirements for each link based on the degree of harm caused by the accident; According to the sequence of emergency response paths, the Markov decision process is used to simulate the state transition in the emergency resource allocation process, and a set of emergency resource allocation strategies that meet the emergency resource needs of each link is generated; Taking the minimum loss as the first objective function, the expression is: , in, , and They are the number of facility categories, the number of production links, and the number of environmental impact factors. is the weight coefficient of the damage cost of the i-th type of facility, is the damage cost suffered by the i-th type of facility in time t, is the weight coefficient of production loss in the jth production link, is the production loss caused by the accident in the jth production link within time t, is the weight coefficient of the damage cost of the kth environmental factor, is the environmental damage cost caused by the accident of the kth environmental factor within time t; Taking the highest disposal efficiency as the second objective function, the expression is: , in, is the number of processing tasks, is the weight coefficient of the success degree of the rth disposal task, is the success rate of the rth task in time t, v is the number of response links, is the weight coefficient of the u-th response link time, is the time spent on the uth response link within time t, x is the number of categories of resource investment, is the weight coefficient of the w-th type of disposal resource cost, is the input cost of the wth type of disposal resource in time t; The same initial weight coefficient is assigned to the first objective function and the second objective function, and the optimal weight coefficient is found through the multi-objective optimization algorithm. Finally, the optimal emergency resource allocation strategy is selected from the emergency resource scheduling strategy set as the emergency response plan.

[0033] In this embodiment, the method for updating the emergency response plan according to production process changes, equipment updates and personnel changes includes: monitoring changes in production processes, equipment status and personnel configuration, modifying the emergency resource allocation strategy set, the first objective function and the second objective function according to the changes, and dynamically adjusting the emergency response plan.

[0034] Reference Figure 2 As shown, the second aspect of the present invention also provides an intelligent generation system for emergency response plans for production safety accidents, including: The spatiotemporal feature extraction module is used to obtain the spatiotemporal features of historical accident cases and calculate the correlation between spatiotemporal features; An accident prediction network module is used to construct an accident prediction network through a graph neural network based on the correlation between the temporal and spatial features, and to add a disturbance term to the accident prediction network to simulate the influence of uncertain factors; The accident development reasoning module is used to calculate the similarity between the current accident and the historical accident cases, determine the prior probability of the current accident based on the similarity, and perform accident development reasoning on the current accident through the prior probability combined with the accident prediction network The cause analysis module is used to trace the degree of accident hazards in different production links based on the reasoning results and obtain the cause chain; A disposal plan generation module is used to use the cause chain as an emergency response path, use Markov decision making to generate an emergency resource allocation strategy set that satisfies the emergency response path, construct an objective function for optimization, and obtain an emergency disposal plan; The plan update module is used to update the emergency response plan according to changes in production processes, equipment updates and personnel changes.

[0035] The above contents are merely examples and explanations of the structure of the present invention. The technicians in this technical field may make various modifications or additions to the specific embodiments described or replace them in a similar manner. As long as they do not deviate from the structure of the invention or exceed the scope defined by the claims, they should all fall within the protection scope of the present invention.

Claims

1. A method for intelligently generating emergency response plans for production safety accidents, characterized in that: The following steps are involved: Obtain the spatiotemporal characteristics of historical accident cases and calculate the correlation between the spatiotemporal characteristics; Based on the correlation between the spatiotemporal features, an accident prediction network is constructed through a graph neural network, and disturbance items are added to the accident prediction network to simulate the influence of uncertain factors; Calculate the similarity between the current accident and the historical accident cases, determine the prior probability of the current accident based on the similarity, and perform accident development reasoning on the current accident by combining the prior probability with the accident prediction network; Based on the inference results, the accident hazard levels of different production links are traced back to obtain the cause chain; Taking the causal chain as the emergency response path, using Markov decision making to generate an emergency resource allocation strategy set that satisfies the emergency response path, constructing an objective function for optimization, and obtaining an emergency response plan; The emergency response plan will be updated according to changes in production processes, equipment upgrades and personnel changes.

