Chemical fire and explosion coupling consequence prediction method and system based on large model

Abstract

The application relates to the field of computing models and industrial safety data processing under new-generation information technology, and discloses a chemical fire and explosion coupling consequence prediction method and system based on a large model, which has the technical scheme as follows: physical mapping of a multi-source heterogeneous environment and construction of a graph node digital twin, evolution calculation of a fluid mechanics source term and triggering of a disaster critical point, deduction of a physical large model thermal radiation field in a restricted environment, analysis and quantification of a "thermal-power" dual coupling tensor of a target device and buckling failure, deduction of a space-time graph attention network with strong intervention of a physical prior probability, tracing of global system vulnerability and closed-loop blocking control of a physical execution end; by using intelligent calculation technologies such as a physical information neural network and a space-time graph attention network, a thermodynamic partial differential equation and a material thermal-power coupling failure criterion are forcedly embedded into a graph deduction framework as a physical prior bias, so that complex disaster consequence prediction is accelerated from "hour level" to "second level" in the traditional CFD.

Description

Technical Field

[0001] This invention relates to the field of computational models and industrial safety data processing technology under the new generation of information technology, and more specifically, to a method and system for predicting the coupled consequences of chemical fires and explosions based on large models. Background Technology

[0002] Currently, dynamic prediction of the consequences of such complex accidents mainly relies on traditional computational fluid dynamics (CFD) numerical simulation or purely data-driven artificial intelligence models.

[0003] However, traditional numerical simulations are extremely computationally cumbersome and time-consuming, taking up to several hours, when dealing with multi-physics coupling problems such as thermal radiation in fire scenes and equipment buckling due to heat. This makes it impossible to meet the timeliness requirements of "second-level" emergency response at disaster sites. Furthermore, purely data-driven AI models are prone to "prediction distortion" under extreme conditions lacking boundary constraints, as they easily violate objective physical laws. Moreover, existing prediction technologies generally separate the local mechanical failure of individual equipment from the domino effect of the entire plant, lacking a deep guidance of the accident evolution network by physical prior mechanisms. This results in limited prediction accuracy, poor ability to trace system vulnerabilities, and difficulty in providing highly reliable automated closed-loop blocking decisions for emergency command in real sudden disasters.

[0004] Therefore, this invention provides a method and system for predicting the coupled consequences of chemical fires and explosions based on a large model, thereby improving the aforementioned technical problems. Summary of the Invention

[0005] This disclosure aims to address the shortcomings of existing technologies by providing a method and system for predicting the coupled consequences of chemical fires and explosions based on large models. The present invention employs large model technologies such as Physical Information Neural Network (PINN) and Spatiotemporal Graph Attention Network (ST-GNN) that integrates physical priors. By forcibly embedding the underlying thermodynamic mechanism and the probability of material buckling failure into the dynamic deduction of the spatial accident chain, it achieves second-level accurate prediction and hardware-level automated emergency blocking of complex domino cascade disasters in chemical industrial parks.

[0006] To achieve the above objectives, the present disclosure proposes the following technical solutions:

[0007] In a first aspect, embodiments of this disclosure propose a method for predicting the coupled consequences of chemical fires and explosions based on a large model, comprising the following steps:

[0008] The real-time physical feature vector of the target storage tank is obtained. The real-time physical feature vector includes at least the real-time internal pressure and the real-time liquid level. The historical maintenance text of the target storage tank is processed using a large language model to obtain the text hidden state vector.

[0009] The real-time physical feature vector and the textual hidden state vector are aligned and concatenated to construct an initial feature vector for the graph nodes of the target storage tank.

[0010] When the real-time internal pressure drop of the target storage tank exceeds a preset threshold, the transient leakage mass flow rate is calculated based on the real-time internal pressure and the real-time liquid level, and the transient leakage mass flow rate is input into a preset diffusion neural network to obtain a spatial concentration field, so as to determine the ignition timestamp when the spatial concentration field reaches the lower explosion limit.

[0011] In response to the ignition timestamp, the effective surface temperature of the flame is obtained by solving the physical information neural network based on the constrained partial differential equation of combustion energy conservation, and the incident heat radiation flux of the target tank to the adjacent tank is calculated based on the solid flame model.

[0012] The transient temperature of the shell of the adjacent storage tank is calculated based on the incident thermal radiation flux, and the equivalent coupling stress is calculated in combination with the real-time liquid level of the adjacent storage tank. The equivalent coupling stress is then mapped to the physical failure probability.

[0013] The physical failure probability is embedded as a prior bias term into the spatiotemporal graph attention network to update the disaster propagation weights between nodes and output the active disaster set.

[0014] Based on the activated disaster set, the critical sensitivity of all network nodes is calculated, and based on the critical sensitivity, an electromechanical control message is sent to the target programmable logic controller to control the corresponding storage tank to perform a physical blocking action.

