Mathematical intelligence learning method for forward and inverse modeling of temperature field and functional damage effect of tunnel fire

Abstract

The present invention discloses a digital intelligence learning method for forward and inverse modeling of a temperature field of a tunnel fire and its functional damage effect. The method trains and optimizes a temperature field neural network through a temperature field data set including a training data set and a verification data set, obtains a temperature field digital intelligence learning model, and realizes digital intelligence learning of the temperature field of a tunnel fire; then, according to the learning result of the digital intelligence learning of the temperature field of a tunnel fire, a tunnel damage data set is constructed, and based on the tunnel damage data set, a forward modeling data set and an inverse modeling data set are obtained, which are respectively substituted into the forward modeling neural network and the inverse neural network, and data learning and iteration are performed to construct a tunnel functional damage forward model and a fire source information inversion model. The method of the present invention has high prediction efficiency, strong generalization ability, and high precision.

Description

Technical Field

[0001] The present invention belongs to the field of data processing technology, and in particular to a digital intelligence learning method for forward and inverse prediction of tunnel fire temperature field and its functional damage effect with high prediction efficiency, strong generalization ability and high precision. Background Art

[0002] The fire temperature field refers to the area of ​​temperature distribution in the space around the fire source due to combustion during the fire process. The temperature of this area varies with time and space, and is an important parameter for evaluating the impact of fire and conducting fire simulation.

[0003] Functional damage effect under fire refers to the damage and impact caused by fire on the function of buildings, facilities or equipment. It is not limited to the destruction of physical structure, but also includes damage to the normal operation ability of buildings or facilities. When a fire occurs, due to factors such as high temperature, flames, smoke and toxic gases, the building structure is damaged, the facilities cannot operate normally, or the equipment loses function. This damage effect includes not only direct damage to the physical structure, but also indirect functional losses caused by fire, such as power supply interruption, communication system failure, etc., which will seriously affect the normal use of buildings or facilities and the safety of personnel.

[0004] The forward and inverse mathematical learning method refers to the use of physics-driven deep learning technology, especially Physics-Informed Neural Networks (PINNs). By combining deep learning with the laws of physics and using data-driven methods, it improves the predictive ability and generalization of the model and solves forward and inverse problems.

[0005] 1. Forward Problem: Forward problem refers to predicting the output or behavior of a system based on known physical models and input conditions. Forward simulation is the process of predicting data from a known model.

[0006] 2. Inverse Problem: Inverse problem refers to the process of inferring the model parameters or state of the system based on the observed data. Inverse simulation is the process of inferring the model from known data.

[0007] 3. Physics-Informed Deep Learning: By embedding physical laws into deep learning models, the models can not only learn patterns in data, but also follow physical laws. For example, physics-informed neural networks (PINNs) improve the prediction accuracy and generalization ability of the model by adding physical equations as constraints during the training process of the neural network.

[0008] 4. Digital Intelligence: Digital Intelligence refers to the combination of digitization and intelligence. Intelligence is achieved through digitization, and digitalization is developed with the help of intelligence. By using technologies such as big data, cloud computing, and artificial intelligence, the system can have the ability to realize state perception, real-time analysis, scientific decision-making, and precise execution.

[0009] Forward and inverse numerical learning of the temperature field of tunnel fires and their functional damage effects is the basis of safety protection design, and is also an important prerequisite for predicting and evaluating the consequences of tunnel fires and formulating emergency response plans.

[0010] At present, as described in the paper "Research on Fire Temperature Field of Hydrogen Energy Vehicles in Tunnels under Multiple Working Conditions" (Liu Dong, Ye Jihong. Fire Science and Technology, 2024, 43(3): 344-350. DOI: 10.3969 / j.issn.1009-0029.2024.03.011.), it discloses a method for predicting the temperature field of a hydrogen energy vehicle after releasing hydrogen in different directions in a tunnel and igniting it, including: by using the computational fluid dynamics numerical simulation method, studying the temperature field generated after the hydrogen is released and immediately ignited, and obtaining the influence of the hydrogen leakage direction and the tunnel ventilation speed on the maximum temperature of the jet flame and the distribution of the high temperature zone; by establishing tunnel and vehicle models, setting boundary conditions such as air entry speed and hydrogen release pressure, and using a virtual nozzle model to simplify the calculation, the result of a stable temperature field is obtained.

[0011] However, the above methods rely on complex simulation systems. High-precision numerical simulation requires a lot of computing resources and time, which limits the scale and complexity of the simulation, greatly increases the computing time and consumes a lot of computing resources. It is very time-consuming and labor-intensive and not efficient enough. On the other hand, the many assumptions made in numerical simulation cannot fully capture the dynamic changes in the development of fires. The research mainly focuses on hydrogen-powered car fires, without considering other types of fires or the occurrence of multiple fires at the same time, and has weak generalization ability.

[0012] Another example is the paper “Building target damage assessment based on finite element simulation results identification” (Wang Jun, Lei Hongyu. [J]. Command Control and Simulation, 2024, 46(4): 150-155. DOI:10.3969 / j.issn.1673-3819.2024.04.020.) discloses a building target damage assessment method based on finite element simulation result identification, which completes the identification of result image change features and index quantification through numerical simulation, and further combines the target function and physical damage level judgment criteria to achieve target damage pre-assessment before the strike; uses the finite element numerical simulation results to extract image change features, including the building top angle, visible surface and volume damage degree; by using the area calculation method based on the change contour, the ratio of the damaged area to the original area is calculated to obtain the surface area damage ratio; by comparing the volume of the changed part of the building with the initial volume, the volume change ratio is calculated to obtain the volume damage ratio, and the corresponding relationship between the damage element, the damage assessment parameter index and the damage level is established to determine the damaged part and damage degree of the building target; combined with historical experience data, the building target function and physical damage level judgment criteria are established, and the index synthesis is performed through the weighted average method to give the functional damage evaluation level of the building target.

[0013] However, the above methods rely on numerical simulation, and the finite element model is highly detailed and complex, occupying a large amount of hard disk space and memory resources during the calculation process; when establishing the model, many structural and material simplifications and assumptions are used, such as ignoring certain material properties, boundary conditions or geometric shapes, which causes the results to deviate from the actual situation; when dealing with complex boundary conditions, such as when encountering nonlinear or non-uniform boundary conditions, special processing is required, which increases the amount and complexity of calculations; the results of finite element analysis are highly dependent on the accuracy of the input data, such as material parameters, geometric shapes and boundary conditions, and any errors in the input data may lead to deviations in the final results.

[0014] For example, the paper "Inversion of Fire Source Location and Heat Release Rate in Tunnel Fire Based on Deep Learning" (Jiang Li, He Tingquan, Guo Xin, et al. [J]. Fire Science and Technology, 2024, 43(9): 1292-1297, 1315. DOI: 10.20168 / j.1009-0029.2024.09.1292.06.) discloses a prediction method for predicting and inverting the fire source location and HRR in tunnel fires by constructing and training a deep learning model and using numerical simulation and experimental data, including: establishing a tunnel fire data set through numerical simulation, and verifying the validity of these data sets with experimental data; generating sensor temperature time series data under different boundary conditions through numerical simulation, and dividing them according to different time input steps and sensor spacing to establish multiple data sets; By using convolutional neural networks (CNN) to extract spatial and local features in sensor data and long short-term memory networks (LSTM) to process time series data, a model is established to improve the inversion effect and realize the relationship mapping between fixed sensor temperature data and fire source location and HRR. By inputting the data set into the deep learning model for training, the inversion results of the fire source location and HRR are obtained, and the inversion effect of the model is evaluated. By comparing the model performance under different time steps and sensor spacings, the influence of time input step and sensor spacing on the model inversion performance is analyzed, and the optimal model parameters are determined.