2. According to claim 1, a method for intelligently generating an emergency response plan for a production safety accident is characterized in that: The method of obtaining the spatiotemporal features of historical accident cases and calculating the correlation between the spatiotemporal features includes: Extract time features from historical accident cases, including production shifts, equipment maintenance cycles, and seasonal periods, and count the accident frequency in each period; extract spatial features, including the functional zoning of the factory, process flow layout, and logistics channel direction; Combine the corresponding time features and spatial features in each historical accident case into a feature vector; Calculate the correlation between spatiotemporal features, the expression is: ， in, and are the feature vectors of the i-th and j-th spatiotemporal features, r, s and t represent the number of accident types, the number of production links and the number of environmental factors, respectively. For the accident type , production link is m and environmental factor is n, the characteristic vector and The probability of simultaneous occurrence, and The accident types are , production link is m and environmental factor is n, the characteristic vector and the eigenvector Probability of occurrence.

3. The method for intelligently generating an emergency response plan for a production safety accident according to claim 1, characterized in that: The method for constructing an accident prediction network through a graph neural network based on the correlation between the temporal and spatial features includes: The spatiotemporal features are used as nodes in the graph, and the correlation degree is used as the connection weight to obtain the accident feature graph; The convolutional layer of the graph neural network is used to extract features and aggregate information from the accident feature graph. Through convolution operations, the feature representation of accident development under different combinations of spatiotemporal features is learned. In the fully connected layer of the graph neural network, a prediction model for the development of an accident is constructed based on the learned feature representation. The initial state information of the accident is used as input, and the development state and trend of the accident at different moments in the future are output, including the severity and scope of the accident. The graph neural network is trained using historical accident case data. By adjusting the network parameters, the error between the prediction results and the accident development situation is minimized to obtain an accident prediction model.

4. The method for intelligently generating an emergency response plan for a production safety accident according to claim 1, characterized in that: The method of adding a disturbance term to the accident prediction network to simulate the influence of uncertain factors includes: Through the probabilistic reasoning model of Bayesian network, the probability distribution of uncertain factors on the development of accidents is quantified; A disturbance term generation model is constructed through generative adversarial networks, which takes historical accident data and the probability distribution of uncertain factors on accident development as input to generate disturbance terms. The generated disturbance term is integrated into the accident prediction network through the attention mechanism. Specifically, the similarity score between the disturbance term and the features of each layer of the accident prediction network is used as the attention weight, the features are weighted summed up, and the disturbance term is integrated.

5. The method for intelligently generating an emergency response plan for a production safety accident according to claim 1, characterized in that: The method of calculating the similarity between the current accident and the historical accident cases and determining the prior probability of the current accident based on the similarity includes: Calculate the similarity between the current accident and the historical accident case. The expression is: ， in, and are the feature vectors of the current accident and historical accident case i respectively, q is the total number of feature dimensions, is the weight coefficient of the jth feature dimension, is the balance coefficient of the Sigmoid function, e is a natural constant, is the slope of the Sigmoid function of the jth feature dimension, is the jth eigenvalue in the eigenvector of the current accident, is the jth eigenvalue in the eigenvector of historical accident case i, is the upper limit of the value range of the jth feature dimension; According to the calculated similarity, the probability distribution fitting method based on kernel density estimation is used to determine the prior probability of the current accident, which is expressed as: ， in, is the conditional probability of an accident developing to state A given the similarity S, p is the number of historical accident cases used to calculate the similarity, h is the smoothing parameter of the probability density function, is the bandwidth parameter of the Gaussian kernel function, S is the value of the given similarity, is the similarity between the current accident and the i-th historical accident case.

6. The method for intelligently generating an emergency response plan for a production safety accident according to claim 1, characterized in that: The method for performing accident development reasoning on the current accident by combining the prior probability with the accident prediction network includes: Assign the prior probability of the current accident as the initial weight to the input layer nodes of the accident prediction network; Using the information fusion module based on the attention mechanism, the prior probability is dynamically fused with the features extracted from each hidden layer in the accident prediction network, so that the network can focus on the feature information with high correlation with the prior probability; The output of the accident prediction network is used as the accident development reasoning result of the current accident.