[0015] As a preferred embodiment of the present invention, the real-time physical feature vector and the textual latent state vector are aligned and concatenated to construct an initial feature vector for the graph nodes of the target storage tank. The specific calculation formula is as follows:

[0016] ;

[0017] in, This represents the initial feature vector of the graph node i corresponding to the target storage tank; Represents a non-linear activation function; This represents the first weight projection matrix; Represents the hidden state vector of the text; Represents the tensor concatenation operator; This represents the second weight projection matrix; The real-time physical feature vector represents the initial monitoring time, and the real-time physical feature vector includes: real-time internal pressure, real-time liquid level, and ambient wind speed; This represents the bias vector.

[0018] As a preferred embodiment of the present invention, the transient leakage mass flow rate is calculated based on the real-time internal pressure and the real-time liquid level, and the specific calculation formula is as follows:

[0019] ;

[0020] in, This represents the transient leakage mass flow rate of the target storage tank at time t; Indicates the emission factor; Indicates the area of ​​physical damage; Indicates the liquid density of the material; Indicates real-time internal pressure; The standard atmospheric pressure constant is represented by g; g represents the gravitational acceleration constant. Indicates the real-time liquid level.

[0021] As a preferred embodiment of the present invention, the combustion energy conservation partial differential equation is used as a loss function term of the physical information neural network. Its expression is:

[0022] ;

[0023] in, This represents the scalar value of the mean squared error loss. Represents partial differential operators; This represents the transient temperature field on the flame surface; Represents a time variable; Indicates the thermal diffusivity; Represents the Laplace operator; This represents the pre-exponential factor constant; Represents the natural constant; Indicates activation energy; Represents the ideal gas constant; This represents the flame tilt angle compensation function; Indicates ambient wind speed; Indicates real-time liquid level; Indicates the total design height of the storage tank; This represents the mean squared error norm.

[0024] As a preferred embodiment of the present invention, the incident heat radiation flux from the target storage tank to adjacent storage tanks is calculated based on a solid flame model. The specific calculation formula is as follows:

[0025] ;

[0026] in, This represents the dynamic incident thermal radiation flux received by the neighboring storage tank j; Indicates atmospheric transmittance under the influence of wind speed; Indicates the emissivity of flame blackness; This represents the Stefan-Boltzmann constant; Indicates the effective surface temperature of the flame; Indicates the physical tilt angle of the flame Reconstructed spatial geometric perspective factors.

[0027] As a preferred technical solution of the present invention, the equivalent coupling stress is calculated by combining the real-time liquid level of the adjacent storage tank, and the equivalent coupling stress is mapped to the physical failure probability. The specific calculation process includes:

[0028] Calculate the thermal expansion stress of the shell of the adjacent storage tank. Circumferential stress of fluid With fluid axial stress :

[0029] ;

[0030] ;

[0031] Based on the fourth strength theory, the thermal expansion stress, the fluid circumferential stress, and the fluid axial stress are aggregated into an equivalent coupled stress. :

[0032] ;

[0033] Extract the material yield strength at the transient absolute temperature of the fire-facing surface of the shell. , building time Dynamic load factor :

[0034] ;

[0035] Substituting the dynamic load factor into the Weibull distribution model, it is mapped to the physical failure probability. :

[0036] ;

[0037] in, Indicates the elastic modulus; Indicates the coefficient of linear expansion; This indicates the transient absolute temperature of the fire-facing surface of the casing; Indicates the initial absolute temperature; Indicates Poisson's ratio; Indicates the density of the liquid; Represents the gravitational acceleration constant; Indicates the real-time liquid level of adjacent storage tanks; Indicates the diameter of the storage tank; Indicates the thickness of the tank wall; This indicates the thermal expansion stress of the shell; Indicates equivalent coupling stress; Indicates the time of integration Equivalent coupling stress at time; This represents the material yield strength at the transient absolute temperature of the fire-facing surface of the shell. Indicates the fire time stamp; t represents the current assessment time; Indicates time Dynamic load factor; Represents the integral time element; Parameters representing the characteristic scale of fatigue life; This represents the shape parameter of the Weibull distribution.

[0038] As a preferred embodiment of the present invention, the physical failure probability is embedded as a priori bias term into the spatiotemporal graph attention network to update the disaster propagation weights between nodes. The specific calculation formula is as follows:

[0039] ;

[0040] in, Indicates the weight of disaster propagation; This represents the modified linear unit activation function with leakage; Represents a learnable weight vector; and Let i and j represent the feature tensors of nodes i and j at time t, respectively; This represents the physical prior penalty amplification factor; This represents the physical failure probability of node j corresponding to the adjacent storage tank; Represents the set of local neighbor nodes; The feature tensor representing the neighbor node k; This represents the physical failure probability of neighbor node k.