[0015] However, the above methods rely on inversion simulation to establish data sets. When they are not fully represented in the training data, they perform poorly in dealing with abnormal fire conditions or atypical fire behaviors, and the simulation accuracy is insufficient, which affects the reliability of subsequent inversion results. Although the effectiveness of numerical simulation is verified by using medium-scale fire tests, it is limited by the test conditions and scale and cannot fully represent the actual situation of full-scale tunnel fires. Limited by the specific tunnel size and fire conditions, the generalization ability of the model to other types of tunnels or fires of different scales is not high. Limited by the sensor layout and data volume used, the inversion ability of the model is limited, and the physical dependence between 0HRR and sensor temperature data is relatively complex. The model may require more physical information or a more complex model structure to more accurately invert HRR. The model combining CNN and LSTM is computationally complex and resource-intensive, which limits the application of the model in resource-constrained environments. There is a delay in processing real-time data, and it cannot achieve rapid response in emergency situations. In short, the problems existing in the prior art are: the forward and inversion methods of tunnel fire temperature field and its functional damage effect are not efficient enough, the generalization ability is not strong enough, and the accuracy is not high enough. Summary of the invention

[0016] The purpose of the present invention is to provide a digital intelligence learning method for forward and inverse modeling of tunnel fire temperature field and its functional damage effect, which has high efficiency, strong generalization ability and high precision.

[0017] The technical solution to achieve the purpose of the present invention is:

[0018] In a first aspect, a digital intelligence learning method for a tunnel fire temperature field is provided, comprising:

[0019] The temperature field neural network is trained and optimized by using a temperature field data set including a training data set and a verification data set to obtain a temperature field digital intelligence learning model, thereby realizing digital intelligence learning of the tunnel fire temperature field; wherein the temperature field neural network includes an encoder, a processor, and a decoder, wherein:

[0020] The encoder includes a Fourier encoder and a long-short time encoder. The weights of the Fourier encoder and the long-short time encoder are obtained by using a feedforward neural network, and the corresponding weights are respectively embedded in the Fourier encoder and the long-short time encoder. The outputs of the Fourier encoder and the long-short time encoder are superimposed to obtain a coding result.

[0021] In the processor, a multi-layer perceptron embedded with a ReLU activation function is used to iterate the encoding result to obtain context vector data;

[0022] In the decoder, a recurrent neural network is used to process the context vector data to generate an output sequence, and then a feedforward neural network is used to process the output sequence to output the decoding result.

[0023] Preferably, geometric physical information is embedded in the temperature field neural network, and a geometric loss term is added to the loss function of the obtained temperature field neural network. and physical loss :

[0024] .

[0025] .

[0026] in, Indicates the number of point cloud transformations, represents the temperature field predicted after the i-th point cloud transformation, represents the real temperature field, Indicates the number of nodes, Representation Node Predicted temperature field, t represents time.

[0027] Preferably, the temperature field data set construction step includes:

[0028] Collect the test data of tunnel fire temperature field, obtain the data of tunnel three-dimensional characteristics, building material characteristics, tunnel fire temperature, fire heat radiation flux and fire source parameters, and realize the fire feature selection; the fire source parameters include fire source location and fire source heat radiation flux, and the building material characteristics include yield stress of building materials, elastic modulus of building materials, and thermal expansion coefficient of building materials;

[0029] The selected fire characteristics are screened and reconstructed. Through missing value completion, amplitude adjustment, and spatial correlation conversion, the tunnel fire temperature, thermal radiation flux, fire source parameters, and building material characteristics are converted into a two-dimensional array form to obtain the processed tunnel fire temperature field test data.

[0030] The processed tunnel fire temperature field test data are combined according to the tunnel fire cases to obtain the tunnel fire temperature field test data set;

[0031] Analyze the test data of tunnel fire temperature field and quantify the fire scenarios and parameters. The fire parameters include at least the probability of vehicle fire in the tunnel, the vehicle type and its corresponding fire scale. The fire scale parameters include the location of the fire source in the tunnel and the heat radiation flux of the fire source.

[0032] With the help of quantified fire scenarios and parameters, the probability of fire occurrence and vehicle type are taken as random variables;

[0033] The specific value of the random variable is obtained through the random number method, and the corresponding fire source location and fire source heat radiation flux are obtained by using the data interpolation method to realize the random process simulation;

[0034] Conduct statistical analysis on the fire source location and fire source heat radiation flux obtained from multiple random process simulations to obtain the distribution law of the fire source location and fire source heat radiation flux, and implement Monte Carlo analysis;

[0035] Through Monte Carlo analysis and digital twin method, we can obtain the twin of the real tunnel and the surrounding environment. Through meshing, we can obtain the numerical mesh data and conduct simulation. With the help of data crawling method, we can obtain the characteristics of building materials. Through data alignment, we can obtain the tunnel fire simulation data set.

[0036] The temperature field dataset of tunnel fire temperature field test and tunnel fire simulation was integrated, nonlinearity and evaluation index were added as new features, time and tunnel fire case number were added, and the temperature field dataset was obtained.

[0037] Preferably, the data pattern in the temperature field data set is [tunnel fire case number, node number, time, node location information, yield stress of building materials, elastic modulus of building materials, thermal expansion coefficient of building materials, fire source location, cosine value of the angle between the node and the fire source location, node fire temperature, node fire thermal radiation flux, fire source thermal radiation flux, nonlinearity, evaluation index].

[0038] Preferably, the nonlinearity and evaluation index formulas are as follows:

[0039]

[0040]

[0041] Where NL represents nonlinearity, MD represents maximum deviation, and FSR represents full-scale range.

[0042] Secondly, a digital intelligence learning method for forward and inverse modeling of tunnel function damage effects is also provided, including:

[0043] Construct a tunnel damage dataset, which consists of tunnel fire case number, node number, time, node location information, yield stress of building materials, elastic modulus of building materials, thermal expansion coefficient of building materials, fire source location information, cosine value of the angle between the node and the fire source location, node thermal stress, node fire temperature, node fire thermal radiation flux, fire source thermal radiation flux, nonlinearity, evaluation index, node physical damage, vertex damage ratio, surface area damage ratio, volume damage ratio, and tunnel functional damage; among them, the tunnel fire temperature field and tunnel fire thermal radiation flux field are predicted by the temperature field digital intelligence learning model as described above;

[0044] Based on the tunnel damage dataset, the forward modeling dataset and the inversion dataset are obtained and substituted into the forward modeling neural network and the inversion neural network respectively for data learning and iteration to construct the tunnel functional damage forward model and the fire source information inversion model.

[0045] Preferably, the step of constructing the tunnel damage data set includes:

[0046] Collect tunnel fire information at the fire scene, including fire source location, fire source heat radiation heat flux, building material characteristics, tunnel positioning, tunnel fire temperature field, tunnel fire heat radiation flux field;

[0047] Using digital twin methods, we can obtain a real tunnel twin;

[0048] The tunnel fire information at the fire scene is processed and converted, combined according to the tunnel fire case, to obtain the real tunnel fire event data, and combined with the real tunnel twin to realize the digital twin of the tunnel fire condition;

[0049] Using the temperature field mathematical intelligence learning model, the tunnel fire temperature field and tunnel fire thermal radiation flux field are predicted;

[0050] Using the relationship equation between temperature change and tunnel thermal stress , to obtain the tunnel stress field; where, represents the thermal stress at node j at time t, represents the elastic modulus of the building material, represents the coefficient of thermal expansion of building materials, Indicates the predicted time The node below The corresponding temperature field is Indicates the predicted time The node below The corresponding temperature field;

[0051] By describing the equations that relate physical damage to temperature fields , obtain the node physical damage; where, is the physical damage of node j at time t, represents the thermal stress at node j at time t, Indicates the yield strength of the material;

[0052] According to the physical damage of the node, the number of vertex damage, the damaged surface area, and the damaged volume are obtained, and then the formula , and the tunnel function is damaged; among them, Indicates that the tunnel function is damaged. represents the vertex damage ratio, represents the surface area damage ratio, represents the volume damage ratio, and A, B, and C represent weights.