7. The method for intelligently generating an emergency response plan for a production safety accident according to claim 1, characterized in that: The method for tracing the degree of accident hazards in different production links based on the inference results includes: Calculate the accident hazard degree of different production links in the reasoning results, the expression is: ， in, is the degree of accident hazard in the i-th production link, is the number of risk factors included in the i-th production link, is the importance of the jth risk factor in the i-th production link, is the probability of occurrence of the jth risk factor in the i-th production link, is the degree of loss caused by the occurrence of the jth risk factor in the i-th production link, is the spatiotemporal adjustment coefficient, and are the time and location information of the occurrence of the jth risk factor in the i-th production link, and They are the time with the highest accident rate and the spatial location of the equipment with the most serious damage in historical accidents; The production links where the degree of accident harm is greater than the average value are regarded as the key links. Starting from the key links, the fault tree analysis method is used, with the accident harm result as the top event, and the factors leading to the result are gradually decomposed downwards; Use Bayesian network to conduct probability analysis on the decomposed factors, determine the causal relationship strength and conditional probability between factors, and construct a complete causal network; The causal path with the largest total accident hazard degree is selected from the causal network as the final causal chain.

8. The method for intelligently generating an emergency response plan for a production safety accident according to claim 1, characterized in that: The method of using Markov decision making to generate an emergency resource allocation strategy set that satisfies the emergency response path and constructing an objective function for optimization includes: Determine the emergency resource requirements for each link based on the degree of harm caused by the accident; According to the sequence of emergency response paths, the Markov decision process is used to simulate the state transition in the emergency resource allocation process, and a set of emergency resource allocation strategies that meet the emergency resource needs of each link is generated; Taking the minimum loss as the first objective function, the expression is: ， in, , and They are the number of facility categories, the number of production links, and the number of environmental impact factors. is the weight coefficient of the damage cost of the i-th type of facility, is the damage cost suffered by the i-th type of facility in time t, is the weight coefficient of production loss in the jth production link, is the production loss caused by the accident in the jth production link within time t, is the weight coefficient of the damage cost of the kth environmental factor, is the environmental damage cost caused by the accident of the kth environmental factor within time t; Taking the highest disposal efficiency as the second objective function, the expression is: ， in, is the number of processing tasks, is the weight coefficient of the success degree of the rth disposal task, is the success rate of the rth task in time t, v is the number of response links, is the weight coefficient of the u-th response link time, is the time spent on the uth response link within time t, x is the number of categories of resource investment, is the weight coefficient of the w-th type of disposal resource cost, is the input cost of the wth type of disposal resource in time t; The same initial weight coefficient is assigned to the first objective function and the second objective function, and the optimal weight coefficient is found through the multi-objective optimization algorithm. Finally, the optimal emergency resource allocation strategy is selected from the emergency resource scheduling strategy set as the emergency response plan.

9. The method for intelligently generating an emergency response plan for a production safety accident according to claim 1, characterized in that: The method for updating the emergency response plan according to production process changes, equipment updates and personnel changes includes: monitoring changes in production processes, equipment status and personnel configuration, modifying the emergency resource allocation strategy set, the first objective function and the second objective function according to the changes, and dynamically adjusting the emergency response plan.

10. A production safety accident emergency response plan intelligent generation system, used to execute a production safety accident emergency response plan intelligent generation method according to any one of claims 1 to 9, characterized in that: The system comprises: The spatiotemporal feature extraction module is used to obtain the spatiotemporal features of historical accident cases and calculate the correlation between the spatiotemporal features; An accident prediction network module is used to construct an accident prediction network through a graph neural network based on the correlation between the temporal and spatial features, and to add a disturbance term to the accident prediction network to simulate the influence of uncertain factors; The accident development reasoning module is used to calculate the similarity between the current accident and the historical accident cases, determine the prior probability of the current accident based on the similarity, and perform accident development reasoning on the current accident through the prior probability combined with the accident prediction network The cause analysis module is used to trace the degree of accident hazards in different production links based on the reasoning results and obtain the cause chain; A disposal plan generation module is used to use the cause chain as an emergency response path, use Markov decision making to generate an emergency resource allocation strategy set that satisfies the emergency response path, construct an objective function for optimization, and obtain an emergency disposal plan; The plan update module is used to update the emergency response plan according to changes in production processes, equipment updates and personnel changes.

Images