[0041] Secondly, this disclosure proposes a chemical fire and explosion coupled consequence prediction system based on a large model, the system comprising: an edge sensing network, a cloud computing center, and a physical execution terminal;

[0042] The edge sensing network is used to collect real-time physical parameters and environmental meteorological data of the target storage tank. The real-time physical parameters include at least real-time internal pressure and real-time liquid level.

[0043] The cloud-based intelligent computing center is communicatively connected to the edge sensing network, and the cloud-based intelligent computing center includes:

[0044] The feature construction submodule is used to process the historical maintenance text of the target storage tank using a large language model to obtain a textual latent state vector, and to align and concatenate the real-time physical parameters with the textual latent state vector to construct the initial feature vector of the graph node.

[0045] The source term calculation submodule is used to calculate the transient leakage mass flow rate and extrapolate the spatial concentration field when the real-time internal pressure drop exceeds a preset threshold in order to determine the ignition timestamp.

[0046] The radiation inference submodule is used to solve for the effective surface temperature of the flame based on the physical information neural network, and to calculate the incident heat radiation flux of the target tank to the adjacent tank.

[0047] The failure assessment submodule is used to solve the transient absolute temperature of the shell fire-facing surface of the adjacent storage tank and calculate the equivalent coupling stress, and map the equivalent coupling stress to the physical failure probability.

[0048] The graph network deduction submodule is used to deduce the activation catastrophe set using a spatiotemporal graph attention network;

[0049] The electromechanical control submodule is used to calculate the critical sensitivity based on the activated disaster set and generate electromechanical control messages;

[0050] The physical execution terminal is communicatively connected to the cloud-based intelligent computing center and is used to receive the electromechanical control messages and drive the emergency equipment of the corresponding storage tank to perform physical blocking actions.

[0051] Thirdly, embodiments of this disclosure provide an electronic device, including: a processor and a memory for storing processor-executable instructions;

[0052] The processor is configured to execute the instructions to implement a method for predicting the coupled consequences of chemical fires and explosions based on a large model.

[0053] Fourthly, embodiments of this disclosure provide a computer-readable storage medium that, when the instructions in the computer-readable storage medium are executed by a processor of an electronic device, enables the electronic device to execute a large-model-based method for predicting the coupled consequences of chemical fires and explosions.

[0054] In summary, the present invention has the following beneficial effects:

[0055] Firstly, existing pure data prediction models are prone to producing "algorithmic illusions" that violate energy conservation under extreme conditions, while traditional CFD simulations are too time-consuming. This invention constructs a Physical Information Neural Network (PINN) to forcibly embed the combustion thermodynamic partial differential equation with crosswind correction into the loss function. This mechanism enables the system to maintain a "second-level" output speed while the thermal radiation field prediction results strictly follow the objective laws of physics, perfectly meeting the stringent requirements of timeliness and accuracy for emergency rescue under sudden disasters.

[0056] Secondly, existing technologies often analyze the heating temperature of a single device in isolation, or rely solely on AI models to blindly predict the propagation path. This invention, through the essence of engineering mechanics, transforms the nonlinear temperature rise and hydrostatic pressure of the target device into a rigorous "thermal-mechanical" equivalent coupled stress, and then quantifies it into the real probability of material buckling failure. The system forcibly injects this physical failure probability as a penalty term into the bottom weights of the spatiotemporal graph attention network (ST-GNN), so that the inference path of artificial intelligence is caught by real physics, breaking through the logical gap between "local equipment mechanical failure" and "plant-wide spatial cascade disaster".

[0057] Thirdly, in response to the problem that traditional systems rely solely on temperature and pressure sensors while neglecting hidden structural defects, this invention introduces a Large Language Model (LLM) to perform semantic dimensionality reduction on unstructured historical maintenance records. By tensor splicing of the implicit physical degradation state extracted from the text with real-time fluid dynamic parameters (such as internal pressure and liquid level-to-air ratio) in a unified topological implicit space, a high-precision holographic mapping of physical risk sources to digital maps is achieved.

[0058] Fourth, most existing disaster prediction systems only output consequence contour lines as a reference, lacking actual blocking capabilities. This invention can not only forward extrapolate the consequences of domino spread, but also use the partial derivative of system collapse calculus to reverse trace the critical sensitivity (CRIT) of nodes. Based on the sensitivity, the system accurately locks the "high-risk hub node" in the accident chain and automatically generates electromechanical control messages to send to the edge-side PLC controller, forcibly executing the physical action of opening the deluge valve or emergency shut-off valve (ESD), thereby completely cutting off the physical evolution path of the disaster within the golden time with minimal emergency rescue costs. Attached Figure Description

[0059] Figure 1 A framework diagram of a chemical fire and explosion coupling consequence prediction system based on a large model provided in an embodiment of the present invention;

[0060] Figure 2 The flowchart illustrates the chemical fire and explosion coupling consequence prediction method based on a large model, as provided in this embodiment of the invention. Detailed Implementation

[0061] The present application will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present application, but do not limit the present application in any way. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present application. These all fall within the protection scope of the present application.