[0053] Preferably, the steps of constructing the tunnel function damage forward model include:

[0054] The fire temperature of each node in the tunnel at each time is taken as input, and the tunnel function damage is taken as output to obtain the forward modeling data set;

[0055] The self-attention memory module is embedded in the convolutional long short-term memory network, and the data is processed by the tanh function combined with the sigmoid function to form a gate, update the memory information, and output the data through gated processing to construct a forward neural network;

[0056] The forward modeling data set is substituted into the forward modeling neural network for data learning and iteration to obtain the tunnel functional damage forward model.

[0057] Preferably, the fire source information inversion model construction step includes:

[0058] Taking the tunnel function damage as input, the fire source location information, the cosine value of the angle between the node and the fire source location, the node thermal stress, the node fire thermal radiation flux, and the fire source thermal radiation flux as output, the inversion data set is obtained;

[0059] Using multi-layer perceptron neural network, according to the temperature control equation, combined with nonlinear activation function and node information aggregation strategy, an inversion neural network is constructed;

[0060] The inversion data set is substituted into the inversion neural network for data learning and iteration to obtain the fire source information inversion model.

[0061] Preferably, physical information is embedded in the inversion neural network, and a physical information constraint loss term is added to the loss function of the obtained inversion neural network. :

[0062]

[0063]

[0064]

[0065]

[0066]

[0067] In the formula, Indicates the number of nodes, Indicates the node number, represents the heat radiation flux of the fire source, Representation Node The distance to the fire source, Representation Node The cosine value of the angle to the fire source, represents the node heat radiation flux, Representation Node Physical damage, Representation Node Heat stress, represents the yield strength of the material, Indicates that the tunnel function is damaged. represents the vertex damage ratio, represents the surface area damage ratio, represents volume damage ratio, and D, E, and F represent weights.

[0068] Compared with the prior art, the present invention has the following significant advantages:

[0069] 1. High prediction efficiency: The graph neural network of the present invention establishes a forward and inverse prediction model for the temperature field of tunnel fire and its functional damage effect. Through the deep learning method of the three-dimensional physical field of the graph neural network, it directly processes the three-dimensional physical field calculation of the forward and inverse of the temperature field of tunnel fire and its functional damage effect, avoiding the traditional numerical format derivation and solution process, improving the calculation efficiency, and achieving higher prediction efficiency;

[0070] 2 Strong generalization ability: The present invention establishes a temperature field digital intelligence learning model and a tunnel function damage forward and inverse model by establishing tunnel temperature field data with random regional characteristics, random facility characteristics, and random fire types, thereby ensuring data diversity and the robustness of the neural network, being able to adapt to different tunnel fire application scenarios and having stronger generalization ability;

[0071] 3. High precision: The present invention combines the non-rigid point cloud registration method (NrtNet) and the physical information embedding method (PINNs) and applies them to the loss function of the model to construct a geometric and physical information embedding method, which improves the prediction accuracy of the model through information constraints in both geometry and physics.

[0072] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 It is the main flow chart of the numerical intelligence learning method for forward and inverse modeling of tunnel fire temperature field and its functional damage effects.

[0074] Figure 2 yes Figure 1 Flowchart of the steps for constructing the temperature field dataset.

[0075] Figure 3 yes Figure 1 Flowchart of the steps for building a digital intelligence learning model for medium temperature field.

[0076] Figure 4 It is a schematic diagram of the temperature field neural network architecture.

[0077] Figure 5 It is a flowchart of the total loss term of the geometric information constraints embedded in the neural network for obtaining the tunnel fire temperature field.

[0078] Figure 6 yes Figure 1 Flowchart of the steps for constructing the building damage dataset.

[0079] Figure 7 yes Figure 1 Flowchart of the steps for constructing the forward and inverse model of the functional damage of the medium tunnel. DETAILED DESCRIPTION

[0080] like Figure 1 As shown, the following steps are included:

[0081] S10. Construction of temperature field data set: Through the digital twin of the environment, the geographic information data of the tunnel interior and the surrounding environment are obtained, and the twin of the real tunnel and the surrounding environment is obtained. The Monte Carlo method is used to analyze the location of the fire source and the heat radiation flux of the fire source in the tunnel. The temperature field data set is obtained by constructing the experimental data set and the simulation data set.

[0082] like Figure 2 As shown, the temperature field data set construction steps include:

[0083] S11. Environmental digital twin: Obtain geographic information data of the real tunnel and surrounding environment through data crawling, and obtain the twin of the real tunnel and surrounding environment through the building three-dimensional digital twin method.

[0084] S12, Monte Carlo analysis: Analyze the tunnel fire temperature field test data, quantify the fire scene and parameters, and implement Monte Carlo analysis through random variable description and random process simulation.

[0085] Through Monte Carlo method analysis, the fire source of the simulation data set is obtained, and the fire source location and heat radiation flux of the tunnel fire in the tunnel fire simulation are analyzed.

[0086] The Monte Carlo analysis steps include:

[0087] S121. Quantification of fire scenarios and parameters: Analyze the tunnel fire temperature field test data, obtain parameters such as the probability of vehicle fires in the tunnel, vehicle types and their corresponding fire scales, and obtain quantified fire scenarios and parameters.

[0088] Fire scenarios and parameters refer to the probability of vehicle fires in tunnels, vehicle types and their corresponding fire scale parameters, where fire scale parameters include the location of the fire source in the tunnel and the heat radiation flux of the fire source, etc. Different numbers in the vehicle type represent different types.

[0089] S122. Description of random variables: With the help of quantified fire scenarios and parameters, the probability of fire occurrence and vehicle type are used as random variables.

[0090] S123. Random process simulation: obtain the specific value of the random variable through the random number method, and use the data interpolation method to obtain the corresponding fire source location and fire source heat radiation flux to achieve random process simulation.

[0091] S124. Monte Carlo analysis: Statistical analysis is performed on the fire source location and fire source heat radiation flux obtained from multiple random process simulations to obtain the distribution law of the fire source location and fire source heat radiation flux in the tunnel, thereby realizing Monte Carlo analysis.

[0092] S124. Statistical analysis methods in Monte Carlo analysis:

[0093] 1. Simple Monte Carlo method: This is a basic Monte Carlo method used to estimate the expected value of a function under a certain probability distribution. The formula can be expressed as:

[0094] (1)

[0095] in, is the expected value obtained by estimation, It is Random samples generated from independent trials, It is for The function value calculation of the sample is is the number of trials.

[0096] 2. Calculation method of estimated expected value in Monte Carlo method: In Monte Carlo method, the expected value of random variable can be estimated by generating a large number of random samples. If the sample obeys a known probability distribution (for example, normal distribution, uniform distribution, etc.), the expected value can be estimated by calculating the sample mean using the formula:

[0097] (2)

[0098] in, The representative is currently The expected value of the estimate is Will follow As the value increases, it gets closer and closer to the theoretical expectation value. The theoretical expectation is shown in the formula:

[0099] (3)

[0100] in, refers to the theoretical expected value, Represents the average level of all values ​​in the data set. A random variable representing a random sample.

[0101] 3. Estimation integral in Monte Carlo method: Assume that the function needs to be estimated In the interval The integral over can be approximated by generating random samples and calculating the corresponding function values:

[0102] (4)

[0103] in, Estimate the integral value. represents the integral variable, indicating the interval The amount of internal change. represents the estimation function, Represents the estimated function in the interval The probability distribution function on .

[0104] S13. Construction of test data set: by collecting tunnel fire temperature field test data and performing data analysis, fire feature selection is completed, the selected fire features are screened and reconstructed, and data combination is performed to complete data processing. By combining the processed data according to the tunnel fire case, the fire temperature field test data set is obtained.