[0062] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0063] Unless otherwise defined, all technical and scientific terms used in this specification have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The term "and / or" as used in this specification includes any and all combinations of one or more of the associated listed items.

[0064] Furthermore, the technical features involved in the various embodiments of this application described below can be combined with each other as long as they do not conflict with each other.

[0065] This disclosure aims to address the problems of time-consuming multiphysics coupling simulations in traditional chemical fire and explosion scenarios, difficulty in handling the dynamic evolution of complex domino chain accidents, and the tendency of purely data-driven artificial intelligence prediction models to produce "physical distortion." In view of this, this disclosure proposes a method and system for predicting the coupled consequences of chemical fires and explosions based on a large model, used to accurately predict the spatiotemporal evolution of complex disasters in chemical industrial parks and execute emergency containment control. This method employs intelligent computing technologies such as Physical Information Neural Networks (PINN) and Spatiotemporal Graph Attention Networks (ST-GNN). By using thermodynamic partial differential equations and material thermo-mechanical coupling failure criteria as a physical prior bias to force-embed a graph deduction framework, it accelerates the prediction of complex disaster consequences from the "hour-level" of traditional CFD to the "second-level," and completely eliminates the possibility of algorithmic predictions violating objective physical laws.

[0066] Please refer to Figure 1 , Figure 1 The diagram illustrates the framework of the chemical fire and explosion coupling consequence prediction system based on a large model according to an embodiment of this disclosure. The architecture includes: an edge sensing network deployed at the front end of the chemical plant area (including a distributed DCS system, pressure transmitters, anemometers, and level gauges), a cloud-based intelligent computing center (deployed with a large language model, PINN network, and graph computing engine), and physical execution terminals (including PLC controllers, fire deluge valve assemblies, and emergency shut-off valves).

[0067] The cloud-based intelligent computing center interacts with the edge sensing network and physical execution terminals via real-time signaling. Please refer to [link / reference needed]. Figure 2 , Figure 2 The flowchart of the chemical fire and explosion coupling consequence prediction method based on a large model according to an embodiment of this disclosure is shown; the overall process mainly includes the following 6 steps:

[0068] S1. Physical mapping of multi-source heterogeneous environments and construction of digital twins of graph nodes.

[0069] In the daily operation of a chemical industrial park, the health status and real-time operating conditions of physical equipment simultaneously determine the initial risk baseline of the system. The cloud-based intelligent computing center sends polling commands to the edge sensing network at 100ms clock cycles. The edge sensing network collects real-time fluid and meteorological environmental parameters of the i-th storage tank and packages them into a digital physical feature vector. :

[0070] ; in, This indicates the transient internal absolute pressure of the storage tank; Indicates the actual liquid level inside the storage tank; Indicates the ambient wind speed; Represents the matrix transpose operator. It represents the set of real numbers, meaning that the vector belongs to the three-dimensional real number column vector space.

[0071] Meanwhile, for hidden physical defects such as "structural fatigue" and "weld corrosion" that cannot be directly measured by hardware sensors, the cloud-based intelligent computing center uses a large language model to perform natural language parsing on the historical maintenance work orders and flaw detection reports of the i-th tank in the Enterprise Asset Management System (EAM). Defect keywords are extracted through the self-attention mechanism of the large language model, and then dimensionality is reduced using a pooling layer to generate a hidden state vector representing the physical degradation state. .

[0072] To uniformly map the aforementioned heterogeneous data onto the topological graph of the digital space, the cloud-based intelligent computing center establishes digital twin graph nodes for the devices. By introducing a learnable projection matrix, the physical feature vector and the text defect vector are aligned to the same latent space dimension and then concatenated to generate the fused feature vector of graph node i at the initial monitoring time (t=0). The specific quantification formula is as follows:

[0073] ;

[0074] in, This represents the feature tensor of graph node i at time t=0, with dimension . ; Represents a non-linear activation function; This represents the method used to map text hidden state vectors to... The first weight projection matrix of dimension , with dimension . ; This represents the device physical degradation hidden state vector output by the large language model, with dimension [missing information]. ; Represents the tensor concatenation operator; This represents the method used to map three-dimensional physical feature vectors to higher dimensions. The second weight projection matrix of dimension is ; This represents the physical feature vector uploaded by the edge sensing network at time t=0, including the internal pressure of the storage tank. Internal liquid level And the external environment side wind speed u, with dimension ; This represents the bias vector, with dimension . .

[0075] After initialization, the cloud-based intelligent computing center will fuse the feature vectors. The data is injected into the GPU memory of the graph computing engine, and the real-time collected physical environment parameters are streamed to step S2 to perform fluid dynamics monitoring.

[0076] S2, Fluid dynamics source term evolution and catastrophe critical point triggering.

[0077] Chemical leaks are a prerequisite for fires and explosions in engineering projects. The cloud-based intelligent computing center continuously monitors the real-time internal pressure transmitted in step S1. When the edge pressure transmitter detects that the rate of decrease in internal pressure over time exceeds the safety threshold, the system determines that the physical storage tank has torn or the flange has failed; at this time, the system calls the real-time liquid level of the storage tank. The generated hydrostatic head data is used to calculate the transient mass leakage rate under the current physical conditions based on the modified Bernoulli equation in fluid dynamics. :

[0078] ;

[0079] in, This represents the physical leakage mass flow rate of the i-th storage tank at time t, driven by a fluid dynamics mechanism. The dimensionless emission coefficient characterizing the fluid contraction effect at the breach; This represents the physical damage area calculated by the system based on the pressure drop gradient inversion. This indicates the actual liquid density of the hazardous chemical material; This represents the transient internal absolute pressure of the storage tank input in step S1; The external standard atmospheric pressure constant is represented by g; g represents the gravitational acceleration constant. This indicates the actual liquid level inside the tank as input in step S1.

[0080] The cloud computing center will then As a boundary source term, it drives a pre-built three-dimensional gas cloud diffusion neural network. The system monitors the diffusion concentration field in real time. Once a concentration patch exceeding the lower explosive limit (LFL) of the material is detected in the three-dimensional spatial grid, a highest-priority hardware interrupt signal is immediately sent to the main control bus, and the current system clock is recorded as the ignition timestamp. And forcefully activate the thermal field calculation process in step S3.

[0081] S3. Thermal radiation field derivation of a large physical model (PINN) under confined conditions.

[0082] Once an accident is triggered, traditional pure data-driven networks, lacking physical boundary constraints, are prone to "algorithmic illusions" that violate the law of conservation of energy when extrapolating the spread of heat from a fire. To address this, the cloud-based intelligent computing center activates a Physical Information Neural Network (PINN). PINN extracts the real-time wind speed u input in step S1 and calculates the dimensionless air-to-height ratio of the storage tank. (Characterizing the "chimney effect" potential of the gas phase space inside the storage tank), the loss function that forcibly embeds the thermodynamic combustion partial differential equation into the network. Iterative convergence is performed during the process:

[0083] ;

[0084] in, This represents the scalar of mean square error loss under partial differential equation constraints. Represents partial differential operators; This represents the transient temperature field on the flame surface calculated by PINN; Represents a time variable; Indicates the physical thermal diffusivity of hazardous chemical materials; Denotes the Laplace space differential operator; The pre-exponential factor constant representing the chemical reaction of combustion; Represents the natural constant; This represents the activation energy required for the combustion reaction; Represents the ideal gas constant; This represents a dimensionless flame deflection attitude compensation function calculated based on the actual crosswind speed and air height ratio, used to correct the actual spatial shape of the flame. This represents the ambient wind speed obtained in step S1; This indicates the actual liquid level height of the storage tank obtained in step S1; Indicates the total design height of the storage tank; This represents the mathematical norm of the mean square error.

[0085] After the model converges, the cloud-based intelligent computing center extracts the effective surface temperature of the flame. Based on the solid flame model of heat transfer, the incident heat radiation flux projected from accident source i onto the surface of the adjacent target tank j in physical space is calculated. :

[0086] ;

[0087] in, This represents the dynamic incident thermal radiation flux received by the neighboring storage tank j; This represents the dimensionless atmospheric transmittance dynamically affected by wind speed u. The dimensionless emissivity of a flame; This represents the Stefan-Boltzmann constant; This indicates the effective surface temperature of the flame output by PINN; Indicates based on the physical tilt angle of the flame And the dimensionless spatial geometric perspective factor for real-time reconstruction of the spatial distance between the two devices (whose physical meaning already includes the spatial distance attenuation effect).

[0088] S4. Analysis of the thermo-mechanical dual-coupling tensor and quantification of buckling failure of the target equipment.