[0105] The test data set construction steps include:

[0106] S131. Fire feature selection: By collecting tunnel fire temperature field test data, conducting test data analysis, environmental digital twin, and data crawling, we can obtain data on tunnel three-dimensional characteristics, building material properties, tunnel fire temperature, thermal radiation flux, and fire source parameters to achieve fire feature selection.

[0107] Fire characteristics include: data on tunnel 3D characteristics, tunnel fire temperature, thermal radiation flux, and fire source parameters;

[0108] Building material characteristics: Building material characteristics are obtained by data crawling. Building material characteristics include yield stress of building materials, elastic modulus of building materials, and thermal expansion coefficient of building materials;

[0109] Tunnel 3D features: Tunnel 3D features are represented by the 3D features of the tunnel entity model obtained by the digital twin of the environment, and are represented by the 3D position information of the entity nodes obtained by the digital twin;

[0110] Fire source parameters include: fire source location, fire source heat radiation flux;

[0111] Test data analysis: Complete data collection by collecting tunnel fire temperature field test data; check data integrity, handle missing values, identify and handle outliers or erroneous data, and standardize data formats; use descriptive statistics to analyze basic data characteristics, such as mean, median, mode, variance, standard deviation, etc.; select appropriate statistical methods, such as hypothesis testing, variance analysis, regression analysis, etc., to complete test data analysis.

[0112] S132, data processing: screen and reconstruct the selected fire characteristics, and convert the tunnel fire temperature, thermal radiation flux, fire source parameters, and building material characteristics into a two-dimensional array form through missing value completion, amplitude adjustment, and spatial correlation conversion to obtain the processed tunnel fire temperature field test data.

[0113] 1. Data screening: Improve the normalization of the collected data, eliminate the influence of different dimensions, compare and process the data at the same scale, and use multi-scale analysis methods to study the temperature field characteristics under tunnel fire conditions and normal operating conditions. For example, use the orthogonal decomposition method to extract low-dimensional basis vectors, decompose the original temperature field into basis functions and linear combinations of basis coefficients to achieve dimensionality reduction, complete feature extraction, and realize data screening of tunnel fire temperature field data.

[0114] 2. Data reconstruction: Using the intrinsic orthogonal decomposition basis and combining it with the tunnel fire temperature field data, the missing data are estimated by solving a set of linear equations. The data reconstruction of the tunnel fire temperature field data is achieved through the intrinsic orthogonal decomposition method with missing data.

[0115] 3. Two-dimensional array form: For the same tunnel fire case, the same array is used to represent it, where a single array consists of multiple node arrays;

[0116] For example, a single tunnel fire case: [[node number 0, node parameters], [node number 1, node parameters]], the node parameters include the tunnel fire temperature, thermal radiation flux, fire source parameters, and building material properties under this node.

[0117] S133, test data set construction: the processed tunnel fire temperature field test data are combined according to the tunnel fire cases to obtain the tunnel fire temperature field test data set.

[0118] S14. Construction of simulation data set: Through the Monte Carlo analysis, the fire source location and the heat radiation flux of the simulated tunnel fire are obtained. Combined with the environmental digital twin method, the twin of the real tunnel and the surrounding environment is obtained. Through grid division, numerical grid data is obtained for simulation. With the help of data crawling method, the characteristics of building materials are obtained. Through data alignment, the tunnel fire simulation data set is obtained.

[0119] Data alignment: Tunnel fire temperature, thermal radiation flux, fire source parameters, building material properties and other characteristics are organized into simulation data sets in the form of two-dimensional arrays according to node numbers and tunnel fire cases to achieve data alignment;

[0120] Different from the experimental data, the node numbers of the simulation data are not obtained by the digital twin, but by the numerical grid data to better construct the data set and conduct subsequent model training.

[0121] S15. Temperature field data set construction: Integrate the test data set and simulation data set, perform data processing, design nonlinearity and evaluation indicators, add nonlinearity and evaluation indicators as new data set features, add time, tunnel fire case number and other parameters in the two-dimensional array, and obtain the tunnel fire temperature field data set.

[0122] 1. Integration: Use time series to align data from different sources to the same time axis and complete the alignment of data sets containing time series information. For example, use interpolation methods to fill in missing values ​​in time series. Extract and select key features, reduce data dimensions, improve the efficiency of data fusion, and complete the integration of simulation data and test data.

[0123] 2. Data processing: Clear outliers and noise in simulation data and test data to ensure data consistency, such as unifying timestamps, units, and formats. Standardize data from different sources to achieve comparison and analysis on the same scale, and complete data cleaning and standardization. According to the reliability and relevance of the data, assign different weights to simulation data and test data to achieve data fusion and complete data processing. For example, if the reliability of the test data is high, a higher weight can be given to achieve weighted fusion of data.

[0124] For example, the tunnel fire temperature field dataset is represented as follows: [tunnel fire case number, node number, time, node location information, yield stress of building materials, elastic modulus of building materials, thermal expansion coefficient of building materials, fire source location, cosine value of the angle between the node and the fire source location, node fire temperature, node fire thermal radiation flux, fire source thermal radiation flux, nonlinearity, evaluation index] = [1, 2500, 50, (70, 10, 8), 20 6 , 30, (20,4,5), 0.99, 50, 205, 8, 30, 1%, 0.8]. This result shows that the yield stress of the building material is 20 under the tunnel fire event with tunnel fire case number 1. 6 Gpa, elastic modulus is 30 Gpa, thermal expansion coefficient is / ℃ in the three-dimensional physical field of the tunnel, located at the (70, 10, 8) node, which is numbered 2500. At the 50th moment, the fire source is located at (20, 4, 5), the cosine value of the angle between the node and the fire source position is 0.99, the node temperature is 205℃, the node thermal radiation flux is 8kW / m², the fire source thermal radiation flux is 30kW / m², the nonlinearity is 1%, and the evaluation index is 0.8.

[0125] Obtaining the cosine value of the angle between the fire source location: The cosine value of the angle between the node and the fire source location is obtained by the following formula (5):

[0126] (5)

[0127] In the formula, It represents the angle between the direction vector between the fire source and the node and the normal direction vector of the node surface. Represents the angle The cosine of the angle is obtained by taking the dot product of the two vectors and the ratio of their moduli. Refers to the coordinate origin of the 3D tunnel model. Refers to the location information of nodes in the 3D tunnel model. Refers to the location information of the fire source in the three-dimensional tunnel model.

[0128] Acquisition of nonlinearity and evaluation index: By determining the full-scale range FSR, measuring the actual output AO and the ideal output IO, calculating the maximum deviation MD, and designing the data nonlinearity NL as shown in the following formula (6), and designing the nonlinear evaluation index EI as shown in the following formula (7), adding the data nonlinearity as one of the data inputs to achieve data preprocessing.

[0129] The data set changes as shown in the formula:

[0130]

[0131]

[0132] Full Scale Range (FSR): The full scale range is the difference between the maximum and minimum values ​​that a sensor or measurement system can measure.

[0133] Actual Output (AO): Actual output refers to the output value of a sensor or measurement system under actual working conditions.

[0134] Ideal Output (IO): Ideal output refers to the output value of a sensor or measurement system under ideal working conditions, usually a linear function.

[0135] Maximum Deviation (MD): Maximum deviation refers to the maximum difference between the actual output and the ideal output.

[0136] Nonlinearity (NL): Nonlinearity refers to the ratio of the maximum deviation between the actual output and the ideal output to the full-scale range.

[0137] Evaluation Index (Evaluation Index, EI): The evaluation index refers to the ratio of nonlinearity to the maximum value of nonlinearity, which is used to evaluate the degree of nonlinearity of data.