[0089] In practical engineering, explosions of adjacent storage tanks are often not caused by simple high-temperature melting, but by the deterioration of the yield strength of the metal shell at high temperatures, making it unable to resist the circumferential expansion caused by the internal liquid hydrostatic pressure, thus resulting in mechanical buckling and tearing; the cloud-based intelligent computing center receives the thermal load transmitted in step S3. Solve for the nonlinear temperature rise of the fire-facing shell of the target storage tank j. The system consults the built-in material physical property library, extracts the shell's elastic modulus at the current temperature, and, based on the objective mechanism of physical compression caused by thermal expansion of the shell, combines the liquid level height of the target tank j in step S1. The cloud-based intelligent computing center constructs the principal stress tensor of the thin-walled cylinder and calculates the thermal expansion stress of the shell. Circumferential stress of fluid With fluid axial stress :

[0090] ;

[0091] ;

[0092] in, A scalar representing the thermal stress caused by restricted thermal expansion; This represents the elastic modulus of the shell steel under temperature-dependent conditions. Indicates the coefficient of linear expansion of the shell material; This indicates the transient absolute temperature of the fire-facing surface of the casing; Indicates the initial absolute temperature; Indicates the Poisson's ratio of the material; and These represent the physical circumferential stress and axial stress caused by the hydrostatic pressure of the liquid level, respectively. and These represent the target tank wall thickness and diameter obtained in step S1, respectively. Represents the gravitational acceleration constant; This indicates the real-time liquid level of a nearby storage tank.

[0093] The system employs the fourth strength theory (Von Mises criterion) to aggregate the aforementioned stresses into an equivalent coupled stress scalar. :

[0094] ;

[0095] Extract the material yield strength at the transient absolute temperature of the fire-facing surface of the shell. , building time Dynamic load factor :

[0096] ;

[0097] in, This indicates the thermal expansion stress of the shell; Indicates the time of integration Equivalent coupling stress at time; This represents the material's yield strength at the transient absolute temperature of the fire-facing surface of the shell.

[0098] The Weibull model is used to map "engineering mechanical parameters" to the physical failure probability of target node j. :

[0099] ;

[0100] in, Scalar represents the failure probability of target node j under thermal double tearing, with a value between [0,1]. This represents the fire ignition timestamp captured in step S2; t represents the absolute time when the system is currently performing a mechanical assessment. Indicates time Dimensionless dynamic load factor; Represents the time-dependent variable in the integration process; Represents the integral time element; Characteristic scale parameters representing the evolution of fatigue life of metal materials; This represents the dimensionless shape parameter of the Weibull distribution.

[0101] Physical failure probability scalar derived from mechanical analysis It is directly configured as the forced physical bias term of the artificial intelligence graph network in step S5.

[0102] S5. Derivation of a Spatiotemporal Graph Attention Network (ST-GNN) with Strong Physical Prior Probability Involvement.

[0103] In response to a large-scale domino effect across the entire plant, the cloud-based intelligent computing center activates a spatiotemporal graph attention network. To ensure that the AI's deductive path does not deviate from the objective laws of real physical development, the system innovatively modifies the traditional data-driven attention mechanism at the physical level. The system reads the high-dimensional feature tensors of the graph nodes cached in step S1. and Attention weight for disaster spread between execution nodes During the calculation, the physical and mechanical failure probability scalar output in step S4 is forcibly applied. It is added as a separate penalty / incentive term inside the activation function:

[0104] ;

[0105] in, The dimensionless attention weight score represents the trend of disaster energy transfer from source node i to target node j. This represents an exponential function with the natural constant e as its base. This represents a nonlinear activation function for a modified linear unit with leakage. This represents a learnable weight vector, whose dimension is... Ensure that the concatenated feature tensor is dimensionality-reduced and mapped to scalar scores, thus ensuring strict self-consistency with the subsequent equation dimension; and This represents the feature tensors of nodes i and j generated in step S1 at time t, with each tensor having dimension 1. ; This represents the tensor concatenation operator along the feature dimension; This represents a pre-defined dimensionless physical prior penalty amplification factor, used to adjust the influence of the mechanical failure probability in the AI ​​model. This represents the scalar physical failure probability of the target node j calculated in step S4. Represents the set of all local neighbor nodes of source node i in physical space. Perform the traversal and summation; The feature tensor representing the neighbor node k; This represents the scalar physical failure probability of neighbor node k.

[0106] By incorporating prior knowledge of materials mechanics into the graph network update described above, the system accurately predicts the queue of catastrophic nodes whose physical bearing capacity is about to be exhausted within a future preset time window, and outputs the set of activated catastrophic nodes. Then submit it to step S6 to perform hardware intervention.

[0107] S6. Global system vulnerability tracing and physical execution closed-loop blocking control.

[0108] The ultimate goal of this invention is to guide engineering intervention using the calculation results; the cloud-based intelligent computing center uses the disaster set generated in step S5 as a basis. Calculate the system-level probability of a catastrophic total collapse in the entire chemical industrial park. To accurately determine the equipment for which physical rescue actions should be prioritized, the system performs reverse calculations on the critical sensitivity scalar of each independent node in the set. :

[0109] ;

[0110] in, This represents the dimensionless critical sensitivity of target node j to the overall system collapse effect; Indicates based on The partial calculus derivative of the total collapse probability of the global domino accident system constructed from the set; The integral partial derivative represents the probability of physical failure of the corresponding target node j. This represents the transient physical failure probability of target node j calculated in step S4; This represents the total probability of system collapse in a global accident.