[0138] S20, construction of digital intelligence learning model of temperature field: Combine Fourier transform (Fourier) and long short-term memory network (LSTM) graph neural network (GNN) to construct temperature field neural network. Use non-rigid point cloud registration and physical information constraint methods to construct geometric physical information constraints and embed them into temperature field neural network. Use S10, the temperature field data set obtained under the action of tunnel fire, to train, evaluate and optimize the constructed tunnel fire temperature field neural network to obtain the digital intelligence learning model of tunnel fire temperature field;

[0139] like Figure 3 As shown, the steps of constructing the temperature field digital intelligence learning model include:

[0140] S21. Temperature field neural network construction: Construct Fourier transform encoder and long short-term memory network encoder to realize encoder construction. Construct temperature field neural network by constructing processor and decoder. The temperature field neural network is as follows: Figure 4 As shown;

[0141] The temperature field neural network construction step includes:

[0142] S212, Fourier encoder construction: by constructing a Fourier operator, realizing Fourier transform, and constructing a Fourier encoder;

[0143] The Fourier operator is obtained by superimposing a multidimensional Fourier layer and a linear layer. The Fourier graph operator matrix stacking is realized by performing matrix multiplication in the frequency domain using the Fourier graph operator. The tunnel temperature field data is sorted according to the time series to obtain the time series data, which is divided into sliding windows according to the time series of the data. The Fourier transform is used to realize the extraction of short-time strong nonlinear features.

[0144] S213, long-short time encoder construction: sort and arrange the data set data in time series, construct the LSTM unit by constructing the input gate, forget gate, and unit gate, realize the update of hidden state and cell state, and obtain the encoded data through the hidden layer output of the last state;

[0145] S214, encoder construction: using a feedforward neural network, according to the data nonlinearity, the encoder weights are obtained, the weights are embedded into the Fourier encoder and the long-short time encoder, and a joint encoder is constructed;

[0146] The encoder weight is positively correlated with the nonlinearity in the data set. The greater the data nonlinearity, the greater the proportion of the corresponding Fourier encoder output in the total output. Conversely, the greater the proportion of the corresponding long-short-time encoder output in the total output.

[0147] S216, processor construction: using a multi-layer perceptron in combination with a RELU activation function, the encoded data output by the encoder is subjected to neural network iteration to obtain updated context vector data.

[0148] S217, decoder construction: using a recurrent neural network to generate an output sequence from the context vector generated by the processor, and using a feedforward neural network to convert the output sequence into decoded data with a time series;

[0149] S218, temperature field neural network construction: The temperature field neural network is constructed by connecting the encoder, the processor, and the decoder in series. It should be noted that the structures of the above processor and the decoder are prior art.

[0150] S22. Geometric and physical information embedding: Geometric and physical information embedding methods are as follows: Figure 5 As shown in the figure, considering the geometric similarity and consistency of physical laws, the non-rigid point cloud registration method is used to transform the point cloud information at different angles to register the loss term of the point cloud about the spatial position information difference. , obtain geometric information constraints and physical information constraints, and construct a new loss function containing geometric and physical information. The loss function includes the corresponding loss of the geometric point set , heat conduction physics equation and original , the physical information constraint loss term and the data loss term are added to the loss term of the geometric physical constraint graph neural network to obtain a tunnel fire temperature field neural network embedded with geometric physical information.

[0151] (8)

[0152] In the formula, represents the geometric and physical information loss term, , , They are the weights corresponding to geometric loss, physical loss, and data loss, and their sum is 1.

[0153] The step of embedding geometric physical information comprises:

[0154] S221, Geometric information constraint: Using the non-rigid point cloud registration method, through the point cloud information transformation at different angles, the geometric information constraint loss term of the point cloud about the spatial position temperature field information difference is registered. , the geometric information constraint loss term is added to the loss term of the geometric physics constraint graph neural network.

[0155] Point cloud information: Point cloud information includes node position and node characteristics. Node characteristics include node temperature and node thermal radiation flux.

[0156] Taking the temperature field as an example, It represents the average value of temperature error under each point cloud transformation obtained by MSE method.

[0157] (9)

[0158] In the formula, Indicates the number of point cloud transformations, represents the predicted temperature field after point cloud transformation, It represents the real temperature field.

[0159] S222. Physical information constraint: Obtain the tunnel fire temperature field parameters through the tunnel fire temperature field control equation, construct the physical information constraint loss term of the three-dimensional tunnel fire temperature field, and add the physical information constraint loss term to the loss term of the geometric physical constraint graph neural network.

[0160] As shown in formula (10), the control equation of tunnel fire temperature field is:

[0161] (10)

[0162] (11)

[0163] In the formula, Indicates temperature, Indicates time, is the thermal diffusivity of the material, is the Laplace operator.

[0164] (12)

[0165] In the formula, represents the number of nodes for the numerical simulation of the tunnel fire temperature field, represents the node number, Represents the predicted temperature of a single node, and xyz represents the direction of the xyz axis.

[0166] S223. Geometric and physical information embedding: Through the mean square error (MSE), a data loss term is constructed, and the data loss term is added to the loss term of the graph neural network to achieve geometric and physical information embedding.

[0167] Physical information embedding is achieved by adding geometric information constraints, physical information constraints and data loss to the loss term of the fire temperature field model. The loss term of the model consists of data loss term, physical information constraint loss term and geometric information constraint loss term. Since the tunnel fire temperature field is finally predicted, the physical information constraint loss term is obtained by the fire temperature field control equation. By using a graph neural network for unsupervised large deformation non-rigid point cloud registration, the correspondence probability between point set pairs is learned by the graph neural network through the principle of self-attention network, so as to achieve registration between point clouds, thereby obtaining the tunnel fire temperature field neural network geometric information constraint loss term.

[0168] S23, temperature field digital intelligence learning model construction: The tunnel fire temperature field data set is constructed from the temperature field data set, and is substituted into the temperature field neural network constructed with embedded geometric physical constraint information to obtain the tunnel fire temperature field digital intelligence learning model. The present invention can not only ensure the model's learning ability for the physical problems of the tunnel fire temperature field, but also ensure the model's learning ability for three-dimensional unstructured data by establishing a geometric physical constraint graph neural network of the fire temperature field, thereby achieving efficient and universal prediction of the tunnel fire temperature field.

[0169] S30. Construction of building damage dataset: Collect tunnel fire information at the fire scene, build a digital twin of the tunnel fire condition, predict the tunnel fire temperature field and tunnel fire thermal radiation flux field through the temperature field digital learning model, obtain the tunnel's physical damage field, and obtain the tunnel's physical damage parameters and functional damage from the physical damage field. Integrate the tunnel's physical damage field, physical damage parameters and functional damage to the corresponding nodes to construct a building damage dataset.

[0170] like Figure 6 As shown in the figure, the steps for constructing the building damage dataset include:

[0171] S31. Digital twin of tunnel fire conditions: collect tunnel fire information at the fire scene, digitally twin the real tunnel twin, process and convert the tunnel fire information at the fire scene, and combine it with the real tunnel twin to realize the digital twin of tunnel fire conditions.

[0172] The tunnel fire condition digital twin steps include:

[0173] S311. Collection of tunnel fire information at the fire scene: Collecting tunnel fire information at the fire scene, including fire source location, fire source heat radiation heat flux, building material characteristics, tunnel positioning, tunnel fire temperature field, and tunnel fire heat radiation flux field;

[0174] S312, digital twin acquisition: according to the tunnel positioning, using the environment digital twin method, constructing a twin of the real tunnel and the environment to acquire a real tunnel twin;

[0175] S313, data processing and conversion: convert and combine the real tunnel twin and the tunnel fire information at the fire scene according to data features, format and standardize according to data types, convert data of different dimensions into dimensionless numerical ranges, complete data conversion, combine data in the form of two-dimensional arrays, and realize data processing and conversion.

[0176] S314, digital twin of tunnel fire conditions: Combine the processed and converted data according to tunnel fire cases to obtain real tunnel fire event data, and combine it with the real tunnel twin to realize the digital twin of tunnel fire conditions.

[0177] S32. Acquisition of physical damage: Through the temperature field digital learning model and with the help of tunnel fire conditions, the tunnel fire temperature field and tunnel fire thermal radiation flux field are acquired. With the help of the relationship equation between temperature change and tunnel thermal stress, the tunnel stress field is acquired. Through the equation describing the relationship between physical damage and temperature field, the node physical damage is acquired, the physical damage field of the tunnel is acquired, the number of tunnel vertex damages is calculated, and the vertex damage ratio, surface area damage ratio, and volume damage ratio are used to describe the physical damage parameters of the tunnel and acquire physical damage.

[0178] As shown in formula (13), by calculating the stress changes in the tunnel under fire, the relationship equation between physical damage and temperature field is constructed, as shown in formula (14). Combined with the yield strength of the building material, it is determined whether the node has failed and the physical damage field of the tunnel is obtained.

[0179] As shown in formula (13), the relationship between node temperature and node thermal stress is:

[0180] (13)

[0181] In the formula, refers to the thermal stress at node j at time t, Refers to the elastic modulus of building materials. Refers to the coefficient of thermal expansion of building materials. It represents the moment obtained through prediction The node below The corresponding temperature, Indicates the same node A moment before this moment The predicted temperature is, j represents the node number, Indicates the current prediction time.

[0182] As shown in formula (14), the relationship between node physical damage and node temperature is described as follows:

[0183] (14)

[0184] In the formula, is the physical damage of node j at time t, refers to the thermal stress at node j at time t, It refers to the yield strength of the material, that is, the stress value at which the material begins to undergo plastic deformation.

[0185] S33. Functional damage acquisition: Based on the physical damage of the node, the physical damage field of the tunnel is obtained. Combined with the structural characteristics of the building, the physical damage parameters of the tunnel are further obtained. The functional damage of the tunnel is obtained through formula (15).

[0186] 0 For example, the paper "Building Target Damage Assessment Based on Finite Element Simulation Results Identification" (Wang Jun, Lei Hongyu. [J]. Command Control and Simulation, 2024, 46 (4): 150-155. DOI: 10.3969 / j.issn.1673-3819.2024.04.020.) discloses a damage assessment index parameter to describe the damage of building targets. By calculating the number of damaged corners of the simplified cube model of the building, the corner damage ratio, surface area damage ratio, and volume damage ratio are obtained to describe the physical damage field of the tunnel. The corner damage ratio, surface area damage ratio, and volume damage ratio indicators given in the paper are used to describe the physical damage field of the tunnel, and then the tunnel function damage index is obtained by formula (15):

[0187] (15)

[0188] The physical damage parameters of the tunnel include top angle damage ratio, surface area damage ratio, and volume damage ratio.

[0189] In the formula, It refers to the top angle damage ratio. It refers to the surface area damage ratio. It refers to the volume damage ratio. Refers to their respective weights, the sum of the three is 1. Combined with the tunnel fire test data, the weight values ​​of the three are 0.2, 0.3, and 0.5 respectively. Refers to the damage to tunnel function.

[0190] The physical damage of the tunnel node is obtained by formula (13). If it is greater than 1, it means that the node of the three-dimensional tunnel model is completely damaged. Then the number of damaged vertex corners, damaged surface area, and damaged volume are obtained.

[0191] By calculating the number of damaged corners, the damage ratio of the corners of the three-dimensional tunnel model is obtained. By calculating the ratio of the damaged surface area of ​​the building to the surface area of ​​the building before damage, the surface area damage ratio of the three-dimensional tunnel model is obtained. By calculating the ratio of the volume of the changed part of the building to the initial volume of the building, the volume damage ratio of the three-dimensional tunnel model is obtained. The tunnel function damage is obtained by formula (15).

[0192] S34. Construction of tunnel damage data set: Integrate the physical damage field, physical damage parameters, and functional damage into corresponding nodes, and combine them according to tunnel fire cases to obtain the tunnel damage data set.

[0193] The tunnel damage dataset consists of tunnel fire case number, node number, time, node location information, yield stress of building materials, elastic modulus of building materials, thermal expansion coefficient of building materials, fire source location information, cosine value of the angle between the node and the fire source location, node thermal stress, node fire temperature, node fire thermal radiation flux, fire source thermal radiation flux, nonlinearity, evaluation index, node physical damage, vertex angle damage ratio, surface area damage ratio, volume damage ratio, and tunnel functional damage. The tunnel damage field dataset is represented as follows: [tunnel fire case number, node number, time, node location information, yield stress of building materials, elastic modulus of building materials, thermal expansion coefficient of building materials, fire source location information, cosine value of the angle between the node and the fire source, node thermal stress, node fire temperature, node fire thermal radiation flux, fire source thermal radiation flux, nonlinearity, evaluation index, node physical damage, vertex angle damage ratio, surface area damage ratio, volume damage ratio, tunnel functional damage] = [1, 2500, 50, (70, 10, 8), 20, 30, , (20,4,5), 0.99, 50, 205, 8, 30, 1%, 0.8, 0.8, 0.4, 0.5, 0.66]. The results show that in the tunnel fire event with fire damage case number 1, the yield stress of the building material is 20Gpa, the elastic modulus is 30Gpa, and the thermal expansion coefficient is / ℃ in the three-dimensional physical field of the tunnel, at the node (70, 10, 8), the node unit with node number 2500, at the 50th moment, the fire source is located at (20, 4, 5), the cosine value of the angle between the node and the fire source is 0.99, the tunnel thermal stress is 50MPa, the node temperature is 205℃, the node thermal radiation flux is 8kW / m², the fire source thermal radiation flux is 30kW / m², the nonlinearity is 1%, the evaluation index is 0.8, the vertex angle damage ratio is 0.8, the surface area damage ratio is 0.4, the volume damage ratio is 0.5, and the tunnel function damage is 0.66.

[0194] S40. Construction of forward and inverse model of tunnel functional damage: Utilize the building damage data set to construct, analyze the input and output data of the forward model, obtain the forward data set, substitute it into the constructed forward neural network to obtain the forward model of tunnel functional damage, and realize the conversion from three-dimensional temperature field to functional damage.

[0195] By using the building damage data set, the input and output data of the inversion model are analyzed to obtain the inversion data set, which is substituted into the constructed inversion neural network to obtain the tunnel functional damage inversion model, realizing the conversion of functional damage to fire source information; by completing the design of the tunnel functional damage forward and inversion parallel system, the tunnel functional damage effect forward and inversion model is constructed to realize the forward and inversion of the functional damage model.

[0196] It should be noted here that the results of the above forward and inverse neural networks are all existing technologies.

[0197] like Figure 7 As shown in the figure, the steps of constructing the forward and inverse model of tunnel functional damage include:

[0198] S41. Construction of forward model of tunnel functional damage: The forward model is constructed using the building damage dataset. The fire temperature of each node in the tunnel at each moment is used as the input of the forward modeling dataset, which is then input into the constructed forward neural network. The tunnel functional damage is used as the model output, and the building damage dataset is brought in. After further iteration, the forward model is obtained.

[0199] The tunnel function damage forward modeling step includes:

[0200] S411. Forward modeling dataset construction: The building damage dataset is used to construct the building damage dataset, and the fire temperature of each node of the tunnel at each time is used as input, and the tunnel function damage is used as output to obtain the forward modeling dataset.

[0201] S412, forward neural network construction: by introducing the self-attention mechanism, constructing the self-attention memory module, embedding the self-attention memory module into the standard convolutional long short-term memory network, using the tanh function combined with sigmoid to process data to form a gate, update the memory information, and output the data through gated processing to construct a forward neural network.

[0202] S413, forward model construction: construct according to the forward modeling data set, obtain the forward modeling data set, substitute it into the forward modeling neural network, perform data learning and iteration, and construct the forward model.

[0203] S42. Construction of tunnel function damage inversion model: Use the building damage dataset to obtain the building damage dataset, take the tunnel function damage as input, and the fire source location information, the cosine value of the angle between the node and the fire source location, the node thermal stress, the node fire thermal radiation flux, and the fire source thermal radiation flux as output to construct an inversion dataset and substitute it into the inversion neural network to obtain the inversion model.