[0111] Cloud-based intelligent computing center The system sorts all equipment in descending order based on the numerical values, identifying the highest-ranking "high-risk hub tank". The system encapsulates the physical logic address of this tank into a control message and sends it to the PLC (Programmable Logic Controller) on site via industrial Ethernet. The PLC responds to the received control message by driving the intermediate relay in the explosion-proof zone of the corresponding tank to engage, physically forcibly opening the fire deluge valve assembly to spray water at a high flow rate for cooling, and simultaneously triggering the emergency shut-off valve (ESD) at the bottom of the tank to cut off the feeding system, thereby completely cutting off the physical evolution path of the accident and completing an automated defense closed loop from environmental perception and large-scale model calculation to electromechanical control.

[0112] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for predicting the coupled consequences of chemical fires and explosions based on a large model, characterized in that, The method includes the following steps: The real-time physical feature vector of the target storage tank is obtained. The real-time physical feature vector includes at least the real-time internal pressure and the real-time liquid level. The historical maintenance text of the target storage tank is processed using a large language model to obtain the text hidden state vector. The real-time physical feature vector is aligned and concatenated with the textual hidden state vector to construct an initial feature vector for the graph nodes of the target storage tank. When the real-time internal pressure drop of the target storage tank exceeds a preset threshold, the transient leakage mass flow rate is calculated based on the real-time internal pressure and the real-time liquid level, and the transient leakage mass flow rate is input into a preset diffusion neural network to obtain a spatial concentration field, so as to determine the ignition timestamp when the spatial concentration field reaches the lower explosion limit. In response to the ignition timestamp, the effective surface temperature of the flame is obtained by solving the physical information neural network based on the constrained partial differential equation of combustion energy conservation, and the incident heat radiation flux of the target tank to the adjacent tank is calculated based on the solid flame model. The transient temperature of the shell of the adjacent storage tank is calculated based on the incident thermal radiation flux, and the equivalent coupling stress is calculated in combination with the real-time liquid level of the adjacent storage tank. The equivalent coupling stress is then mapped to the physical failure probability. The physical failure probability is embedded as a prior bias term into the spatiotemporal graph attention network to update the disaster propagation weights between nodes and output the active disaster set. Based on the activated disaster set, the critical sensitivity of all network nodes is calculated, and based on the critical sensitivity, an electromechanical control message is sent to the target programmable logic controller to control the corresponding storage tank to perform a physical blocking action; The partial differential equation for the conservation of combustion energy serves as the loss function term of the physical information neural network. Its expression is: ； in, This represents the scalar value of the mean squared error loss. Represents partial differential operators; This represents the transient temperature field on the flame surface; Represents a time variable; Indicates the thermal diffusivity; Represents the Laplace operator; This represents the pre-exponential factor constant; Represents the natural constant; Indicates activation energy; Represents the ideal gas constant; This represents the flame tilt angle compensation function; Indicates ambient wind speed; Indicates real-time liquid level; Indicates the total design height of the storage tank; This represents the mean squared error norm; The equivalent coupling stress is calculated by combining the real-time liquid level of the adjacent storage tank, and the equivalent coupling stress is mapped to the physical failure probability. The specific calculation process includes: Calculate the thermal expansion stress of the shell of the adjacent storage tank. Circumferential stress of fluid With fluid axial stress : ； Based on the fourth strength theory, the thermal expansion stress, the fluid circumferential stress, and the fluid axial stress are aggregated into an equivalent coupled stress. : ； Extract the material yield strength at the transient absolute temperature of the fire-facing surface of the shell. , building time Dynamic load factor : ； Substituting the dynamic load factor into the Weibull distribution model, it is mapped to the physical failure probability. : ； in, Indicates the elastic modulus; Indicates the coefficient of linear expansion; This indicates the transient absolute temperature of the fire-facing surface of the casing; Indicates the initial absolute temperature; Indicates Poisson's ratio; Indicates the density of the liquid; Represents the gravitational acceleration constant; Indicates the real-time liquid level of adjacent storage tanks; Indicates the diameter of the storage tank; Indicates the thickness of the tank wall; This indicates the thermal expansion stress of the shell; Indicates equivalent coupling stress; Indicates the time of integration Equivalent coupling stress at time; This represents the material yield strength at the transient absolute temperature of the fire-facing surface of the shell. Indicates the fire time stamp; t represents the current assessment time; Indicates time Dynamic load factor; Represents the integral time element; Parameters representing the characteristic scale of fatigue life; This represents the shape parameter of the Weibull distribution.