[0204] The tunnel function damage inversion step comprises:

[0205] S421. Inversion dataset construction: The building damage dataset is constructed to obtain the building damage dataset, with tunnel functional damage as input, fire source location information, cosine value of the angle between the node and the fire source location, node thermal stress, node fire thermal radiation flux, and fire source thermal radiation flux as output to obtain the inversion dataset.

[0206] 0S422. Inversion neural network construction: Using the multi-layer perceptron neural network, according to the temperature control equation, combined with the nonlinear activation function and the node information aggregation strategy, the fire source information inversion neural network is constructed, and the inversion neural network is obtained by embedding the physical information.

[0207] The inversion neural network construction step includes:

[0208] S4221. Construction of a neural network for the fire source information inversion model: Utilize a multi-layer perceptron neural network, combined with nonlinear activation functions and node information aggregation strategies, to construct edge information and point information processors, design a strategy for updating edge information with point information, and construct a neural network for the fire source information inversion model for unstructured information prediction.

[0209] S4222, Physical Information Embedding: Construct a new loss function for physical information, which includes the thermal radiation physics equation and original , the physical information constraint loss term and the data loss term are added to the loss term to obtain an inversion model neural network embedded with physical information.

[0210] (16)

[0211] In the formula, represents the physical information loss term, , They are the weights corresponding to physical loss and data loss respectively, and their sum is 1.

[0212] The point source model is used to deal with the thermal radiation in tunnel fire. In the point source model, the radiant heat is emitted from the center point of the central axis of the fire source where the flame is concentrated. The thermal radiation flux received by the node is obtained by formula (18):

[0213] (17)

[0214] In the formula, Represents the fire source location information. Represents the location information of the tunnel node. Refers to the distance between the tunnel node and the fire source, in units of .

[0215] (18)

[0216] In the formula, Refers to the node thermal radiation flux, in units of . Refers to the heat radiation flux of the fire source, in units of . It refers to the angle between the node and the fire source. It refers to the cosine value of the angle between the node and the fire source.

[0217] Tunnel fire case number, node number, time, node location information, yield stress of building materials, elastic modulus of building materials, thermal expansion coefficient of building materials, fire source location, cosine value of the angle between the node and the fire source location, node fire temperature, node fire thermal radiation flux, fire source thermal radiation flux, nonlinearity, and evaluation indicators.

[0218] The node thermal stress is obtained by formula (13), the relationship between physical damage and temperature is obtained by formula (14), and the relationship between tunnel physical damage parameters and functional damage is obtained by formula (15). The relationship between node thermal radiation flux, fire source position and fire source thermal radiation flux is obtained by formula (18). Combining the obtained tunnel functional damage, node thermal strain and fire source thermal radiation flux, the physical information constraint loss term of the temperature field is obtained by formula (19); the physical information constraint loss term of the damage field is obtained by formula (18), and the physical information constraint loss term of the damage field is added to the loss term of the inversion model neural network;

[0219] The temperature field physical information constraint loss term of the fire source information inversion model neural network under tunnel fire events is obtained from the following formula (20):

[0220] (19)

[0221] In the formula, Represents the predicted fire source location information. Represents node location information.

[0222] (20)

[0223] In the formula, The number of nodes representing the tunnel fire temperature field, Indicates the node number, Represents the output fire source heat radiation flux, Representation Node The distance to the fire source, Represents the cosine value of the angle between the node and the fire source.

[0224] According to the following formula (21) and formula (22), the physical information constraint loss term of the damage field of the fire source information inversion model neural network under the tunnel fire event is obtained: :

[0225] (twenty one)

[0226] In the formula, represents the number of nodes in the tunnel fire temperature field, represents the node number, Indicates that the node is physically damaged. Represents the node thermal stress predicted by the tunnel functional damage inversion model.

[0227] (twenty two)

[0228] In the formula, represents the number of nodes in the tunnel fire temperature field, represents the node number, It refers to the damage to the tunnel function. It refers to the top angle damage ratio. It refers to the surface area damage ratio. It refers to the volume damage ratio. Refers to their respective weights, the sum of the three is 1. Combined with the test data, the weight values ​​of the three are 0.2, 0.3, and 0.5 respectively.

[0229] As can be seen from formula (23), the physical information constraint loss term of the damage field of the fire source information inversion model neural network under tunnel fire events is The loss term of tunnel fire temperature field and physical damage field And the tunnel fire physical damage field and functional damage loss items composition:

[0230] (twenty three)

[0231] Through formula (23), the physical information constraint loss term of the neural network function damage embedded in the fire source information inversion model under tunnel fire events is obtained.

[0232] As can be seen from formula (24), the total physical information constraint loss term The loss term is constrained by the physical information of the temperature field and the physical information constraint loss term of the damage field Adding them together gives:

[0233] (twenty four)

[0234] Through formula (24), the physical information constraint loss term of the neural network embedded in the fire source information inversion model under the tunnel fire event is obtained: , combined with the original data loss term , using formula (16), the total physical information loss term of the neural network embedded in the fire source information inversion model under the tunnel fire event is obtained.

[0235] S4223. Inversion model neural network construction: construct a data loss term through mean square error (MSE), add the data loss term to the loss term of the fire source information inversion model neural network, realize physical information embedding, and obtain the inversion model neural network.

[0236] S423, inversion model construction: according to the inversion data set construction, an inversion data set is obtained, and the inversion model neural network obtained by S422, inversion model neural network construction, is used to perform data learning and iteration to realize a fire source information inversion model that inverts from temperature field data to obtain corresponding fire source information data.

[0237] S43. Construction of forward and inverse model of tunnel function damage: By designing a parallel system of forward and inverse models of tunnel function damage, a forward and inverse model of tunnel function damage effect is constructed;

[0238] The tunnel function damage forward and inverse model construction steps include:

[0239] S431. Design of parallel system of forward and inverse model for damage to tunnel function: According to the forward model, the forward model is obtained; according to the inverse model, the inverse model is obtained. Through real-time synchronization of forward and inverse model data, the forward and inverse forms can be freely selected, the forward and inverse iterative calculations can be completed, the forward model and the inverse model can be combined in parallel, and the design of parallel system of forward and inverse model for damage to tunnel function can be completed.

[0240] S432, constructing forward and inverse models of tunnel function damage: by using the obtained forward model and inversion model, according to the parallel system design of the forward and inverse models of tunnel function damage, the forward model and the inversion model are connected in parallel to obtain the forward and inverse model of tunnel function damage.

[0241] Through the temperature field digital learning model, the temperature field data under the tunnel fire condition at a certain moment is obtained, and the forward model is combined with the forward model to obtain the forward model, which can obtain the high-temperature functional damage of the tunnel under the tunnel fire condition and realize the forward modeling of tunnel functional damage; through the temperature field digital learning model, the temperature field data under the tunnel fire condition at a certain moment is obtained, and the inversion model is combined with the inversion model to obtain the node fire source position information and fire source thermal radiation flux data under the predicted tunnel condition to realize the inversion of tunnel functional damage, and integrate the temperature field data inversion model and the fire source information inversion model to form a complete inversion system, realize the conversion of functional damage to fire source information, and complete the inversion of tunnel functional damage.

[0242] The above embodiments are only for illustrating the technical idea of ​​the present invention, and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the present invention.

Claims

1. A digital intelligence learning method for tunnel fire temperature field, characterized in that: include: The temperature field neural network is trained and optimized by using a temperature field data set including a training data set and a verification data set to obtain a temperature field digital intelligence learning model, thereby realizing digital intelligence learning of the tunnel fire temperature field; wherein the temperature field neural network includes an encoder, a processor, and a decoder, wherein: The encoder includes a Fourier encoder and a long-short time encoder. The weights of the Fourier encoder and the long-short time encoder are obtained by using a feedforward neural network, and the corresponding weights are respectively embedded in the Fourier encoder and the long-short time encoder. The outputs of the Fourier encoder and the long-short time encoder are superimposed to obtain a coding result. In the processor, a multi-layer perceptron embedded with a ReLU activation function is used to iterate the encoding result to obtain context vector data; In the decoder, the context vector data is processed using a recurrent neural network to generate an output sequence, and then the output sequence is processed using a feedforward neural network to output the decoding result; Geometric physics information is embedded in the temperature field neural network, and the geometric loss term is added to the loss function of the obtained temperature field neural network. and physical loss : ， ， in, Indicates the number of point cloud transformations, represents the temperature field predicted after the i-th point cloud transformation, represents the real temperature field, Indicates the number of nodes, Representation Node Predicted temperature field, t represents time.

2. The digital intelligence learning method for tunnel fire temperature field according to claim 1 is characterized in that: The steps to construct the temperature field dataset include: Collect the test data of tunnel fire temperature field, obtain the data of tunnel three-dimensional characteristics, building material characteristics, tunnel fire temperature, fire heat radiation flux and fire source parameters, and realize the fire feature selection; the fire source parameters include fire source location and fire source heat radiation flux, and the building material characteristics include yield stress of building materials, elastic modulus of building materials, and thermal expansion coefficient of building materials; The selected fire characteristics are screened and reconstructed. Through missing value completion, amplitude adjustment, and spatial correlation conversion, the tunnel fire temperature, thermal radiation flux, fire source parameters, and building material characteristics are converted into a two-dimensional array form to obtain the processed tunnel fire temperature field test data. The processed tunnel fire temperature field test data are combined according to the tunnel fire cases to obtain the tunnel fire temperature field test data set; Analyze the test data of tunnel fire temperature field and quantify the fire scenarios and parameters. The fire parameters include at least the probability of vehicle fire in the tunnel, the vehicle type and its corresponding fire scale. The fire scale parameters include the location of the fire source in the tunnel and the heat radiation flux of the fire source. With the help of quantified fire scenarios and parameters, the probability of fire occurrence and vehicle type are taken as random variables; The specific value of the random variable is obtained through the random number method, and the corresponding fire source location and fire source heat radiation flux are obtained by using the data interpolation method to realize the random process simulation; Conduct statistical analysis on the fire source location and fire source heat radiation flux obtained from multiple random process simulations to obtain the distribution law of the fire source location and fire source heat radiation flux, and implement Monte Carlo analysis; Through Monte Carlo analysis and digital twin method, we can obtain the twin of the real tunnel and the surrounding environment. Through meshing, we can obtain the numerical mesh data and conduct simulation. With the help of data crawling method, we can obtain the characteristics of building materials. Through data alignment, we can obtain the tunnel fire simulation data set. The temperature field dataset of tunnel fire temperature field test and tunnel fire simulation was integrated, nonlinearity and evaluation index were added as new features, time and tunnel fire case number were added, and the temperature field dataset was obtained.

3. The digital intelligence learning method for tunnel fire temperature field according to claim 2 is characterized in that: The data pattern in the temperature field dataset is [tunnel fire case number, node number, time, node location information, yield stress of building materials, elastic modulus of building materials, thermal expansion coefficient of building materials, fire source location, cosine value of the angle between the node and the fire source location, node fire temperature, node fire thermal radiation flux, fire source thermal radiation flux, nonlinearity, evaluation index].

4. The digital intelligence learning method for tunnel fire temperature field according to claim 2 is characterized in that: The nonlinearity and evaluation index formulas are as follows: ， ， Wherein, NL represents nonlinearity, MD represents maximum deviation, FSR represents full scale range, and EI represents evaluation index.

5. A digital intelligence learning method for forward and inverse modeling of tunnel function damage effects, characterized in that: include: Construct a tunnel damage data set, which consists of tunnel fire case number, node number, time, node location information, yield stress of building materials, elastic modulus of building materials, thermal expansion coefficient of building materials, fire source location information, cosine value of the angle between the node and the fire source location, node thermal stress, node fire temperature, node fire thermal radiation flux, fire source thermal radiation flux, nonlinearity, evaluation index, node physical damage, vertex damage ratio, surface area damage ratio, volume damage ratio, and tunnel function damage; wherein, the tunnel fire temperature field and the tunnel fire thermal radiation flux field are predicted by the temperature field digital intelligence learning model as described in any one of claims 1 to 4; Based on the tunnel damage dataset, the forward modeling dataset and the inversion dataset are obtained and substituted into the forward modeling neural network and the inversion neural network respectively for data learning and iteration to construct the tunnel functional damage forward model and the fire source information inversion model.

6. The digital intelligence learning method for forward and inverse modeling of tunnel function damage effects according to claim 5 is characterized in that: The steps to construct the tunnel damage dataset include: Collect tunnel fire information at the fire scene, including fire source location, fire source heat radiation heat flux, building material characteristics, tunnel positioning, tunnel fire temperature field, tunnel fire heat radiation flux field; Using digital twin methods, we can obtain a real tunnel twin; The tunnel fire information at the fire scene is processed and converted, combined according to the tunnel fire case, to obtain the real tunnel fire event data, and combined with the real tunnel twin to realize the digital twin of the tunnel fire condition; Using the temperature field mathematical intelligence learning model, the tunnel fire temperature field and tunnel fire thermal radiation flux field are predicted; Using the relationship equation between temperature change and tunnel thermal stress , to obtain the tunnel stress field; where, represents the thermal stress at node j at time t, represents the elastic modulus of the building material, represents the coefficient of thermal expansion of building materials, Indicates the predicted time The node below The corresponding temperature field is Indicates the predicted time The node below The corresponding temperature field; By describing the equations that relate physical damage to temperature fields , obtain the node physical damage; where, is the physical damage of node j at time t, represents the thermal stress at node j at time t, Indicates the yield strength of the material; According to the physical damage of the node, the number of vertex damage, the damaged surface area, and the damaged volume are obtained, and then the formula , and the tunnel function is damaged; among them, Indicates that the tunnel function is damaged. represents the vertex damage ratio, represents the surface area damage ratio, represents the volume damage ratio, and A, B, and C represent weights.

7. The method for forward and inverse numerical learning of tunnel function damage effects according to claim 5 is characterized in that: The steps for constructing the tunnel functional damage forward model include: The fire temperature of each node in the tunnel at each time is taken as input, and the tunnel function damage is taken as output to obtain the forward modeling data set; A forward modeling neural network is constructed, and the forward modeling data set is substituted into the forward modeling neural network for data learning and iteration to obtain a tunnel functional damage forward model.

8. The method for forward and inverse numerical learning of tunnel function damage effects according to claim 5 is characterized in that: The steps of building the fire source information inversion model include: Taking the tunnel function damage as input, the fire source location information, the cosine value of the angle between the node and the fire source location, the node thermal stress, the node fire thermal radiation flux, and the fire source thermal radiation flux as output, the inversion data set is obtained; An inversion neural network is constructed, and the inversion data set is substituted into the inversion neural network for data learning and iteration to obtain a fire source information inversion model.

9. The method for forward and inverse numerical learning of tunnel function damage effects according to claim 8 is characterized in that: Physical information is embedded in the inversion neural network, and the loss function of the obtained inversion neural network adds a physical information constraint loss term : ， ， ， ， ， In the formula, Indicates the number of nodes, Indicates the node number, represents the heat radiation flux of the fire source, Representation Node The distance to the fire source, Representation Node The cosine value of the angle to the fire source, Representation Node Thermal radiation flux, Representation Node Physical damage, Representation Node Heat stress, represents the yield strength of the material, Indicates that the tunnel function is damaged. represents the vertex damage ratio, represents the surface area damage ratio, represents volume damage ratio, and D, E, and F represent weights.

Images