2. The method for predicting the coupled consequences of chemical fires and explosions based on a large model according to claim 1, characterized in that, The real-time physical feature vector is aligned and concatenated with the textual latent state vector to construct the initial feature vector of the graph nodes for the target storage tank. The specific calculation formula is as follows: ； in, This represents the initial feature vector of the graph node i corresponding to the target storage tank; Represents a non-linear activation function; This represents the first weight projection matrix; Represents the hidden state vector of the text; Represents the tensor splicing operator; This represents the second weight projection matrix; The real-time physical feature vector represents the initial monitoring time, and the real-time physical feature vector includes: real-time internal pressure, real-time liquid level, and ambient wind speed; This represents the bias vector.

3. The method for predicting the coupled consequences of chemical fires and explosions based on a large model according to claim 1, characterized in that, The transient leakage mass flow rate is calculated based on the real-time internal pressure and the real-time liquid level. The specific calculation formula is as follows: ； in, This represents the transient leakage mass flow rate of the target storage tank at time t; Indicates the emission factor; Indicates the area of ​​physical damage; Indicates the liquid density of the material; Indicates real-time internal pressure; The standard atmospheric pressure constant is represented by g; g represents the gravitational acceleration constant. Indicates the real-time liquid level.

4. The method for predicting the coupled consequences of chemical fires and explosions based on a large model according to claim 1, characterized in that, The incident heat radiation flux from the target tank to adjacent tanks is calculated based on a solid flame model. The specific calculation formula is as follows: ； in, This represents the dynamic incident thermal radiation flux received by the neighboring storage tank j; Indicates atmospheric transmittance under the influence of wind speed; Indicates the emissivity of flame blackness; This represents the Stefan-Boltzmann constant; Indicates the effective surface temperature of the flame; Indicates the physical tilt angle of the flame Reconstructed spatial geometric perspective factors.

5. The method for predicting the coupled consequences of chemical fires and explosions based on a large model according to claim 1, characterized in that, The physical failure probability is embedded as a priori bias term into the spatiotemporal graph attention network to update the disaster propagation weights between nodes. The specific calculation formula is as follows: ； in, Indicates the weight of disaster propagation; This represents the modified linear unit activation function with leakage; Represents a learnable weight vector; and Let i and j represent the feature tensors of nodes i and j at time t, respectively; This represents the physical prior penalty amplification factor; This represents the physical failure probability of node j corresponding to the adjacent storage tank; Represents the set of local neighbor nodes; The feature tensor representing the neighbor node k; This represents the physical failure probability of neighbor node k.

6. A chemical fire and explosion coupled consequence prediction system based on a large model, characterized in that, The system is used to implement the chemical fire and explosion coupling consequence prediction method based on a large model as described in any one of claims 1 to 5. The system includes: an edge sensing network, a cloud computing center, and a physical execution terminal. The edge sensing network is used to collect real-time physical parameters and environmental meteorological data of the target storage tank. The real-time physical parameters include at least real-time internal pressure and real-time liquid level. The cloud-based intelligent computing center is communicatively connected to the edge sensing network, and the cloud-based intelligent computing center includes: The feature construction submodule is used to process the historical maintenance text of the target storage tank using a large language model to obtain a textual latent state vector, and to align and concatenate the real-time physical parameters with the textual latent state vector to construct the initial feature vector of the graph node. The source term calculation submodule is used to calculate the transient leakage mass flow rate and extrapolate the spatial concentration field when the real-time internal pressure drop exceeds a preset threshold in order to determine the ignition timestamp. The radiation inference submodule is used to solve for the effective surface temperature of the flame based on the physical information neural network, and to calculate the incident heat radiation flux of the target tank to the adjacent tank. The failure assessment submodule is used to solve the transient absolute temperature of the shell fire-facing surface of the adjacent storage tank and calculate the equivalent coupling stress, and map the equivalent coupling stress to the physical failure probability. The graph network deduction submodule is used to deduce the activation catastrophe set using a spatiotemporal graph attention network; The electromechanical control submodule is used to calculate the critical sensitivity based on the activated disaster set and generate electromechanical control messages; The physical execution terminal is communicatively connected to the cloud-based intelligent computing center and is used to receive the electromechanical control messages and drive the emergency equipment of the corresponding storage tank to perform physical blocking actions.

7. An electronic device, characterized in that, include: A processor and a memory for storing processor-executable instructions; The processor is configured to execute the instructions to implement the large-model-based method for predicting the coupled consequences of chemical fires and explosions as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, When the instructions in the computer-readable storage medium are executed by the processor of the electronic device, the electronic device is able to perform the large-model-based method for predicting the coupled consequences of chemical fires and explosions as described in any one of claims 1 to 5.

Images